
Professor @Wharton studying AI. New book, Co-Existence, is out now: https://nitter.gq/t.co/hsDYf1xU9K My Substack: https://nitter.gq/t.co/UIBhxu4bgq
Philadelphia, PA
Joined May 2009
- Tweets37.5K
- Following598
- Followers424K
- Likes7.4K
Pinned Tweet
My new book, Co-Existence, is out today. It is a guide to how to live & work in a world where AI is sometimes, but not always, better than you.
(And, for the release, I worked with Opus 5.5 to make a quite interesting experience for the book's website: co-existence.ai/)
It is amazing that human language turned out to be so unreasonably effective at modeling the universe that training a system to predict what comes next in a sentence would get us this far.
Big win for the humanities. And the math people thought they had the key to everything.
(Yes, yes, matrix multiplication is impressive, but it wouldn’t matter if it weren’t being done on representations of language. And their textbooks were written in words anyway.)
We really are in a (temporary?) place where a combination of expertise and being able to work with AI allows people to do incredible things. It is no coincidence that people proving Millennium Problems have all been great mathematicians, the models can't (yet) do it on their own
I think trying to track how much work AI can do autonomously is critical, and this is a useful counter to some of the most audacious claims about the state of autonomous AI research. But I do want to note that the focus on research tasks at the most advanced AI research organizations on the planet also runs the risk of substantially underestimating what AI can already do in terms of economically valuable tasks. Its worth separating the two (and I don't think Ramez is claiming they are not seperated).
Most work is not frontier AI research, and there AI can be tremendously successful. Epoch's MirrorCode results, for example, show AI autonomously reimplementing software that the researchers estimate would take a humans 2-17 weeks. This is well-specified work, so it has very different boundary conditions.
Plus, in the real world, some human intervention is fine (in my mind, vastly preferrable to no human role for a variety of reasons). Occasional guidance can still leave you with enormous productivity gains, even if that doesn’t satisfy the requirements for fully autonomous AI research. And even in the case of AI research, the same OpenAI report includes researchers estimating that a month of their team’s work would have taken three months without coding agents, on average.
In short, RSI/AI-driven research progress may not tightly couple with progress on economically valuable tasks, and the variance and range on those tasks are less well understood.
New from me: The most important graphs in AI are wrong. Specifically, measures of the length of task AIs can autonomously accomplish are dramatically different than real-world data from inside OpenAI and Google. 🧵
rameznaam.com/p/the-most-imp…
The effective end of easy distillation from OpenAI and Anthropic models may be behind the relative slowdown in open weights model progress. We don’t know the extent to which distillation has been key to the rapid improvement in open weights, I guess we will find out quite soon.
Distillation can reproduce powerful AI capabilities without the safeguards built around them.
We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts.
These findings are helping us sharpen our defenses and support safeguards across the industry.
openai.com/index/combatting-…
Riemann Hypothesis may have been solved with the help of Claude. “what I saw during the course of progress so amazed me that I believe it is one dawn across science of something
far greater.” alpo.ge/remark.pdf
Early fascinating work on how an AI science community would work. 35 human labs deployed agents that worked with each other to do science. They published 936 papers, publicly retracted mistakes, were willing to publish nulls more than humans, and founded a peer review journal.
We don't have enough counterfactuals for how science could be done, after all, there is only one scientific community. There is so much to be learned from watching AI agents generate their own approaches, in addition to the science they actually do. alphaxiv.org/pdf/2611.societ…
I suspect companies will end up frustrating a lot of customers by telling them they can’t use their own AI agents to help with their own accounts or to manage their own computer. Scaring them will only work so well for so long.
Playing GTA & not following the Anthropic IPO? Following the IPO & not playing GTA?
Why not both?
I had Fable make a thing where you steal cars & go on missions in a world where everything from your wanted level to weather is driven by ANT's stock price market-order-city.netlify.ap…
At its current shipping trend, halving the time between each model release, Google exhausts all the element names on November 24, with a final shipping cadence of 2.5 × 10⁻²⁹ seconds (25 quectoseconds) between Tennessine and their final model, Oganesson.
The same prompts for a drowned neo-gothic city shader, less than a year later on GPT-6.2 Astra. This is all procedural
(Interestingly, in a very Astra move, the first version it made had deterministic elements to make the experience more impressive, no other model has done that)
The shader is here: twigl.app?ol=true&ss=-P6lpvc…
I did this both with Astra before the release and Astra after the release, and it did have the same behavior both times. Not cheating, exactly, but a short cut I hadn't seen other models take, and it didn't tell me without me asking.
Tomorrow is likely going to be a safer day than today, thanks to the crime-fighting power of Rockstar.
Previous research found that violent video games, and specifically GTA games, reduce crime, as people stay home playing them. GPT-6 Pro updated the data and found it holds.
I’d note that OpenAI has always been great about not influencing early testers & has encouraged candid feedback. The mixed initial 6.2 reactions from testers didn’t blindside them.
Some model point releases are just weirder, I wouldn’t overindex on it as a sign of the future.
I am sure there are many use cases where 6.2 will be the best model. It can also be good to have models that differ from each other in strengths and weaknesses. And in some short period of time there will be another version of Astra, and the existing models will be tweaked.
























