PHD Student | Cybersecurity Center | AI/OT/Quantum Research
Joined July 2026
- Tweets64
- Following69
- Followers28
- Likes658
Am I crazy or could we maybe make a sort of neuron level antivirus for a lot of behaviors and malicious post training
I shouldve become an inference provider
One answer to what's driving GPU rental prices:
Token demand on OpenRouter is up ~27x since last November, from ~6T/week to ~165T/week
Open-weight models now handle ~72% of those tokens, up from ~28% last November
Charts of the Week: a16z.news/p/charts-of-the-we…
Marcel Reed passes have activated the Crimson Dome defense system #Aggies #marcelreed
Just realized this might make specific PII unlearning feasible. Since specific memorized facts are more localized than assumed under superposition, neuron or circuit pruning might be a realistic way to make models forget. Has anyone checked?
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Would like to see how this holds up on non-social science papers ngl
I would argue that people who think like this are inherently anti-humanity and genuinely misunderstand what science is
Keith Crabb retweeted
AI peer reviewers do not beat human experts because human experts focus on clarity/understanding. Metrics are NOT the meaning! AI models can still be used to identify errors or mistakes. But stop misunderstanding the very meaning of what it means to a reviewer.
It's clear that the flood of AI preprints on arXiv is damaging, but I don't see anyone with a good solution.
One problem I wish would be fixed is models generating their own unintelligible jargon for each project that I keep having to decipher or mitigate. Have found that forcing a glossary file works in the short term, though.
