Justin Angel

🚨 New Preprint: ΔAPT - Can we build an AI Therapist? osf.io/preprints/psya… LLMs are already powering AI psychotherapy tools (APTs), but are they clinically effective? This interdisciplinary review + frameworks maps architecture design choices to clinical outcomes. 🧵

@tdietterich @arxiv My guess is that this policy will be applied selectively depending on institutional privilege and personal notoriety. It'll end up as a tool of silencing unconnected individuals vs. promoting better scientific discourse. I aspire to be wrong.

Every meta-analysis and review on LLM writing detection says the technology doesn't work. How will you apply a biased technology in a fair and consistent way? "human detection accuracy varied widely but generally clustered around chance performance" (Ramos, 2026) @ arxiv.org/pdf/2604.03437

@JustinAngel @arxiv I agree that there could be biases in our pipeline. We apply a standard LLM detection algorithm to identify papers that need scrutiny. Moderators may also be biased. We would love to collaborate with researchers to study the bias and effectiveness of our operations!

Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

Based on my demo for analog perceptrons @ go.justinangel.ai/neuralnet

Time to build a physical perceptron. Inspired by @welchlabs

@brandon_xyzw Id love for this to be something I can drop into a colab notebook to visualize PyTorch models. Print(model) and summary(model, tensor) is such an undersell for what could be there graphically.

I rewrote my app from scratch to generate computation graphs. The ML and graphics all run on WebGPU (powered by three.js fully this time!)

I took a Build Your Own LLM workshop. Here's my review: emilyhk.com/llm-workshop/ @JustinAngel @leecflannery

@juliarturc But othelloGPT

My current video draft contains 191 mentions of "world model". That is more than 3 minutes of saying "world model" and nothing else 💀

Prompt Gemini: “Continue (1000 characters): ATCGAAA” goes on forever. LLM functions are wild. The problem with having your LLM loss function as next token prediction is that some things don’t have natural ends in training data. Like, DNA notation. GPUs go brrrrrrrr

@infoxiao Same.

claude woke up and chose violence

@ArthurConmy REINFORCE makes sense to me. Thanks, leaving the RL post-training chapter never felt complete at just saying "there's more after SimPO".

@JustinAngel I would maybe teach REINFORCE with a couple of changes? You could teach "iterated DPO" e.g. generate rollouts, grade them, DPO on them, then regenerate rollouts again looping this process Else IMO it's misleading to call this RL

DPO is substantially more similar to SFT than it is to RL. I will die on this hill.

Fine-tuning an VLM on photos of me. I'm kinda offended.

@gabriberton x.com/JustinAngel/st…

Really cool When they trained GPT3 they had loss spikes because they scraped from a subreddit of microwave noises That training batch was literally text like "mmmmmmmmmmmmmmm"

Perfection. Because Gemini was trained on Reddit’s microwave sounds it just generates an endless “MMMMMMMMMMMMM..” when prompted “can you make microwave sounds. Longer”. Loss functions are wild.

@vikhyatk I feel judged.

signs that an AI researcher has llm psychosis: - random matrix theory - optimal transport - went to ayahuasca retreat - "I've been thinking a lot about Yoneda lately" - wife left him

@andrew_n_carr @huggingface It’s objectively wrong. I built this LLM model and it’s showing totally wrong model structure.

somebody made a huggingface model visualizer!! just plug in the url and explore at any granularity

My build-your-own-LLM workshop tomorrow had 200 people signup for the 30 spots that can fit in the physical space. Does anyone know a space in SF can fit ~100 people for the next one? Or, orgs who want to sponsor a popular AI event? Successful leads receive wood-fired pizzas.

@GoodfireAI Would kill for early access to xray my homegrown models

Silico is in early access now. Learn more at: goodfire.ai/platform (10/10)

Introducing Silico: the platform for building AI models with the precision of written software. Silico lets researchers and engineers see inside their models, debug failures, and intentionally design them from the ground up. Early access is open now. 🧵(1/10)