fuzzy (@1ffuzzy) header
fuzzy (@1ffuzzy) profile picture

do android dream of electric sheep

Joined December 2024
Photo by fuzzy (@1ffuzzy)Photo by fuzzy (@1ffuzzy)Photo by fuzzy (@1ffuzzy)
if you value intelligence above all other human qualities, you’re gonna have a bad time
898
3,524
1,184
24,367
9,972,473
Experimenting with procedural generated words in @threejs
62
31
5
806
49,165
This is really 2019 pure science fiction.
20
82
8
2,019
238,965
Actual zero day vulnerability researcher and exploit developer here. Top-of-the-field skills across Browser full-chain, hypervisor VM escape, OS kernel and baseband on record since before AI existed. And a proud solo owner of a zero day intelligence company. Anthropic just dropped the mic on my industry. Where do I begin…………
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software. It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. anthropic.com/glasswing
65
36
5
739
247,924
Diffusion (stochastic SDE sampler): erratic Brownian trajectories zigzagging through noise. Flow Matching (deterministic ODE integrator): clean, straight-line paths to the data modes. Same start, radically different dynamics.
14
282
9
2,610
331,171
the malaria parasite has IRON CRYSTALS inside it that spin nonstop like in a washing machine. someone recently figured out they’re powered by the same chemical reaction that launches rockets into orbit. H2O2 decomposition. A PARASITE BUILT A ROCKET ENGINE
36
113
13
1,070
31,842
My dear front-end developers (and anyone who’s interested in the future of interfaces): I have crawled through depths of hell to bring you, for the foreseeable years, one of the more important foundational pieces of UI engineering (if not in implementation then certainly at least in concept): Fast, accurate and comprehensive userland text measurement algorithm in pure TypeScript, usable for laying out entire web pages without CSS, bypassing DOM measurements and reflow
1,308
7,985
2,663
64,316
24,606,488
Recently, I was teaching the Law of Large Numbers and diversification, so I decided to share this illustration here. As you average more and more independent, identically noisy copies of an image, the noise cancels out and the original emerges. Code in the comments.
46
197
16
2,521
243,039
fuzzy retweeted
Aint no way this is the same @bcherny that created claudecode.
34
64
15
2,003
336,550
Russian math is on another level. In a rural classroom, barefoot peasant children were solving—mentally—problems that would slow down today’s MIT PhDs and quant researchers.
35
152
5
2,238
110,886
A self-taught “high school dropout” @gabriel1 ends up at @OpenAI Researcher. The new “degree” is proof-of-work 💼 In 2026, anyone can learn to do AI research by starting small. I’m curating a mini tutorial series on the most fundamental but underrated frontier-lab P0 task: training stability 📉 I’ll use small toy setups to show: - what triggers the instability - how to diagnose it by reading the signals - the specific techniques to stabilize the run. If you can understand and solve the problem in the toy environment, you’re halfway to solving it at scale. 👉Read / follow along here: orchestra-research.com/intro… All the tutorials are built with @orch_research
10
32
2
392
24,539
fuzzy retweeted
Notice how software became 10x easier to code overnight and absolutely nothing changed? Fake ass industry 😭
251
321
71
13,578
538,097
fuzzy retweeted
software development in 2026 is going to require some to loosen up a little code doesn't have to be as perfectly crafted the way we did it pre-ai call it slop if you want, but if you're still demanding perfection on every pr while your competitors are shipping "slop" that works... you're fighting from a disadvantaged position shipping velocity matters more than perfection
769
107
551
1,883
1,367,579
playing with fluid particle simulation 🌊🖐️
112
806
52
10,295
235,794
The most capable individuals are learning and shipping 1000x faster using AI. I sat down with Gabriel Petersson (@gabriel1) - a high school dropout who self-taught Math & ML with AI and is now an AI Researcher at OpenAI. We deep dive into the most important traits to thrive in the age of infinite leverage: – How to learn hard skills insanely fast with AI – Dropping out in Sweden → America on the O-1 @extraordinary visa - How to get hired by the best when you are a nobody – How to turn prompting into a daily habit – Why America is still the launchpad for ambitious builders - Breaking down Gabriel's most iconic tweets Curiosity + Agency + AI will take you further than any legacy credential Timestamps: 01:12 - Swedish high-school dropout → OpenAI research scientist 03:04 - First startup & sales hacks: door-knocking, instant A/B tests 05:28 - Couch-life grind & learning to code under pressure 08:16 - Learning by doing: top-down vs bottom-up; what schools miss about AI 12:10 - Recursive learning with ChatGPT: ELI5, intuition, gap-filling 16:20 - Prompting habits, make AI your always-on tutor 22:16 - Research workflow: papers, code inserts; “shortcuts to foundations” 27:13 - Feedback obsession, elite code reviews, AI as reviewer 33:12 - Path to the US🇺🇸 Midjourney, getting the O-1 @extraordinary visa, shipping demos that signal skill 39:29 - bypass recruiters via proof of work, risk-free trials & academia hot takes 55:06 - why San Francisco; visas & how to move; final pep talk
136
371
82
3,443
1,304,820
this take is pure BS and misses how deep tech innovation actually works Ilya has a PhD in CS from Toronto under Geoff Hinton where he co-invented AlexNet & literally helped birth the modern DL revolution before founding OpenAI Adam has degrees in CS and Mathematics & built PyTorch during research internships at FAIR with some of the best systems researchers in the world the Cursor team are MIT grads who went through CSAIL & OpenAI’s accelerator before building their stack these aren’t people who just decided to do things and figured it out, they spent years building foundational knowledge in optimization theory & systems architecture & distributed computing before they had the domain expertise to even identify the right problems to solve the real insight is that credentials don’t matter but deep technical fluency absolutely does & that fluency comes from thousands of hours immersed in the mathematical foundations & implementation details whether that’s in a PhD program or grinding through papers and codebases on your own what separates great engineers from people who just ship code is understanding the loss landscape well enough to know when you’re stuck in a local minimum VS when you need to completely rethink your architecture you can’t build a novel neural architecture without understanding information theory & backpropagation from first principles & you can’t optimize distributed training without reasoning from the ground up about communication overhead & gradient synchronization YES the path doesn’t matter but the depth does & there’s no shortcut to internalizing how systems actually compose
The creator of GPT doesn’t have a PhD. The creator of PyTorch doesn’t have a PhD. The research lead at Cursor dropped out of NEU. You don’t need a PhD or a top school to become a great researcher or engineer. You can just do things!
97
474
29
4,651
478,577
Andrej Karpathy says you should learn AI depthwise, not breadthwise. Most education is breadthwise: watch lectures, memorize formulas, and trust you'll need it later. Karpathy flips this by learning "depthwise, on demand." What this means: Pick a project, start building, and learn exactly when you hit a wall. When he created a tutorial on transformers (the architecture behind ChatGPT), he didn't start by explaining attention mechanisms or complex architectures. Instead, he started with the simplest possible thing: a lookup table that predicts the next word. You build that first. Then you try to make it handle more complex patterns. And it breaks. Only then, when you've felt the limitation, does he introduce the next concept. Each piece solves a problem you've actually encountered. As he puts it: "It's a dick move to present the solution before I give you a shot to try it yourself." When you attempt the problem first, the solution actually makes sense. Teaching forces you to learn. "If I don't really understand something, I can't explain it." When you try to explain and stumble, you've found the gaps in your understanding. ... Build a project that gives you a reward. Hit a wall. Learn just enough to solve it. Then explain it to someone else. Don't consume content. Build the code. That's how you actually learn.
66
562
47
4,908
404,167
Excited to release new repo: nanochat! (it's among the most unhinged I've written). Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single, dependency-minimal codebase. You boot up a cloud GPU box, run a single script and in as little as 4 hours later you can talk to your own LLM in a ChatGPT-like web UI. It weighs ~8,000 lines of imo quite clean code to: - Train the tokenizer using a new Rust implementation - Pretrain a Transformer LLM on FineWeb, evaluate CORE score across a number of metrics - Midtrain on user-assistant conversations from SmolTalk, multiple choice questions, tool use. - SFT, evaluate the chat model on world knowledge multiple choice (ARC-E/C, MMLU), math (GSM8K), code (HumanEval) - RL the model optionally on GSM8K with "GRPO" - Efficient inference the model in an Engine with KV cache, simple prefill/decode, tool use (Python interpreter in a lightweight sandbox), talk to it over CLI or ChatGPT-like WebUI. - Write a single markdown report card, summarizing and gamifying the whole thing. Even for as low as ~$100 in cost (~4 hours on an 8XH100 node), you can train a little ChatGPT clone that you can kind of talk to, and which can write stories/poems, answer simple questions. About ~12 hours surpasses GPT-2 CORE metric. As you further scale up towards ~$1000 (~41.6 hours of training), it quickly becomes a lot more coherent and can solve simple math/code problems and take multiple choice tests. E.g. a depth 30 model trained for 24 hours (this is about equal to FLOPs of GPT-3 Small 125M and 1/1000th of GPT-3) gets into 40s on MMLU and 70s on ARC-Easy, 20s on GSM8K, etc. My goal is to get the full "strong baseline" stack into one cohesive, minimal, readable, hackable, maximally forkable repo. nanochat will be the capstone project of LLM101n (which is still being developed). I think it also has potential to grow into a research harness, or a benchmark, similar to nanoGPT before it. It is by no means finished, tuned or optimized (actually I think there's likely quite a bit of low-hanging fruit), but I think it's at a place where the overall skeleton is ok enough that it can go up on GitHub where all the parts of it can be improved. Link to repo and a detailed walkthrough of the nanochat speedrun is in the reply.
669
3,305
731
23,898
5,824,234