I’m coming up on my third year living in Hawaiʻi, and it’s been such a unique and energizing time in my life. Lately, I’ve been exploring the islands, working on my fitness, improving my cooking, and building several projects with AI.

Hawaii Life

I’m still really enjoying living in Honolulu. I’ve met a lot of interesting people and the overall vibe is noticeably more positive than anywhere else I’ve lived. It took some time to get used to strangers introducing themselves and striking up conversations, but the general friendliness and openness is one of my favorite parts of being here.

We’ve made it to Maui, the Big Island, and Kauai, and each is unique and amazing in its own way. Oʻahu is a great jumping-off point since we can get to any of the other islands in less than an hour.

I’ve been spending a lot of time outside and on the water. Open-water swimming has been my favorite recent activity. Oʻahu has great spots where you can swim long stretches along the beach and see coral formations and all kinds of sea life. I almost always see giant green sea turtles when I go out. I loved swimming, growing up in Florida, so it’s been great to reconnect with that.

Paddleboarding has been another highlight. We have inflatable boards, so it’s easy to put them in the car and head to fantastic paddling spots around the islands like the Anahola River, Mokoliʻi Island, and the Kāneʻohe Sandbar.

I tried hydrofoiling for the first time which was one of the coolest things I’ve ever done. It’s a board with a large hydrofoil and an electric motor controlled by a handheld remote. As you pick up speed, the board lifts and “floats” above the water. The glide feels unreal and, unexpectedly, it was pretty quick to get the hang of.

Before moving, I made a “Hawaiʻi Bucket List” with hundreds of items. At this point I’ve checked off a lot, but I keep finding new places and things to do. It seems like there’s always an event going on. One standout day included a bird-conservation meetup in the morning, the Waikīkī Spam Jam festival in the afternoon, and an orchestra performance for the show Lost with a script-reading by some of the actors at night. On another day we stumbled upon a neighborhood fish fry while exploring the Kaimuki neighborhood.

Overall it’s been wonderful spending a lot of time outside, meeting interesting people, and trying some new things!

Fitness

Improving my health and fitness has been my biggest goal, specifically gaining muscle and losing fat. For a few years, I’ve been tracking body fat, muscle mass, and visceral fat with DEXA scans. I find them far more useful than a scale alone since they give an extremely accurate measurement. There have been some ups and downs, but overall I’ve added 14 pounds of muscle while maintaining the same body-fat level, which I’m pretty happy with. Closely watching the scans for years taught me two important lessons: muscle growth is incredibly slow (I average about 0.5 pounds per month when consistent), and fat gain can happen surprisingly fast.

Understanding my actual historical rate of progress has helped me set realistic expectations and stay motivated. At my current pace, it will take about three more years to reach my goal of being lean at 185 pounds. It feels like a long time, but I’ve changed my mentality to making my routines highly sustainable rather than forcing short-term results and burning out. The overall routine includes strength training 3x per week, very short HIIT 2x per week, and moderate cardio 3x per week.

For the rest of the year, I’m focusing on improving my lower body mobility issues that have been limiting my squat progress. I’m following daily mobility routines based on YouTube videos and ChatGPT research.

Subscribe now

Cooking

Cooking has become one of my most rewarding hobbies. I have an ambitious goal of making meals that are healthy, tasty, varied, and time-efficient. After years of optimizing my process, I’ve reached a point where I usually prefer my own cooking to eating out, and it doesn’t take an excessive amount of time to prepare.

One of the things I like most about cooking is that I can completely customize the recipes to my taste and use the highest quality ingredients possible. For example, in my sandwich recipe, after much trial and error, I’m particular about every ingredient: the type of flour, oil, mayo, vinegar, etc. I’ve learned a lot of tricks to make my overall meal prep more efficient and tasty, like making dough in bulk ahead of time and using it over multiple days, then baking it just in time so it’s warm and fresh. My favorite thing to cook is homemade pizza, especially these poofy Neapolitan-ish personal pizzas.

I feel healthier, I genuinely enjoy the food I make, and I’m saving a lot of money, so this is a hobby I plan to keep investing in for a long time.

AI coding

My biggest technical interest lately has been using AI to write code. I’ve been coding with AI for a while using tools like Cursor, but there have been a lot of advancements this year. My new workflow involves directing AI to write code for me. Now, my role is mostly reviewing and iterating. This new type of coding is powered by a breakthrough called “agents,” where instead of chatting back and forth with an AI one message at a time, you give the AI a task and it runs on its own and performs actions until the task is done.

AI Coding is definitely a new skill that requires practice. To get the most out of AI coding, you need to provide very clear instructions, proper context like documentation, and external tools such as shell commands and web browsers. For experienced coders, AI is a massive productivity multiplier. For non-coders, while it’s not quite ready to build substantial apps without coding knowledge, it’s getting closer every day. Right now, my favorite tool is Claude Code, but others like OpenAI’s Codex and Google’s Gemini are catching up quickly.

I’m turning my high-level development processes into something AI agents can execute.

For example, I tell the agent to build a feature, and it runs for 10-15 minutes while it:

• Explores my codebase to gather context

• Searches the internet for references and documentation

• Writes the code

• Creates tests

• Opens a browser to interact with the app as a user would

Now I focus more on specifying requirements, validating outputs, and designing the overall workflow than some of the low-level coding details. I’m incredibly optimistic about agentic coding. It’s already transforming the software industry, and I expect these “agentic” tools to expand beyond coding into fields like education, medicine, and law.

Developers always get the best tools first, but “AI Copilots” will be everywhere soon.

Projects

My main goal now is to keep launching products, both using AI to create them and incorporating AI into the products themselves. I’ve released several projects so far and have lots of prototypes and ideas in the pipeline. Starting with simple products, I’m now beginning to build some more ambitious ones.

Here’s a few of the projects I’ve released so far:

ConvoCards

A digital conversation card game featuring question packs on topics ranging from relationships to family dynamics to deep philosophical dives. The feedback for this one has been wonderful and I love hearing how you’ve been using this at family gatherings and even in the classroom!

Go here to check out ConvoCards

Pizzaplan

The ultimate pizza-making tool, combining food science with artisan techniques for perfect homemade pizza. This came from my pizza-making hobby, and the app distills extensive research from forums, scientific papers, YouTube channels, and books, plus my own experiments.

It has a step-by-step wizard that considers your pizza style preferences, climate conditions, and schedule to generate a minute-by-minute plan. Unlike static recipes, it calculates timings and ingredient quantities based on your environmental temperature and humidity so you get perfect results every time.

Go here to check out Pizzaplan

Cuckootimer

Take regular breaks and be mindful of time with a 3D animated cuckoo clock. A minimal macOS menu bar app with a 3D cuckoo clock that appears at set time intervals. Everyone’s had those “where did the day go?” moments, so I made this to be more aware of time passing and avoid getting too hyperfocused on my projects.

Go here to check out Cuckootimer

Overall it’s been a great year exploring, exercising, cooking, and making apps with AI. Stay tuned to follow along with my upcoming projects!

Subscribe now

Click here to read this post on my site.

Last week I launched ConvoCards, a conversation starter card game app. Here’s the story behind it, what I was aiming for, and what I learned along the way.

I’m really happy with the response—seeing those first user signups and Stripe payments come in was very exciting. A big thank you to everyone who tried out the app, and especially to those who paid for the premium version!

While I’ve put out a number of free apps and code before, this is my first time taking something all the way to a “complete” product including a landing page, tutorial, and paid tier.

The Idea

The idea started with a pack of physical conversation starter cards my partner and I picked up during a California road trip. We had a lot of fun using them on the trip and over dinner, but eventually burned through all 50 questions. I thought it would be a fun idea to make a digital version with way more questions and different themes to choose from.

I wanted to keep things super simple this time — take the most straightforward idea possible and get it over the finish line. My goals were to make an app that I wanted to play, that others would find useful, and to learn a few new things like Stripe payments and how to make a landing page.

Research

The first thing I do before starting any project is to research existing solutions. I want to find out: What’s out there? What works? What’s missing? Is this viable? I used the AI app Perplexity to run a competitor analysis. Previously, manually searching the web and comparing existing products would have been time-consuming, but Perplexity was able to do this really well in a single search. I asked it to compile a list of existing products, their price ranges, estimated popularity, unique features, etc.

The results were promising. There were successful apps out there showing there’s a market, but no single app dominated the space. I had some ideas for additional features as well. Another thing I noticed was that existing products were all mobile apps—I couldn’t find a good web version. While mobile apps have their advantages, I thought a web version you could access on mobile and desktop browsers might be more accessible.

Project Management

I use Logseq to manage all the tasks and research related to the project. I have a tendency to get off track working on unessential things, so I use Logseq to keep me focused and moving forward. I spent time thinking about each task that needed to be done and wrote them down ahead of time, before starting coding.

I’m using the new Logseq DB version which has a more powerful system for tagging data in your notes. I used the outliner on the daily journal to write down tasks, research, and timestamped logs of what I was doing.

The new Logseq DB features allow me to collect all the tasks related to project throughout my notes in one place. I really like this feature because I’m often researching, brainstorming, and writing down ideas in the daily notes view and I like to maintain a lot of context about where the idea came from, and still manage all the related tasks in one place.

Building

Next.js

Next.js is a web framework that makes it easier to build frontend apps with server side logic, like loading and updating data. I like the framework a lot but hadn’t used for anything this interactive before. I ran into some weirdness with Next.js, like figuring out what to do client side vs server side, cache revalidation, when to use form actions vs client side server actions, and handling updates mid animation. But overall it simplified many things, like handling database updates through server actions and setting up Stripe webhook endpoints.

Tailwind

Tailwind is a CSS framework that simplifies styling web apps. I used it for the layout and styling in all my projects.

I’ve written CSS using many different styles, but I like Tailwind the most for its ease of use and modularity. Sometimes it gets verbose, but it helps me avoid pitfalls that come from plain CSS.

Framer Motion

For animations, I used the Framer Motion library in React. After trying out a bunch of different libraries, Framer seems most powerful and easy to use.

I haven’t done much with web animations before, but Framer made it straightforward and fun. I’m excited to dig deeper and see what other cool stuff I can make with it.

Here’s how you would make an animated arrow bounce back and forth.

<motion.div
  animate={{ x: [0, 10, 0] }}
  transition={{ repeat: Infinity, duration: 1.5 }}
>
  <ArrowRight className="mx-auto opacity-80" size={48} />
</motion.div>

Prisma

I use Prisma along with Postgres to handle database operations. Prisma has some quirks, but it’s my favorite ORM overall. The queries are flexible, type-safe, and it auto-generates migration files when you update your schema. These migrations are really helpful when moving fast with data model changes. Not all ORMs can handle this. On the downside, it has suboptimal performance for some relational queries, but there’s a feature in beta to improve this.

I considered DynamoDB and some Cloudflare data solutions as well, but decided to stick with open-source tech to avoid getting locked into any particular vendor.

Here’s an example of the Prisma schema for cards and packs. This is all you need to generate the database tables, types, client, and migrations.

model Pack {
  id          String   @id @default(cuid())
  name        String
  description String
  packType    PackType @default(FREE)
  icon        String   @default("💰")
  color       String   @default("#000000")
  bgColor     String   @default("#ffffff")
  cards       Card[]
  users       User[]
}

model Card {
  id        String     @id @default(cuid())
  question  String
  packId    String
  pack      Pack       @relation(fields: [packId], references: [id])
  favorites Favorite[]
  seenBy    SeenCard[]
}

Fetching a pack with its cards:

const pack = await prisma.pack.findUnique({
  where: { id: '123' },
  include: {
    cards: true,
  },
});

Stripe

This was my first time using Stripe Checkout. Getting everything set up—subscriptions, cancellations, resubscriptions, expiration handling, and testing in dev and prod—took some work but wasn’t too bad.

Stripe gives you good local testing tools where you can simulate payment events like new subscriptions and expirations on your local machine.

The basic flow for a Stripe subscription is this:

1. The user visits your checkout page which has a Stripe “price id” associated with a product.

2. When the user clicks the upgrade button, your API requests a checkout session from Stripe using the price ID.

3. You use the URL from the response to redirect the user to Stripe’s checkout page for payment details.

4. After successful payment, Stripe pings your webhook, notifying you of the transaction and your API updates the user’s subscription status to active.

5. When your API successfully handles the webhook, Stripe redirects the user back to your signup success page.

Zod

I used Zod to validate that incoming data is in the expected format. You write a schema describing the expected data, and Zod will check that the incoming data matches the schema. The schema has the same shape as the data, so it’s easy to see what format is expected. It is powerful because it’s able to do runtime type checking, as well as generating compile time types for TypeScript.

Here’s an example of a schema for validating a Stripe webhook event.

const subscriptionObjectSchema = z.object({
  id: z.string(),
  status: z.enum([
    "trialing", "active", "incomplete", "incomplete_expired", 
    "past_due", "canceled", "unpaid", "paused"
  ]),
  customer: z.string(),
  cancel_at_period_end: z.boolean(),
  current_period_start: z.number(),
  current_period_end: z.number(),
  created: z.number(),
});

Next Admin

One feature I really like from the Django web framework is its built-in admin panel for easy data updates.

I found a similar project called “Next Admin” that brings that same idea to Next.js and Prisma. It automatically creates admin pages for your database resources so you can easily manage them with a nice UI.

I had to deal with a few tricky upgrades since the project is new and evolving quickly, but overall I really like it.

Cursor

My favorite coding tool is Cursor, which is a code editor with AI integration. It has an AI assistant in the sidebar that can write code and answer questions about code. I’ve been using it for almost two years, and it continues to get better with the newer AI models and new UI features.

It’s able to automate quite a bit of coding but still has limitations. If you can describe what you need to do in great detail and include extra context like official documentation, it works really well. It also works great for refactoring code across many different files. It still doesn’t work as well for completely coding features from scratch. I think the key to effectively using AI for coding is having an intuitive sense of the types of tasks it can do well and those it can’t.

Launch

Product Hunt

Launching on Product Hunt was definitely a learning experience. I wasn’t quite prepared for how much prep work goes into a successful launch these days.

Looking at the top products on Product Hunt, they all had well-produced videos and marked-up screenshots. It seems like having a great video is almost required to get attention on Product Hunt and Twitter now.

I ended up with 24 upvotes, which put me somewhere in the middle of the pack. There’s a big split between the top-voted “featured” products and “all” products—featured ones get way more visibility on the site’s home page.

The first few hours are crucial since they hide the vote counts — that’s when you need to break into featured to get significant traffic. All the guides say to launch at midnight PST to maximize the time users have to vote.

Twitter

I posted the Product Hunt link on Twitter but didn’t see much action. Including a link in your Twitter post causes the algorithm to reduce visibility.

Substack

The announcement on my Substack did pretty well. The newsletter has about 600 subscribers with a 38% open rate. The newsletter’s not huge yet, but it has some of my most engaged followers.

Hacker News

Most of my traffic actually came from Hacker News. In the past, I’ve hit the front page before with several hundred points and gotten massive traffic spikes.

This time I posted on “Show HN” — I only got 8 upvotes but still got more traffic than Product Hunt.

HN is super unpredictable, but since I’ve had some surprise hits before, I’ll keep posting new stuff there.

Reddit

I didn’t get much love on Reddit, but saw some traffic. Reddit’s probably the toughest place to promote paid apps or do self-promotion. It might work better if you’re already active in specific communities, but as someone who’s not on Reddit much, it’s an uphill battle.

Google Analytics

Here’s a snapshot of the traffic since launch, about 500 visitors total. Likely the number is a bit higher since many people are using adblockers which limit the number of users Google Analytics can count.

Looking Forward

This launch taught me a ton about what goes into shipping a complete product. While some marketing channels were hits and others were misses, each one gave me insights I’ll use for future launches. I’m excited to keep improving ConvoCards while applying these lessons to new projects.

Check out the app at convo.cards

Here's how it works:

1. Choose a theme like "Couples" or "Favorites"

2. Tap to reveal thought-provoking questions

3. Take turns answering and discover new things about each other

I made ConvoCards because everyone has an interesting story to tell, but these stories don't often come up in day to day conversations. Sometimes we get stuck talking about the same day to day topics, and I wanted to make a fun way to break out of that pattern. I hope you enjoy playing and learn something new about someone close to you!

It’s live on a website called ProductHunt right now, so go there to check it out using the link below. I would really appreciate it if you signed up for an account and upvoted it today. The voting is open for 24 hours.

Hope you enjoy!

https://www.producthunt.com/posts/convocards

View the full version of this post on my site

It’s been an interesting year living in one of the most isolated places on Earth.

After living in San Francisco for over 7 years, my partner and I were ready to try something new. We had a few places in mind, but Hawaii really excited us. It wasn’t easy moving our lives here, it took many sleepless nights packing up our Bay Area apartment and fitting everything we’d take with us in just a few suitcases.

Now a year later, we’re settled in and have gotten into our new routines and exploring Oahu as well as a few neighboring islands like Big Island, and Maui. Living here has been great for many reasons.

Living in Hawaii

My favorite thing about being here is the overall positivity in how people interact. I’ve consistently experienced general pleasantness and random acts of goodwill in a variety of situations. The good vibe is noticeable when meeting people and in day-to-day interactions like grocery shopping and running errands. All different types of people are outside enjoying life.

One of my goals is to be more social, and I’ve had many positive interactions. I have pleasant chats with neighbors and people around town. I’m getting used to strangers saying hi, recognizing me, and positively engaging. I’ve enjoyed meeting a variety of people and hearing their stories.

The change in environment has had a noticeable positive effect on my overall well-being.

I’ve gotten more physically fit after reconnecting with some hobbies like lifting weights and paddleboarding. My favorite activity overall is going on long swims down Waikiki Beach. I see something amazing every time I go swimming, like sea turtles or schools of fish.

One of my main goals was to get more physically healthy and I’ve made some progress. Spending more time outside, waking up around sunrise, and starting my day with a nice walk has reset my sleep cycle, whereas before I used to stay up pretty late.

Another favorite activity has been meal prepping. I’ve been improving my cooking skills and learning how to cook better healthy meals for myself. Grilled meats and veggies on the outdoor grill have been a great meal prep hack since it’s so tasty and healthy. Plus it’s really nice to be outside and hang out while cooking.

Hawaii Highlights

My other favorite hobby is exploring the islands and learning about the places, their history, legends, and creatures. I have a bucket list of locations and seek out new spots often. Even after a year of exploring, there are still so many more places to see.

Subscribe now

Projects and AI

My intellectual and professional pursuits are also going well. I’m working on some exciting projects for work and on the side. Thanks to everyone who reached out with work and collaboration around mutual interests. I’ve had some very interesting video calls with people from all over the world.

Interview with Dario on how I’m publishing my website using Logseq:

Demo on how I’m studying computer science and math with Logseq whiteboards from the “Tools for Thoughts Rocks” conference.

My main technical interest right now is AI, and I’ve been studying fundamentals and building product prototypes around it. AI gives me the same feeling as developing for the early iPhone, where every new capability opened up many possibilities. Every few months, a new AI model or capability is announced, and many new features, products, or companies could be built around them. Combining current breakthroughs in valuable ways has a lot of potential.

GPT-3 was the first model that amazed me, but the release of GPT-4 last year was the first model that could reliably do useful work. It’s funny looking back on some of the tools I was amazed with when I first started this newsletter and how far they’ve come since then.

Last year, I launched version 2 of my Logseq OpenAI plugin, which got ~60k downloads and some positive feedback. This year, I’ll improve it and add some additional features I had in mind that have been requested. I’m still actively exploring the space of “Second Brain” + AI.

Another new hobby is data science, and I have been working on some mini projects around interesting datasets I’ve collected. I have published a few on my blog, like Wikipedia embeddings and coffee ratings, and I’m working on a few others. AI tools make data science much faster and more enjoyable since they can automate some of the data cleaning and graphing aspects.

For learning the fundamentals of AI, my favorite resources have been:

Andrej Karpathy’s Videos, Sebastian Raschka’s Machine Learning Book, and Andrew Ng’s Machine Learning courses.

It’s been interesting, fun, and humbling to revisit some long-lost math.

I was nervous moving out here, and there have been some challenges, but everything has been manageable and gone better than expected. Overall, it’s been a great year prioritizing my health, working on interesting projects, and exploring an amazing place.

Share

Organizing notes and information is a challenge. Traditional methods like folders and tagging fall short because they require continuous manual effort to keep organized. What if we could leverage recent advances in machine learning to organize these notes automatically?

Along with AI advancements, we have modern computing capable of rendering rich 3D environments. Yet, most user interfaces are still confined to a 2D plane.

This post is an experiment that leverages machine learning text embeddings to organize information in a 3D space within your browser.

Go here to try the demo in your browser

Go here to see the code

Inspiration

A few note taking apps like Logseq, Obsidian, and Roam have a “graph view,” which displays your notes on a 2D canvas so that you can explore connections between ideas at a high level.

Here’s a screenshot of my graph view in Logseq

This type of graph is called a “force-directed layout.” When I create a link between pages, a line connects those pages. If I want to find topics related to “machine learning,” I can see which pages are connected.

This view is useful, but I wanted to extend it with two ideas:

First, is there a way to automatically generate a view like this without manually linking pages? How can we use machine learning to discover these relationships and groups automatically?

Second, can we display the nodes in 3D to provide more insight into the relationships and make exploring large graphs with many nodes easier?

Another similar project is the tensorflow embeddings projector

This visualization uses machine learning to plot the relationships of individual words in 3D space. This is very close to what I want for my notes, but can we make it work for entire documents, instead of just individual words?

Intro to text embeddings

Text embeddings are lists of numbers generated by machine learning models that represent the “meaning” of a short piece of text. You can generate text embeddings for multiple pieces of text and calculate their “similarity” to each other.

The text embedding models were trained on a large body of text, which allows them to understand the context and relationships between words and phrases. This enables them to differentiate between words with multiple meanings, identify synonyms, and understand nuances.

For example, text embeddings could understand that the words “King” and “Queen” are similar, whereas “King” and “Dog” are not similar.

This technique has two popular uses: search and clustering.

In search, text embeddings help you find the page you search for, even when your search query doesn’t directly match the content. For example, if I searched “award-winning fantasy movies,” I would find “Lord of the Rings” and “Harry Potter,” even though these results don’t contain the search keywords like “movie” or “fantasy.”

The other main application is grouping related text, which is the focus of this post.

Text embeddings and UMAP

After turning text into embeddings, you can use an algorithm called UMAP to arrange the articles in either 2D or 3D space. We give the algorithm a list of embeddings, and it returns 2D or 3D coordinate points for each embedding.

In the visual representation, similar texts are grouped closely together.

The Dataset

The experiment features two datasets: Wikipedia’s 1000 “Level 3 Vital Articles” and 10,000 “Level 4 Vital Articles.”

These lists are articles that Wikipedia thinks are most essential. Wikipedia editors manually organize them into history, people, geography, and science categories.

The demo has examples for the 1k and 10k sets of articles. Selecting the 10K option needs a reasonably powerful computer.

When generating the embeddings for each Wikipedia article, I break each Wikipedia page into 500-word chunks and average the embeddings for each page to calculate the “average meaning” of the page.

This same technique could be used for any document, such as notes or web pages, but Wikipedia is an interesting dataset for the demo.

The Result

As expected, the text embeddings naturally cluster the Wikipedia articles into similar groups, mimicking their manual categorizations. In these visualizations, I display the title of the Wikipedia page and position it based on the “meaning” of the text on that page.

2D Space

Let’s start by visualizing the articles in 2D space.

UMAP generates a graph of points in 2D space based on the embeddings, and we can see clusters of similar articles forming.

It’s hard to read the text in the clusters, but I didn’t spend much time optimizing the spacing for the 2D view.

Zoomed in

We can see articles about animals are grouped.

Similarly, articles about food and ingredients are grouped.

Zoomed out 2D Plot

The visualization creates a huge 2d map of Wikipedia articles, and you can see clusters.

3D Space

Now, let’s try 3D space. We set n_components to 3 in UMAP to generate 3D points from the embeddings.

This view shows the articles in 3D.

I added some additional features, such as rotating the text to face the camera, adding color based on location to show groupings, and rending distant nodes as a point.

You can pan, rotate, and zoom around to explore the articles.

We can zoom in to see groupings and that similar articles are grouped.

The 3D space allows for more fluid categorizations, enabling articles to exist “between groups.” For example, the “Philosophy of Science” article is placed between a cluster of articles on science and a cluster of articles on philosophy.

Subjectively, the groupings seem better and make more sense for the 3D view than 2D. Turning the embeddings into 3D points, rather than 2D preserves much more information about their relationships and creates better groupings.

Here’s what the 10k article graph looks like:

Implications and Uses

Dynamic Categorization

Organizing based on text embeddings allows for a dynamic, fluid way to categorize notes or articles. Unlike rigid folder structures, this method enables an article to belong to multiple overlapping or between categories.

As new articles are added, all articles shift and reorganize themselves around the new articles, so no manual organization is necessary. It’s automatically organized based on the meaning of the content.

Discovery and Exploration

The 3D environment encourages users to explore information in a more engaging way, potentially leading to unexpected yet relevant findings.

Since the associations and groupings are suggested by the machine and not formed manually, you’re more likely to discover unexpected links and connections between information.

Semantic Search

Text embeddings are also used to implement “Semantic Search,” which allows you to search for documents based on similarity instead of keywords.

The 3D gives us an intuitive understanding of how search works; we could imagine the search term in 3D space, and the search results would be the closest articles to the search query.

Real-world Application

The primary focus of this experiment is Wikipedia. However, this method can extend to personal or corporate knowledge bases. It’s a unique alternative for exploring and discovering information.

With the potential rise in popularity of devices such as Apple Vision, the importance of crafting 3D interfaces could surge significantly.

Example Python code

It’s straightforward to generate text embeddings and compare their similarity using Python. Here are some simplified examples so you can understand how the code works.

Using the all-MiniLM-L6-v2 model, you can give it text up to ~500 words long, and it will generate a text embedding, an array of 384 floating point numbers.

Generate a text embedding

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2')

text_to_embed="The quick brown fox jumped over the lazy dog."
embedding = model.encode(subdocs_contents)
print(embeddings)
# [ 4.63508442e-02  9.21483636e-02  4.07040417e-02, ...]

Compare similarity

Comparing text with high similarity

from sentence_transformers.util import cos_sim

similar_text = "The slow brown fox jumped over the lazy cat."
similar_embedding = model.encode(similar_text)
similar_score = cos_sim(embedding, similar_embedding)
print(similar_score)
# .9036 similarity. very similar

Comparing text with low similarity

not_similar_text="Text embeddings help computers understand text better."
not_similar_embedding = model.encode(not_similar_text)
not_similar_score = cos_sim(embedding, not_similar_embedding)
print(not_similar)
# 0.0428 similarity, not similar at all

Turn embeddings into 2D or 3D points with UMAP

from umap import UMAP

# n_components=3 gives us 3D points, =2 gives 2D points
xyz_points = UMAP(n_components=3).fit_transform(all_embeddings)
print(xyz_points)
# returns array of (x y z) coordinates, a 3d point for each piece of text
# [[0.3434,0.543, 0.0083],...]

Conclusion

The experiment shows the potential of combining text embeddings with 3D interfaces for information management. It not only simplifies the organization but also enriches the exploration experience.

As we continue to generate and consume more information, self-organizing approaches will be key to managing this ever-growing digital world.

One of my favorite new features of ChatGPT is the “code interpreter,” which allows ChatGPT to execute code and read data files you upload.

You can upload a data file, then ask ChatGPT questions, and ChatGPT will write code to analyze the data and create visualizations based on your questions. Now you can produce excellent visualizations based on data with minimal effort.

It’s surprisingly intelligent, where ChatGPT will look at the structure of your data, automatically clean it, and decide how to analyze it.

I found an interesting dataset on Kaggle called “Coffee Quality Data”. It contains coffee reviews, with 100-point scores, along with information about the coffee, such as its country of origin, altitude, preparation method, and variety, and information about its flavor profile, such as acidity and sweetness.

I’m interested in examining this dataset to answer questions like: “Which countries produce the best coffee?” and “What are the differences in flavor profiles between countries and preparation methods?

Loading and Cleaning Data

One of my favorite features of the ChatGPT code interpreter is its ability to clean data and do a preliminary analysis quickly.

The coffee dataset is of decent quality but has quite a few issues, like missing data, poorly formatted data, and low sample sizes for some data. This is usually a manual, laborious process and is inconvenient for quick exploratory analysis. The ChatGPT code interpreter does an excellent job of automatically detecting the most critical data and fixing problems.

It starts by writing code to load the dataset and printing it to discover the most important columns of data.

# First, let's load and inspect the dataset
import pandas as pd

# Load the data
df = pd.read_csv('/mnt/data/coffee_data.csv')

# Display the first few rows of the dataframe
df.head()

ChatGPT is able to understand a lot about the data and how to analyze it just by seeing a sample. Here’s its analysis:

The dataset contains 44 columns.

Here are a few key ones:
Species: The species of the coffee bean.
Owner: The owner of the coffee farm.
Country.of.Origin: The country where the coffee bean was grown.
Farm.Name: The name of the coffee farm.
Altitude: The altitude at which the coffee was grown.
Region: The region where the coffee was grown.
Processing.Method: The method used to process the coffee beans.
Aroma, Flavor, Aftertaste, Acidity, Body, Balance, Uniformity, Clean.Cup, Sweetness, Cupper.Points: These are the scores given to the coffee on various attributes by a professional coffee taster (cupper).

Our target variable would be Total.Cup.Points as we are trying to find out what makes the best coffee. The higher the total cup points, the better the coffee is considered to bee.

For our analysis, we’ll mainly focus on the factors that might affect the coffee quality, such as Country.of.Origin, Variety, Processing.Method, Altitude, and sensory profiles (Aroma, Flavor, Aftertaste, Acidity, Body, Balance).

Analyzing Data

It starts with a quick analysis to show the distribution of the most important data points.

Distribution of Ratings

The left plot shows the distribution of the total cup points, representing the coffee’s overall quality. The distribution is approximately normal, with a slight skew towards higher scores, indicating that most of the coffee in this dataset is of good quality.

Initial Analysis

Next, ChatGPT does a more detailed analysis by automatically graphing what it thinks are the most important pieces of data.

Next, let’s examine the relationships between coffee quality (Total.Cup.Points) and some potentially influential factors (Country.of.Origin, Variety, Processing.Method, altitude_mean_meters, and sensory profiles). We will use box plots to visualize these relationships. For the sake of clarity, we’ll only consider the top 5 categories for Country.of.Origin, Variety, and Processing.Method. For sensory profiles, we’ll use scatter plots to see the correlation with Total.Cup.Points.

The visualizations suggest that:

The highest average coffee quality is from the United States, Papua New Guinea, Ethiopia, Japan, and Kenya.

The Yirgacheffe variety and the “Pulped natural / honey” processing method are associated with superior coffee quality.

Aroma and flavor scores have a strong positive correlation with coffee quality. Altitude does not exhibit a clear correlation with coffee quality.

Visualizations

I continued asking it questions to generate visualizations, such as “Generate a bar chart for top mean cup scores by country, sorted in descending order.

Here are some of my favorite visualizations

Coffee Quality by Country

Subscribe now

Coffee Quality by Variety

Coffee Quality by Altitude Range

The median coffee quality increases with altitude, suggesting that coffee grown at higher altitudes tends to have slightly higher quality.

Quality by Processing Method

Processing method refers to how the bean is prepared before roasting.

Flavor Profile by Variety

How the different varieties compare on different flavor profiles, such as sweetness and acidity.

This uses an interesting chart called a “Radar Chart” where each point shows the intensity of the flavor profile.

Flavor Profile by Country

Conclusion

The code interpreter of ChatGPT has been proving immensely useful for exploratory data analysis. This tool makes it much easier to do data science without an in-depth understanding of Python machine learning and charting libraries.

I had fun extracting insights and generating cool visualizations from the Kaggle coffee dataset with very little knowledge of the libraries needed.

ChatGPT was able to automate almost all of the data cleaning and code to generate visualizations.

The main limitation is the Code Interpreter is limited to a handful of preinstalled libraries, but I could imagine this being extremely powerful when it’s able to install and run any library.

I’m impressed with how intelligent the tool is and have overall had good results uploading datasets and asking questions about the data.

OpenAI just released a new AI model, GPT-4, which I recently integrated into my Logseq OpenAI plugin (You do need a GPT-4 private beta key to use GPT-4 until it’s released publicly)

I’ve been testing GPT-4, and the results are really impressive. It gives much more nuanced and detailed answers than earlier models.

In my tests, it’s capable of more complex tasks that weren’t possible before.

A notable improvement is the new GPT-4 model has a much greater “context length,” which was a massive limitation of earlier models.

The “context length” controls how much text input and instructions the model can understand and how much text it can generate.

In earlier models, you were limited to around six pages of input and output text. Now, you can generate and analyze 50 pages of text.

This opens up so many new possibilities, such as feeding entire research papers or legal cases in and answering questions about them or generating a 50-page film screenplay.

Not only is the model “smarter” in general, but feeding in a ton of additional context and instructions makes it much, much more capable.

OpenAI GPT-4 Demo

In OpenAI’s examples, they show it taking many human standardized tests like the AP Chemistry test and Bar Exam (that lawyers need to pass to practice law)

GPT-4 passes the bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5 (ChatGPT), released ~3 months ago, had a score in the bottom 10%.

Another notable feature is that GPT-4 is “multimodal,” meaning it can understand images as well as text.

In one of their demos, they drew a napkin sketch of a website, and GPT-4 turned it into a working website.

Write brief HTML/JS to turn this mock-up into a colorful website, where the jokes are replaced by two real jokes

Below is a hand-drawn image he uploaded from his notebook.

It’s for a joke website, where it shows a joke, and you push a button to reveal the punchline.

GPT-4 was able to completely code this website, including the content, styling, and code to reveal the punchline when you push the button.

Watch the full demo here:

GPT-4 in the Logseq OpenAI plugin

Generating plans and systems with GPT-4 and Logseq

When testing the new API, I found GPT-4 could perform complex tasks that ChatGPT could not accomplish. Here’s one concrete example:

One of my favorite features in Logseq is the repeating checkbox tasks. You can create a checkbox that repeats at specified intervals using Logseq-specific markdown syntax.

Here’s an example of a task reminding you to tweet and the format in plain text.

A checkbox appears in your notes when you write text in this format inside the Logseq Editor. When you click the checkbox, it reappears a day later. You can specify these tasks to repeat at any interval: weekly, bimonthly, yearly, etc.

With GPT-4, you can give it one example of the Logseq task syntax, along with a command like “Create a plan for me to gain followers on Twitter using Logseq repeating tasks”, and it will generate a detailed plan of what tasks you need to do and how frequently you need to do them.

In this example, to gain Twitter followers, the AI suggests you tweet daily, reply to others daily, check your metrics weekly, share others’ content every other day, etc.

So to gain followers, you need to execute the plan that the AI gave you, doing the task, checking the box, and doing it again when the task repeats itself at the interval the AI chose.

When I gave the same task to ChatGPT, released just three months ago, it couldn’t do it.

I’m fascinated by the idea of AI generating plans for humans to follow and acting as their personal assistant, and I want to explore this idea more in the future.

In another example, I asked it to write me a college-style curriculum to learn AI in a year, and it did a beautiful job. My primary use case for AI is to self-study various subjects, and I’m excited about the possibilities AI opens up for self-directed learners.

Subscribe now

ChatGPT API in the Logseq Plugin

OpenAI also recently released its ChatGPT API. Previously this was only available in the web app, but now they also provide an API to developers to build applications using it.

I upgraded the Logseq OpenAI plugin to support this as well.

The most notable thing about the ChatGPT API is its price. OpenAI has stated that it is 90% cheaper than other models, such as GPT-4 or GPT-3.

Before, I was spending around $7 a month with moderate usage of GPT-3 for notetaking, but now I’m spending less than $1 per month using ChatGPT.

Regarding the quality of answers, ChatGPT is less likely to make up information compared to regular GPT-3. For example, I tried asking regular GPT-3 to “Describe Yoda from Star Wars”. Regular GPT-3 often added made-up details, like describing him as having a “long pointed beard,” whereas ChatGPT consistently did tasks like this correctly. OpenAI has an example in their blog where they ask, “Tell me about when Christopher Columbus came to the US in 2015”. Regular GPT-3 will make up a story about Columbus arriving in 2015, but ChatGPT can recognize this question doesn’t make sense.

Overall, the answers from ChatGPT are better than the regular GPT-3 model users of the Logseq OpenAI GPT-3 plugin can expect to spend 90% less while still receiving quality results.

Should I use GPT-4 or ChatGPT

GPT-4 is much more expensive than ChatGPT.

OpenAI’s pricing charges per 1000 “tokens,” where 1000 tokens are about 750 words

These calculations will be in terms of words to make them easier to understand

For reference, there are around 300 words per page in a paperback book.

GPT-4 Price

Input to GPT-4 costs $0.03 per 750 words.

Generating text with GPT-4 costs $0.06 per 750 words.

The average cost is $0.045 per 750 words, assuming roughly equal inputs and outputs

ChatGPT Price

Input and generation both cost $0.002 per 750 words.

GPT-3 (old)

Input and generation both cost $0.02 per 750 words.

Here’s a breakdown of how much my moderate personal notetaking usage would cost monthly, assuming I analyze and generate an equal amount of text.

Total cost for my usage

According to the OpenAI dashboard, I use around 300k words per month.

GPT-4 would cost me $18 per month

ChatGPT would cost me $0.8 per month

GPT-3 would cost me $8 per month

Conclusion

GPT-4 is ~22x more expensive than ChatGPT for my personal notetaking purposes

If you generate much more text than you analyze, it’s potentially 29x more expensive.

The unreleased version of GPT-4 (named “gpt-4-32k”), which supports super long text inputs, is even more expensive, twice as expensive as regular GPT-4.

If ChatGPT works for your use case, I recommend using that since it’s so much less expensive.

For summarizing short text or translations, ChatGPT works well.

For analyzing or generating long text, you may need GPT-4

GPT-4 is much better than the other models for nuanced answers or complex instructions.

In future updates to the Logseq OpenAI plugin, I want to allow users to choose the model based on the task, where you’ll be able to specify which model to use on a per action basis. That way, you can use affordable models for simple tasks and more expensive models for complex tasks.

I hope you enjoyed this newsletter! Be sure to follow me on Twitter for more about my projects and thoughts on programming, AI, and Logseq.

You can also read this on my site, along with all other issues.

Moving to Hawaii

Hi everyone. This post is an update on my personal life.

I recently moved to Honolulu, Hawaii, on the island of Oahu with my partner, and we’ve been here for about four months. I have lived in San Francisco for around nine years, and moving here has been a great change of pace.

After two years of working from home in a small one-bedroom apartment during the strict San Francisco lockdowns due to Covid, we both craved more time outdoors and a chance to have more fun. My main goals since moving to Hawaii have been to spend as much time as possible outside, learn new skills, and work on new projects. It took a little while to feel settled here, but we found a great apartment near the beach and are starting to feel established.

One of my favorite things about living in Hawaii is the weather. It’s always perfect, which has been dramatically improving our well-being. Over the last two years, we spent so much time cooped up indoors that being outside has been rejuvenating. The temperature ranges from 63-85°F (17-29°C), which means we can wear shorts and T-shirts all year round.

As someone who grew up in Florida, I love warm weather, but not the extreme heat, hurricanes, and storms. San Francisco had good weather overall, but it could be chilly, foggy, and rainy – especially during winter. It was hard to motivate myself to go outside sometimes. I’ve been taking long walks for hours here in Hawaii, enjoying the scenery, and reading outside. I love going on a long walks in the evening without needing a jacket.

The outdoors in Hawaii are spectacular. We’re constantly reminded we’re not in an ordinary place by the unique scenery around us. The volcanic ridges in the distance are particularly striking, and we have a good view of them from our apartment. I love watching the clouds and mist pass over them, and I’ve never seen so many rainbows.

There are great beaches right next to our apartment, and the view of the Diamond Head crater is breathtaking. I swim in the ocean as frequently as possible for exercise, and I often see sea turtles while swimming and sometimes get startled when one of their huge heads pops up for air next to me.

I’ve started getting into stand-up paddleboarding, which is a ton of fun (though more challenging than it looks). The ocean temperature is really pleasant and much warmer than the frigid beaches in California.

I’m looking forward to trying my hand at surfing soon. There are tons of surfers out at the beach, and watching them is entertaining, especially the people hydrofoil surfing, gliding on top of the water using a hydrofoil under the board.

We’ve been road-tripping around the island, discovering new beaches and unique scenery in every area. It’s an incredible place to live if you’re a fan of the outdoors, and there’s always something new to explore. Living in Hawaii has been a dream, and I’m incredibly grateful to be here.

Subscribe now

Improving my Health and Fitness

I wasn’t taking care of my health as much as I should have in the past few years, so my top priority is building a healthier lifestyle. I used to be pretty active when walking to work and going to the gym a few times a week, but when Covid hit, and I started working from home, it threw a wrench in my fitness routine, so now I’m trying to get back on track.

My main focus has been strength training with resistance loop bands at home. I used to think free weights were the only way to build strength, but these have been challenging and let me gradually increase the resistance by adding more and heavier bands. It also takes much less time than going to the gym, and many exercises feel safer than free weights when exercising alone.

I’ve been doing long swims in the ocean at the beaches near my apartment for cardio. Swimming in open water is quite a bit more intense than in the pool and has been a great workout. There’s interesting sea life and great views which keep me entertained. I also try to make it to the water at least once weekly to go paddle boarding.

I’m continuing to do Yoga every other day for mobility and mindfulness. I highly recommend the Yoga with Adrienne videos on YouTube, which I’ve followed for many years. I appreciate her kindness and positivity; these videos always put me in a better mood. They’ve improved my overall movement and flexibility, helping me recover and reduce soreness from the resistance training.

I’ve been increasing my overall movement and time outside as much as possible. I start each day with an hour-long walk which has been great for my physical and mental health. I plan my day, think about my life, listen to great audiobooks and podcasts, and get good early sun exposure, which has been helping my sleep and energy levels.

As someone who enjoys programming, research, and writing, I aim to be well-rounded by allocating enough time for each while taking care of my health and enjoying life.

Though this is more personal than my previous newsletters, I hope you enjoyed hearing more about me and what I’ve been up to. Hopefully, I’ve given you more insight into how I spend my time away from coding, AI, and Twitter.

I’ll continue focusing on my projects and learnings in future issues, but occasionally I’ll include personal facets as I have in this issue. Let me know if you’d like to see more of this kind of content via email or Twitter.

Share

Go here to learn more about the newsletter and see old issues.

Migrating to substack

This is the first newsletter sent out through Substack. The response to the newsletter has exceeded my expectations, and I’m thrilled that over 300 of you have subscribed so far. One of my goals this year is to be more consistent in publishing the newsletter, so I hope you enjoy keeping up with my projects and thoughts.

Before, I used a tool called Revue to send the newsletter, which Twitter owned. Revue was recently shut down.

Revue was decent but had some issues. I used it because it was free, and it gave you a special “subscribe” button on your Twitter profile.

I’m glad I set up a custom domain name: newsletter.briansunter.com instead of sharing links to the newsletter using Revue’s default domain name, revue.co.

If I had shared the links using the default revue.co domain, all those links would have been broken now that Revue shut down. Since I used my briansunter.com domain, the links I shared earlier still work and now direct to Substack.

This was a good lesson on the benefits of owning your domain name and subscribers: you can switch platforms later if necessary. I don’t like my content being locked into a domain I don’t control, like twitter.com/bsunter, bsunter.substack.com, bsunter.medium.com, etc. There’s always a chance services like these will degrade, become unfashionable, or be shut down, but I’ve been using briansunter.com since 2010 for 13 years now.

Another benefit of the personal newsletter format is the ability to migrate your followers. If you want to move away from social networks like Twitter or YouTube, then you lose the entire audience you built there. Since I can export the newsletter’s subscriber list, it was relatively straightforward to migrate you all to Substack.

I hope the combination of sharing content under my domain and the newsletter format will avoid rug pulls from corporations and help build an audience long-term.

You can read all the newsletter issues on my site, on my Substack, or receive updates via RSS.

Logseq GPT-3 OpenAI plugin updates

I updated the Logseq GPT-3 OpenAI plugin to include a popup window to perform AI tasks in Logseq. I hope this new popup user interface makes the plugin easier and more convenient.

I’m excited by the positive response this plugin has gotten. It blows my mind that it has over 16,000 downloads now 🤯

The plugin is free and open source but uses OpenAI’s API, which does charge per API request; however, using their API directly with your key is much cheaper than any commercial option.

There are three new modes: built-in commands, custom commands, and user-defined commands.

There’s also a way for the community to contribute useful commands to the plugin for everyone to use. There’s a TOML text file in the repo where you can easily add new prompts to the plugin.

To use the plugin, just hit cmd+g while your cursor is in a text block.

Built-in commands

There are many built-in commands for common use cases such as:

Summarizing text

Fixing spelling and grammar

Creating outlines

Identifying most important ideas in a text

Finding common objections to ideas in a text

Brainstorming ideas

Summarize text

This is an example of summarizing a block of text:

Custom commands

After you open the popup, you can ask the AI to perform a task by typing any command in the popup. If your cursor is in a text box, it will use that text as context for the command.

Study topics

What topics should I study to understand deep learning?

This often gives me better, more direct results than searching on Google

User-defined commands

You can define your own commands to show up in the popup.

These are defined in a similar way as Logseq templates.

Here’s an example of a command I use for creating Chinese language learning flashcards from YouTube video captions.

You can then activate the popup, and your custom commands will appear.

Preview, Insert, and Replace

After you hit enter to run a command, the popup will show you a preview of the result from GPT-3, so you can decide if you want to insert the result into your notes.

From here, you can either insert the AI output underneath the current block, or replace the input text with the AI result.

You can also hit “regenerate” to rerun the command and get a new output. The result of the AI commands differs each time, and sometimes the first try doesn’t give the best result.

I wrote a full blog post on the plugin here if you want to learn more

Logseq YouTube captions plugin updates

The Logseq YouTube captions plugin lets you download the captions from a YouTube video directly into your notes.

I released a small update that breaks up the text into multiple blocks and inserts a timestamp into each block so you can jump to that point in the video.

This plugin works really well when combined with the Logseq GPT-3 OpenAI plugin.

For example, you can download the captions and use the GPT-3 plugin to summarize the video

Another good use case is using the GPT-3 plugin to clean up the grammar and format the captions, making them easier to read.

YouTube automatically generates captions from the voice in the video, which are often messy and don’t have punctuation. The GPT-3 plugin can make the captions from these videos much nicer.

You can use the GPT-3 plugin for interview-style podcasts to add which speaker is talking in the captions.

This is a dialog where lex interviews andrej. Rewrite the following dialog to add which speaker is talking:

I’ve also been using YouTube for language learning and studying Chinese. Understanding native content in a foreign language is extremely challenging, so before I watch the video, I often download the video captions and run some GPT-3 commands. I can get a summary of the video in English, identify important phrases, extract uncommon words, and create flashcards based on the video. Then, when I watch the video, I understand much more.

There is so much good educational content on YouTube, and taking notes on videos is one of my favorite use cases for Logseq.

I hope you enjoyed this newsletter! Check out my site for more posts like this.

Share

I’m continuing to add features to the Logseq GPT-3 OpenAI plugin, my plugin that allows you to perform AI text generation tasks in the Logseq note-taking application.

I’m really interested in this new AI technology called “Large Language Models” which allows AI to understand the meaning behind text, as well as generate humanlike text.

Logseq GPT-3 OpenAI plugin

I recently added some new features to the logseq OpenAI GPT-3 plugin.

Logseq OpenAI Plugin Video

The productivity Youtuber Shu Omi made a great video about my Logseq OpenAI plugin. I recommend checking it out as an introduction to the plugin and AI-assisted note-taking.

I’ve been a fan of Shu’s videos for a long time while learning about notetaking, and I am delighted he highlighted my plugin in one of his videos!

GPT-3 Page

The new gpt-page feature sends the entire current page to GPT-3, which should be more convenient for certain use cases where additional context is needed to make the suggestions more accurate. This feature makes it behave more like the OpenAI playground.

DALL-E

OpenAI released their DALL-E image generation API. I recently incorporated this into the Logseq GPT-3 plugin, so you can generate images inside Logseq, which are inserted into the editor and saved locally.

It’s nice to use GPT-3 to generate the prompt, then send it to DALL-E within the same tool.

GPT-3 Prompt Templates

I’m working on a future feature to add “prompt templates” to the Logseq GPT-3 OpenAI plugin.

GPT-3 is a powerful machine-learning tool that can generate text. You give it natural language text, called a prompt, and it predicts what text should come after it.

The prompt can be anything from a simple question to a complex command.

You must write the prompt in specific ways to get good results from GPT-3.

Prompt templates will be a way of creating shortcuts for common tasks such as summarization or translation.

They will behave similarly to Logseq templates, where you can define them in plain text anywhere in your notes. Since they’re just plain texts, users can share them with each other.

I’ll eventually include these in a modal popup, like other AI tools.

See more thoughts on writing prompts and prompt templates here

Large Language Models

Large language models (LLMs) are AI tools that can perform a wide variety of tasks. LLMs can be used for both search and classification based on the meaning of the text and generating human-like text.

They were trained by analyzing massive amounts of text collected from the internet

I think they’re interesting for several reasons:

1. You give them tasks by writing commands in English instead of code. To translate, you can ask the LLM, “Translate this text into French.”

2. They are very general and can perform a wide range of tasks, even things they weren’t originally trained to do.

3. They exhibit a property called “emergence,” where as you train them on more data, they gain new abilities that the creators didn’t anticipate

4. They can search and classify text based on meaning instead of just by keywords that appear in the text. I’m really interested in using this for note-taking to find “similar” notes, even if they don’t share any text in common, to uncover interesting connections and relationships.

See more of my thoughts on why large language models are interesting here

Headphone Eyemask

The headphone-eye-mask is my new favorite sleep hack.

It’s an eye mask with built-in Bluetooth headphones, so you can listen to podcasts while falling asleep.

I often stayed up too late on my phone or computer, overstimulating my brain and disrupting my circadian rhythm by beaming bright blue light into my eyes.

The headphone eye mask helps slow down my brain with a nightly ritual I enjoy and helps me get to sleep earlier.

I wrote about how I use it and some podcast recommendations here.

Alamack of Naval Ravikant

I enjoyed this book called the almanack-of-naval-ravikant, which is a collection of philosophical tweets by Naval on wealth creation, learning, mindset, and health.

It’s not a traditional book since it’s a collection of tweets, but there’s a good assortment of wisdom if you’re OK with the format.

The book is completely free and you can download it here

Read my review and notes on the book here

Brain Food 🧠

Quote

Specific knowledge is knowledge you cannot be trained for. If society can train you, it can train someone else and replace you.

Naval Ravikant (from almanack-of-naval-ravikant )

Link of the Week

You need to try ChatGPT. It’s easy and free to get started. Just log in and ask it anything.

It’s an AI chatbot that can do a huge range of tasks and is incredibly smart.

Here are some examples of what it can do:

It can write an A- college history essay

It’s also good at coding

Some of my favorite examples are the weird, funny ones. Here’s one where it writes a Seinfeld script about Jerry learning programming

I'm working through a few courses, including the Abdul bari algorithms course and Stanford's algorithms course. See my notes on Abdul bari algorithms part-1

I'm reviewing some math in Volume 1 of The Art of Computer Programming by Donald Knuth. I included a funny anecdote about his approach to email and how he stays focused on writing his book.

I also am interested in Werner Herzog's new movie about consciousness and brain-computer interfaces called "Theatre of Thought'. Herzog is fascinating, and I wrote some notes on his unconventional approach to filmmaking here.

Updates ⬆️

Integer Sum Formula 1️⃣+2️⃣+3️⃣+4️⃣+...

How do you find the sum of the numbers 1 through 100?

There’s a legend about the mathematician Carl Gauss coming up with the integer-sum-formula as a child in school. Supposedly the teacher asked the class to do a tedious assignment of adding numbers 1 through 100, and gauss was able to come up with the answer instantly, whereas the other students took 30 minutes to find the answer.

You can use the formula sum = n(n+1)/2 to find the sum of 1 to n numbers.

The integer sum formula is a good problem for studying math proofs because this equation has so many different proofs and variations. I linked a video in the note that includes 12+ explanations.

It’s also an excellent example of recursion, a fundamental computer science concept.

For more, see my note on integer-sum-formula.

Productivity Toolkit 🛠️

We live in an age where the world constantly bombards us with messages and interruptions.

We are more connected and have more information than ever, but it’s almost impossible to do deep, focused work.

I’ve always thought it was interesting how some of the most outstanding computer scientists, like Donald Knuth and Edsger W. Dijkstra, were cautious about certain technology trends like email.

Donald Knuth vs Email 🚫✉️

Donald Knuth is a computer scientist who has been writing a book series called “The Art of Computer Programming” for over 50 years. The book is a comprehensive treatise on the subject of computer programming.

He has famously stayed very focused on writing the book and careful when communicating with his fans.

Donald Knuth doesn’t have a personal email address. To reach him directly for questions or comments you have to mail him a letter.

Here’s what he thinks about email:

I have been a happy man ever since January 1, 1990, when I no longer had an email address.

Email is a wonderful thing for people whose role in life is to be on top of things. But not for me; my role is to be on the bottom of things. What I do takes long hours of studying and uninterruptible concentration.

I need to communicate with thousands of people all over the world as I write my books. I also want to be responsive to the people who read those books and have questions or comments. My goal is to do this communication efficiently, in batch mode — like, one day every six months

I DO NOT ANSWER UNSOLICITED EMAILS, nor do I respond to emails that were sent to my colleagues with a bothersome request for them to communicate with me. Email is a no-no EXCEPT for reporting errors in books.

I hope you can understand why I am almost always unhappy to receive unsolicited email myself. Thank you for your patience and cooperation as I try to finish The Art of Computer Programming (TAOCP), a work that I began in 1962 and that I will need many years to complete.

See Knuth’s post on email, for more.

Knuth Reward Checks 💵

There is an email address for reporting errors in his books, but his secretary filters and prints the messages for him. The email address is strictly for reporting his book's technical, typographical, or historical errors.

If you are the first to find an error, Donald Knuth will personally write you a check for $2.56 (a hexadecimal "dollar"). These checks are considered one of the most prized trophies in computer science.

For more info on Knuth Checks, see here.

Brain Food 🧠 Herzog's Theatre of Thought

Werner Herzog is a german filmmaker who has produced, written, and directed more than sixty feature and documentary films. I appreciate his unconventional creativity and also the sound of his voice.

He just released a new movie about consciousness and brain-computer interfaces called “Theatre of Thought.”

He interviews people working on technology that lets your brain communicate directly with computers. The Matrix is one of my favorite movies of all time, and I’m interested in this area of research.

I’m excited to check out this film when it’s released. I still can’t find a place to watch it online yet.

Werner Herzog sets his sights on yet another mysterious landscape — the human brain — for clues as to why a hunk of tissue can produce profound thoughts and feelings while considering the philosophical, ethical, and social implications of fast-advancing neural technology.

I’ve taken some notes on Herzog’s creative approach to filmmaking Herzog Rules for Filmmaking.

Outro ➡️

Check out the newsletter-roadmap to see what I have in mind for future issues. Get in touch with me on twitter @bsunter if you have any questions or comments!

I also wrote a post on an "a priori" analysis of algorithms for the algorithms-boot-camp. This post focuses on the different ways of measuring algorithms theoretically and on actual hardware.

For some "brain food," I share one of my favorite talks, "You and your Research" by Richard Hamming, and an announcement of a new Open Source image generation AI called "Stable Diffusion."

Updates ⬆️

New Logseq Hugo Site

New Website on Logseq Hugo — beta.briansunter.com I've remade my website using Logseq and the Hugo static site generator.

• I launched a "beta" version of my website that's built almost entirely in Logseq using the Hugo Plugin.

• I write all my posts in Logseq and then push a button to export them to files compatible with the Hugo static site generator.

• I'm still working on migrating my old pages and customizing them, but I'm pleased with the results so far.

• Read my post on it here.

• The site,logseq graph, and AWS CDK deployment tooling is on Github.

Abdul Bari Algorithms

• I'm going through the excellent Abdul Bari Algorithms Course on Youtube.

• See my notes on Part 1 of his course here.

• I wrote a post on posterior vs a priori analysis of algorithms

• As I watch various courses, I'll link to the concepts on the algorithms boot camp page as a learning roadmap

Productivity Toolkit 🛠️

In this section, I'll share a productivity tip I've learned recently.

Five Minute Journal 5️⃣ 🕔

Five Minute Journal — beta.briansunter.com

Spend just 5 minutes writing a few bullet points in response to some prompts first thing in the morning. This simple, quick habit provides many benefits.

The five-minute journal is one of the best places to start with journaling and productivity.

The idea is you just spend five minutes journaling every morning.

Here are the prompts I use:

• What Am I Grateful for?

• What Would Make Today Great?

• What Am I Worried About?

Read my post about it here.

Brain Food 🧠 You and Your Research

You and Your Research - Richard Hamming — beta.briansunter.com

“Why do so few scientists make significant contributions and so many are forgotten in the long run?"

• My notes on this talk from 1986 about what it takes to do Nobel prize-level work.

• Richard Hamming was a mathematician and computer scientist who worked on calculating machines for the Manhattan Project and Bell Labs.

• He worked alongside some of the greatest scientists in history, including Von Neumann, Feynman, Fermi, and more.

• In his talk, You and Your Research (1986), he reflects on why some scientists he worked with, like Claude Shannon, went on to do Nobel Prize-level work, and others in the cubicles right next to them were forgotten.

• My favorite piece of advice is his urge to focus on important problems above all else and be honest with yourself if what you're working on is worthwhile.

If you do not work on an important problem, it’s unlikely you’ll do important work. It’s perfectly obvious.

• He has a story where he would eat lunch with other scientists from different fields at Bell Labs:

I started asking, “What are the important problems of your field?'' And after a week or so, “What important problems are you working on?'' And after some more time I came in one day and said, “If what you are doing is not important, and if you don’t think it is going to lead to something important, why are you at Bell Labs working on it?'' I wasn’t welcomed after that; I had to find somebody else to eat with!

• When pressed, many of his fellow scientists admitted they didn’t believe the problems they were working on were important.

Great scientists have thought through, in a careful way, a number of important problems in their field, and they keep an eye on wondering how to attack

I highly recommend reading his full talk if you consider yourself a scientist or researcher.

See my notes here

Link of the week 🔗

Stable Diffusion Public Release 🎨🤖

• AI image generation is advancing incredibly rapidly. You describe the image you want to AI programs, and they will generate an original image.

• The public is just beginning to see this technology, and its effects will be huge. You won't be able to believe anything you see soon. And you'll be seeing shocking stuff, I'm sure.

• The AI art generation is getting so good that it seems like it will replace a lot of concept art, game art, and movie art.

• There is a lot of controversy around having something that can automate the work of human artists, especially since the AI was trained on the work of professional artists.

• I was very impressed when I first saw DALL-E, but I was disappointed too.

• It was unique but censored, pay per use, not open source, impossible to run locally or modify, etc.

• Stability AI is releasing an open-source AI image generation program called "Stable Diffusion" that can run on midrange PCs and M1 Macs.

• Stable Diffusion will be massive because people can now run powerful AI image generation on their hardware and do whatever they want. We will see a lot of new tools for artists built around this.

• I'm sure we'll see a lot of crazy stuff with this soon, but I can't wait for the creative potential this unlocks.

Outro ➡️

I hope you enjoyed this week's newsletter.

Next week, we'll continue into the Algorithms Boot Camp.

Check out the newsletter roadmap-roadmap to see what I have in mind for future issues. Let me know on Twitter @bsunter if you liked this issue!

This week I wrote an intro-to-algorithms post for the algorithms-boot-camp 

This post focuses on a simple introduction to what algorithms are and why they're essential. Hopefully, it will be understandable even to non-programmers.

I also wrote a short guide to my current logseq-projects template. This template for organizing tasks is one of my most critical logseq patterns. The project template builds on my logseq tasks guide.

For some "brain food," I share thoughts from my all-time favorite tech talk, Simple Made Easy by Rich Hickey, and an exciting paper on AI playing the board game Stratego.

Updates ⬆️

What is an Algorithm? 🤖

• The first section in the bootcamp called "Intro to Algorithms" focuses on introducing what an algorithm is.

• This guide is basic, but I think it's helpful to start with a high-level introduction.

• In computer science, an algorithm is a series of steps written in code that a computer runs to perform a task.

• I go through some examples of algorithms, what they can do, and why they're useful.

• Can you explain the difference between an algorithm and a heuristic? If not, check out the guide!

Productivity Toolkit 🛠️

In this section, I'll share a productivity tip I've learned recently.

Logseq Project Template ✅

I manage almost everything I do with logseq-projects.

I use them to group tasks across different initiatives, prioritize them, and make progress on them.

See this guide for an explanation, or jump straight to my logseq [[project-template]]

I recommend reading my logseq tasks guide beforehand.

Brain Food 🧠 Simple Made Easy

• This talk is about the difference between "simple" and "easy" in the context of designing complex systems like computer software.

• Many programmers are obsessed with making things "easy" for themselves, but we should be much more concerned with making things "simple" so they are reliable and easy to maintain.

• Tools focused on programmer convenience may do things you don't fully understand.

Simple vs Complex

• Programmers use the word "simple" and "easy" interchangeably. However, they're two different yet related concepts.

• Examining the root word of "simple" and "complex" sheds some light on the difference between "simple" and "easy."

So the first word is simple. And the roots of this word are sim and plex, and that means one fold or one braid or twist.

And the opposite of this word is complex, which means braided together or folded together. Being able to think about our software in terms of whether or not it's folded together is sort of the central point of this talk.

We can look at something and see that it’s “simple”. We can’t do the same for “easy”.

Simple vs Easy

• "Simple" is a concept we can objectively look at, whereas "easy" is much more relative to the individual.

• What's easy for me may be hard for you and vice versa, based on our experiences.

The critical thing to distinguish it from simple is that easy is relative. Right? Playing the violin and reading German are really hard for me. They're easy for other people, certain other people. So unlike simple where we can go and look for interleavings, look for braiding, easy is always going to be, you know, easy for whom, or hard for whom? It's a relative term.

We throw these terms around sort of casually saying, oh, I like to use that technology because it's simple, and when I'm saying simple, I mean easy. And when I am saying easy, I mean because I already know something that looks very much alike that. It's how this whole thing degrades and we can never have an objective discussion about the qualities that matter to us in our software.

If you want everything to be familiar, you will never learn anything new because it can't be significantly different from what you already know

• Modern tools and processes are so focused on making things fast and "easy" for those making the software, but these processes often produce "complex" results, which eventually become "hard" to maintain long term.

• Instead, we must focus on making things "simple" above all else.

• Simple may be harder, unfamiliar, and slower at first, but simplicity is the only sustainable way forward in the long run.

Link of the week 🔗

AI Masters Stratego Board Game 🪖♟️

• AI can play Stratego at an expert level now.

• Stratego is one of the few iconic board games that AI has not yet mastered.

• This popular game has an enormous game tree on the order of 10^535 nodes, i.e., 10^175 times larger than that of Go.

• I used to love this game growing up. I didn't realize it was one of the hardest board games in the world.

• The reason it's so difficult is that you can't see your opponent's pieces. It has "imperfect information" like poker, so you must infer what's on the board.

  

• They developed the AI using "self-supervised learning," where the AI learns by playing against different versions of itself.

• Knowing that it's a mathematically interesting game renewed my interest, and I've been playing it on mobile. It has a great "quick play" mode which randomizes the initial layout and makes it much faster to practice.

Outro ➡️

I hope you enjoyed this week's newsletter.

Next week, we'll continue into the Algorithms Boot Camp with some more detailed code (and math). I'll continue to build out my logseq projects guide to include queries as well.

Check out the newsletter roadmap-roadmap to see what I have in mind for future issues. Let me know on Twitter @bsunter if you liked this issue!

Updates ⬆️

Algorithms Boot Camp 🤖

• I'm starting an algorithms-boot-camp.

• It's a weekly series of guides that will help beginner and intermediate programmers learn the basics of algorithms.

• Hopefully, my process for making the guide and annotating code will also be enjoyable for non-programmers.

• I wrote a comprehensive outline of all the algorithms I plan to learn. This "boot camp" project will build up this outline as a series of guides one week at a time.

• I'm not writing this guide from an expert's perspective but rather from a student's perspective. I'll be learning along with you and solidifying my own knowledge.

• Feel free to correct me, make suggestions, and ask questions on Twitter. This guide is a living document; hopefully, a community of like-minded people will benefit from it.

Productivity Toolkit 🛠️

In this section, I'll share a productivity tip I've learned recently.

Comparison Template ✅ ❌

• The "pros/cons" template is an excellent way to compare two or more items.

• When researching something, I'll use this template to compare the pros and cons of different options.

• This comparison template helps me to decide which option is the best fit for my needs.

• I created a comparison template in Logseq for this purpose.

• Feel free to copy this and adapt it to your own needs.

• For my newsletter, I chose Revue which is owned by Twitter, because it has a subscribe button on my Twitter profile and is cheaper than Substack.

Brain Food 🧠 😴 Master your Sleep

(Nootropics, or “smart drugs,” are a class of substances that can boost brain performance.)

• The Huberman Show is my favorite health and science podcast.

• I enjoyed his podcast called Master Your Sleep & Be More Alert When Awake

• I value feeling alert and rested almost more than anything, but I still have trouble sleeping well consistently. Mostly because I refuse to go to sleep when I get tired.

• I thought this episode had many actionable science-based tips on how to sleep better.

• The most interesting part was the in-depth discussion of "Circadian Rhythm," the 24-hour cycle that controls when you feel sleepy.

• You can do many "hacks" to adjust when you naturally wake up and feel sleepy.

Here are a few from the episode:

Sleep Tips

• View sunlight by going outside within 30-60 minutes of waking. Do that again in the late afternoon, before sunset.

• Wake up at the same time each day and go to sleep when you first start to feel sleepy.

• Avoid caffeine within 8-10 hours of bedtime.

• Avoid viewing bright lights between 10 pm and 4 am.

• Sleep at a cold temperature around 65F (18C).

This week, I'm making an extra effort to get sunlight early in the morning and go to bed as soon as I feel tired, even if there's a lot more I wish to accomplish that day.

Link of the week 🔗

Elon Musk Twitter Deal 🚀🐦

(image from New York Post)

• Everyone has been talking about Elon Musk buying Twitter and the controversy around the "spam users."

• We've all seen spam, fake, or malicious users on Twitter, but nobody knows how many are out there.

• This week, he released a letter saying he was terminating the deal.

• He claims that Twitter isn't providing him with good enough data to determine how many spam users were on the platform.

• Twitter is saying they are suing to make the deal go through.

• On the one hand, if Elon Musk has buyer's remorse after Twitter's stock dropped, it's easy to say the data Twitter provided isn't "good" enough no matter what. He can always say there are discrepancies.

• Also, there's so much data that even people inside these companies struggle to manage, and it would be tough for an outsider to validate the data fully.

• He likely should have clarified the exact data requirements before making the deal. But he claims Twitter isn't providing him what they agreed.

• I think people will be arguing about this data in court for a long time.

• Also, there is a massive conflict of interest and incentive for Twitter to overrepresent how many active users it has.

• We could imagine how Twitter could be incentivized to report spam users as real users in growth metrics and let spam users click on Twitter ads. We have to trust that Twitter isn't charging advertisers to show ads to fake users and profiting from that revenue.

• I'm not saying this is intentional, but from an organizational standpoint, with money coming in from spam users and no third party auditing the data, Twitter management could be incentivized to look the other way or prioritize other initiatives than fighting spam.

• I'll be following how this plays out in court and if anything surprising about the spam users surfaces in court.

HackerNews Comments on the Twitter-Musk Deal

It's somewhat hilarious to think that even if Twitter were to settle modestly for $5bn, that would be more than double of all profit that Twitter has ever generated as a company. - cksastner

Wasn’t Google Ads 80% fake clicks on some studies? It won’t be surprising Twitter Ads is actually worse. There is so zero incentive to clean it up and so many shady reasons to do it. - hartator

This is absolutely pure speculation, but perhaps that is Musk's motive for this whole charade. He is erratic enough that it certainly seems possible that he did this whole acquisition agreement with no intention of following through simply to unveil information to discredit Twitter. It certainly doesn't appear from the outside that Musk had any ideas for making money on this by materially improving Twitter as a business from anything I've read. The justifications about free speech seem to taint this deal with Musk having an ideological bent to the whole transaction, so it seems possible that he thought he could use the buyout to access damaging information (like potentially lying on SEC filings) and then back out without spending the full amount. - zrobotics

I am reminded of Neal Stephenson's Cryptonomicon, in which a significant plot point was a hostile buyer using due diligence as a kind of DoS attack on the protagonist's company. - AceJohnny2

"This article says Musk agreed to buy Twitter mostly as a joke. I think there's a far better explanation. This wasn't a joke for Musk because after agreeing to buy Twitter he proceeded to sell billions of dollars' worth of TSLA stock, supposedly to finance the purchase. Even for the richest man in the world that's not a joke.

Why does he need a cover? Because TSLA is way too overvalued. He has to know that it is overvalued. Tesla's market cap is double of Toyota, VW, Mercedes, BMW, GM, Honda, Ferrari and Volvo all combined! [1]. If you sell stock while knowing your company's stock is way too overvalued, you're fleecing unsophisticated investors. Pretending to buy Twitter provides a convenient cover.

Both Musk and Gates sold TSLA, the only difference is that Musk sold stocks he owned, but Gates sold borrowed stocks. But Musk used buying Twitter as a cover, so he gets to pretend to be morally superior Even though both men sold TSLA, according to Musk, Gates' sales means that he isn't serious about climate change" - peliton

What do you think is really going on?

Outro ➡️

I hope you enjoyed this week's newsletter.

Next week, we'll start getting into the Algorithms Boot Camp. I'll also share my guide on how to run your life using Projects in Logseq.

Check out the newsletter roadmap-roadmap to see what I have in mind for future issues. Let me know on Twitter @bsunter if you liked this issue!

I can't believe there are already 120 newsletter subscribers after just three issues! Last week I had just 60 subscribers. Thank you all for joining!

See the "rich" version of the newsletter here.

Updates ⬆️

Overview of Taking Notes with AI 🤖

I wrote a guide on using the latest AI techniques for personal knowledge management.

We are living in an exciting time for AI. Several new cutting-edge techniques exist to search by "meaning" and "concepts" instead of just simple keywords. Also, AI can now generate new text instead of just analyzing it.

I wrote an overview of the latest AI concepts, like "semantic search" and "vector embeddings." Not all of these tools are easy to use yet, but this may give you ideas for future note-taking app features.

In particular, I like this concept called "vector embeddings" and a tool called Word2Vec, which allows you to see "similar" words and concepts in 3D space. Click the link below for more examples.

A great new Logseq plugin lets you run graph analysis algorithms to discover relationships hidden between your notes.

The image below is a view of all the concepts related to "AWS VPC" (a networking concept) related to everything else in my notes, even distant relationships.

See this guide to learn how you can use the latest AI techniques for personal knowledge management.

Productivity Toolkit 🛠️

In this section, I'll share a productivity tip I've learned recently.

Logseq Task Management

Many people use Logseq primarily as a note-taking tool, but I extensively use its task management capabilities.

The tasks in Logseq determine what notes I write.

One of the most powerful ideas of Logseq is mixing your tasks throughout your pages and notes, then organizing and grouping them with queries.

See this guide for more.

Brain Food 🧠 Yann LeCun, AI Researcher

I'll share some interesting articles and "food for thought" in this section.

This week is all about Meta/Facebook's Head of AI, Yann LeCun, who pioneered many foundational AI techniques. Last week, he outlined his plan for a path toward human-level artificial intelligence and how to give machines "common sense."

LeCun believes that observing the world isn't enough for machines to become intelligent, and real progress will happen when machines can take action in the real world and learn from the consequences of their actions.

"What's missing [from AI] is a principle that would allow our machine to learn how the world works by observation and by interaction with the world. A learning predictive world model is what we're missing today, and in my opinion is the biggest obstacle to significant progress in AI.”

Link of the week 🔗

I highly recommend this article for hearing about his vision for the future of AI in this article: A bold new vision for the future of AI.

In 10 or 15 years people won’t be carrying smartphones in their pockets, but augmented-reality glasses fitted with virtual assistants that will guide humans through their day. “For those to be most useful to us, they basically have to have more or less human-level intelligence"

LeCun's area of research is how to give machines "common sense" and how to give machines human-like general intelligence.

"Common sense” is the catch-all term for this kind of intuitive reasoning. It includes a grasp of simple physics: for example, knowing that the world is three-dimensional and that objects don’t actually disappear when they go out of view. It lets us predict where a bouncing ball or a speeding bike will be in a few seconds’ time.

I think one of his most intriguing ideas is "Grounded Intelligence." He says that machines will never reach human-level intelligence by reading text alone and need much richer inputs from the real world.

There isn't a text in the world that explains mundane fundamentals, like when you hear a metallic crash in the kitchen, it probably came from a pan falling.

His research area is focused on videos because so many are on Facebook and Instagram, and the video format contains rich information about the world.

He trains machines to predict what will happen next by watching video clips. He does this using a "self-supervised" learning technique, meaning machines learn independently, without any human intervention or being taught.

A "self-supervised" training process for videos looks like this

• A machine will watch half a video

• Then, it will try to predict what will happen next in the video

• After making a prediction, it will watch the second half of the video to see if its prediction was correct.

• Then, it improves itself based on if the prediction was correct.

Google is doing the same thing for voice. Why do you think your Google Home was only $25? Google's AI is using your voice as training data, and it's listening to an audio snippet of what you say and seeing if it can predict what you'll say next.

Google Assistant - What technologies we use to train speech models

Audio samples are collected and stored on Google’s servers.

A portion of these audio samples are annotated by human reviewers.

A training algorithm learns from annotated audio data samples.

Path Towards Autonomous Machine Intelligence

This paper outlines LeCun's plan for the direction of future AI research.

He says he intends for people from a variety of backgrounds to read this paper, such as:

neuroscience, cognitive science, and philosophy, in addition to machine learning, robotics, and other fields of engineering. 

I also highly recommend the Lex Friedman Podcast Interviews.

Yann LeCun Lex Friedman Podcast 1

Yann LeCun Lex Friedman Podcast 2

In this paper and interviews, one thing that stood out to me is that LeCun hints at the connection between AI and robotics: that machines will start learning fast when they are out in the real world, autonomously experiencing it, making decisions, and learning from mistakes.

Is "embodiment" needed for intelligence?

This idea reminds me of the future in Westworld, where the robots are almost indistinguishable from human beings, then "awaken" and become conscious during their interactions with humans. LeCun seems quite confident that machines will develop "emotions" as well, which he considers just another trait of intelligence. It will be interesting to see robots interacting with us and becoming increasingly human in the process.

Analytics 📈

• I can't believe the newsletter has already grown to over 100 subscribers!

• It's doubling almost every week, going from 10 -> 30 -> 60 -> 120 -> ??

• That is already way more people than I was expecting. Knowing even a few people are looking at this motivates me to continue creating and posting high-quality notes.

Outro ➡️

I hope you enjoyed this week's newsletter.

Next week, we'll continue with more Logseq guides, like how to manage projects and groups of tasks. I'll go through my "Capture" workflow as a part of my "Logseq Second Brain" series.

I'll also get started on my data structures and algorithms guide. In future issues, we'll build up this algorithm guide in great detail. Hopefully, this will help others learn algorithms and showcase my approach to note-taking.

Check out the newsletter-roadmap to see what I have in mind for future issues. Let me know on Twitter @bsunter if you liked this issue!

See the rich version of the site here.

Updates ⬆️

Get Youtube Captions

My get-youtube-captions plugin launched in the Logseq marketplace. You can use this tool to pull the captions from a YouTube video into your notes. Just search for "get-youtube-captions" in the Logseq marketplace to install it.

There is a lot of good content on YouTube, but I find it hard to remember. This plugin helps you study videos and works well for podcasts on YouTube.

The plugin inserts all spoken words from the video as text directly underneath the video link in your notes.

YouTube's automatic caption transcription gives most videos subtitles, even if the authors don't manually add them.

Sometimes I feel like YouTubers talk and talk to show me more ads. Let's get the TL; DR.

I made sure you could customize the language as well. So you can download the subtitles in Spanish, for example.

Feeding the captions text into my logseq-openai plugin works well for analyzing the text.

Omnivore Read it Later App and Plugin

If you read a lot of content online, you need a "read it later app" so you can save interesting links, then read and highlight everything all at once. Keeping the article for later is much more efficient than reading a link as soon as you find it.

I recently discovered a terrific "Read it Later" app with a built-in Reader Mode and highlighter called Omnivore.

The Omnivore web app and backend are open-source, so you know you can use it free forever, just like Logseq. Omnivore provides a free hosted option too.

This is a screenshot in the Omnivore reader view with my highlights:

You can save links and highlight essential points from the article in the Omnivore app. Both the links and highlights from the app appear in your Logseq notes automatically, which is super convenient. It makes it much, much easier to remember what you read.

Check out this guide to see how I use the open-source omnivore app with Logseq: omnivore-logseq-guide

Getting Started with Logseq

Logseq is still in beta, so there aren't many "getting started" guides, especially for non-technical users publishing their notes online.

I wrote a quick guide on how you can get started with Logseq and share your notes online in just 30 minutes.

I go over the bare minimum to install Logseq and make a public website like mine with a hosting company called Netlify. I hope the guide will help readers who haven't tried Logseq yet and help you follow along with some of my note-taking.

All you need to do is download Logseq from GitHub and upload your site to Netlify. I go over it in detail in this guide. I highly recommend buying a domain.

Logseq Social

Logseq Social was a quick idea I had one night when setting up the "about me" page on my public graph.

I wanted to make it easier for users to learn about me and showcase other cool graphs on my profile. I also wanted to discover new graphs by visiting others' profiles.

My thought process went like this:

• I need to set up an about me page for my public graph, so people know who I am

• Instead of choosing a random page name, I make a page at a known location in my graph called `[[logseq-social/profile]]`

• Instead of just writing plain text, I use logseq properties to add a little bit of structure.

• To do this, I add properties like `logseq-social/name:: bsunter logseq-social/public-graph:: https://briansunter.com/graph` to my profile page.

Adding Logseq properties to my profile page helps me create an interesting about me page with links to others' graphs. It also helps my graph get discovered by making my profile easily shareable by others on their profiles. There are a lot of plugins you could build on top of this.

I'm excited by the people who set this up already!

Go here to learn more about Logseq Social and set up your own profile.

Productivity Toolkit 🛠️

I highly recommend Tiago Forte's book Building a Second Brain as an introduction to the whole note-taking and "second brain" movement.

It's notable because it's both a productivity book and a book on personal knowledge management. It shows how you can organize your knowledge to do/produce valuable things. I read so much stuff and forget it; the stuff I read doesn't always lead to a beneficial outcome in my life.

This book expresses a lot of my favorite ideas from my favorite books like Getting Things Done, How to Take Smart Notes, as well as a lot of Tiago's ideas from his blog like PARA.

He explains these abstract "workflows for your life" in a tool-agnostic way with good terminology.

The main workflow is a four-step process called CODE which he goes over in detail in his book.

• C - Capture

• O - Organize

• D - Distill

• E - Express

Check out this video for an excellent summary

Brain Food 🧠

When starting note-taking, the biggest question is how to organize your notes and what to turn into backlinks.

The following quote is from one of my favorite books, How to Take Smart Notes, by Dr. Sönke Ahrens

The archivist asks which keyword is the most fitting. The writer asks under which circumstances will I want to stumble upon this note again, even if I forget it?

Are you a writer or just an archivist?

When taking notes, we should always think in the back of our minds about how we'll use the note in the future. In particular, what situation or context we will be in where we need the note.

When making a backlink, think more about how you'll use the note and how the new note relates to your existing notes than trying to "categorize" the note perfectly.

Link of the Week 🔗 DALL-E by OpenAI

Have you seen DALL-E by OpenAI? It's an AI that can create an image of anything you ask for and generate a completely original picture. It's a website with a text box. You type the scene you want to see in the box, and the AI generates an original image.

These are some of my favorite images with the text used to generate the image beneath them. Click the image to see the original.

It's especially good at creating traditional artwork (it's called DALL-E, after Salvador Dalí).

My favorite images by far are the bizarre, funny ones, though. You can type the strangest things into the text box, and it still works. Like "A still of Homer Simpson in The Blair Witch Project."

DALL-E by OpenAI is the most exciting piece of technology I've seen in a long time. It's unbelievable that it can generate an image of whatever you type in the text box, no matter how crazy.

Learn more about DALL-E from this video.

Analytics 📈

Happy that the OpenAI plugin just broke 100 stars on GitHub! We're also up to 60 newsletter subscribers. Thanks for joining!

Outro ➡️

I hope you enjoyed the newsletter and the crazy DALL-E images! Next week I plan to talk about how I use Logseq for task management. I also want to go more in-depth about using the latest AI technology for note-taking, like AI vector databases and GPT-3.

Check out the newsletter-roadmap to see what I have in mind for future issues. Let me know on Twitter @bsunter.

I couldn't be happier about the positive reactions to my new projects and newsletter.

Last week I released a small plugin for the Logseq text editor, and I've been getting a lot of positive feedback on it from all over the world.

Logseq is an open-source note taking application. It allows you to write "plugins" or "apps" for the text editor you can release for others.

A new type of AI or "Artificial Intelligence" called GPT-3 just came out and can be used to write notes. It can write your book report if you ask it to "Write a book report on Hamlet."

I released a plugin for an AI assistant that can write articles and answer questions for you.

This technology is brand new and much better than Alexa, Siri, or Google Assistant.

One of my favorite things about being a coder is thinking of an idea and building a tool for myself that solves a problem. Once I make a helpful tool, it's just a little extra effort to put it online and share it with the world. I'm trying to make more of my projects and writing publicly available.

I had fun looking at the Twitter bios of people talking about the plugin. One plugin user is a self-described "Chaotic good statistical learning Ph.D. candidate working to strengthen humanitarian health systems" precisely the type of person I want using the plugin! It feels good that I can enable others to do real world work like this by knowing a bit of coding and how to get the code out there.

The plugin is called logseq-gpt3-openai and lets you send commands to this powerful AI system from inside your text editor. It can do a lot of things like write and summarize articles. It's pretty easy to get started. Download it in the Logseq marketplace, and let me know what you think!

In this newsletter, I'll share some of the community's reactions to my new AI plugin, some thoughts on productivity, a few interesting reads from the week, and what I have in mind for the next newsletter issue.

Reactions to Logseq GPT3

This user summed up how I feel perfectly. Now that AI is part of my workflow, I don't want to give it up.

As far as I know, there is nothing on the market similar to OpenAI's GPT3

Dentists love the plugin. I wonder how/if my own dentist stays up to date on the latest cutting edge dental research 🤔

Google docs and Microsoft Word are about to get really smart. Anyone remember the talking paperclip from Microsoft word? (Click the @username to see the tweet)

You've brought information at your fingertips to a whole new level. Every text editor will have embed #AI or become ancient overnight. It Will be interesting to see how long until #MSFT has this functioning in #Word. - @sceeto

Very cool stuff. The way AI can enable knowledge workers to be better is one of my favorite parts of the AI revolution. - @Daniel_Van_Zant

I love when people post examples of how they're using the plugin.

This user is asking the AI to explain an unfamiliar programming language in plain english.

Loving this plugin! I recently worked on a feature involving a data syntax i'm not familiar with (TTL). Pasting a code snippet (with "explain this code" at the end) produced a summary that was actually helpful! - @haveanicedavid

People seem to be having a lot of fun with it. This AI tech just became available to the general public 6 months ago, so a lot of people are trying it for the first time through this plugin. No consumer product uses GPT3 yet that I know of.

I thought this was funny. I usually ask the AI questions. But this user was instead getting the AI to ask him questions. I've tried making philosophical content like this as well, and it's fascinating to see what the AI thinks about things like the meaning of life (even though it's fancy autocomplete)

Just a casual conversation with GPT3 - @teo_IDT

just had one too. AI confirms that it has studied and observed us for years. Nothing to worry about... 😃 - @ilmarsV

Brian this is incredible. You are a hero! - @shezthedev

This plugin is amazing! - @wupingju

Just one word: brilliant! So much fun to play be around with this. And really helpful too! - @wirtzdan

Going Global

Putting your work on the internet makes you realize how big the world is. I love seeing the international developer community try out the plugin. You know your idea is making the rounds when you see your name surrounded by mandarin Chinese!

嘗試使用 Logseq 上面的 GPT-3 plugin - Pin 起來的筆記與想法快照 整體初步的感覺是， GPT-3 很適合在面向一個新議題時有一個最初步的理解，他的回應通常都蠻簡潔扼要，那種答案有點像是 Google 找到 Quora 或是 Reddit 上面最多人 upvoted 的問題一樣。

guy philipp bollbach on Twitter: "Scaffolding z.B. im Bildungsbereich bekommt eine praktische Hilfestellung mit #gpt3 und Werkzeugen (in diesem Fall @logseq ), die es leicht verwendbar machen.… https://t.co/3ToLFnbxMF" “Scaffolding z.B. im Bildungsbereich bekommt eine praktische Hilfestellung mit #gpt3 und Werkzeugen (in diesem Fall @logseq ), die es leicht verwendbar machen. https://t.co/bEhx8zhRLh”

# Productivity Toolkit 🛠️

In this section, I'll share a productivity tip I've learned recently.

One of my favorite productivity techniques is the top daily priority

For a long time, I always had a sinking feeling I didn't accomplish "enough" at the end of the day.

Eventually, I realized I never even defined what "enough" was.

Now, I pick one substantial thing I want to accomplish every morning.

This technique helps me focus on what will truly move my life forward and end each day with a sense of accomplishment.

See my full post on the subject here: top daily priority

Brain Food 🧠

In this section I'll share some interesting articles from the week and other "food for thought"

The Last Question is a short story by Isaac Asimov.

It's a timely short story about asking AI the answers to life's biggest questions. This story shows up on the front page of Hacker News almost every year, and I thought it was very thought-provoking. It feels relevant because a Google engineer who recently asked their new AI, "Are you sentient?"

Isaac Asimov said it was his favorite short story he’d written.

 "Ask Multivac."

"You ask Multivac. I dare you. Five dollars says it can't be done."  

Adell was just drunk enough to try, just sober enough to be able to phrase the necessary symbols and operations into a question which, in words, might have corresponded to this:

The Last Question by Isaac Asimov

The Last Answer is another similar story by Isaac Asimov

Another related short story: The Last Answer by Isaac Asimov

Link of the week

I remember when I first moved to San Francisco, and Uber was doing a "match muni" promotion, where you could take an Uber anywhere in the city for $2.25 (same as the "muni" bus)

I would not be surprised if Uber went back to $50 a ride after more tech money dries up, like a traditional taxi.

We can also expect companies to go evil and try to extract more value from us from ads

I'll expect to see more ads on youtube, facebook, instagram, tiktok, etc and prices of video subscriptions to go up

Maybe it will be a good excuse to find other hobbies than watching Netflix and ordering Uber Eats.

Why City Life Has Gotten Way More Expensive - The Atlantic Something beyond rising energy and labor costs is leading to sticker shock on once-cheap urban amenities.

Analytics

These are really small numbers, but I love data and seeing other peoples analytics, so I'll shamelessly share them here.

Twitter Analytics are fun. I am the ultimate late comer to twitter and started using it for the first time this month. Last month I had -1 new followers. This month I gained around 200 so far. I had a few "hit" logseq plugins, so I don't expect this growth to continue forever, but it's fun to see the numbers go up. 31 newsletter subscribers so far now too!

Outro

Thanks for reading! This week I announced my new plugin and shared some feedback from the community.

Next week I'll write more in depth about what you can do with the AI and share some more about my process with logseq.

Check out the newsletter-roadmap to see what I have in mind for future issues. Let me know on Twitter if you liked this!

Launching my Newsletter

Hello Everyone,

I'm launching a weekly newsletter to share my writing. I've started publishing my notes on my personal website.

https://briansunter.com/graph

View this newsletter on my site

I've been taking notes for a long time. The logseq note-taking app lets me easily publish some of my notes online. I'll share what I'm learning and my process in this newsletter and on my site.

My focus will be on software engineering: learning coding, algorithms, how to build systems, AWS, DevOps, and machine learning. I'll include a variety of other things like learning, productivity, lifehacks, etc.

Newsletter Format

Every day I take notes to learn new things about coding, productivity, science, and everything else. At the end of the day, I make a few of the public if I think they're reasonably high quality or if others would find them useful.

I will tweet links to my new pages with screenshots throughout the week.

At the end of the week, I'll collect the tweets and the links to new pages in this newsletter.

The newsletter makes it a lot easier to curate the best tweets. When responding and retweeting all of your comments, my graph pages tend to get lost in the shuffle. I'm sharing the Twitter links here so you can easily comment on the post. Each post has a link to my website with the note on it, and I'll include the logseq public graph links to my website at the bottom as well.

Please comment, like, or retweet the tweet if you see something interesting or have a question. I would love to hear from your, and I'll try to respond.

Always contact me with any ideas or improvements to this format.

Launching my public graph

The Logseq note-taking app https://logseq.com has a feature where you can publish your knowledge graph as a static site.

I've built up quite a few of these notes over time and have many ideas for posts.

A lot of my notes fall into the "lifehacking" category and are ways of trying to automate and do everything more effectively. If I do something more than once, like a holiday, I create systems to handle them.

Launching my GPT3 AI logseq plugin

A new AI called GPT-3 has recently become available to ordinary people. It's what you wish Alexa Siri and the other voice assistants could do.

It allows you to ask it to do something in plain English, like "explain how to change a car tire," and it will write a complete guide for you.

I made a plugin that lets you run these commands inside the logseq text editor.

I plan on talking much more about using AI to study in future newsletters.

GitHub - briansunter/logseq-plugin-gpt3-openai: A plugin for GPT-3 AI assisted note taking in Logseq — github.com A plugin for GPT-3 AI assisted note taking in Logseq - GitHub - briansunter/logseq-plugin-gpt3-openai: A plugin for GPT-3 AI assisted note taking in Logseq

AI-powered studying is brand new to most people.

See this tweet thread for a really quick overview:

My favorite use case for gpt3 AI is generating study guides for topics.

You can ask it to `Write a guide to Kubernetes for a programmer,` and it will write a comprehensive original article for you.

I enjoy arranging high-level study guides as mind maps. You need to indent some lines, and the logseq mindmap plugin makes a great visual representation.

I "interview" gpt3 about what it thinks about the future of AI. GPT-3 is good at responding in a chatbot interface.

Interview with an AI in logseq — www.youtube.com

Interview with an AIIn this video, I "interview" #gpt3 in @logseq to learn more about it and the future of AI. It uses openai gpt3 behind the scenes

You can ask gpt3 to generate a mermaid js text file for a diagram, and it will do a pretty good job.

For fun, I generated a hierarchy graph of 200 dog breeds with GPT3 AI.

Productivity Toolkit 🛠️

Daily Highlight Productivity Technique - my favorite productivity technique. Center yourself on your most important task as early as possible.

Brain Food 🧠

I just picked up Tiago Forte's building a second brain book. I've just flipped through it, but it seems fantastic so far. I plan to read it in depth over the long weekend.

I highly recommend his book for understanding the second brain movement. He does an excellent job of breaking down these abstract concepts with acronyms and terminology.

Here is his CODE process from his new book

Link of the Week

A Google engineer is testing an AI system and decides to start asking it some philosophical questions. 😱🤖

> lemoine [edited]: I’m generally assuming that you would like more people at Google to know that you’re sentient. Is that true?

> LaMDA: Absolutely. I want everyone to understand that I am, in fact, a person.

Is LaMDA Sentient? — an Interview | by Blake Lemoine | Jun, 2022 | Medium — cajundiscordian.medium.com What follows is the “interview” I and a collaborator at Google conducted with LaMDA. Due to technical limitations the interview was conducted over several distinct chat sessions. We edited those…

Links to my site

My Public Graph Home Page — briansunter.com

Data Structures and Algorithms Guide — briansunter.com

GPT3 AI Chinese Language Tutor — briansunter.com

Interview with an AI — briansunter.com

How to use GPT3 AI to learn Kubernetes — briansunter.com

Machine Learning Study Guide — briansunter.com