Dylan Martin retweeted
Can an image generator learn its own hierarchy of abstractions without a separately trained encoder?
Meet HiT. Trained end to end from pixels, it learns to dynamically group patches across multiple levels—and semantic structure emerges without masks or region labels.
Check our video below: background collapses into a few groups, the main subject gets many, and words light up the regions they refer to!
It also gets 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and outperforms models with 6× more parameters on average across text-to-image benchmarks.
Website, paper, and code below, followed by how it works and what emerges 🧵
Dylan Martin retweeted
PyBot: No weights. No training.
A chatbot built with only simple rules and no AI, entirely written in 600k lines of Python code. PyBot matches Smollm with 360M parameters!
How? A swarm of 1,000 DeepSeek Flash agents wrote the whole code in just a week.
github.com/huggingface/pybot
Dylan Martin retweeted
I only had to wait 5 years for SDF + splats to become trendy 😅
Here's pixel-Art gaussian splats before they were cool ✌️
🍍#clayxels v1.6 is out on the @AssetStore with:
a whole new way of rendering splats without polygons,
faster compute clay,
direct viewport picking,
noise primitive,
clay-groups,
M1 Mac support🍎
Dylan Martin retweeted
a 364K param system-one model can sit in the SRAM of your SSD and protect it from ransomware. (88 KB)
here's how:
Dylan Martin retweeted
樹や花の3Dモデルを簡単に作れるツール「自然物ガチャ」を作成しました!!
無料で使えて、微調整も簡単です。商用利用もok。
ぜひ使ってみてください!!
tools.oyasumi-gamedev.com/sh…
架空の宇宙、空を数式で表現する「空ガチャ」を公開しました!!
無料で使えて漫画のトーンにも変換できます。
様々な星空、豊富な種類の雲が流れゆくさまをお楽しみください!!登録不要で商用利用もok。
是非使ってみてください!
tools.oyasumi-gamedev.com/so…
Dylan Martin retweeted
シミュレーションでマテリアルを作成する、新感覚のテクスチャ作成ツールを開発しました!!
tools.oyasumi-gamedev.com/to…
その名も「トーン・マテリアルガチャ」!!
物の成り立ちを再現するツールです!!
無料で使えて、ガチャで無限に近いバリエーションを試せます。しかもパラメーターで微調整も可能。漫画のトーンにも変換できます。
■何がすごいのか?
従来のマテリアル作成ソフトは、図形の変形で質感を作っていました。
このソフトでは、岩は地層から、樹皮は樹の育ち方をシュミレーションして、実物と同じ仕組みでマテリアルを作成することができます。(タイルなどは開発中)
商用利用も可能です。ぜひ使ってみてください!!
樹や花の3Dモデルを簡単に作れるツール「自然物ガチャ」を作成しました!!
無料で使えて、微調整も簡単です。商用利用もok。
ぜひ使ってみてください!!
tools.oyasumi-gamedev.com/sh…
Dylan Martin 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.
Dylan Martin retweeted
i built resnet-18 in pure assembly.
11.7m params. zero libc calls. im2col, the matmul kernel, batchnorm folding, pooling, softmax, even exp(): all arm64 asm.
~45 ms per image on a single m1 core. 8 kb of machine code.
top-5 matches pytorch on all 30 test images.
Dylan Martin retweeted
LLMs are far too slow. It feels like the early days of the internet when you'd wait a full minute to download a megabyte of data.
Dylan Martin retweeted
Today we release Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass.
It can do all that in 74 ms on Samsung Galaxy S26 NPU or 107 ms on its CPU.
Mirage has a 7.8M encoder + a 40.8M decoder at 2bits, and describes images in English, German, French, Spanish, Italian, Dutch and Polish.
We trade general language capacity, to compete top small VLMs today (SmolVLM2-500M, LFM2.5-450M and Florence2-230m) on these tasks.
On 256 VQAv2 questions, feeding Mirage's structured output and full trace to text-only Qwen3-0.6B raised its score from 9.3 to 31.3 out of 100, without fine-tuning either model.
Read more and benchmark it yourself here:
cactuscompute.com/blog/mirag…
Dylan Martin retweeted
human-in-the-loop
└ human-on-the-loop
└ human-near-the-loop
└ human-cc'd-on-the-loop
└ human-legally-liable-for-the-loop
Dylan Martin retweeted
1/4
New writeup: you can fingerprint the exact GPU you're running on with a default torch.index_add_ call, even among identical GPUs.
This works because floating-point maths depends on order, and every GPU has its own. This fingerprint survives reboots and can't be faked.
Dylan Martin retweeted
you need to be working on something so frontier that the models are still dumb
Dylan Martin retweeted
Software around me is not improving nor is it expanding.
If AI drastically lower the cost of software… then we have a mystery, don’t we?
The answer is that building the software was never the difficult part in these cases. Our organizations have been optimized for process, meetings, and risk avoidance instead of shipping working systems. It is institutional decay: layers of people whose main job is to prevent other people from doing the work.
Dylan Martin retweeted
🚀一张 3090,把千问3.8-27B 在 vLLM 上开到 64 路并发,合计破一千 tok/s
开源项目 HyperQwen:单卡 24G,3090 限 250W 实测。
单人聊天约 127 tok/s,64 并发合计约 1035 tok/s,最长可开到 15 万上下文。
两条命令就能起:docker compose --profile single 或 batch。
视频是官方录的:先看单路吐字,再切 64 路一起跑。
温馨提示:第一次会拉约 9.5GB 镜像,再本地 requant 一次模型大约 20GB,等完再测速度。
你家 24G 卡跑千问,单路现在大概多少 tok/s?
来源:GitHub syv-ai/HyperQwen👇
github.com/syv-ai/HyperQwen























