Bourke Floyd, IV

Reading #runnindownadream by @bgurley with my 17yr daughter then 9yr old son starting this weekend 📖 My kids dream big already 👾 I’m hopeful this will help navigate some of the world ahead of them ❤️🚀

I have created github.com/unclebob/AIR-J so you can see what's going on with this language. I have no idea where this will lead, but I'm having one hell of a good time getting wherever that might be.

It blows my mind that codex built an AIR-J (codex chose that name) compiler from scratch that compiles down to JVM bytecode and interops with the JVM libraries. At some point I'm going to have to run some benchmarks on token consumption between Clojure and AIR-J.

@travisk Great vision 🚀 GL 👍

Atoms. atoms.co/vision

@SeloSlav Find the fun core loop. Make that great everything else is gravy 👍👾

Selo Empire is coming along... Should this be a cozy farming sim/city builder or an real-time strategy game with warfare? Maybe I'll just build out the engine first and let players decide. Would you play this?

My 9 year old son Charles V decided to make @OpenAI voice count one to a million for fun. Reminds me of @MrBeast challenges 😆

@unclebobmartin In the age of AI this is technically possible now 😆😎

I am simultaneously working on: 1. A simple wiki demonstration. 2. Improvements to the Empire game. 3. An AI only language for the JVM. All testing, crapping, and mutating.

@unclebobmartin That is a genuine ask to define it @grok or you’ll be decimated

Why doesn’t anyone know what decimate means?

@unitygames Link to MCP please 🙏😎

Our Unity AI Beta will roll out with integration with the Profiler, so you can now analyze captures with AI to understand what's causing performance issues in your game.

@chefgoyardi Well done 👍

We built PudgyWorld from scratch. No traditional game engine. No downloads. Just open your browser and you're in. Here's what it took to get here: We created custom world-building tools using open-source web technology, giving us a lightweight editor built for speed and rapid iteration. Our asset pipeline lets artists work in Maya, Cinema4D, or Blender while custom Houdini scripts automatically convert everything into a web-optimized format. Creative freedom without compromise. We engineered physics specifically for the browser. Snappy movement, parkour, fluid navigation, and high frame rates even on lower-end devices. Our intelligent camera system follows the player automatically, making the experience intuitive for everyone, including people who have never picked up a game before. It detects collisions, prevents clipping, and shifts focus during key moments. We developed the "Pudgy" aesthetic using pre-computed shadows and custom outlines to nail the iconic look while keeping GPU draw calls low. No real-time lighting overhead. Just a premium, polished feel. Every visual detail, from interactive snow to flowing water, runs on custom shaders and dynamic particles built to load instantly on the web. Multiplayer is live. Hundreds of players can share the same space in real time, express themselves through emotes, and yes, throw snowballs at each other. And it all scales. Whether you're on a Mac Studio or an old laptop, the experience stays fluid and inclusive. This is what it takes to build a world that anyone can walk into. No barriers. No installs. Just play. Excited for you guys to try it out.

@TimSweeneyEpic Kind of hilarious that I completely randomly ran into you at GDC on the first day at B last night 🚀 Not expecting that at all 😆

@TimSweeneyEpic Will you make it? 👾

This is the week of the Game Developers Conference in San Francisco. Since lots of game makers are talking about the Epic v Google settlement, here is the (partially redacted) term sheet from the public court filing: cdn2.unrealengine.com/epic-v--google…

@onirenaud Funny was doing the same last week but with terrain first. Working pretty well so far. I think we will want a open source streamable approach for the web tbh because will become mainstream at some point

After porting the Spawn renderer from WebGL to WebGPU, we decided to implement Nanite for AI-generated models. The first step is almost done: hierarchical clustered LOD based on NVIDIA vk_lod_clusters for meshlets. Currently too slow for sparse geometries. Continuing tomorrow.

@unclebobmartin 💯

AI agents have vastly changed the build vs buy calculus. The vast majority of tools can be built at virtually no cost.

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

@OfficialLoganK Also #gdc 👾

Going to be a fun week of launches : )

@VitalikButerin @0xMiden @zengjiajun_eth tbh it’s a little embarrassing how long it took to take privacy seriously for these primary chains. Still work todo but look forward to actual privacy besides anon-ish 👍

Crypto privacy is needed if you want to make API calls without compromising the information of your access patterns. eg. even with a local AI agent, you can learn a lot about what someone is doing if you see all of their search engine calls first-order solution to that is to make those calls through mixnet but then (or in fact, even without the mixnet) the providers will get DoSed, and they will demand an anti-DoS mechanism, and realistically payment per call by default that will be credit card or some corposlop "yeah we'll get to the privacy later" stablecoin thing so we need crypto privacy But yes, for privacy you have to think full stack. Local AI agent layer is very important. It is like longevity: if there are 10 things damaging your body, curing one of them increases your longevity by 11%, curing two by 25%, and curing three by 42% (1 / (1 - 0.3) minus 100% base). Risks from data leakage are similar, and so mitigations similarly compound super-additively.

Few people care about crypto privacy. But everyone I talk to care about their ai agent privacy and safety. That’s a huge reason why Mac mini are sold out everywhere.

@altcap @TrumpAccounts @altcap any use of this for a kid turning 18 in a few months? Seems not to me

The number of kids by state, year of birth and $$ amount that have waiting for them in their @TrumpAccounts. Kids over 10 don’t get seed $$ but still get a free investment acct that parents & business will likely add to if they claim it. 🇺🇸🚀

@SteveLovesAmmo @LimitingThe Cool idea 💡 I would do it 😎👍

Dear Elon and Nikita, Is there a way we can verify Veterans on X with a green checkmark? ID.me could work. I have an active account there. When world conflicts happen, it would be prudent for service members that have served, especially in the Middle East to have a higher voice and impact here on X. .@elonmusk .@nikitabier