like many genius people, gpt-5 (Auto) is displaying high variance.

Deep research systems can't handle questions involving dozens of documents. Let me show you why this is (still) true 🧵and what does it all have to do with Grace Kelly? (1/)

Some users can't be helped...

⎿  Read 20 lines (ctrl+r to expand) ⏺ Perfect! I now have a clear understanding of the codebase.

Neurosama (LLM Vtuber) was just called a cl*ank*r and in response brute forced her filter to send death threats in response and then got stuck in a loop about wanting to stop existing

Yeah, it's called doing a PhD

for someone getting into RL, what are some good seeds?

just remove the following string from claude-code cli.js and it will always just read full files: offset:h.number().optional().describe("The line number to start reading from. Only provide if the file is too large to read at once"),limit:h.number().optional().describe("The

Most takes on RL environments are bad. 1. There are hardly any high-quality RL environments and evals available. Most agentic environments and evals are flawed when you look at the details. It’s a crisis: and no one is talking about it because they’re being hoodwinked by labs

I can't state how amazing of an idea this UE8MO is. Fuck negative numbers Fuck the mantissa All you need is exponent

why dspy usually wastes your time (and when it doesn't) the question: "should i use dspy for prompt optimization? it seems like the perfect tool for improving my rag system." the answer: dspy is great for very specific, well-defined tasks. but for most rag systems, it's a

this figure needs updating

If only there was a benchmark that tests this capability

Secret model nerfing paranoia will never recover from this

Neil Lawrence I have a blanket ban on "rm" and only let my LLM use "rip"! github.com/MilesCranmer/r…

If only he could use it at work lol

Anyone else feeling the same? It's a bit annoying that existing benchmarks like SWEBench don't capture this reward-hacking.

I don't understand why people keep saying that men don't see therapists, all the girls I know who study psychology are seeing someone

Which LM is better at agentic coding? We have a bunch of useful academic benchmarks like SWE-Bench, but we don't have a good comparison of agentic coding LMs *in the wild*. To solve this, we released PR Arena: github.com/neulab/pr-arena

Unfortunate reality: most open-source LLM servers (e.g. Together) don’t offer cache-hit discounts, while closed providers like OpenAI do. DeepSeek does discount, but most third-party servers don't. Closed models can end up much cheaper than open ones :(