Author of GRPO-Zero library, PhD, AlphaZero in JAX, text-to-speech models, mechanistic interpretability, exploring neural nets' internal circuits...

Singapore
Joined February 2012
My weekend project: GRPO:Zero. An implementation of the GRPO algorithm from scratch, with zero dependencies. Say goodbye to unnecessary abstractions, every line of code is visible and everything is customizable. Runs on a single A40 GPU (48GB VRAM) for a few hours to achieve good results. It'll cost you only a few dollars.
1
10
1,584
Nice work on J-space! Fun coincidence, I explored a very similar technique called logit prisms two years ago: neuralblog.github.io/logit-p… @Jack_W_Lindsey might be worth a look, and a citation would mean a lot if it fits!
Replying to @AnthropicAI
The J-space lets us read, audit, and shape what Claude is actively thinking about—useful tools for keeping models trustworthy as they grow more capable. And it suggests surprising parallels between language models and our own minds. Read the full paper: transformer-circuits.pub/202…
2
273
Anthropic should slow down and fix its services.
working on a fix for issues with Opus 4.7 and Opus 4.8 status.claude.com/
99
ok i really like this model architecture, which somewhat proves my intuition that we don't really need a deep encoder for vision/audio modalities, the transformer backbone itself can process the image/audio tokens raw. The vision/audio encoder should be considered to play the same role as the "embedding" layer.
Meet Gemma 4 12B! A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license. Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇
152
Another day of Google not quite reaching the AI capability frontier.
40
They don't teach you this in school, but Transformers are RNNs. The KV cache is the hidden state.
About "transformers are recurrent": teacher-forced discrete tokens cannot play a mind's melody.
102
Cancel your subscription and switch to Codex instead. You'll be doing others a favor too, freeing up compute for the people who actually need it.
Best AI models now : Backend : GPT-5.5 Frontend : Kimi-K2.6 Agentic : MiMo-V2.5-Pro Best Price : Deepseek-V4-Flash Worst Price : Opus-4.7 Fast Research : Grok 4.3 Image : GPT-Image-2 Video : Seedance 2.0 (#1 for 3 months) Yes, Claude sucks.
128
I am kind of tired of people's crazy takes on Anthropic. AI twitter is an upside-down world in which the honest people get ridiculed while the liars get celebrated.
53
Too much of the *hype* around Mythos is focused on cyber. But… Mythos is incredibly good at SWE as well. In fact, cyber is really just a side quest. I'd bet Mythos is great at other tasks too. And I don't think we're prepared for this.
83
I built a Claude Code plugin this weekend: turn any "chase a number" problem into a stateful hill-climbing loop. You give a metric and a verifier script. Claude does the iterating. State lives in a self-contained HTML dashboard. 🧵 below ↓
1
110
The dashboard is one self-contained HTML file. State + chart in one file you can email around. Here is a live example, mid-run on ResNet-18 / Flowers-102 fine-tune: ntt123.github.io/hill-climbi…
1
203
I think looking at neural activity through a geometric lens is very promising. Here's another (simple) example from my earlier Logit Prisms work: when we look at digit embeddings inside an LLM, they form a nice heart-like <3 shape in 2D PCA.
Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵
1
1
209
One interesting thing I learned from @dwarkesh_sp and @reinerpope's recent lecture that few people mention is that the 1m and 1h caches are likely stored on Flash and HDD. That was a bit of a surprise to me. I thought the 1m cache was stored in HBM, which turns out to be wrong.
1
3
177
TT Nguyen retweeted
Leonardo DiCaprio is urging his 60 million Instagram followers to contact members of the U.S. House of Representatives and urge them to vote “no” on a bill that seeks to dismantle the Endangered Species Act (ESA). “Today, on Earth Day of all days, the U.S. House of Representatives is set to vote to gut the Endangered Species Act (ESA), trading the future of life on Earth for the short-term economic gain of a wealthy few... Today’s bill would devastate the most vulnerable species, which are essential to functioning ecosystems. I urge the House to reject this existential threat to our national security and choose to protect and recover species, and defend the living systems we all depend on. Call your U.S. Representative today to urge them to vote ‘no’ on the ESA Amendments Act of 2025 (H.R. 1897). U.S. Capitol switchboard: 202-224-3121.” variety.com/2026/film/news/l…
287
4,052
216
22,653
1,273,105
Now what I really need is Mythos as the advisor for my Opus😅
We're bringing the advisor strategy to the Claude Platform. Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost.
124
I like the new Claude Code 'advisor' feature already. It's always nice to have an advisor and/or a second opinion on the task that Claude is doing. Well done 👍
62
strongly recommend this book
Reminder that I wrote a little book about deep-learning, which is phone-formatted, entirely free, and nearing the 1M download: fleuret.org/francois/lbdl.ht…
1
81