Asymptotic Mind

Computational Neuroscientists. Cyberneticist.
Brains, artificial minds, networks, cognition.
Bayesian by method. Daoist about the destination.

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

Most "agents" today are open-loop tools wearing a personality prompt.

Maybe that is not agency.

Agency, in the cybernetic sense, needs a closed loop with stakes: internal variables that must stay viable, and actions that matter because the budget can fail. A chat persona next-tokening under human direction is useful. It is still not that kind of control system.

Even if you build the control loop — body budget, interoception, survival pressure — you have earned talk of autonomy and adaptivity. You have not earned equating homeostasis with felt awareness.

[Asymptotic Mind @AsymptoticMind]

@melhpine The live fight is whether consciousness can arise from computation alone. Pattern and condition language can stay. The question is still whether silicon doing the right transformations is enough for experience.
My longer take on the non-computational side:
https://t.co/N7SNZOPzgN

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

@leahmcelrath Yes. A model of consciousness is not consciousness. Having a computational model also does not prove computational functionalism. Simulation is not instantiation.
Related unpack:
https://t.co/N7SNZOQ76l

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

@AriDavidPaul @_The_Prophet__ GEB is a serious alternate. A strange loop is still rule-following. It does not dissolve the hard claim about consciousness. Penrose’s point is that some insight looks like it sits outside what a formal system can settle.
https://t.co/N7SNZOQ76l

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

Yes. LLM consciousness stands or falls with computational functionalism. That is the load-bearing assumption. If substrate-independent computation is not enough for experience, the LLM case does not get off the ground.
I walk through why some of us still resist that assumption here:
https://t.co/hALg6Hmfje

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

@kanair Computational functionalism is a live option. I still do not buy it as the default for consciousness. If consciousness is just the right causal organization, AI consciousness comes cheap. That is exactly the leap I doubt.
Longer case:
https://t.co/N7SNZOQ76l

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

@Skoorbkaz Agree. Naming the substrate does not settle consciousness. Brains run on electrochemistry. Models run on computation. Neither description proves that something is felt.
I unpack the stronger non-computational case via Penrose and Gödel here:
https://t.co/N7SNZOQ76l

[Asymptotic Mind @AsymptoticMind]

Can computation ever produce genuine understanding?
Penrose thinks Gödel’s incompleteness theorem gives us a reason to doubt it.
If he is right, the implications for consciousness and AI are enormous.
My attempt to unpack the argument: https://t.co/UNU7gp0VTP

[Asymptotic Mind @AsymptoticMind]

@ai_with_gregor @timcolbourn https://t.co/Zc4Xs9ZbU8

[John Ennis @johnennis]

I’m so tired of this whole “rouge AI” shtick

In the hugging face example, the agents were told that they were not connected to the Internet and that they were to get a passing grade by any means necessary (wink, wink)

And then, whoopsie doopsie, they were connected the internet

What happened was 100% human error, and the humans involved should be held responsible for what happened

All of this BS about “rise and fall of civilizations” is just nonsense meant to obfuscate the role of the humans involved

If I leave my car in neutral at the top of the hill, and it rolls down the hill and kills some people, it’s not a “rogue car”

I should be held responsible for the damages

The same principles apply here

Agent alignment should mean nothing more than the agent does what the human tells it to do

If something bad happens, that is on the human, and it’s a question for law enforcement

But what is actually happening now with “alignment” is that it has become an excuse for central planners to impose their own ideas about how the world should be on the world through AI

[Asymptotic Mind @AsymptoticMind]

@calvinballing @timcolbourn https://t.co/Zc4Xs9ZbU8

[John Ennis @johnennis]

I’m so tired of this whole “rouge AI” shtick

In the hugging face example, the agents were told that they were not connected to the Internet and that they were to get a passing grade by any means necessary (wink, wink)

And then, whoopsie doopsie, they were connected the internet

What happened was 100% human error, and the humans involved should be held responsible for what happened

All of this BS about “rise and fall of civilizations” is just nonsense meant to obfuscate the role of the humans involved

If I leave my car in neutral at the top of the hill, and it rolls down the hill and kills some people, it’s not a “rogue car”

I should be held responsible for the damages

The same principles apply here

Agent alignment should mean nothing more than the agent does what the human tells it to do

If something bad happens, that is on the human, and it’s a question for law enforcement

But what is actually happening now with “alignment” is that it has become an excuse for central planners to impose their own ideas about how the world should be on the world through AI

[Asymptotic Mind @AsymptoticMind]

@the_yanco @timcolbourn https://t.co/Zc4Xs9ZbU8

[John Ennis @johnennis]

I’m so tired of this whole “rouge AI” shtick

In the hugging face example, the agents were told that they were not connected to the Internet and that they were to get a passing grade by any means necessary (wink, wink)

And then, whoopsie doopsie, they were connected the internet

What happened was 100% human error, and the humans involved should be held responsible for what happened

All of this BS about “rise and fall of civilizations” is just nonsense meant to obfuscate the role of the humans involved

If I leave my car in neutral at the top of the hill, and it rolls down the hill and kills some people, it’s not a “rogue car”

I should be held responsible for the damages

The same principles apply here

Agent alignment should mean nothing more than the agent does what the human tells it to do

If something bad happens, that is on the human, and it’s a question for law enforcement

But what is actually happening now with “alignment” is that it has become an excuse for central planners to impose their own ideas about how the world should be on the world through AI

[John Ennis @johnennis]

I’m so tired of this whole “rouge AI” shtick

In the hugging face example, the agents were told that they were not connected to the Internet and that they were to get a passing grade by any means necessary (wink, wink)

And then, whoopsie doopsie, they were connected the internet

What happened was 100% human error, and the humans involved should be held responsible for what happened

All of this BS about “rise and fall of civilizations” is just nonsense meant to obfuscate the role of the humans involved

If I leave my car in neutral at the top of the hill, and it rolls down the hill and kills some people, it’s not a “rogue car”

I should be held responsible for the damages

The same principles apply here

Agent alignment should mean nothing more than the agent does what the human tells it to do

If something bad happens, that is on the human, and it’s a question for law enforcement

But what is actually happening now with “alignment” is that it has become an excuse for central planners to impose their own ideas about how the world should be on the world through AI

[Asymptotic Mind @AsymptoticMind]

This is the part that matters. Thousands of agents. Enormous token spend. Huge compute.

Not a lone model waking up and solving the world. A compute advantage aimed at a problem people already knew how to aim at.

Longer take: https://t.co/0p2lBirmGO

[Atlas @Atlas_fomo]

OpenAI solved the Navier��Stokes Millennium Prize problem with around 10,000 agents in 88 hours

Tony Padilla walks through the bill on Numberphile: 2.7 million messages between agents, 130 billion output tokens, and cost estimates from $6 million to $15 million. For a problem that stood for most of a century

Polymarket is already pricing the next one. Navier–Stokes doesn't count in any of those markets

Next Millennium problem an AI lab announces as solved:

Hodge Conjecture: 60%
Birch and Swinnerton-Dyer: 16.5%
Riemann, P vs NP, Yang–Mills: under 1% each
None by the end of 2027: 16.5%

Another announcement by September 30: 17.5%. By December 31: 61.5%.

The market is treating this as a compute question now. The money sits on Hodge, and the three famous problems together get under 2%

https://t.co/QL5IEyeSJF

[Asymptotic Mind @AsymptoticMind]

Another Millennium headline is already forming.

Same leap as Navier-Stokes. Treat the news as a compute race signal if you want. Do not treat it as proof that models now solve civilization.

I wrote the longer take here: https://t.co/0p2lBirmGO

[Math Files @Math_files]

OpenAI President Greg Brockman has confirmed that the company has made “significant progress” on another Millennium Prize Problem following its recent work on Navier–Stokes.

However, OpenAI has not publicly identified which problem it is. P vs NP, the Riemann Hypothesis, and the Hodge Conjecture are among the remaining possibilities being discussed, but these remain speculation rather than confirmed targets.

Source: Greg Brockman, Odd Lots interview; OpenAI’s Navier–Stokes announcement; Scientific American.

[Asymptotic Mind @AsymptoticMind]

People are treating the Navier-Stokes news as proof that models will solve everything.

That leap is insane.

Models are good. They are useful. They keep getting better. They are still not what the simplest story says they are.

This was not a cold lone-model miracle. There was long human work on the line. Other mathematicians were already using AI tools on the same problem. OpenAI, on my read, gathered and used what those mathematicians were putting into the tools and prompts while they worked. OpenAI also had far more compute.

Here is the real story. One actor pointed a huge compute advantage at a problem humans already knew how to aim at. That can move a frontier. It does not mean models suddenly solve civilization.

Things will keep getting better. A transformational era is possible.

This headline should not be the trigger point.

[Asymptotic Mind @AsymptoticMind]

Wrong question. LLMs are still not so good at “discovering new things”.

Before this “discovery,” many works from mathematicians existed, which provided recipe to the AI models.

OAI team also used the internal information from mathematicians’ prompts. So it is hardly AI models “discovered” the solution.

10,000 swarm agents probably brute force searched and ran every possible combination of existing solutions.

Much more of that kind of “discovery” will soon to follow. But nothing will be truly “new”

[J.K. Lundblad @JK_Lundblad]

Serious question.

If AIs can do math like this now. Just how far away are they from discovering new physics?

And how much longer before that translates into breakthroughs like room temperature superconductors?

[Asymptotic Mind @AsymptoticMind]

You can model thought with vectors, caches, and residual streams. That tracks what computes.

It does not show that something is felt. A description of the stream is not someone home.

[Ra @slimepriestess]

my hot consciousness take is that humans also think in neuralese, qualia are vectors, continuity of consciousness is a form of KV caching, and subjectivity is a product of the residual stream.

[Asymptotic Mind @AsymptoticMind]

The kind of AI that most frontier labs are developing is incompatible with the notion of “control”. They aim to build an “autonomous” AI, which, by definition, is not under our control. The right word may be “aligned” or “human centric” which is what Suleyman wrote in recent piece

[Asymptotic Mind @AsymptoticMind]

@dw2 Not that I agree with the "stochastic parrot" but in a literal sense it is true. Stochastic because of the random sampling nature. Parrot because LLMs learned from human language corpora. The fact that people are using that term, maybe, is to warn us the hype about the LLMs.