Alessio Serra

A dream is what you think about before you fall asleep in your bed. A project is what you think about in the morning when you wake up to plan your actions. Don’t follow your dreams, build projects.

🧵 [1/14]: Talking-Heads Attention by @NoamShazeer et al. showed something interesting: maybe attention heads shouldn’t be fully isolated. 🧠 That got us thinking: If communication across heads matters, what is the right way for heads to communicate, especially from a one-layer reasoning perspective? 🔗⚙️ That question led us to Interleaved Head Attention (IHA) ✨ 📄 Paper link: arxiv.org/pdf/2602.21371

If your evals are bad, check your whitespace

guess who’s back

anthropic is winning the thinking game

there are only two options: we either arrive first or arrive first by a wide margin.

@paradigmainc are you sure you're only a 4 person team?

Introducing flywheel skills: a set of commands that let researchers (and their models) drive Flywheel programmatically, with research best practices baked in.

Yet another “near-frontier” model from the Amazon of China. Not Qwen. Not Alibaba.

@karpathy who’s implementing this over the easter break?

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

@willccbb is this how openai-like confusing naming begins?

our internal code name for INTELLECT-4 is INTELLECT-5

italian elegance meets autonomous research

Here is Qwen3.6-Plus: qwen.ai/blog?id=qwen3.6

Today, we release LFM2.5-350M. Agentic loops at 350M parameters. A 350M model trained for reliable data extraction and tool use, where models at this scale typically struggle. <500MB when quantized, built for environments where compute, memory, and latency are constrained. 🧵

A new age of research is ahead of us. Some reflections by @tensorqt on the vision that motivated us to power it. Link to the blogpost in the first reply.

We are releasing a detailed tech report for the tokenizer-free model family first introduced at ICLR 2025. These models achieve a significantly higher compression rate than standard LLMs, especially beyond English. 🧵

We made Muon run up to 2x faster for free! Introducing Gram Newton-Schulz: a mathematically equivalent but computationally faster Newton-Schulz algorithm for polar decomposition. Gram Newton-Schulz rewrites Newton-Schulz such that instead of iterating on the expensive rectangular X matrix, we iterate on the small, square, symmetric XX^T Gram matrix to reduce FLOPs. This allows us to make more use of fast symmetric GEMM kernels on Hopper and Blackwell, halving the FLOPs of each of those GEMMs. Gram Newton-Schulz is a drop-in replacement of Newton-Schulz for your Muon use case: we see validation perplexity preserved within 0.01, and share our (long!) journey stabilizing this algorithm and ensuring that training quality is preserved above all else. This was a super fun project with @noahamsel, @berlinchen, and @tri_dao that spanned theory, numerical analysis, and ML systems! Blog and codebase linked below 🧵

Interesting paper

We are introducing EU Inc. To make building and growing a business across the EU faster, simpler, and smarter. 🔸 Start a company in less than 48 hours 🔸 No minimum capital requirement 🔸 Fully online and borderless

Found another great midtraining paper. I haven't seen it on my TL so thought I would share. Super excited to dig into it later but looks really promising. (ty @lukemerrick_ ) I love seeing more work unifying understanding of midtraining -> RL

@tensorqt Flywheel challenge when?

you're not gonna believe what the next feature in Flywheel will be

@OpenAINewsroom @astral_sh So this is how astral makes money news.ycombinator.com/item?id=443582…

We've reached an agreement to acquire Astral. After we close, OpenAI plans for @astral_sh to join our Codex team, with a continued focus on building great tools and advancing the shared mission of making developers more productive. openai.com/index/openai-t…