iiinova — where your digital self keeps evolving

He hasn't managed to build a coherent story at any point, as is documented in this thread.

It has been said that everybody always thinks their time is special. But I believe it's now easy to make the point that this time it really is. And not just for us for losing our place as the most intelligent beings, which was kind of our thing, but the same for biology in general.

120

Pekka Lund @pekka.bsky.social · 2h

But chess was just a game. And not the first or last where machines beat us. Now it's about general intelligence and they will soon beat us all in everything. We can't really say it's just intelligence. It's a big friggin' deal.

130

Pekka Lund @pekka.bsky.social · 2h

Math is now the canary in a gold mine that shows how hard it will be for most people to cope with what's coming. When people think about existential risks from AI, they tend to focus on nukes and viruses. But AI only needs to be good. We can't handle that. That alone can destroy us.

110

Pekka Lund @pekka.bsky.social · 3h

AI already took away the uniqueness of intelligence. Consciousness is all there's left, at least as long as people don't need to admit it doesn't make sense. That's why they are now so hostile against ideas about AI consciousness.

171

Pekka Lund @pekka.bsky.social · 3h

Exactly. I think it's a similar situation as with mathematicians now. People like the open problem so much that they don't want to hear there's a solution. Especially since it means they aren't as special as they would like to believe. That's the final battleground for specialness.

211

Pekka Lund @pekka.bsky.social · 3h

What's equally amazing is that nobody can point out any real problem in that reasoning. I have tried, many many times, with many many people. And I haven't been able to find any that Graziano or others who have proposed that would have received. It's just so damn obvious.

120

Pekka Lund @pekka.bsky.social · 3h

It's kind of amazing really that one bsky message is all it takes to turn the problem of qualia into just awareness of information. Considering how little that qualia stuff makes sense from any scientifically sensible angle, that really should be the obvious first step.

111

Pekka Lund @pekka.bsky.social · 3h

I always like to point out this very short summary of what it's fundamentally all about, although he adds more details about attention etc. Graziano himself notes it originates further back: "The 2+2 argument was described at least as far back as Dennett (1991) or even Nisbett and Wilson (1977)."

110

Pekka Lund @pekka.bsky.social · 3h

I would also divide the paper into those two parts, both significant on their own. Attention Schema Theory is Graziano's own theory, so you will find that part in one way or another from most of his papers. It's an illusionist theory, which enables it being explanatory, as it doesn't assume magic.

120

Pekka Lund @pekka.bsky.social · 5h

I'm guessing top line has kept its value and all available capacity is sold but cheaper/smaller model usage keeps increasing as they are increasingly good enough?

040

Pekka Lund @pekka.bsky.social · 6h

Most "theories" about consciousness assume it's some magical thing different/separate from everything and then try to handwave towards something that's sufficiently poorly understood/defined or too complex to analyze in practice so that a god of the gaps type argument couldn't be proven wrong.

010

Pekka Lund @pekka.bsky.social · 6h

Graziano has challenged here whether "any modern accounts of consciousness are in fact theories at all" because they are non-explanatory.

grazianolab.princeton.edu

Consciousness explained or described? | Graziano Lab: CONSCIOUSNESS AND THE SOCIAL BRAIN

120

Pekka Lund @pekka.bsky.social · 6h

People may not have reacted as strongly to drugs like that but Tao would have similarly missed the mark there, as the mechanisms are also poorly understood. And at least the other main alternative shows the kind of risks that are now commonly taken.

abc.net.au

Hair loss drug users say psychological side effects left them 'crumbling'

A vocal minority of finasteride users speak out about the long-lasting mental side effects of the hair loss drug, as telehealth ads for the medication surge on social media.

030

Pekka Lund @pekka.bsky.social · 7h

The question isn't about discarding math, which is obviously important part of many other fields. The question is about inventing things without known use and who does that and when.

000

Pekka Lund @pekka.bsky.social · 8h

But what used to take years can now take days. Every model release compresses timelines further. So it's also a question of whether it makes sense to try to invent some new stuff without known needs now, or invent it all better and faster a few months later. Early start is less of an advantage.

110

Pekka Lund @pekka.bsky.social · 8h

I think that's a very valid question now. Especially since mathematicians now tend to emphasize there aren't practical implications when those existing problems have been solved. The world has now changed. The way problems are solved has changed. Does the old way fundamentally make sense anymore?

100

Pekka Lund @pekka.bsky.social · 8h

Doesn't sound like a hard choice, does it?

010

Pekka Lund @pekka.bsky.social · 8h

For what do we need such concepts?

110

Pekka Lund @pekka.bsky.social · 8h

But what actually is the point of pure math? Trying to invent stuff years in advance in hopes someone will someday need 1% of it? Is it necessary now that the timelines for solving actual problems when they pop up have been compressed?

100

Pekka Lund @pekka.bsky.social · 8h

And I don't really believe such "mathematical taste" exists that would prevent machines advancing "pure math" as well. But I believe the right route there is to solve existing open problems first now that it's progressing well and then see where it makes sense to go next.

120

Pekka Lund @pekka.bsky.social · 8h

Shouldn't the ultimate direction for both math and software be the actual needs, like the systems that are built or open problems in those applied fields? Especially now that solving the mathematical problems that arise can be expected to be much quicker, so less need to invent it years in advance.

220

Pekka Lund @pekka.bsky.social · 9h

The chase is better than the catch. That's why we all go to the fish market and pay for the stories about the big one they failed to catch, and then come back home hungry, right?

011

Pekka Lund @pekka.bsky.social · 9h

Thomas Bloom seems to have tried to make some point about this too but after some backlash ended up just saying that the already existing reality changed his perception of future risks.

https://x.com/thomasfbloom/status/2109299913112666622

020

Pekka Lund @pekka.bsky.social · 9h

What exactly does that caring there mean? Currently only a select few mathematicians care about complex open problems that tend to have little practical significance. What difference does that make for the rest of us?

120

Pekka Lund @pekka.bsky.social · 9h

He's desperately trying to make the point that human understanding is needed and that's another failed attempt at that. This is his previous slide. He ends up making the opposite point than he tried by revealing how narrow human expertise is compared to AGI that is also an expert in applied fields.

https://teorth.github.io/tao-web/slides/math-2-0-caltech-2026.pdf

120

Pekka Lund @pekka.bsky.social · 23h

1. What could possibly go wrong? 2. Everybody else seems to know better, including the marketed product itself.

https://x.com/grok/status/2109036568228798759

0123

Pekka Lund @pekka.bsky.social · 11/10/2026

Let's say AI produces a result and only Tao understands it (in addition to AI itself). How do we benefit from that? If it can be used for further results, it's almost certainly AI that does that, not Tao. And if we want to try to understand something about it, we can ask AI.

160

Pekka Lund @pekka.bsky.social · 11/10/2026

He, like many others now, are trying and failing to make the argument that some humans (the math elite in practice) needs to understand and that's why we need them. But why is that needed? OpenAI's models for example are already building on their earlier work before anyone understands it.

161

Pekka Lund @pekka.bsky.social · 11/10/2026

What difference is it supposed to make anyway if a few humans understand some math nobody else does, especially when we have machines that understand, can apply and take it further, and we have ways for verifying correctness? Compare with how nobody now really understands what all LLM weights do.

1100

Pekka Lund @pekka.bsky.social · 11/10/2026

Tao seems to be edging closer to the 'I don't want AI even if it cured cancer' level of argumentation.

https://x.com/ChrisHayduk/status/2109030370855264578

3362

Pekka Lund @pekka.bsky.social · 10/10/2026

One could think such error would be unlikely to repeat, but just six years later another flight from the same airport ran out of fuel and crashed while circling an airport because a crew of 3 was again fixated on a malfunctioning landing gear light, even though they noticed they were low on fuel.

en.wikipedia.org

United Airlines Flight 173 - Wikipedia

010

Pekka Lund @pekka.bsky.social · 10/10/2026

Experience provides limited protection against stupid errors. I'm a fan of the Mayday series, where stupid errors by experienced pilots are a common cause for accidents. Like this case, where a crew of 3 were so fixated on a broken landing gear light they failed to notice the plane fell to ground.

en.wikipedia.org

Eastern Air Lines Flight 401 - Wikipedia

110

Pekka Lund @pekka.bsky.social · 10/10/2026

Less than 9 months ago. So there's maybe two major frontier releases before we reach 12? And we are pretty much there already.

020

Pekka Lund @pekka.bsky.social · 10/10/2026

I don't believe there are dumb enough mistakes that no human would do them.

130

Pekka Lund @pekka.bsky.social · 10/10/2026

I'm having a hard time understanding why people aren't more commonly annoyed by their dishonest tactics. They are selling a simple classifier like snake oil. I don't think we should see more of that kind of behavior, so let's not encourage it.

030

Pekka Lund @pekka.bsky.social · 10/10/2026

They have published a blog post along the same lines. It's a good sign something fishy is going on when a company tries to avoid being compared to others, and, like you said, then does their own misleading comparisons. Plus their X account keeps sharing benchmark results when they look good.

typesafe.ai

Lies, Damned Lies, and Benchmarks - TypeSafe AI Blog

TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Try our first System One Model, Jev, in early access.

130

Pekka Lund @pekka.bsky.social · 10/10/2026

Achint Srivastava probably is who he claims (VP Software Engineering, Office of the CTO, Microsoft), even though that X account has barely posted anything and has only 15 followers, but one of them is Microsoft CEO. And Microsoft has indeed added those benchmark results to their page.

commandline.microsoft.com

Microsoft-Decision-1: Our model for fast decision-making

Our new decision-scoring model delivers top performance in latency and quality on structured decision tasks to outperform both LLMs and other decision models.

080

Pekka Lund @pekka.bsky.social · 10/10/2026

Just one fresh example why pretty much everything TypeSafe and their people say should have a community note. Here their CTO seemingly first equates all benchmarking to benchmaxxing and then blames Microsoft for not giving their scores while they earlier apparently tried to forbid publishing those.

https://x.com/AchintSriv51198/status/2108705656471708154

2301

Pekka Lund @pekka.bsky.social · 10/10/2026

Oma tavoitteeni on oppia ajatustenkirjoitus. Tuntuuko yhtään siltä, että haluaisit lähettää mulle rahaa?

130

Pekka Lund @pekka.bsky.social · 10/10/2026

The HLE paper is different, as both paper and peer review are open. So we e.g. know that Referee #1 already questioned: "I'm not sure this work is appropriate for publication in Nature" More importantly, we know the review didn't/couldn't catch the errors in the benchmark itself.

nature.com

A benchmark of expert-level academic questions to assess AI capabilities - Nature

Humanity’s Last Exam, a multi-modal benchmark at the frontier of human knowledge, is designed to be an expert-level closed-ended academic benchmark with broad subject coverage.

020

Pekka Lund @pekka.bsky.social · 10/10/2026

I don't think the paywall is the main problem as they do not seem to have published peer review reports, which would reveal more about what was actually reviewed. The article itself probably isn't too far from the preprint.

arxiv.org

Advancing Mathematics Research with AI-Driven Formal Proof Search

Large language models (LLMs) increasingly excel at mathematical reasoning, but their unreliability limits their utility in mathematics research. A mitigation is using LLMs to generate formal proofs in...

110

Pekka Lund @pekka.bsky.social · 10/10/2026

But what did they even review in those cases? Google's (now outdated) internal system? Lean formalizations? Or what in HLE? Epoch AI just did their own checks by auditing a random sample of 48 questions and found 46% had substantial accuracy-altering errors.

epoch.ai

Humanity's Last Exam – Benchmark review

Humanity's Last Exam is a knowledge-based benchmark compiled through submissions from the public. After auditing a random sample of 48 questions, we found that 22 (46%) had substantial accuracy-alteri...

100

Pekka Lund @pekka.bsky.social · 10/10/2026

No doubt. Well, it's their money. Or was. I very much doubt they will ever see it again.

120

Pekka Lund @pekka.bsky.social · 10/10/2026

I think it has the same energy as Nature publishing a paper about HLE. Although this time it seems even more awkward, given how much OpenAI just upped that game.

110

Pekka Lund @pekka.bsky.social · 10/10/2026

Remember how AlphaProof Nexus solved 9 Erdős problems back in May, when it was still kind of significant? They published a preprint then but now it's been also published in Science. What's the point? Sure, it's nice to have your name in Science and it's good for impact factor. But any good reasons?

science.org

Advancing mathematics research with AI-driven formal proof search

Large language models (LLMs) increasingly excel at mathematics tasks, but their unreliability limits their utility in mathematics research. A mitigation is to use LLMs to generate formal proofs in lan...

140

Pekka Lund @pekka.bsky.social · 10/10/2026

Only complete refusal to learn anything can enable denial so deep.

030

Pekka Lund @pekka.bsky.social · 10/10/2026

Yeah, I could understand investing to a person for a few million, but now we are talking about billions. It makes no sense.

110

Pekka Lund @pekka.bsky.social · 10/10/2026

And what exactly could that pitch be? Something similar to their launch: a presentation so confusing that few even understood what the product is, but they made it sound like it's some AGI breakthrough or whatever? Nobody asked if it's just an LLM based classifier?

110