Sharing my hobby project, penpusher: AI-powered form filling! Give it files/images/previously filled forms, and penpusher will fill out new ones for you. Imo it's one of the best LLM use cases -- it gets useful work done. Looking for a few interested beta testers!

Woah, remember Bard?

Service fees and paying for invoices. Plus token costs are higher than using providers directly, so you pay:.- higher per-token cost.- percentage of credits service fee to load credits.- (optionally) percentage of credits invoice fee

Not only that, but speed/quality feels pretty variable depending on which provider your request gets routed to for open models.

Surprised by how little I like OpenRouter. All the extra fees, feels like DoorDash for LLM inference.

JAX ecosystem names are so delightful:. $ uv add jax flax optax.

Hmmm, seems like a red flag in an NLP paper. (A ROUGE flag, perhaps?)

The more I read, the more compelling I find the idea of keeping a commonplace book for meaningful quotes. Does anyone do something similar?.

RT @psresnik: To be presented at ACL 2025: Large Language Models Are Biased Because They Are Large Language Models. Article: https://t.co/….

Quick python question -- why do asyncio's gather(*tasks) and as_completed(tasks) take different inputs (positional args vs a sequence)?.

Question: Huggingface's ViT interpolates position encodings at runtime (for every batch recompute the interpolated position encoding). Is there an advantage to this over initializing a new model with interpolated encodings and training from there like normal?. cc @mervenoyann.

Link:.

In which I argue that LLM-produced bounding boxes aren't good enough for document tasks (yet).

🤔ask it to "generate white plays d4" and things get even weirder.

4o-generated LiChess starting board. Far better than any diffusion model I've seen, still weird errors.

I am begging people not to do this. Science isn’t “accepted conference papers,” it’s a collective, human, theory-building endeavor.

Gemma 3 Pan & Scan results show just how important resolution is when processing documents. Using higher res than 896x896 gives pretty substantial gains when processing documents and infographics.

I always forget, is Max higher than Ultra?.

+34 elo vs the ChatGPT 4o model (if that's what they used for comparison) is _a little_ lower than you'd expect with these win rates, but not far off!

someone who is good at the economy please help me budget this

git checkout -b claude_code_sandbox.