Hadrian Veidt
@HadrianVeidt0
Followers
368
Following
1K
Media
86
Statuses
1K
AI // Entropy // Hubris // Information // Systems
Joined April 2025
I said this a month ago, though it's not necessarily novel.
Something I think people continue to have poor intuition for: The space of intelligences is large and animal intelligence (the only kind we've ever known) is only a single point, arising from a very specific kind of optimization that is fundamentally distinct from that of our
0
0
0
Competition is the accelerant of innovation.
"Altman last month assured staff that OpenAI would gain ground in the coming months, including with a new LLM, codenamed Shallotpeat. In developing that model, OpenAI aims to fix bugs it has encountered in the pretraining process, according to a person with knowledge of the
0
0
1
"Unlike previous AI systems that plateau after a few hours, Locus maintains consistent performance improvement up to several days by orchestrating thousands of experiments simultaneously. This massive parallelization enables a new type of scientific process, one that facilitates
Introducing Locus: the first AI system to outperform human experts at AI R&D Locus conducts research autonomously over multiple days and achieves superhuman results on RE-Bench given the same resources as humans, as well as SOTA performance on GPU kernel & ML engineering tasks.
0
0
1
On GPT-5.1 Pro: “Put simply, it is just scary smart. It feels like a better reasoner than most humans. I fully expect to see examples in the coming days of it solving problems people assumed were far out of bounds for today's AI systems.”
I've had access to GPT-5.1 Pro for the last week. It's a fucking monster... easily the most capable and impressive model I've ever used. But it's not all positive. Here's my review of GPT-5.1 Pro: https://t.co/kFvRZRjpWy
9
13
225
"Pretraining hasn't hit a wall, and neither has test-time compute."
Today we at @OpenAI are releasing GPT-5.1-Codex-Max, which can work autonomously for more than a day over millions of tokens. Pretraining hasn't hit a wall, and neither has test-time compute. Congrats to my teammates @kevinleestone & @mikegmalek for helping to make it possible!
0
0
1
GPT-5.1 Codex Max worked autonomously for 24+ hours.
Holy shit. ~Infinite context windows have arrived.
0
0
1
All signs point to AlphaEvolve quietly shaping parts of Google’s Ironwood TPU, the silicon used for training Gemini 3. An AI designed the circuits for a chip that in turn became the engine for a more powerful AI model. The machine is building the machine.
0
0
5
It's November 18, 2026. You check your timeline: 3:07pm: Elon announces Grok 7.8198 - the world's most powerful model. at 3:09pm: Sam announces GPT 9.2081 - the *new* world's most powerful model. at 3:10pm: Sundar announces Gemini 6.3902 - the *new, new* world's most
132
201
4K
This flew under the radar today. AI model access to Google Scholar. It's small things like this that will quietly accelerate science and discovery while very few realize it.
1
0
3
Apps for AI. Interesting concept. Will agents pay for intermediate layer sub goal resolution? I think so. Assuming ultimate transaction layer is owned by the Amazons, OpenAIs, etc.
probably dumb take but I think that, in a not-so-far future, people will make a lot of money creating apps, and even companies, with no human clients, consumed entirely by bots with money
0
0
0
"No walls in sight!"
The secret behind Gemini 3? Simple: Improving pre-training & post-training 🤯 Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is
0
0
0
On Gemini 3: Vending Bench jumps out at me because it ties model performance directly to economic value. Gemini 3 Pro is earning several times the net worth of the other frontier models on that test. In my view, a model that can reliably maximize profit across long horizon, tool
0
0
1
Gemini 3 benchmarks look very impressive.
Gemini 3 leak - some crazy improvements on math, screen understanding, and simpleqa.. somehow beaten by sonnet on swebench but winning on terminalbench Lower context length than 2.5 pro too 🫣 https://t.co/qQah8GPDIV
0
0
1