[27183]

I used the example of 1G constant acceleration space flight in another thread which got downvoted to oblivion, but I think it's a good one. That's a technology we know how to build. We just need superconducting electronics and miniaturized fusion reactors, or a ship which is built like Project Orion to use nuclear bombs for propulsion.

Now write down a blueprint for superintelligence.

So I've given you two impossible engineering challenges, but one of them is feasible in principle because we at least have the tools to begin to tackle the theoretical calculations and therefore we can do engineering. We cannot do engineering on the superintelligence problem yet.

In my view it would be insane to believe we can build something that we can't even reliably imagine yet.

[derektank]

As early as the late 19th century, Louis Pasteur’s work had inspired a belief in the scientific community that it must, in principle be possible to selectively exterminate bacteria. The German physician Paul Ehrlich expounded on this in greatest detail in 1907 when he described his “magic bullet” (or Zauberkugel) theory for effectively targeting pathogens without harming the human host, similar to the immune system.

However, if you had had demanded someone for a blueprint in 1925 of how to design such a magic bullet, especially a magic bullet that targeted virtually all forms of bacteria, it would have sounded ludicrous. Yet, 20 years later, the world was manufacturing 6-7 trillion units of penicillin a year, capable of treating 3-6 million people. And that’s in spite of the fact that Fleming’s work sat mostly untouched for a decade before Howard Florey and Ernst Chain seriously set about to isolate and purify the substance.

You can quibble and say that penicillin was discovered, not designed, which is certainly true. But I would ask you to consider, does current AI development look more like design or discovery? Does it look more like analytical engineering or evolutionary selection? I would say on both counts the latter, in which case, we should prepare to be surprised how long it might take to make revolutionary advances. And that’s on both sides of the ledger, we might find ourselves stuck in the current paradigm for a long time. But, we might not be.

[27183]

Yes I think the drug discovery analogy is apt. I've spent a bunch of time playing with evolutionary algorithms, they're great fun. And when they work they can do surprising things! [edit] I think the drug discovery analogy does have some limits though. Drug discovery isn't a blind search through fitness space, it's informed by physics, chemistry, biology, and medicine. We have many guiding lights to illuminate the space and identify regions (still high-dimensional infinite regions!) that are likely to be productive. There are fewer lights to guide the way on a search for fitness in intelligence. Hell, we don't even know how to write down a decent objective function.

I wouldn't bet on evolving an intelligent, sentient being-in-a-box on a computer any time soon though. I'm of course prepared to be pleasantly surprised.

That said, I think it's pretty clear that LLMs are not going to get us there.

[davemp]

I don’t think people are arguing to stop researching AGI. Moreso against sales people trying to use the concept of AGI to sell products that are very much not AGI. Or devoting so many of our resources into such a pursuit that it causes harm to real people.

This is obviously complicated by the fact that LLMs/Agents are useful by themselves, but that’s not really the topic at hand.

[ThrowawayR2]

The parent poster argument boils down to "[something] is theoretically possible, therefore 1) it is guaranteed to practically implementable 2) in the reasonably near future". Both are simply prima facie false; one can ask an LLM to explain why if there's any doubt.

[techpression]

And now we’re desperately trying to ”upgrade” penicillin (and friends) because it doesn’t work any more in many cases. Do you think we can repeat the process or do we need something completely different?

This is why biological comparisons are weak, we talk about a few agents verifying and checking LLMs, meanwhile the world consists of almost an infinite number of the same, just operating on different time scales. I agree that with we don’t know the timescale, and we definitely don’t know if long term it will continue to work ”adding more of the same”. Throwing more penicillin at the problem sure as hell didn’t, but it looked great initially. And I’m obviously not arguing the human benefits of penicillin, just that what we thought would work forever quickly didn’t.

[27183]

To quote the great Dr. Malcolm:

> Life, uh, finds a way.

[stalfie]

No one imagined LLMs in their current format, it was simply a result of discovering that scaling compute and tokens produced better and better results with the Transformer architecture. The inventors of the Transformer architecture were working on better translation, and probably did not imagine that their architecture would lead to modern LLMs.

Imagining something in advance is not necessary at all for scientific advancement. This is particularily true in AI, and no one expects to imagine what superintelligence is until after it is created. You set up your datasets, your architecture tweaks, and measure the results on some set of benchmarks. There never was a blueprint, no plan beyond the experiment itself. We're not even close to understanding the things we have already created, and yet we created them. So why expect anything else for the next step?

[iugtmkbdfil834]

<< No one imagined LLMs in their current format

That is simply not accurate. There are examples of scifi novels, novellas and other media that dealt with it. We can argue over whether it was that exact format, implementation and so on, but that 'shape' ( to use a common llm term ) of technological advances was very much explored.

[stalfie]

I was referring more to the fact that no one predicted that next token predicting Transformers would go so far. Not about "AI" in general.

[27183]

> Imagining something in advance is not necessary at all for scientific advancement. This is particularily true in AI, and no one expects to imagine what superintelligence is until after it is created.

Then why does anyone expect to create it? I'll take a stab at an answer: they think an LLM is some kind of "incremental improvement" and therefore a step along the inevitable path to discovering AI. But that seems delusional to me. I can't imagine anyone sound of mind who knows how an LLM works thinks it's actually intelligent. So in what sense is it an "advancement" on the path to AI?

The concept of an incremental improvement in an objectiveless search in a high dimensional space is.. absurd.

[kalkin]

> actually intelligent

It's reasonable to doubt that LLMs are a path to AGI, but I don't understand how this is still a matter of dispute in 2026. What's your definition of intelligence that doesn't cover an entity that can translate fluently between dozens of languages and also solve open problems in mathematics? And be real-if you have one, is it a definition you or anyone would have given a decade ago, or are we doing "god of the gaps"?

[27183]

I can't give you or your sibling a better answer than "you'll know it when you see it". Some people see it now. I think they're wrong, because it seems like the results you're describing are easily explained by fuzzy search in the space of embeddings and then forming strings of plausible tokens related to the resulting region of embeddings space. In other words, the things we know LLMs actually do.

That's more or less looking for interesting patterns in a jpeg or another lossy compression result. It's interesting that the models seem to be able to (fairly) reliably return relevant chunks of the image. Even more interestingly, they seem to be able to invent plausible chunks of image that aren't even there. That doesn't meet my bar for intelligence though. I'd need to see it learn and adapt. I'd need to see it be clever, not merely "knowledgeable". I'd need to see it capably analyze itself. I'd need to see it reasonably estimate uncertainty and know itself in the sense that it has some idea how right or wrong it is about something. I'd need to see it exercise judgment.

I don't think I'd give a different answer a decade ago but who knows.

[edit] For all we know, one of the salient features of intelligence is that intelligent beings are incapable of precisely defining it. I'm not sure how productive it is to attempt to do so.

[kalkin]

I appreciate the straightforwardness, but you probably understand that's pretty unsatisfying.

Actually, stronger - it's valid in some circumstances to say something is infeasible to precisely to define and you'll just know it when you see it. But I don't think it's reasonable to take that stance and then assert that "anyone sound of mind who knows how an LLM works" must agree with what you see. You gotta pick between striving for rigor and denying your opponents' soundness of mind.

[yusufozkan]

What is your definition of "actually intelligent"? I believe LLM's are more intelligent than the average human in a lot of ways according to the Legg/Hutter definition of intelligence: "Intelligence measures an agent's ability to achieve goals in a wide range of environments".

[stalfie]

No one knows how LLMs work. We know how the architecture works, but almost nothing about why. Saying "statistical next token prediction" tells you about as much about LLMs as saying "action potential thresholds" tells you about the brain. A true fact that explains very little.

And I'm sorry, but you're not up to date about interpretability literature, or for that matter philosophical discourse, if you think you have to be delusional to question whether LLMs are "intelligent", whatever you define that word to mean. The [latest publication](https://www.anthropic.com/research/global-workspace) from anthropics interpretability team purports that they see structures in Claude akin to those we think are associated with human consciousnesses experience in the brain. Are you going to dismiss the whole team as not being of "sound mind"?

Intelligence is a word with a somewhat unclear meaning to begin with, but you have move the goalposts pretty damn far to exclude LLMs at this point. They are certainly still lacking in some regards, but whether that disqualifies them for intelligence is very much a matter of debate.

[nadir_ishiguro]

"And I'm sorry, but you're not up to date about interpretability literature, or for that matter philosophical discourse,"

I see no citations referring to the current philsophical discourse, unless you mean to imply anthropic's paid people are to be considered to be part of that.

That they "see structures in Claude akin to those we think are associated with human consciousnesses experience in the brain" is if anything discrediting.

[27183]

+1 I can't imagine how any corporate entity could be credible in this financial environment. Nothing they say can be reliably considered as anything but marketing copy. This situation is exactly what academic publishing is for. Although that institution has also been degrading.

[stalfie]

Sure, here are some citations:

https://arxiv.org/abs/2401.03910?utm_source=chatgpt.com https://www.frontiersin.org/journals/psychology/articles/10.... https://arxiv.org/pdf/2408.04666 https://arxiv.org/abs/2402.00901 https://ar5iv.labs.arxiv.org/html/2407.11015 https://ar5iv.labs.arxiv.org/html/2202.05262 https://www.sciencedirect.com/science/article/abs/pii/S13646...

My point was not to argue one point or another about LLM intelligence, but to push against the notion that you have to be "delusional" to even argue that it is possible that LLMs can qualify as intelligent (although the op prefaced it with "actually"). That's mainly what irked me about the original comment, the arrogance of dismissing everyone even having the discussion as insane, as if there's no legitimate argument to be made.

And this was mainly the point I was arguing. However since we're on the topic, I also happen to think it's intellectually lazy to dismiss the Anthropic interpretability teams work as "delusional" simply because they have a conflict of interest. Of course, that is not an irrelevant fact, but much of their original work has since been replicated by independent entites (eg. https://arxiv.org/abs/2510.01246). Until they publish something that turns out to be fraudulent, I think it's reasonable to consider Anthropic's paid people a very relevant, and in fact excellent part of interpretability discourse.

Dismissing all opinions where there is a perceived conflict of interest is a pleasant cognitive bias to have, but reality is often more nuanced than that.

[layla5alive]

I guess you don't really know how an LLM works then..?

[0xDEAFBEAD]

In this very thread I am being told that Fable is nothing but a bit of scale and refinement on well-known neural network techniques. And next I am told that we can't even imagine how to build superintelligence. Which is it folks?

[layla5alive]

Indeed, the real fun signal here is the conflict between how intelligent humans think they are and the different story told by much of human behavior, including within this very thread!

When an LLM makes a cognitive mistake, humans jeer "dumb stochastic parrot!" - when a human does the same, many are blind to it. Like a weird auto-Gell-Mann Amnesia - relative to reflecting on our own perceived cognitive "strengths" versus others weaknesses.