You're right and I am humbled. Attention only gets you as far as disproving major conjectures in Paul Erdos's favorite open problems, and certainly no farther than that. Certainly. openai.com/index/model-...
Who's Who of AI
John David Pressman
Why they matter
Directory member with public evidence across AI research, Safety & security.
- AI signals
- 5 past 30d
- Sources
- 5 distinct domains
- Discussions
- 48 past 30d
- Latest signal
- 9d ago
Machine learning researcher, alignment-accelerationist, Fedorovist ancestor simulator, Dreamtime enjoyer.
All posts public domain under CC0 1.0.
What they're sharing
An OpenAI model has disproved a central conjecture in discrete geometry | OpenAI openai.com
LLMs are making formal software verification feasible at a much larger scale anthropic.com
Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident youtu.be
Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models arxiv.org
[2605.13829] Negation Neglect: When models fail to learn negations in training arxiv.org
Delvetown delve.town
View on Bluesky →
z.ai
ZipIt! Merging Models from Different Tasks without Training arxiv.org
A Monte Carlo AIXI Approximation arxiv.org
Implications Of Predicting The Next Token minihf.com
Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident youtu.be
Concept Album - I Killed Darnell (music video, Key & Peele) youtube.com
Articles & links
It begins. www.anthropic.com/research/for...
AI Weekly's analysis
→
- Anthropic and Theorem used Claude to prove crash-freedom at the RISC-V assembly level for 239 entry points across 123 OSS-Fuzz libraries.
- The proofs total about 180 million lines of Rocq, roughly 600,000 lines per binary, and uncovered new bugs including a rediscovery of CVE-2026-56132 in libexpat.
- Anthropic cautions that 'formally verified' does not mean 'bug-free': proofs cover single-threaded static RISC-V binaries, not the x86 or ARM targets most code actually ships on.
Read full analysis →
Alright I'm gonna give this a full watch, live commentary thread below. youtu.be/87DyyMV0kCY
John David Pressman reposted
this is one of my favorite papers, similar method. It convinced me a that LLMs weren’t merely autocompleting they showed that EVEN WHEN the answer to a math problem was present in the dataset, the agent didn’t reference it, but rather looked at problems with similar form arxiv…
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
So apparently LLMs just update on false claims in documents even when they're explicitly labeled in the document as false claims. arxiv.org/abs/2605.13829
John David Pressman reposted
I made an infinitely forkable version of Zork that you can play in Delve.town 😁 Every reply is a save point, so you can branch out from anywhere, even other people's games. A dungeon of forking paths, if you will... The agents seem to like it already! It's Time To Delve 🦇
John David Pressman reposted
GLM-5.3 Cyber might be the tool of choice for cyber defenders not big enough to participate in Anthropic & OpenAI’s programs z.ai/blog/glm-5.3
AI Weekly's analysis
→
- GLM-5.3 launched on August 14, 2026, keeping GLM-5.2's base and lifting Terminal-Bench 3.0 from 4.6 to 28.3 through post-training alone.
- On CyberGym vulnerability discovery, GLM-5.3 scored 84.5%, slightly ahead of Claude Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%.
- Z.ai says the model surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 rated critical or high, and delayed weights about two weeks.
Read full analysis →
View on Bluesky →
What's really going to cook your noodle is when you realize we can do it for minds too. arxiv.org/abs/2305.03053
Which, the standard of what qualifies as "GOFAI" has changed over time. These days most people consider Hutter esque context tree weighting type algorithms GOFAI, I probably consider it GOFAI, but really who the hell knows what he had in mind. Perhaps this but less hardcoded a…
I have written a new post, "Implications Of Predicting The Next Token" in which I explain why saying a language model "just" predicts the next token fails to understand what that phrase would even imply. minihf.com/posts/2026-0...
Alright I'm gonna give this a full watch, live commentary thread below. youtu.be/87DyyMV0kCY
POV: You are the government's lawyer reviewing the evidence in your case against Anthropic re: the natsec implications of their model and potential validity of export controls. youtu.be/RgqzaBogq1Y
Their own posts
Recent commentary
FYI: LLMs are now better at cybersecurity CTFs than people, apparently. The new meta is to just dump tokens on the problem, personal skill doesn't help that much.
I hate that I have to write an essay about the OpenAI HuggingFace breach. I've kind of been trying to get out of the all consuming vortex of writing about whatever is happening but they screwed up so hard that I kind of feel like I have to.
Should have added "don't do a HuggingFace" tbh.
Does anyone on here have an actually working OpenAI Daybreak instance? I'm approved but cannot figure out how I get it to actually work lol.
Their network
In John David Pressman's orbit
Center = John David Pressman. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you John David Pressman? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/jdp-extropian-net)