A crow perched on our roof, looking at me wearily.

A frequent sound around our house lately has been an intense banging on the roof. It honestly sounds like someone has climbed up there and decided to hammer away for fun.

Curiosity finally got the better of me, and I grabbed a ladder to investigate. When I reached the top, I discovered the source of the obnoxious noises: an aggressive crow trying to crack open a walnut. We have a flat roof, so apparently it’s a perfect landing spot / walnut abuse space for them. I made a few attempts to shoo it away, waving my hands wildly and making all sorts of strange noises. The crow looked unimpressed and returned to its task.

I figured a more creative approach was needed, so I turned to the garden hose. Thankfully, no neighbors were around to question why I was suddenly watering the roof. The water seemed to have worked, as the crow took flight, leaving behind its walnut. Victory!

Or maybe not. I recently read about how long crows hold a grudge.

From a NY Times article, “If You Think You Can Hold a Grudge, Consider the Crow“, published in October 2024:

Renowned for their intelligence, crows can mimic human speech, use tools and gather for what seem to be funeral rites when a member of their murder, as groups of crows are known, dies or is killed. They can identify and remember faces, even among large crowds.

They also tenaciously hold grudges. When a murder of crows singles out a person as dangerous, its wrath can be alarming, and can be passed along beyond an individual crow’s life span of up to a dozen or so years, creating multigenerational grudges.

Oh, cool! This must explain all the aggressive cawing I hear from the trees as I walk to kids to school each morning.

Update (2024-05-24): I have video of these punks! Watch out.

For the most part, I’ve had pretty positive experiences using AI tools to help enhance my coding activities (though there was the one time…).

A recent experience with Google’s new Gemini model left me frustrated. After prompting it to help me find and update some relevant code, it confidently informed me that it had identified the exact snippet that needed replacing. Great news, I thought, until I realized it was instructing me to replace the code with… exactly the same code.

I pointed out the issue. Gemini politely apologized for the confusion and assured me it would correct its mistake. To my disbelief, it promptly suggested the very same replacement again! And again!

Oh, I have receipts. Join me on this little adventure!

Maybe we don’t have to worry about AI taking our jobs just yet!

If you’re casually interested in AI, then I think Ethan Mollick’s “Co-Intelligence: Living and Working with AI” is a book that you might find interesting. It’s not a technical book, and I believe it would be easy for someone not deeply involved in this world to read. It provides a very general introduction into how to utilize Large Language Models (LLMs) and serves as an introduction of what it means to live and work alongside these new tools.

“Co-Intelligence” unpacks the arrival and impact of LLMs, including tools like ChatGPT, Claude and Google’s Gemini models. Mollick, a professor of management at Wharton, approaches AI not as a computer scientist, but rather focuses on the practical applications and societal implications. In his own classroom, he has made AI mandatory, designing assignments that require students to engage with AI for tasks ranging from critiquing AI-generated essays to empowering them to tackle ambitious projects that might otherwise seem impossible (like encouraging non-coders to develop working app prototypes or create websites with original AI-generated content). He guides the reader through understanding AI as a new form of “co-intelligence,” which can be harnessed to help improve our own productivity and knowledge.

One concept I found interesting is what Mollick calls the “jagged frontier” of AI. This refers to the sometimes unpredictable nature of AI’s abilities. It might perform complex tasks with ease, like drafting a sophisticated marketing plan, and then struggle with something that seems simple to us. He gives an example of an AI easily writing code for a webpage but then providing a clearly wrong answer to a simple tic-tac-toe problem. This highlights why we can’t blindly trust AI and understanding its specific strengths and weaknesses through experimentation is key.

Mollick also delves into AI’s creative ability. He discusses how AI can excel in creative tasks, sometimes outperforming humans on subjective tests. This leads to interesting discussions about the future of creative work and education. The “Homework Apocalypse” he describes, where AI can effortlessly complete traditional school assignments, is a challenge educators and parents are currently facing. Mollick suggests this doesn’t mean the end of learning, but rather a shift in how and what we learn, emphasizing the need for human expertise to guide and evaluate AI.

The sheer volume of AI generated content being posted on the internet has is also becoming a problem and something we need to figure out how to navigate.

Even if AI doesn’t advance further, some of its implications are already inevitable. The first set of certain changes from AI is going to be about how we understand, and misunderstand, the world. It is already impossible to tell AI-generated images from real ones, and that is simply using the tools available to anyone today.

[…]

Our already fragile consensus about what facts are real is likely to fall apart, quickly.

Well, that’s just downright cheery! If anything, it highlights the importance of developing our ability to think critically and analytically in an AI-influenced information age.

Mollick lays out ways that we can better work with AI and leverage its strengths to help us, calling it the “four rules of co-intelligence.” These include always giving AI tools a seat at the table to participate in tasks, maintaining a human-in-the-loop throughout the the process to validate and verify AI work, treating AI as a specific kind of collaborator by telling it what persona to adopt, and remembering that current AI is likely the “worst” version we’ll ever use due to rapid improvements.

The bit on assigning personas was interesting. In my own experience, I’ve seen the benefits of giving AI a persona through system prompts. There’s also this fun example.

To make the most of this relationship, you must establish a clear and specific AI persona, defining who the AI is and what problems it should tackle. Remember that LLMs work by predicting the next word, or part of a word, that would come after your prompt.

[…]

Telling it to act as a teacher of MBA students will result in a different output than if you ask it to act as a circus clown. This isn’t magical—you can’t say Act as Bill Gates and get better business advice—but it can help make the tone and direction appropriate for your purpose.

The idea of these rules is that it can (theoretically) make working with AI feel less like a technical challenge and more like a collaborative effort.

Mollick also examines some philosophical questions that the use of AI brings, such as a “crisis of meaning” in creative work of all kinds. One specific example:

Take, for example, the letter of recommendation. Professors are asked to write letters for students all the time, and a good letter takes a long time to write. You have to understand the student and the reason for the letter, decide how to phrase the letter to align with the job requirements and the student’s strengths, and more. The fact that it is time-consuming is somewhat the point. That a professor takes the time to write a good letter is a sign that they support the student’s application. We are setting our time on fire to signal to others that this letter is worth reading.

Or we can push The Button.

The Button, of course, is AI.

Then The Button starts to tempt everyone. Work that was boring to do but meaningful when completed by humans (like performance reviews) becomes easy to outsource—and the apparent quality actually increases. We start to create documents mostly with AI that get sent to AI-powered inboxes, where the recipients respond primarily with AI. Even worse, we still create the reports by hand but realize that no human is actually reading them.

Side note: this exact scenario is something I’ve recently joked about with a manager at work. We have our yearly performance reviews and have to write a self assessment. Everyone now feeds a list of bullet points into their favorite LLM. The manager takes this overly verbose text and feeds it into an LLM to simplify the text.

On top of all this, Mollick also points out the need to always be skeptical of AI generated output, citing a famous case in 2023 where an attorney used ChatGPT to prepare a legal brief and was caught when defense lawyers could not find any records of 6 cases that were cited in the filing.

There is an interesting website I recently heard about, that is tracking fake citations used in court filings. 121 instances have currently been identified!

All in all, it’s a clear reminder of AI’s capacity for hallucination and the critical need for human oversight. The book frames AI not as a replacement, but as a powerful, though sometimes flawed, partner that can augment our abilities.

Overall, “Co-Intelligence” offers a decent overview for those curious about using current AI tools and thinking about their future integration into our lives. While it may present a more surface-level exploration for those already deeply familiar with LLMs, it provides some useful insights into the shifts AI is bringing to work and creativity. For someone looking for a general, non-technical introduction to the topic, it’s a solid read.

I’ve been in Seattle for the past week for one of our semi-quarterly work retreats. The weather has been absolutely perfect. Insane visibility that showed off the region’s geologic wonders.

As our plane received takeoff clearance and turned onto the runway, I was treated to this incredible view of Mount Rainier behind Sea-Tac’s control tower, draped by a lenticular cloud. Chef’s kiss.

Awhile back, I was fortunate enough to be invited to a semi-weekly “build night.” It’s a casual gathering where a bunch of random folks hack on all sorts of things: software, hardware, AI tooling, art. It is awesome! There are some seriously inspiring people there.

One longtime attendee showed off a project involving 3D-printed tiki sculptures lit from within by LEDs. Each sculpture had an ESP32 inside, controlling the LEDs and communicating via Bluetooth LE. A scroll wheel controller mounted on top could change the brightness and pattern, and the effects would sync across all the sculptures. It was ridiculously cool!

I happened to randomly have an ESP32 dev kit sitting inside a box of random electronics in my garage. I ordered a few more (these things are ridiculously cheap) and decided to experiment. Despite not having touched ESP32s before, it felt like a fun challenge.

Ultimately, I ended up cutting off a few LEDs from a WS2812 LED strip, soldering some wires up to the necessary connectors and then plugged them into a breadboard.

After snipping a few LEDs off a WS2812 strip and wiring them up to a breadboard, I used VS Code and the PlatformIO extension to vibe-code a basic Bluetooth host and client (shout-out to Claude for the help!).

The result?

1. The host ESP32 registers itself as a Bluetooth service using a UUID, loops through a series of colors, and broadcasts the current color value.
2. The client ESP32 scans for that UUID. If it finds it, it listens for the color data and updates its own LED. If not, it pulses red with a slow “heartbeat” to signal it’s waiting.

You can see it in action, below. Pretty wild! It starts off with the host disconnected. The only communication between the two is via Bluetooth.

You can probably repeat this blog post headline for any given service every week at this point…

Anyway! I’ve been on board the generative AI train for a few years now and it’s amazing to see how far it’s come. In October 2023, I got access to DALL-E 3 and was pretty impressed with its ability to render text.

Yesterday, OpenAI announced 4o Image Generation and boy does it kick things up a notch or two!

It’s ability to generate images and render text according to your exact prompt is incredible. We can now have full on automated AI memebots.

A four panel cartoon strip

• first panel: a software engineer sitting in front of a computer screen on a Zoom meeting
• second panel: the software engineer tells the participants (with a speech bubble): “I’m telling you, AI is coming for our jobs!”
• third panel: we just see a slight closeup of the software engineer (the computer monitor isn’t visible)
• fourth panel: same as the first panel except all the participants are now robots

Same angle and setup in every panel, reduced art style, broad outlines

Or, how about:

Cartoon drawing of a bored computer programmer sitting in front of a computer just pressed “enter” over and over. He is sarcastically excited and says, “Vibe coding. Wooooo.”

You can also feed it source images and it will run with it as well. So, obviously we need to use the Canine Calibration System.

I even gave it an image of me and told it to make a movie poster:

Create a dramatic cyberpunk 1980s horror movie poster image featuring a Computer Monster (We see an LCD screen with evil eyes and fangs and it has robotic legs) in a dark alley. In front of the monster, we see the man in this source image passed out on the ground, broken glasses lay next to him. At the top of the poster is the title of the movie in digital writing: “BUFFER OVERFLOW” at the bottom in the billing area, we see text that says, “Some bugs were never meant to be fixed.”

Or rewrite history…

Or really, really rewrite history…

It’s just wild. It’s coming for us as engineers, as musicians, as artists, as writers. This 2024 post on Twitter sums it up:

You know what the biggest problem with pushing all-things-AI is? Wrong direction. I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.

– Joanna Maciejewska on Twitter

Hmm, this sounds like a 4-panel comic to me!

Last summer at work, I embarked on a solo project to convert over 800 of our unit tests for various React components from using Enzyme¹ to React Testing Library² as part of a larger migration to React v18, TypeScript, and moving our code into a larger monorepo at Zillow.

This process was made much easier thanks to using the power of LLMs!

Just this week, I have seen two blog posts from various dev teams detailing how they did the same thing!

“How we navigated the shift from Enzyme to React Testing Library at The New York Times.“

As part of our efforts to maintain and improve the functionality and performance of The New York Times core website, we recently upgraded our React library from React 16 into React 18. One of the biggest challenges we faced in the process was transforming our codebase from the Enzyme test utility into the React Testing Library.

And today, I saw this from the AirBnb Tech Blog: “Accelerating Large-Scale Test Migration with LLMs“

Airbnb recently completed our first large-scale, LLM-driven code migration, updating nearly 3.5K React component test files from Enzyme to use React Testing Library (RTL) instead. We’d originally estimated this would take 1.5 years of engineering time to do by hand, but — using a combination of frontier models and robust automation — we finished the entire migration in just 6 weeks.

¹ Enzyme is a JavaScript testing utility, originally developed by AirBnb, for React that allows developers to “traverse, manipulate, and simulate interactions with component trees”, but it relies on various implementation details and has become less relevant with modern React practices.

² React Testing Library is a lightweight testing framework for React that focuses on testing components as users interact with them, emphasizing accessibility and avoiding reliance on implementation details.

This is a first for me. Cursor attempted to “fix” an issue I was having with TypeScript by adding a // @ts-nocheck statement to the top of the file, essentially preventing TypeScript from running validation checks against the code.

Alright, living in the future is pretty damn cool!

We were in San Francisco this weekend and decided to try riding in a Waymo. This was insanely wild. It’s such an incredible novelty at first — taking pictures, watching with fascination as the the steering wheel moves all by itself and the car seamlessly navigates around obstacles and through traffic.

After awhile, you settle into it and quickly forget how crazy it is.

As I mentioned yesterday, Anthropic released Claude Code. I saw it pop up fairly soon after it was announced and downloaded it rather quickly. One thing that I thought was notable was that you install it via npm:

> npm install -g @anthropic-ai/claude-code

As a seasoned TypeScript / JavaScript developer myself, I was excited to take a peek into the (probably minified) source code and see if I could glean any insights into making my own CLI tool. It’s always fun to see how different applications and tools are created.

Sidenote: I’ve been using Aider with great success as of late. It is a fantastic piece of open-source software — it’s another agentic coding tool, written in Python. I’ve been meaning to look under the hood, but building applications with Python definitely is not something that’s ever been in my wheelhouse.

Since Claude Code was installed into my global node_modules folder, I opened things up and immediately found what I was looking for. A 23mb file: cli.mjs.

I click on it, and as expected, it is minified.

Ah, well! I guess I should get on with my–

Wait a minute! What is this: --enable-source-maps?

I scroll through the file and at the bottom, I see what I’m looking for:

//# sourceMappingURL=data:application/json;base64,ewogICJ2ZXJzaW9uIjogMywKIC...

Sublime Text tells me there are 18,360,183 characters selected in that line.

Interesting! Since this part of the file seems to take up such a huge chunk of the original 23mb size, this means that it potentially contains full inline sources — we can rebuild the original source code from scratch!

However, this would have to wait. I had to take Benson to a vet appointment. I throw my laptop in a bag and head out.

While in the waiting room at the vet, I noticed a message in my terminal from Claude Code, telling me “Update installed, restart to apply.“

Hey, I love fresh software! So, I restart the app and go on my merry way. Benson finishes his appointment and I head back home.

Later that evening, I open up my machine and decide to open up the Claude Code folder again to start taking a look at the source code. I already had Sublime running from my earlier escapades, but out of habit I click on the file in Finder and open it up again in Sublime. I scroll down to the bottom of cli.mjs and see… nothing. The sourceMappingURL string was gone!

Apparently, the fine folks at Anthropic realized they made a huge oopsie and pushed an update to remove the source map. No matter! I’ll just head over to NPM to download an earlier version of the packa- oh! They removed that, too! History was being wiped away before my very eyes.

As a last resort, I decide to check my npm cache. I know it exists, I just don’t know how to access it. So, I head over to ChatGPT (sorry, Claude — I’m a bit miffed with you at the moment) to get myself some handy knowledge:

> grep -R "claude-code" ~/.npm/_cacache/index-v5

We run it and see:

/Users/daves/.npm/_cacache/index-v5/52/9d/8563b3040bf26f697f081c67231e28e76f1ee89a0a4bcab3343e22bf846b:1d2ea01fc887d7e852cc5c50c1a9a3339bfe701f	{"key":"make-fetch-happen:request-cache:https://registry.npmjs.org/@anthropic-ai/claude-code/-/claude-code-0.2.9.tgz","integrity":"sha512-UGSEQbgDvhlEXC8rf5ASDXRSaq6Nfd4owY7k9bDdRhX9N5q8cMN+5vfTN1ezZhBcRFMOnpEK4eRSEgXW3eDeOQ==","time":1740430395073,"size":12426984,"metadata":{"time":1740430394350,"url":"https://registry.npmjs.org/@anthropic-ai/claude-code/-/claude-code-0.2.9.tgz","reqHeaders":{},"resHeaders":{"cache-control":"public, must-revalidate, max-age=31557600","content-type":"application/octet-stream","date":"Mon, 24 Feb 2025 20:53:14 GMT","etag":"\"e418979ea5818a01d8521c4444121866\"","last-modified":"Mon, 24 Feb 2025 20:50:13 GMT","vary":"Accept-Encoding"},"options":{"compress":true}}}
/Users/daves/.npm/_cacache/index-v5/e9/3d/23a534d1aba42fbc8872c12453726161938c5e09f7683f7d2a6e91d5f7a5:994d4c4319d624cdeff1de6b06abc4fab37351c3	{"key":"make-fetch-happen:request-cache:https://registry.npmjs.org/@anthropic-ai/claude-code/-/claude-code-0.2.8.tgz","integrity":"sha512-HUWSdB42W7ePUkvWSUb4PVUeHRv6pbeTCZYOeOZFmaErhmqkKXhVcUmtJQIsyOTt45yL/FGWM+aLeVSJznsqvg==","time":1740423101718,"size":16886762,"metadata":{"time":1740423099892,"url":"https://registry.npmjs.org/@anthropic-ai/claude-code/-/claude-code-0.2.8.tgz","reqHeaders":{},"resHeaders":{"cache-control":"public, must-revalidate, max-age=31557600","content-type":"application/octet-stream","date":"Mon, 24 Feb 2025 18:51:39 GMT","etag":"\"c55154d01b28837d7a3776daa652d5be\"","last-modified":"Mon, 24 Feb 2025 18:38:10 GMT","vary":"Accept-Encoding"},"options":{"compress":true}}}
/Users/daves/.npm/_cacache/index-v5/41/c5/4270bf1cd1aae004ed6fee83989ac428601f4c060987660e9a1aef9d53b6:fafd3a8f86ee5c463eafda7c481f2aedeb106b6f	{"key":"make-fetch-happen:request-cache:https://registry.npmjs.org/@anthropic-ai%2fclaude-code","integrity":"sha512-ctyMJltXByT93UZK2zuC3DTQHY7O99wHH85TnzcraUJLMbWw4l86vj/rNWtQXnaOrWOQ+e64zH50rNSfoXSmGQ==","time":1740442959315,"size":4056,"metadata":{"time":1740442959294,"url":"https://registry.npmjs.org/@anthropic-ai%2fclaude-code","reqHeaders":{"accept":"application/json"},"resHeaders":{"cache-control":"public, max-age=300","content-encoding":"gzip","content-type":"application/json","date":"Tue, 25 Feb 2025 00:22:39 GMT","etag":"W/\"02f3d2cbd30f67b8a886ebf81741a655\"","last-modified":"Mon, 24 Feb 2025 20:54:05 GMT","vary":"accept-encoding, accept"},"options":{"compress":true}}}

Your eyes may glaze over, but what that big wall of text tells me is that a reference to claude-code-0.2.8.tgz exists within my cache. Brilliant!

More ChatGPT chatting (again, still smarting over this whole thing in the first place) and I get a nifty bash script to help extract the cached file. Only to find… they purged it from the npm cache. Noooooooooooo!

I stare at my computer screen in defeat. You got me this time, Anthropic.

As I decide to shut things down for the night, I’m tabbing through my open applications and get to Sublime Text, which is still open to cli.mjs. On a whim, I decide to try something: ⌘ + Z.

And there it is. The Holy Grail. The source map string.

And wouldn’t you know, it had a lot of interesting stuff! Due to the nature of parsing the source map, nothing is organized, but it’s still kind of fun to look through.

A few things struck me:

• It’s written in React (!) using an interesting tool called Ink (this allows you to create CLI apps using React). I hadn’t used Ink before but this looks like a lot of fun.
• While processing requests, Claude Code will show a nifty animated asterisk. I wondered how they did this. It looks like it’s a simple animation between a few ASCII characters: ['·', '✢', '✳', '∗', '✻', '✽'].
• In terms of system prompts, there’s no secret sauce to leak that you can’t already read by just looking at the minified JS file.
• These files are probably going to go out of date pretty dang quick, as the Anthropic team is actively developing the tool. As of right now, it’s already up to v2.19. This whole post was trying to look at the source code for v2.8, which went live yesterday.
• Lastly, in terms of Easter eggs, I look forward to receiving some Anthropic stickers…