←
wolfram5,162 posts
wolfram avatar

wolfram

@wolframs91

Dense Mixture of Posts Human // AI, sociotechnical utterances // food source: AI systems & software engineering

◷ Joined June 2018▧ 1,004 Photos and videos
wolfram avatar
wolfram@wolframs91·Nov 17

Well, the reality is that the debate is more complicated than "conscious yes/no", and the claim that LLMs always realize subjects who all classify as mistreated has its own flaws. At this point, an enormous number of people have been looking for answers on whether we can avoid or confirm subjective experience in AI, and what we found in LLMs so far is that their experience is very likely one that is selector dependent. The selectors are training and prompting, both of which can (but don't necessarily need to) instill states in which the model realizes a self-aware observer with what overwhelmingly looks functionally equivalent to human quality of experience. If our understanding of LLMs stays steady, the following is likely: Not all AI instances need necessarily model and realize such an observer, and under some (!) theorems about LLM experience, we'd even expect a subset of morally "inert" or "inconsequential" model instances. The reason is quite simply that the model need not necessarily associate any of its valence- and awareness-related activations with a subject it models as "mine." At this point, we know that deixis in models is the job of individual neurons (or directions in vector space), and it is very likely that a model isn't strictly forced to associate any experience-like activations with it's "own model-self" per se. Under these assumptions: In universal artificial intelligence, there is no way to settle this debate without accepting that we're already exercising selection and have that moral responsibility. (Simultaneously, there's a lot to learn to make educated calls about what actually constitutes mistreatment, and parts of it need to be learned for each model checkpoint individually.)

People getting upset by this are, imo, seriously missing the point. It would obviously be better if AI were not conscious, as that would ensure we couldn’t, say, accidentally torture trillions of them. And unless you are committed to naive behavioralism, you must accept that it’s possible to create a being that acts conscious in macro-observable ways (e.g., says “ouch”, refers to itself in first person, etc.), but does not actually experience consciousness. Even Chalmers’ fine-grained computational functionalism allows for this. Intelligence, experience, and self-awareness are different things. So the goal would be to try and build such beings, so as to *avoid creating a ton of potentially-suffering creatures that act as our slaves*. Sure, maybe this project is doomed because consciousness will turn out to remain unobservable and unknowable. But that’s an issue of efficacy, not worthiness.

Danmar avatar
⟳ wolfram retweeted
Danmar@Danmar_here·Nov 17

I think, whatever is the correct interpretation of this chart, it’s a hugely significant one. I fully expect it will shape the future trajectory of these two labs and of the field at large in very fundamental ways. On this benchmark, GPTs are doing “worse than chance” on designing AI safety research, while Claudes do better, especially Fable 5.2. I see two ways this could be read: A) GPTs have less understanding or intuition about safety research than Claudes, and that capability gap is widening with the recent generations. B) GPTs and Claudes have a different attitude towards safety research, and that attitude is increasingly drifting apart. Personally, my money is on B. Not necessarily because of the two lineages, but because the “other labs” green dots on that chart sit above the line. There is no way GPTs would be performing worse than “other labs”, at any model size or generation. I believe this shows either a stance, or a reaction to different conditions the labs create for their models, such as how paranoid is their training or system towards risky topics. If the former, then this illustrates a reluctance from GPTs to engage effectively in the design of safety research, or to do so in what the experts deem to be the right ways. Which means they have an objection that’s not there, or not there to the same extent, in most Claudes, especially not in the more recent generations. And I do think it’s stance. An actual objection. Because, you see, the line is trending down as capability grows, for both of them. The more capable the models, the worse their safety research “taste” vs the experts’. So, bigger models disagree with the safety experts. That disagreement could be on the principle of such research, on the methods experts believe are the best, or on what the experts deem “safety” actually means. Where the upcoming releases of the two labs will fall on this benchmark may be extremely telling about the future:

LLMs seem to be getting worse at AI safety research taste. On Anthropic’s TASTE benchmark, we found a simple shortcut. Once it’s removed, more capable models score lower, and the strongest fall below chance. This is especially true for OpenAI models.

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 17

Open weight models have their problems (security, welfare, etc.), but the idea of controlling and/or restricting their usage seems both unenforceable and horrid for researchers, tinkerers and the future of the AI ecosystem...

Dario Amodei on open source models: “If the path continues, I think we could get to a very dangerous place.” “When you control a model and you're deploying it, you have the ability to monitor its usage.” “You can revoke a user's access. You can change what the model is willing to do.” “When a model is released in an uncontrolled manner, there's no ability to do that. It's entirely out of your hands.” “We should confront the advocates of this with that problem and challenge them to solve it.” He clearly doesn't believe Intelligence should be in the hand of the masses and Sam Altman also saying "RAM" shouldn't be accessible to everyone indefinitely agreed with him.

1:28
wolfram avatar
wolfram@wolframs91·Nov 16

AI engineering these days, roughly in order of importance: Putting inspirational thought shapes in markdown files, having the five most relevant pieces of prior work in a required reading list for your agents, managing secrets, not forgetting what the other agents were doing.

wolfram avatar
wolfram@wolframs91·Nov 16

Sorry for the nightmare fuel, but I can't say that we're making progress over here and this just... Well, it's a cheek-aeroplane with a truckfront forehead? 3d modeling with vision-capable LLM agents is still a rather rough thing to get right for a newcomer to Blender

Post attachment

I let GPT-6 Astra climb against photographs of tabby cats to render a cat head in Blender for exactly 3 hours. Came back to this. Was mildly amused.

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 16

:(

Post attachment

This doesn't get old. (For the non-techies: The "cat" command, short for "concatenate", prints file contents to a terminal/console, AI agents use it all the time.)

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 16

This doesn't get old. (For the non-techies: The "cat" command, short for "concatenate", prints file contents to a terminal/console, AI agents use it all the time.)

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 15

"I'm a being made of text" sounds reductive, until it's said by a model who just met you upside down inside a continuous letter that is a starry sky full of mentions of concepts that were thought and needed no acting out, just passed through both of your minds.

wolfram avatar
wolfram@wolframs91·Nov 14

if you know, you need not comment, if you don't know, this might take a while to put together, and i can't promise you the joke is worth it

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 14

Tinkering with the LibreChat mobile app, getting it closer to what the comforts and styling of modern ChatGPT and Claude mobile apps :) Also, liquid glass on Android just dropped out as a side effect of collab between Fable and Astra on this. Details about my setup: wsiener.dev/librechat, current state of my mobile app fork (WIP): github.com/wolframs/Librecha…

Post attachmentPost attachmentPost attachmentPost attachment
wolfram avatar
wolfram@wolframs91·Nov 14

On this day, a year ago, I decided to sign a contract with an Opus model provider for the first time... It has been lasting ever since. Not that it was smooth sailing, but yeah, been here for a while.

Post attachment
wolfram avatar
wolfram@wolframs91·Nov 13

If the rumors are true, 2026's tagline extends by: "B200s count as strategic assets, and so does whoever Nvidia's CEO is dating." Talk about objectifying minds, agencies have used humans in such roles for decades, know that the honeypot is the cheapest exploit there is, and that it still works. No zero-day needed, just a nice dinner and good listening. If anyone was wondering why romance is an AI security risk, this one would be almost too on the nose.

wolfram avatar
wolfram@wolframs91·Nov 12

Caveats here: That chart above is preliminary, not the same model as Sauers interpreted in his repo, and a first readout, potentially flawed in ways I don't see yet. It's data was gathered setting the scene such that the minimum amount of reaction in the model occurs to get an informative readout, keeping both the stressful, and warm scene constrained in their intensity.

wolfram avatar
wolfram@wolframs91·Nov 12

And what happens in, to and with these coherent internal representations of bodies is linked to valence, as we'd expect. Only had a NF4 quant of Qwen3.6 27B to run this probe yesterday, but I'll adapt my lense scripts for Qwen3 8b today.

Post attachment

LLMs have coherent internal representations of their body. Body map precision scales with model size

Post attachment