gmarik 🇺🇦 (@gmarik) header
gmarik 🇺🇦 (@gmarik) profile picture

trial and error

Ukraine
Joined December 2008
Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)Photo by gmarik 🇺🇦 (@gmarik)
gmarik 🇺🇦 retweeted
"You need 8 hrs" was reductionist sleep advice
115
66
21
2,742
583,174
gmarik 🇺🇦 retweeted
🎉 Introducing SiliconForest: a new open operating system built on Darwin and XNU, the open-source core of macOS, running natively on an M1 MacBook Pro. Here’s the interesting part: it runs unmodified Mac apps. This first version runs Apple’s Terminal and Activity Monitor, copied straight from a macOS install, plus the stock Mac build of Firefox. Working today on a 13" M1 MacBook Pro: GPU acceleration (Vulkan 1.4), hardware video decoding, Wi-Fi, sound, keyboard and trackpad, and the Touch Bar. It’s Apple’s open-source kernel, with the parts Apple doesn’t publish reimplemented from scratch. I’ve also brought my Wayland compositor, Chonkstep, over to Darwin, and the results so far are fantastic. The idea: a free, hackable OS that can run native Mac apps. It’s early. Big apps like Logic Pro and Photoshop need a lot more (Metal, for a start), but that’s the direction. It stands on great work: PureDarwin (which we fork), the Asahi Linux GPU and hardware drivers, Omacom Mesa, and the Finch frameworks project. Source is coming soon. Right now it runs on one machine, the 13" M1 MacBook Pro. If you’d like to help bring it to other M-series Macs, get in touch. github.com/chriskearney/sili…
118
221
47
2,066
126,183
gmarik 🇺🇦 retweeted
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap. Here’s how the Claude Code team thinks about verification: claude.dev/blog/how-the-clau…
373
897
262
14,949
1,822,112
gmarik 🇺🇦 retweeted
Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached. The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network. This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc. I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?
208
517
127
4,321
477,677
gmarik 🇺🇦 retweeted
Some folks asked me what this means (fair), so here goes. The core question is "in a replicated storage system with three copies, how much does repair time affect durability?" We start with the top line: a Markov model. Think about this like a casino game.
Sneaking some Markov models into my reInvent slides.
3
6
97
8,069
gmarik 🇺🇦 retweeted
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1,822
6,424
2,209
41,254
6,788,086
gmarik 🇺🇦 retweeted
SRE has changed a lot in 10 years. We're kicking off a special 2-part series with the minds behind "Site Reliability Engineering, 2nd Edition"! 📖 Tune into Part 1 (Season 7, Episode 5) now! 🎧 sre.google/prodcast/#season7… google.smh.re/5ikY
10
222
7
1,712
156,033
gmarik 🇺🇦 retweeted
The Nvidia CUDA moat is gone. Anthropic sees all compute as fungible and it only took Mythos 3 days to figure out how to run itself on AMD hardware.
133
634
294
9,122
1,288,633
gmarik 🇺🇦 retweeted
Starting to roll out today: Claude Whiteboards, Claude Sheets and Claude Data. Think a plan through on a board, build a tracker in a sheet, and tell a story with your data.
1,549
5,284
3,877
64,405
12,558,221
gmarik 🇺🇦 retweeted
The web rewrite is working. On a slow connection, the old X web took more than 22 seconds to show the first post on a profile. The new X web does it in 2.7 seconds. Here's how we did it, and how we're rolling it out one country at a time. blog.x.com/en/engineering/re…
292
363
164
5,061
1,071,945
gmarik 🇺🇦 retweeted
NEW: Apple is cutting employee AI spending budgets even as it pushes workers to use the technology more. Engineers have seen daily limits cut from $300 to $200, with further reductions to $100 planned. Other teams could see limits fall to $25. bloomberg.com/news/articles/…
334
279
366
4,472
2,138,623
gmarik 🇺🇦 retweeted
I just released eggshell! A bare metal Go unikernel, implementing the Linux ABI, capable of running unmodified POSIX compliant C. github.com/usbarmory/eggshel…
18
70
5
488
24,076
gmarik 🇺🇦 retweeted
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations. Basically one-shotted a full NN that’s better than training it
257
372
235
7,767
911,064
gmarik 🇺🇦 retweeted
(wild) predictions after today's MIT Weight-Space Programming work: 1. Model weights become a programming language that AI can understand, but humans can't. 2. Training and inference will merge. Like human brain, AI could continuously rewrite its own weights while thinking. 3. A new path to RSI: AIs will directly edit each other's weights, evolving together without backprop.
155
248
46
2,522
99,975
gmarik 🇺🇦 retweeted
Seeing this makes the end-state clearer to me: code that can't be read by humans, run on machines that humans can't access, via inaccessible APIs, trained on data that humans didn't create. LLMs are starting to get really good at manipulating mostly-opaque-to-humans code (see: their improved ability to reverse engineer machine code/assembly). Additionally, folks are accessing more-and-more of their LLMs in cloud-based environments without ever actually interacting with an LLM API (myself included). I think it's going to become increasingly likely that the end state is one in which you ask an agentic company to produce some result and it does it via [some system completely opaque to humans] and produces [some output in a format that humans are unable to interpret]. It makes me wonder if our current state of competing LLMs cross-training on each other will eventually become literally impossible (not just logistically so) when they become fully self-recursive and are producing output that even other LLMs can't understand, let alone humans. I think we're in something of a golden period here where humans are still capable of understanding how things work and how things are run (shell commands! python! node.js!) but it's going to quickly become irrelevant. We're already past the previous high-point where we were able to see reasoning traces for the top-level LLMs. I'm skeptical that open models will ever get to this point as the economics are working against them (I suspect that they will always need human-generated training materials in one form or another). But it does make me wonder if the upper-echelon companies will end up converging into this high-level "black box that produces outputs" singularity of sorts. This has been hitting home for me lately as I've been using Claude Projects in a cloud-based environment where I can't see transcripts, look at the machine it's running on, nor even know what commands it's running. I don't even know what languages its writing its code in. We're just a couple steps away from code being completely irrelevant.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
42
36
10
283
48,210
gmarik 🇺🇦 retweeted
Pando is a Git backend which can scale to very large monorepos. We took the hard road of continuing to use Git, which means we could enable Pando with no change to devex. Considering providing this to others, let me know ([email protected]) if you'd be interested.
We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git. Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage.
210
90
24
1,713
176,015
gmarik 🇺🇦 retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,367
7,254
5,755
55,910
10,092,936
gmarik 🇺🇦 retweeted
BREAKING: This morning’s cyberattack against supermarket fridges and freezers has now spread to Belgium, Sweden, Luxembourg and Estonia. The hackers increased the temperature up to 50C. Food worth millions of euros has been spoiled. Experts warn that such cyberattacks on supply chains and logistics can constitute a threat to national food security in many Western states.
897
3,056
979
12,734
2,159,635