Switched off memory (in Claude) immediately, not even tempted to try.
Claude’s memory architecture is the opposite of ChatGPT’s
81–90 of 240 posts
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#82Earlier quoted context omitted.
What is "actual intelligence" and how are you different from a Markov chain?
What I mean is that the current generation of LLMs don’t understand how concepts relate to one another. Which is why they’re so bad at maths for instance. Markov chains can’t deduce anything logically. I can.
> What I mean is that the current generation of LLMs don’t understand how concepts relate to one another.
They must be able to do this implicitly; otherwise why are their answers related to the questions you ask them, instead of being completely offtopic?
https://phillipi.github.io/prh/
A consequence of this is that you can steal a black box model by sampling enough answers from its API because you can reconstruct the original model distribution.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#83Earlier quoted context omitted.
Roughly, actual intelligence needs to maintain a world model in its internal representation, not merely an embedding of language, which is a very different data structure and probably will be learned in a very different way. This includes things like: - a map of the world, or concept space, or a codebase, etc - causality - "factoring" which breaks down systems or interactions into predictable parts Language alone is…
> Roughly, actual intelligence needs to maintain a world model in its internal representation And how's that not like stored information (memories) and weighted links between each and/or between groups of them?
It probably is a lot like that! I imagine it's a matter of specializing the networks and learning algorithms to converge to world-model-like-structures rather than language-like-ones. All these models do is approximate the underlying manifold structure, just, the manifold structure of a causal world is different from that of language.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#84The link to the breakdown of ChatGPT's memory implementation is broken, the correct link is: https://www.shloked.com/writing/chatgpt-memory-bitter-lesson This is really cool, I was wondering how memory had been implemented in ChatGPT. Very interesting to see the completely different approaches. It seems to me like Claude's is better suited for solving technical tasks while ChatGPT's is more suited to improving casual…
You don't want an AGI. How do you make it obey?
How do you make your own children obey?
(Meta-question: since they don't do this, why does it turn out not to be a problem?)
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#85I love Claude's memory implementation, but I turned memory off in ChatGPT. I use ChatGPT for too many disparate things and it was weird when it was making associations across things that aren't actually associated in my life.
I turned it off because it seemed to only remember previous hallucinations it'd made and bring them up again.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#86Earlier quoted context omitted.
Roughly, actual intelligence needs to maintain a world model in its internal representation, not merely an embedding of language, which is a very different data structure and probably will be learned in a very different way. This includes things like: - a map of the world, or concept space, or a codebase, etc - causality - "factoring" which breaks down systems or interactions into predictable parts Language alone is…
> Roughly, actual intelligence needs to maintain a world model in its internal representation This is GOFAI metaphor-based development, which never once produced anything useful. They just sat around saying things like "people have world models" and then decided if they programmed something and called it a "world model" they'd get intelligence, it didn't work out, but then they still just went around claiming people…
I see no reason to believe an effective LLM-scale "world-modeling" model would look anything like the kinds of things previous generations of AI researchers were doing. It will probably look a lot more like a transformer architecture--big and compute intensive and with a fairly simple structure--but with a learning process which is different in some key way that make different manifold structures fall out.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#87The difference is implementation comes down to business goals more than anything. There is a clear directionality for ChatGPT. At some point they will monetize by ads and affiliate links. Their memory implementation is aimed at creating a user profile. Claude's memory implementation feels more oriented towards the long term goal of accessing abstractions and past interactions. It's very close to how humans access mem…
why do you see a "clear directionality" leading to ads? this is not obvious to me. chatgpt is not social media, they do not have to monetize in the same way they are making plenty of money from subscriptions, not to count enterprise, business and API
The router introduced in gpt-5 is probably the biggest signal. A router, while determining which model to route query, can determine how much $$ a query is worth. (Query here is conversation). This helps decide the amount of compute openai should spend on it. High value queries -> more chances of affiliate links + in context ads.
Then, the way memory profile is stored is a clear way to mirror personalization. Ads work best when they are personalized as opposed to contextual or generic. (Google ads are personalized based on your profile and context). And then the change in branding from being the intelligent agent to being a companion app. (and hiring of fidji sumo). There are more things here, i just cited a very high level overview, but people have written detailed blogs on it. I personally think affiliate links they can earn from aligns the incentive for everyone. They are a kind of ads, and thats the direction they are marching towards .
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#88"Claude recalls by only referring to your raw conversation history. There are no AI-generated summaries or compressed profiles—just real-time searches through your actual past chats."
AKA, Claude is doing vector search. Instead of asking it about "Chandni Chowk", ask it about "my coworker I was having issues with" and it will miss. Hard. No summaries or built up profiles, no knowledge graphs. This isn't an expert feature, this means it just doesn't work very well.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#89Earlier quoted context omitted.
Understanding how they work in the sense that permits people to invent and implement them, that provides the exact steps to compute every weight and output, is not "meaningful"? There is a lot left to learn about the behaviour of LLMs, higher-level conceptual models to be formed to help us predict specific outcomes and design improved systems, but this meme that "nobody knows how LLMs work" is out of control.
None of that is inherent, and vanishingly few of Anthropic's users invented LLMs.
What is "inherent" supposed to mean here?
LLMs are understood to the extent that they can be built from the ground up. Literally every single aspect of their operation is understood so thoroughly that we can capture it in code.
If you achieved an understanding of how the human brain works at that level of detail, completeness and certainty, a Nobel prize wouldn't be anywhere near enough. They'd have to invent some sort of Giganobel prize and erect a giant golden statue of you in every neuroscience department in the world.
But if you feel happier treating LLMs as fairy magic, I've better things to do than argue.
Re: Claude’s memory architecture is the opposite of ChatGPT’s
#90Earlier quoted context omitted.
> It may actually be the final breakthrough we need for AGI. I disagree. As I understand them, LLMs right now don’t understand concepts. They actually don’t understand, period. They’re basically Markov chains on steroids. There is no intelligence in this, and in my opinion actual intelligence is a prerequisite for AGI.
I don’t understand the argument “AI is just XYZ mechanism, therefore it cannot be intelligent”. Does the mechanism really disqualify it from intelligence if behaviorally, you cannot distinguish it from “real” intelligence? I’m not saying that LLMs have certainly surpassed the “cannot distinguish from real intelligence” threshold, but saying there’s not even a little bit of intelligence in a system that can solve more…
i dont see the need to focus on "intelligent" compared to "it can solve these problems well, and cant solve these other problems"
whats the benefit of calling something "intelligent" ?