Can LLMs simply tell us about unwanted behaviors they’ve picked up in training? We train a single Introspection Adapter (IA) that makes fine-tuned models describe their behaviors. It generalizes to detecting hidden misalignment, backdoors and safeguard removal.

NVIDIA has solved the biggest trade-off in LLMs. And it delivers a 6x speed boost without losing a single point of quality. Every AI you use today (GPT-4, Claude, Gemini) is "Autoregressive." This means the model is forced to think in a straight line, one token at a time, left-to-right. It’s like a genius writer who can only type with one finger. The hardware under the hood, your massive GPU, is actually sitting idle 90% of the time, waiting for that one finger to hit the next key. NVIDIA published a paper that changes the math. They figured out how to make the AI do two things at once in a single forward pass. 1. The "Talk" (AR): The model handles the immediate next word with perfect logical precision. 2. The "Think" (Diffusion): While it's talking, it uses its "idle" brainpower to parallel-draft the next 10–20 words in advance. It’s a hybrid brain. The results are a massive wake-up call for the industry: - 6x Speedup: It delivers nearly 600% more tokens per second than standard models. - Zero Quality Loss: Unlike previous "fast" models that get "blurry" or hallucinate, TiDAR matches the quality of the world’s best LLMs. - GPU Efficiency: It finally stops wasting the expensive compute power big tech is burning billions on. We’ve spent years trying to make AI smarter by making it bigger. But this paper proves that the real bottleneck wasn't the size of the brain, it was how the brain was scheduled. Paper: TiDAR - Think in Diffusion, Talk in Autoregression, 2025

Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature. Read the paper:

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.

New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

if you reply, i’ll follow you. looking for more people building with ai/llms: agents, evals, rag, coding tools, automation, weird workflows, model routing, infra, anything useful. small accounts welcome.