Each stack frame has its own isolated context. This pushes the token pressure down the stack. The top level conversation can go on for days in this arrangement. There is no need for summarization or other tricks.
Is this related to the paper on Recursive Language Models? I remember it mentioned something similar about "symbolic recursion", but the way you describe it makes it sound too simple, why is there an entire paper about it?
The RLM paper did inspire me to try it. This is where the term comes from. "Symbolic" should be taken to mean "deterministic" or "out of band" in this context. A lot of other recursive LLM schemes rely on the recursion being in the token stream (i.e.. "make believe you have a call stack and work through this problem recursively"). Clearly this pales in comparison to actual recursion with a real stack.
I mean developing taste is what RLHF was supposed to give you. I don't think it's actually a technical problem so much as a social one. The average person _wants_ slop, they don't want to read new yorker articles, they'd much rather read romcoms. A model trained to produce tasteful writing would almost surely have less engagement from the public (considering that engagement and lmarena-maxing is what led to the characteristic punchy style in the first place)
Anthropic's definition of "safe AI" precludes open-source AI. This is clear if you listen to what he says in interviews, I think he might even prefer OpenAI's closed source models winning to having open-source AI (because at least in the former it's not a free-for-all)
The entire thing is a joy to read, you should really set aside some time to cleanse your palette in this age of LLM prose. I mean just look at this juxtaposition
>Altman continued touting OpenAI’s commitment to safety, especially when potential recruits were within earshot. In late 2022, four computer scientists published a paper motivated in part by concerns about “deceptive alignment,” in which sufficiently advanced models might pretend to behave well during testing and then, once deployed, pursue their own goals.
(plus it finally resolves the mystery of "what Ilya saw" that day)
Also since it wasn't stated clearly
>“the breach” in India. Altman, during many hours of briefing with the board, had neglected to mention that Microsoft had released an early version of ChatGPT in India
I think it's similar to the case of counterfactuals, hypotheticals, or steelmaning and how well you can handle them. ("Can you accept that there can be a function named multiplyBy5 that does something else instead").
But I think if someone already is comfortable with working with abstractions such as "function" the thing is trivial, so it's a bit of a weird litmus test.
> "Can you accept that there can be a function named multiplyBy5 that does something else instead").
I think anyone that can understand a function can understand this, but one might not be happy accepting it's the case, and endeavor to change it.
I think it can be easy to lose sight of that distinction, and eagerness to fix it can be confused with not accepting it could be, but is also probably wrong.
What do you mean by "themselves" here? Grok is RL'd to behave like a Grok, so it trivially knows the qualities that define Grok better than Gemini does, which can only go by second hand sources.
> so it trivially knows the qualities that define Grok
How does it know? Where did he get that knowledge from? Did they train Grok, check it's qualities and included them in next training set? Was his source code and summarization of weights included, or maybe he has access to them for "introspection"?
reply