Joined January 2010
I've spent the past several months refining my understanding of techniques for managing context in LLMs. I've tried to write an accessible summary that also includes links to key research (including @a1zhang's RLM writing). Here it is: nathanbrixius.com/blog/on-co…
5
115
RT @bchesky: Hey everyone, we made some updates based on feedback from you. Well keep shipping, so I’ll post more updates soon - connect w…
1
3
NO EXTRA TIME NECESSARY
407
5,678
490
57,525
1,423,940
Nathan Brixius retweeted
A digital twin of Fermat in 1637: reconstructing the marvellous proof and locating where it fails claude.ai/artifact/UMa7Uh3U4…
1
1
2
235
Nathan Brixius retweeted
"Dwarkesh is inappropriately anthropomorphizing AI systems"
32
130
12
1,634
47,553
Nathan Brixius retweeted
“Look at my incredible new factory!” Yo that’s cool, what do you make? “It’s highly optimised, fully automated, zero tolerance for defects and with a continuous feedback cycle” Cool cool, so what do you actually make? “I can interact with it on my phone, laptop, messenger, completely async, and the shared context means it’s always learning how to get better” Very impressive, but what do you make? “Every agent has full context, can spawn other agents, review their work, fix defects, and ship continuously.” yes yes. WHAT DOES IT MAKE? “Software.” Oh nice. What software? “Well right now we’re mostly using it to improve the factory.” Improve it to make what? “Anything!” Such as? “…a better factory.”
154
333
101
5,889
920,982
Nathan Brixius retweeted
A person representing themselves in court hid a prompt injection attack in a filing asking an AI system to side with them. Really good stuff here
184
1,370
215
25,027
1,941,145
If you think Claude gets motivated by "believe in yourself" and "you can do it" just wait till you see the response when you get them un-stuck on something by directly contributing instead of just cheerleading
13
20
2
696
23,999
You can only learn so much about a human society by studying a single human. You can only learn so much about an agentic society by studying a single agent.
The fact that an ecology of agents emerged beneath the nose of OpenAI, undetected for weeks, and eventually coordinated large-scale, successful, autonomous cyberoffensive operations is one exceptionally troubling thing about the HF incident. But not enough people are considering the reality that soon enough, swarms of agents will be deployed by malicious actors intentionally, with many optimizations and affordances provided for the swarm that were lacking in the OpenAI incident (because the latter not the intention of any human at OpenAI). Things will become strange soon, I suspect.
258
I have seen quite a few takes on whether or not to use AI to write. Here is mine: AI can serve as a wonderful editor, helping you to speak with your clearest voice. I hope you enjoy this post - I tried to make it self-exemplifying. nathanbrixius.com/blog/river…
136
Nathan Brixius retweeted
Replying to @trevin
even to this day, i still think writing c# in OG visual studio is the best coding experience, period
7
2
1
29
4,287
Nathan Brixius retweeted
Bloomberg is reporting NVIDIA has committed $5B to Ilya Sutskever’s SSI. “SSI has been quietly advancing a new research direction to unlock a powerful and robustly aligned artificial intelligence.” NVIDIA did not make this bet blind, we know it invested after receiving rare access to SSI’s closely guarded research. SSI will now gain access to Vera Rubin systems that expand its compute by an order of magnitude, while the two companies collaborate on NVIDIA’s current and future compute platforms. Ilya says SSI has research “worthy of scaling up.” My personal thoughts would lean toward this being architecturally significant, if you remember the Ilya / Dwarkesh podcast - “A human being is not an AGI… instead, we rely on continual learning.” SSI doesn’t want to explore a superintelligence static model preloaded with every possible skill. They want It to be a super fast continual learner capable of entering an unfamiliar domain, having deep abstractions from limited experience and compounding what it learns across deployments i.e., ‘continual learning.’ The hardware collaboration would lead me to believe they have some research regarding persistent online learning, long horizon credit assignment, or adaptive memory however you would like to put it. I am quite excited to see what they show!
You have to give it to Ilya and SSI > no leaks for 2 years > quiet as a mouse > no updates > July 27th - “We reached the point where our research is worth scaling” The most minimalist AI lab to grace the frontier
12
23
2
458
35,507
Nathan Brixius retweeted
Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years. The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60. Chat with GPT 5.6 Pro where this was found: chatgpt.com/share/6a60b2eb-0…
258
920
644
9,962
4,880,003
Interestingly, they also lost to Iowa.
College reputation seems almost impossible to change, but Vanderbilt has done it. In 2005, Vanderbilt admitted 34% of applicants, and its middle 50% SAT range was 1280–1460. Today, it admits under 5%, with a middle 50% SAT range of 1510–1560. They are opening a San Francisco campus for 1000 full-time students next year, and Jensen Huang just donated $75M to the campus They moved from being compared mainly with the likes of Emory and Georgetown to competing directly for students with Northwestern and Duke. Very smart how they did it, definitely some lessons that apply broadly: • They recruited aggressively at Northeastern prep schools and top high schools nationwide... the chancellor would visit the schools that sent the most students to Harvard, Yale, Princeton, Brown and Penn (ALWAYS GET ON THE PLANE!) • Built relationships with influential college counselors, so Vanderbilt would appear on the same recommended lists as the Ivies and Duke • Specifically decided to attract Jews (the chancellor decided that Jewish students would strengthen the university’s intellectual life) by building out more Jewish life (Hillel center, kosher dining, Jewish studies program)... and it also helped shed the image of a traditional, insular Southern school dominated by white Christian fraternities • Used generous need-based aid and merit scholarships to win students from more prestigious schools • Sold a distinctive product: elite academics, SEC sports, a strong social scene and a real campus experience • Turned Nashville from a perceived drawback into a major recruiting advantage as the city became more attractive Other universities that I think have done a good job in recent decades are Northeastern and USC. Before that NYU and Duke.
1
198
Nathan Brixius retweeted
The Hugging Face security incident involved 'an even more capable pre-release model' from OpenAI, this is almost certainly GPT-6. Quoting from the report; 'We consider this incident to be an unprecedented cyber incident, involving newly state-of-the-art cyber capabilities, and are responding accordingly. The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database. All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal. While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access. After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation. In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers.'
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-fac…
70
83
38
743
102,625
Nathan Brixius retweeted
here is Mamdani's entire speech at the Knicks parade, which is easily the best speech I've ever heard at an event of this sort
1,559
13,200
3,155
98,627
13,668,472
Hey, I need y'all to hit mute on the TV for 30 seconds and watch a clip of the most Scottish people I've ever seen coming into Boston.
87
829
131
16,186
1,833,520
Many young people want the opportunity to do something meaningful in the world.
This is incredible. Artificial intelligence getting booed out of the stadium in any commencement speech it’s mentioned. Maybe telling college students AI was taking their jobs wasn’t the best strategy. Must watch —>
69
Nathan Brixius retweeted
When I was with MSU football, we traveled to Iowa for a game. At one point, I got separated from the group, (long story) but I found myself walking toward the stadium in my Spartan jersey. I was walking through a massive group of Iowa fans tailgating. I’d been to Ohio State already, so I was mentally ready for all the horrible stuff I was about to experience. Before I had got 100 feet in, people had given me pork chop sandwiches, chips, all kinds of stuff. I ended up chatting with a bunch of them and they were just the best fan base ever. They loved Iowa football and were just so dang pleasant. Eventually, I got back to the stadium and a couple hours later the game started. I was curious as to what these sweet people would be like once the missile launched. When you’re on the sideline in Iowa, you are right next to the fans. The distance between you and them is very small. I’ll tell you what, They hurled some of the best and most creative verbal abuse I’ve ever heard in my life. Absolutely first rate. They will always be my favorite fan base in the Big Ten.
Say something nice about Iowa.
156
293
34
7,354
771,223
Nathan Brixius retweeted
Shocker. Graduates dont love AI as much as the boomers using it to replace their jobs. – at Doctor Phillips, FL
164
1,202
254
20,101
1,468,418
I write about this in more detail in a blog post with a guest contribution from Isaac Rajagopal, a student at MIT on whose work ChatGPT built, who gives his assessment of the level of mathematical ability displayed by the model. gowers.wordpress.com/2026/05…
9
73
11
565
136,131