Will be hanging out at the Agentic Summit this Saturday. Happy to meet up and talk agent observability, evals, and deployment for cyber security. rdi.berkeley.edu/events/agentic…

there is truly no social media-ism that makes me stop reading faster than 'let that sink in'. i would rather see "it's not x, it's y" it's beyond self parody! give it up! the sentence can draw attention to itself if you write it correctly!

Wrote about evals at Dreadnode. This one is for hackers getting up to speed on agents for their use cases. How do you go from PoC to prod? Don't wait for a lab to build benchmarks that measure what you care about. Do it yourself. Here's how:

time to copy edit the paper (you no longer remember a time before editing latex tables)

So far even in highly verifiable settings, RL with judge seems to work better than pure verifiable.

before posting writing online, it's important to read it out loud to yourself. that way every time you hit a difficult to read sentence you can get really mad and delete the whole piece, preventing anyone online from suffering

YC Fall 2025 Startup Requests: We Are Demanding The Torment Nexus Immediately

damn nobody ever talks about how embarrassing it is to ask around for an arxiv endorsement.

Recipe gets confirmed: drop pure verifiable, take the most performant judge that can fit on GPU (latency not a bit issue, so long as batches per step are small) and ask for soft critique.

all 52.8s are equal, but some 52.8s are more equal than others

The room is filling up!