Pinned Tweet
Constructive Specifications: A Short Story
blog.neysofu.me/constructive…
Without getting into too much detail about the conversations I’ve had with my friends at Anthropic, it’s become obvious to me that their long-term ambition is to continue running these small-to-large experiments while scaling experiment throughput with AI agents.
Now, researchers already use smaller models to test architectures and optimizers. However, what I’m hearing is that they want to do this on a much bigger scale.
Think millions of smaller, cheaper experiments with models that have a billion parameters, with the most promising discoveries being validated at 10 billion, then 100 billion, and so on and so forth.
The idea is to have autonomous research agents explore this enormous hypothesis space and identify which architectures and training methods scale best.
Now, this probably sounds obvious to a lot of people already in the AI field. However, from what I’m hearing, there’s a lot of success happening right now, and they’re continuing to scale up this experimentation, which is quite exciting to hear.
Filippo retweeted
we married on the 11/11 and @calcom just hit $11.11M in ARR 🔥
Filippo retweeted
>engineer says an agent swarm secured his software
>ask if it’s formal methods or 8 million lines of smoke tests
>he doesn’t understand
>pull out illustrated diagram explaining formal methods and smoke tests
>he laughs and says “it's good security sir”
>check
>it smoke tests
Filippo retweeted
Composability really is the name of the game. We need to build a culture and ecosystem of open specs, as opposed to monolithic verification efforts.
💭 unpopular opinion: type theories of Lean / Rocq are not complex and heavyweight enough. we need low-level, type-checked control over runtime *and* programs and proofs in the same lang
maybe dependently typed lang with OxCaml's modes. even better: graded modal dependent types!
Filippo retweeted
The bull case for formal verification is "Here is a reusable library of properties/specs you can import in your code" not "Here is a few million LOC of Lean for you to go over, but the model passes!"
prediction: victor will find and showcase a decent intelligence primitive within the next year, that's not matmul
I'm (ofc) super interested in studying systems that do NOT have attention anywhere (nor anything like it), yet that show signs of useful intelligence (like answering English questions). I'm collecting a list of these. Appreciate pointers. Could be anything
85
Filippo retweeted
Replying to @SeanSankerJr
Current programming languages were designed with human ergonomics in mind. Humans do not write code anymore. So we should throw away all old languages and design new ones, geared toward the machines to write code in.
5 years ago
I'm not sure how useful this thing will actually be in my day to day programming activities, but A for effort. #Rust #GitHubCopilot
1
90


























