iamwil

What happened to those retro frontier models only trained on text pre-1900s? Let’s see if they can make discoveries in chemistry and discover quantum physics as a possibility. Where do they miss? Are there alternative tracks to the tech tree that we missed? x.com/emollick/st...

What have I been saying ya'll. Listen to @danluu danluu.com/ambitious/

x.comWil Chung (@iamwil) on XSo for software engineers that delighted in syntax and arcana, sorry that world was simply a necessary evil that wasted everyone's time, and is (rightfully) going away. But now that we potentially…

x.com/iamwil/stat...

x.comWil Chung (@iamwil) on XBut most awesome is the perspective shift. @_chenglou uses new engineering techniques that were simply impossible/irrelevant w/o agents, and out of anyone I follow, lives in the future of engineer…

x.com/iamwil/stat...

What have I been saying ya'll. Listen to @danluu danluu.com/ambitious/

x.comWil Chung (@iamwil) on XThe most interesting part of this is actually how he leveraged AI to do the tedious manual work of verifying the Typescript code against what the browser would report. Coding agents should make u…

x.com/iamwil/stat...

x.comHeatbit (@heatbit_com) on XIntroducing Bitwater: the water heater element that mines Bitcoin. It screws into the electric tank you already have. Physically decentralized mining that runs no matter what 🧵 Heat water. Min…

Honestly, I'd been thinking about how water heaters should have an API and should be more accessible to in-home local agents. So it shouldn't have been a leap to make a joke about "someone should do this". But alas, the people actually making stuff beat me to the punch. x.com/heatbit_com...

If GPU and memory prices keep holding steady or keeps going up, finance people will start to invent financial instruments backed by hardware and trade them. But not if crypto people get there first, if they hadn't already. 😆

One crucial role of art in society is to help everyone process the emotions of change. It's therapy for an ever-changing world. Artists that long for influence and renown, but don't deeply understand the technology nor the medium are missing out on their chance to speak to all. x.com/anabology/s...

westernelectric.comOriginal Western Electric 300B Vacuum Tube — Western Electric - Maker of electron tubes and high fidelityThe legendary 300B triode is manufactured in the USA using original, 1938 specifications and historic Western Electric quality standards. Our latest manufacturing facility was built for continuous production of the 300B, housing a blend of modern tech and restored machinery. Learn more about purchas

I had no idea you could still buy vacuum tubes in this day and age. www.westernelectric.com/300b

x.comSolomon Hykes (@solomonstre) on XI used all caps in my prompt and got banned by Claude. AMA.

Waiting for someone to point out that “how your date treats the waiter” and “how they treat AI in chat” are the same test. x.com/solomonstre...

x.comanabology (@anabology) on XClaude isn't a great musician (yet) But this kind of video would be unfathomable to make even 6 months ago with previous LLMs "Claudia Explains: Euclid"

I’m a technology optimist. However, technology claiming to accelerate learning always gives me pause. I think this is great as a gateway to learning, but if you just stop there, it’s edutainment. There seems no way around struggling when actually learning. x.com/anabology/s...

names.color.pizzaColor Names31,935 colors, each named by a person: wander through them, hanging where their color is.

Clearing out tabs (5): So much random garbage in my tabs. With AI, I'm starting to see a Cambrian Explosion of fun custom websites--after the crunch of everyone having a medium/fb/substack. names.color.pizza/?view=pixel

Nice annotation tool to directly talk about what you're talking about with an agent. x.com/kitlangton/...

Previous: x.com/iamwil/stat...

names.color.pizzaColor Names31,935 colors, each named by a person: wander through them, hanging where their color is.

Not a neurologist, so an explainer helps. For me, it's the unspoken context around the original tweet thread (since I'm not the audience) that made it more difficult to understand. So having AI fill in that context for a lay person is helpful. x.com/Azadux/stat...

youtube.comExamining 'The Witness' Introduction: The Castle Garden (spoilers in subsequent episodes!)A lot of the online discussion about The Witness has focused excess...

I'm currently working my way through Order of the Sinking Star. When I played through The Witness, much of the underlying philosophy was lost on me. There is no meaning if there are no references. Again, filling in context. www.youtube.com/watch?v=6vh...

names.color.pizzaColor Names31,935 colors, each named by a person: wander through them, hanging where their color is.

Aside: best example I've read as an example of designing with types is "Build Systems a la Carte" paper. It's quite readable, so you should give it a shot. dl.acm.org/doi/pdf/10....

luma.comBond AI - San Francisco and Bay Area · Events CalendarView and follow events from Bond AI - San Francisco and Bay Area on Luma. Bond AI - the largest in-person AI events community (150k+ members) 📢 Add your event submit.bondcommunity.ai 📤 Subscribe to the weekly events email 🌎 Explore other cities bondcommunity.ai/#join

Engineers are notoriously hard to get on the phone, so it's easier to just go to meet ups. I finally found a pretty active calendar for agentic coding in the Bay Area. luma.com/genai-sf

names.color.pizzaColor Names31,935 colors, each named by a person: wander through them, hanging where their color is.

It's a skill tree for tiny humans. Many will immediately think of A Young Lady's Enchiridion, and I want to leverage it, but not exactly sure how yet. withmarble.com/curriculum/

youtube.comTypes Are Worth the Typing - Filip Sodic - NDC Oslo 2026This talk was recorded at NDC Oslo in Oslo, Norway. #ndcoslo #ndcco...

It's crazy to me that this is news. Types aren't just getting in your way that you parry with `any`, or for autocomplete. You can design with types, by encoding your domain knowledge. Classic declarative programming. www.youtube.com/watch?v=Vhj...

names.color.pizzaColor Names31,935 colors, each named by a person: wander through them, hanging where their color is.

already-clean.goodfire.comNeural Networks Are Already CleanA working hypothesis about neurons, generalization, and the ecology of circuits.

It's hard to have a counterargument against AI doomers since superintelligence is unbounded and hypothetical. Still working my way through this one, but it seems like good news for mech interp. Alignment might be achievable. already-clean.goodfire.com/

Things keep changing every 2 months when it comes to agentic coding, that scrap blog posts. Forum questions compel me to write something down in the open. Here are four points I can stand by today below: news.ycombinator.com/item?id=501...

Still getting the hang of building my own verifier. Importantly, verifiers get me out of the coding loop. When the agent can check for itself, I don't have to stay up at 2am begging it, "you fixed this, but broken that. pls fix."

I've been asking engineers if it's still important to understand the code they're shipping. The opinions are split for different reasons, and it's also a matter of degree. What's your take? Is it still important to understand what you're shipping?

I also do a form of tactical simulation game via @ejames_c where agents and I will discuss an explainer and discuss the pros and cons of options generated by subagents for the design decision under question.

Verifiers and Guardrails: Guardrails don't change with the feature, like typecheckers, linters, and pre-commit hooks. Verifiers are different for different features, such as unit/integration tests, formal method proofs, design verifiers.

Make the decisions that matter: Not every decision matters in a product. Agents are fast at writing code and pretty good at some of the nitty gritty details, so I let it make those decisions. I make decisions on architecture, approach, and stack that will help differentiate.

Generating explainers, quizzes, and simulations: Asking for explainers helps quite a bit, though I have a custom skill for it. Agents aren't great at it out of the box. I also have it quiz me to help retain the concepts I learned from explainers.

1. Interrogation/interview-based design and specification 2. Make the decisions that matter and leave the rest to the agent. 3. Use agent to generate explainers, quizzes, and simulations. 4. Use verifiers and guardrails in the coding loop.

Interview-based design and spec: When I tell the agent what I want to build or extend, there's naturally ambiguity. But the agent knows what is ambiguous to it, so it can ask me. Closing that specification gap is how you get better results from your agents.

x.comAlexey Fateev (@superalesha) on XMy RTX 3090 made a melancholic music video. Song, pictures, video, upscale. Not a single frame or note from a cloud model. About 11 hours of pure generation time, including failed takes. Video:…

Nice. I didn't think it was possible yet with local models. Frontiers can overshoot certain niches, which is good for us, I think. x.com/superalesha...

x.comTim Sweeney (@TimSweeneyEpic) on X@AndrewCMyers Yes, the models seem to go through phases: hallucinating with confidence, then sycophancy, then ClaudeSpeak, now a sort of gradual decoherence of consistency as conversations grow. T…

This reminds me that I don't think AI models will just get monotonically better across all domains. It also reminds me that system evals are important and I don't spend enough time doing what I preach. x.com/TimSweeneyE...

gist.github.comOptChat - HOW TO REPLICATE MY SETUPOptChat - HOW TO REPLICATE MY SETUP. GitHub Gist: instantly share code, notes, and snippets.

Clearing out tabs (4): Different way to store memory. Instead of chat history being linear, make it hierarchical, branched by summaries. It purports to work well. I haven't tried it yet. gist.github.com/VictorTaeli...

aaronsonoracle.comThe Aaronson OracleA program that guesses which key you'll press next, and is right far more often than chance.

An interactive explainer on simple predictors. It makes me wonder how branch predictors in CPUs work, and if they're anything like this....and apparently, they already are n-gram predictors! (but they don't do backoff, presumably due to latency requirements) aaronsonoracle.com/

Issue 3: x.com/iamwil/stat...

gist.github.comOptChat - HOW TO REPLICATE MY SETUPOptChat - HOW TO REPLICATE MY SETUP. GitHub Gist: instantly share code, notes, and snippets.

I disagree in that you can't learn and explore the state and design space at the spec writing phase. You can't explore every nook and cranny, it's not a silver bullet. But you can learn a surprising amount you couldn't before!

Gosh, this is so awesome, I wish I thought of it. I was investigating in that neighborhood too a couple years ago, but I was too blinded by the goal of incremental computation that I missed this possibility completely. x.com/_chenglou/s...

gist.github.comOptChat - HOW TO REPLICATE MY SETUPOptChat - HOW TO REPLICATE MY SETUP. GitHub Gist: instantly share code, notes, and snippets.

Let's be specific about what spec-driven dev is. I write a spec with an agent. But the spec is small enough to iterate, big enough to take advantage of how fast agents are and one-shot the feature. www.youtube.com/watch?v=F8f...

I also don't sync the specs. I keep an append-only log of specs, just like RFCs. How I work isn't waterfall (what Kent is thinking), nor is it what people commonly refer to now as Spec-driven. But I do agree what's important is the learning and understanding.

gist.github.comOptChat - HOW TO REPLICATE MY SETUPOptChat - HOW TO REPLICATE MY SETUP. GitHub Gist: instantly share code, notes, and snippets.

So then my question: do you actually need the implementation to come to the same conclusions? Before agents, it was unequivocally yes. Now, I'm not so sure. Some things you can get away simulating with the agent, or build a prototype without the impl. What are the conditions?

danwang.co2025 letter | Dan WangCorgis, compute, Cold War; Ecclesiastes; ties; Stendhal; humor; Pascal's Wager; deep infrastructure; Germanic obedience; Texas State Fair

Dan Wang's 2025 was a nice, if long read, mostly on the US, China, Europe comparisons on the industrial base and AI. It takes mostly a fair stab at each, and is a nice broad overview. However, I'm unclear what my own takeaway is, other than ambience. danwang.co/2025-letter/

gist.github.comOptChat - HOW TO REPLICATE MY SETUPOptChat - HOW TO REPLICATE MY SETUP. GitHub Gist: instantly share code, notes, and snippets.

github.comGitHub - elstongun/leviathan: **Deep memory for agents over large datasets.** Leviathan is a single static binary that turns your records (JSONL, JSON, CSV/TSV, SQLite, or anything a database CLI can export) into a ranked full-text index.**Deep memory for agents over large datasets.** Leviathan is a single static binary that turns your records (JSONL, JSON, CSV/TSV, SQLite, or anything a database CLI can export) into a ranked ful...

Haven't tried it, but I need to swing back to working the search feature. This library is for deep memory search for agents over large datasets. If someone tries it and can give a review, I'd love that. github.com/elstongun/l...

registerspill.thorstenball.comBuilding Software Is LearningAn internal note to the Amp team on feedback and shipping faster

I found this on some thread where people were making fun of using specs for agentic coding. Even though I write specs with agents, this is rather similar to how I think I'm working: exploring the state space and design space. registerspill.thorstenball.com/p/building-...