For those that don't know about this. Phi was announced with a paper called "Textbooks are all you need". What they did was use GPT 3.5 and created synthetic textbook chapters and exercises.
They also did some more interesting work like showing very small models can be coherent as long as you have very simple children's book style training data (TinyStories is pretty famous).
Lots of these ideas are still used. Learning facts at scale with active reading is an ICLR 2026 paper from Meta AI that does a lot of similar work.
Have you given any thought to a webui, and a "warp" server I can install on a VPS and interact with it via the webui? I believe this paradigm/approach is the real future.
That's exactly why we complain about the Rs in strawberry. We can get funny-stupid human interpretations all day long. What we can't get is cold facts, and isn't that what was promised by AI (at least, before ChatGPT was released in 2022)?
Unfortunately we fed this current iteration of AI with human behaviour (not only that: human behaviour on the Internet...)
> What we can't get is cold facts, and isn't that what was promised by AI (at least, before ChatGPT was released in 2022)?
Not that I know of. An entity dealing only in cold facts is not intelligent, it's a theorem prover- extremely narrow, rigid and incapable of interpretation and insight- basically of bridging the smallest gap of knowledge. That's exactly what intelligence isn't.
Because I see this becoming bigger than me and a separate organization made sense. There is a super thin backend component right now too. There is the potential to also add in some extra features that require a server/db. I'm kind of inspired by the atuin model of things.
I've got a LONG list of features I'd like to implement over time.
You said factual. But what is factual for you and I may not be for someone else. There are a lot of recollections in the article where sama remembers one version or doesn't remember at all and the other party remembers something else. Combine that with the nature of the article and the legal issues considering egos and sums involved. To top all of that New Yorker is known for fact checking that is exhaustive to the point of paranoia.
I am just speculating but if @ronanfarrow is still checking the discussion here, it would be amazing to hear the actual reasons.
> There are a lot of recollections in the article where sama remembers one version or doesn't remember at all and the other party remembers something else.
Every good journalist knows that when you say potentially damaging stuff about very powerful people, you must be able to prove everything you say. So, you don't say "Sam Altman is a fraudster" because you (likely) can't prove it in court. Instead you say "Person A said that Sam Altman promised X and then reneged" and if anyone challenges that you show "here's proof that person A did tell me that".
If you want to be extra cautious, you only report where you have multiple witnesses agreeing on it.
reply