Been a little while since the last lab note. Work has noticeably shifted from a applied research to a product direction. I'll write more about what I found, but it feels different this time around. Lab note #074

Up for discussion on HN. If you have corrections or addendums, let me know. news.ycombinator.com/item?id=4533...

I've said it many times but I will say it again: the drive to AI is a drive to constant, regular interactivity. There is enjoyment in spending an hour being interacted with -- that is the actual primary driver of so much of this behavior. The product of the interaction is secondary at best.

It's interesting to see the influence of your influences. I had no idea that Moebius underlies what we accept as the visual language of scifi epics and anime. Akira, Naussica, Dune, Blade Runner, Star Wars, just to name a few. But I think we need a new vision of the future.

I wrote about Claude Code, which to my absolute astonishment is quite good at theorem proving. For people who don't know theorem proving, this is like spending your whole life building F1 engines and getting lapped by a Tesco's shopping trolley www.galois.com/articles/cla...

What's Algebraic about Algebraic Effects? Most descriptions explain them as "resumable exceptions", but never really tackle the "algebraic" part. After looking into this for a couple weeks last spring, my own answer is: composition with structure 👇

OpenAI with chartjunk again. The first bar adds up to 5.9%, not 6%, and the x-axis categories adds up to 100.1%. The widths are suppose to adjust for the % in each category, to make the areas indicative of overall fraction, but well-known that areas are hard to eye-ball compare x.com/omooretweet...

I have a tendency to think Enshittification is the result of Seeing like a State. If Clay Christensen's framework of Disruptive/Sustaining/Efficiency innovation aligns with reality, we have too many people doing Efficiency Innovation. www.cubbyathome.com/boxed-cake-...

I finally updated the "about" page on my blog. I can now say this is the scope of the stuff that I'll write about. I too, want my blog to be "a very long and complex search query to find fascinating people and make them route interesting stuff to my inbox" What I write about: 👇

How come we tend to think of point to point messaging as green?

Man, @Aron_Adler was the only person to tell me I had spinach in my teeth. And we don't even know each other! Anyway, if you had thought about subscribing to the blog before and was thwarted, please try again. interjectedfuture.com/ Thanks!

Spotify PMs: What good does a podcast transcript do me if I can't search it and I can't copy and paste it into something that can search?

"@physical_int robotic models can work given only 1 second context length" That's surprising and unintuitive to me as well. I feel like we're finding out the principles of intelligence without imitation, just like we found out principles of flight without flapping wings. x.com/dwarkesh_sp...

There's something about the @daltonc + @mwseibel comedy duo that makes me want to be in the same room. Probably the same thing that draws people to want to have their own Friends or Sex in the City clique. But I like candor and tackling of broad topics. www.youtube.com/watch?v=Cty...

Really enjoyed the post about how there's a lack of art direction in current apps to make the apps where you spend the most time to feel like a place rather than a utility. But I think this can easily be misinterpreted. Where do you draw the line? Re: Microsoft Bob. x.com/danshipper/...

Bret Victor's got his work cut out for him. The world keeps waiting. sethpurcell.com/writing/scr...

Devs are divided on whether you can vibe code in production. The answer is yes. But lots don't know that it requires a lot of hidden scaffolding that no one talks about. Vibe coding in production is a task for experts, not beginners. So what did work? 👇

Dipping toes in Lean with Claude Code. I can ask it to write a tutorial for a proof I care about: that ZSets with add_merge forms an abelian group. Way better than the stock tutorial to maintain interest! And if I need the tutorial to have more detail, I can ask it to update!

Standing up for ideals will always come at a personal cost, one that we should be prepared to make if we’re at all cognizant of what we could lose and what was so hard-won.

This is honestly the worst thing to happen to computing in the age of AI. When LLMs output text, it's in a universal format that can be read by almost anyone. It helps commoditize them, which means values accrue elsewhere! www.anthropic.com/news/create...

I wonder if boutique coffee shops with excellent coffee should have a subscription business model, where working on a laptop needs a monthly subscription. You can work on your laptop a long as you want, the price of coffee is 20% off, w/ call booths.

From the author of Pandoc — "I'll go through the six features of Markdown that I think have created the most difficulties" Beyond Markdown: johnmacfarlane.net/beyond-markd...

Ever since I watched @martin.kleppmann.com's talk, I've felt the same way about databases. Web devs keep writing bug prone versions of transactions and incremental view maintenance in user land. x.com/nicbarkerag...

It's better to have rich people with a vision, than rich people without. At least those people make cool stuff. They're people with an artist's heart that doesn't live like one who make the world wealthier. x.com/dhh/status/...

It's been a long time since I'd visited YC. Lots changed, but some things stay the same. No matter the batch, YC alum say: - The batch experience was better when we did it. - The current batch founders are better than we were back then. - The current valuations are way too high.

Did you know your MacBook has a sensor that knows the exact angle of the screen hinge? It’s not exposed as a public API, but I figured out a way to read it and make it sound like an old wooden door.

A new Sarah Paine lecture just dropped on @dwarkesh_sp! She does talks on historical geopolitics on Russia and China. Been enjoying her talks and even attended two live. Recommend! I even applied concepts in her talks to LLM prompting. 😮 whaaa. 👇 www.youtube.com/watch?v=SdM...

Both @jungle.bsky.social and @seezatnap are at the forefront of discovering native possibilities of this medium--rather than knee-jerk anti-AI sentiments. Artists take note. x.com/seezatnap/s...

Fascinatingly, this was very popular on Mastodon. Many of them hadn’t seen it before even though it made the rounds on twitter and Bsky already. Shows me there’s info arbitrage available, as certain segments are bifurcating hard. bsky.app/profile/int...

Most eval stacks are overengineered: pipelines, dashboards, infra. But what if the simplest tools—markdown and a code agent—were enough? The core idea: Treat your eval as a document. Each step is written in plain language, then executed directly by an AI agent.

"Vibe Coding Done Right is actually an Expert’s Task." It turns out there are two models for vibe coding. First is for ephemeral apps. This is the second. www.stochasticlifestyle.com/a-guide-to-...

Garry actually said, "don't rawdog your prompts". x.com/lennysan/st... As it happens, @sridatta and I wrote an ebook on system evals! You can buy it here at forestfriends.tech

Finally got around to reading this. I didn't quite understand differential dataflow the first time around a couple of years back. Now, circling back to it, I'm a bit surprised that DD establishes a partial order on versions of data. materialize.com/blog/differ...

I want to really like Automerge, as I love what Ink and Switch does. But the docs leave lots to be desired. Organizationally, it's unclear what is porcelain and what is plumbing API. It's hard to find the exact class doc of common obj you work with, like Repo, Doc, & DocHandle.

This was the dream for people touting NFTs as transferable game assets 8 years ago. Wasn't possible then. But now, you should be able to generate in-game assets for NFTs with drastically less effort. x.com/dylan_ebert...

You can leverage AI to learn lots of skills if you first learn to hold it right. x.com/fofrAI/stat...

This is one of those things that would nerdsnipe me so hard, I have to stay away from it. I’d want to build one to run home automation, as if my home was a starship Enterprise. x.com/theahmadosm...

Not if AI gets good at customer development. Maybe AGI is not just science and math reasoning, but be able to discern and tease out human needs and wants. With that super power, they wouldn't build things that no one wants--the bane of many an entrepreneur. x.com/lulumeserve...

I enjoyed this. And what an infectious laugh. ❤️ "Reporter left speechless after witnessing Japan's new $70 million Maglev train in action at 310 mph"

Model evals are like tests to gauge how well a new model does across a range of general tasks. System evals are like tests for a specific task your app does. It's a way to quantify "what good looks like." We wrote an entire ebook on system evals if you want to check it out: forestfriends.tech

Elm should have had algebraic effects. The simple uni-direction data flow is great, until a single user action generates a message that needs to be handled with multiple side effects. This results in cognitive overhead when tracing control flow in these cases!

Trying out Lean, but my workhorse is @zeddotdev. Ran right into the problem of typing Unicode. GPT-5 thought for nearly 7mins to solve this problem. But that's more time than it usually gives to solve my architectural design problems that I pose it. Config is hard.

Really looked at Chrome's performance tool today. It's nice! Never got time before cuz at startups, the existential threat is PMF, rather than speed of app. That said, I lament how easy it is to make web apps slow. More web devs should talk about this than the latest frameworks.

Oh, that's exciting: Lean compiles to Javascript and WASM. That might be useful in writing distributed systems software for the browser. At least, that's the hunch.

I too think this quote about product differentiation is great. But it’s also fun to see the retro prediction. Turns out iPads dominate as a checkout register and a way to placate children at restaurants. I also didn’t predict this. x.com/lennysan/st...

Engineers will bifurcate as they always do when there’s a new viable layer. Most will come to live at the top layer and solve larger scoped problems. Of those that stay in the older layer, they make tools for the top layer and it will seem magical. bsky.app/profile/jas...

I too think this quote about product differentiation is great. But it’s also fun to see the retro prediction. Turns out iPads dominate as a checkout register and a way to placate children at restaurants. I also didn’t predict this. x.com/lennysan/st...

Was thinking that a touchpad or mouse should have better haptics feedback when browsing the web. For example, as confirmation that your action was noted by the web app and on its way. Or instead of spinners (or in addition to spinners), you feel gears clicking on the touchpad.

As more and more information presents itself in video format, I like the fact that I can use LLMs to extract that information quickly--like reading a web page. I was interested in why Disco Elysium decided to use a chat box side bar, rather than interstitial text.I asked Gemini!

While doing a cursory research on dialog systems, I came across how Katana Zero uses interruptions as a way to engage players in cut-scenes, and how text animations can help convey emotions. If chat is the UI for AI, then maybe we need some of that. www.youtube.com/watch?v=40G...