Achievements of Artificial Intelligence (AI)

[neurophusion]

The stakes are very high. Some of those who want AI to fail are greatly concerned about mass job losses, the large-scale violations of copyrighted material with the impact on artists, the enhancement of electronic surveillance and potential for autonomous robots used for a police state, the flood of slopaganda on social media, the escalating security threats, and the erosion of cognitive abilities in children.

Companies have been trying to use AI to replace workers, but in many cases it has not gone well because available models of AI were overhyped or too expensive to use. Others are deluged in 'AI slop' in emails. Companies have also been using AI as a cover for what they are actually doing: offshoring to India to boost profits.

I suspect that what is happening to AI will somewhat resemble the dot com crash and rebound. Overhyped claims lead to over-investment and a big correction. But from that earlier phase rose the greatest achievement in information technology the world has seen in all human history. The internet has since grown beyond the hype of the late 1990s even if somewhat different than imagined.

AI is here to stay in some form or another and keeps improving. The transhumanist claims about a technological singularity might be overblown, but AI is clearly the next step in the evolution of information technology. Regulations might be required as with any other technology or industry, but those regulations will not stop it from coming in the same manner people are hoping for with an AI equivalent of the non-proliferation treaty of nuclear weapons.

New technologies have often brought socioeconomic upheaval followed by a period of adjustment and AI is no different but it is early days and what happens now will set the stage for years to come.

 

[ChronicallyOverIt]

I do believe this guy touches on the societal impacts quite well, especially the general unrest for loss of jobs

 

[Sean]

Have not read it, so make no comment on quality and relevance.

I will say that from what I have seen of AI summaries of the science on ME/CFS I do not think it has so far demonstrated a sufficiently robust grasp of the critical core issue of methodology and its constraints.

AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices

Gary King, Steven Worthington
October 22, 2026

ABSTRACT
We show that AI-generated article reviews are now better, by every measure we could construct, than reviews written by well-compensated human experts and judged to be of unusually high quality by journal editors and authors. We then discuss how this result may inform journal policy choices, with concerning downsides as well as potential paths to increase publication speed and quality; reduce adversarial interactions; improve the experience of editors, reviewers, authors, and readers; and advance science.

 

[Utsikt]

The corpus includes every eligible submission to RapidPeer, including 124 scholarly paper submissions and 134 human reviews (10 submissions received two independent reviews) collected between 2019 and 2026 and drawn from a wide range of research, in- cluding statistical analyses of empirical data, formal mathematical modeling, and purely qualitative studies. Near as we can tell, all reviews were written by humans: many pre- dated ChatGPT; neither the quality nor style changed over time; and reviewers were com- pensated and responsible for their contribution.

I think it’s reasonable to expect the quality of the AI reviews to vary substantially between different fields of research.

We already know AI can perform very well in maths, which by extension also means it can perform well in statistics. Many of the recent maths proofs are supposedly terribly written, but still correct.

Qualitative research is riddled with questionable practices despite having the most easily understood limitations, and a good LLM model (or structure of agentic models) that’s told to follow the rules will have an easy time performing well.

I would expect a good model to be able to evaluate an RCT, it’s usually quite straight forward. Although assessing the appropriateness of an outcome might still be a challenge.

There’s also always the possibility that AI reviews will be used to redefine what constitutes «good» science, and it might not be trivial to pick up on its bias. That’s been going on for ages with humans, but we can’t just assume that AI will be neutral. Especially when we’ve already seem many examples of LLM models trying to hide its track from its assessors.

 

[rvallee]

A good measure of this is that they used Opus 4.8, which was the first model capable of generating code that isn't slop, but wasn't very good at it. Much less capable than Opus 5.5, and the agents used by OpenAI for the math problems and Anthropic for the software vulnerabilities are not available yet, at least one generation better.

It would pretty much compare to a smart student in their first semester, Opus 4.8, compared to a graduate doing a post-doc, Opus 5.5. Huge difference, even if only a few months out. Unreleased models are likely already capable of reviewing most of the published literature, at least from before 2020 (before the explosion).

What's interesting is that it will become obvious that a lot of papers straight up need to be unpublished, are pure slop, so it will pretty much go along with it that all of psychosomatic ideology and most of clinical psychology will be marked as unfit and recommended for unpublication.

Doing new research will happen a bit after, but reviews are pretty much solved already, it just hasn't been implemented. And this will be a huge help for us, probably as much as new findings.

 

[Utsikt]

What's interesting is that it will become obvious that a lot of papers straight up need to be unpublished, are pure slop, so it will pretty much go along with it that all of psychosomatic ideology and most of clinical psychology will be marked as unfit and recommended for unpublication.

That’s a good point. Just getting rid of the obviously flawed papers will help. How to make it happen is a different issue, it’s already really difficult to get the people in charge to take down egregious examples of fraud.

 

[hotblack]

Outside of the world of science two things that are quite incredible are the impacts on retro gaming and music tech

People are no longer using emulators to play old games but are making recompilations using AI. So there’s now native ports of PlayStation games like Gran Turismo and dozens of others playable on platforms they were never intended for, including browsers https://github.com/dubrovskiy-yevhen-stakelogic/gt-2-pc
There’s many many more https://allrecomps.com/ https://www.recompiled.games/

And there’s a flurry of activity around things like this portable synth. So people who never could before are now anke to create new firmware to do all sorts of stuff. The comments over on Ars are interesting as they are not a pro AI tech community but are really 8 to the possibilities here https://arstechnica.com/gadgets/2026/10/from-impulse-purchase-to-pocket-studio-the-m-vave-fm-1/

 

[Utsikt]

I wonder how long time it will take until GTA VI is recreated like that..

 

[bicentennial]

from what I have seen of AI summaries of the science on ME/CFS I do not think it has so far demonstrated a sufficiently robust grasp of the critical core issue of methodology and its constraints.

Maybe the executive order will help with super-intelligence, as notified in the bio-technical (medical) reference library: Until slowed down to reproduce this flag, I was instantly convinced the agencies now use super-intelligent technology in place of previous intelligencing technology. On slowing down, I admit this might just be another substitution in word-only:

NIH Websites Are Changing

• Get the Latest Updates and FAQs on the transition
• Per Executive Order 14434, federal agencies are now using "Super Intelligence" (SI) in place of "Artificial Intelligence" (AI). You may see both terms across NIH content during this transition."

However, the paper referenced in this bio-technical library, was:

The Role of NMDA Receptor Partial Antagonist, Carbamathione, as a Therapeutic Agent for Transient Global Ischemia

What concerns me is the ability of a bio-technical collaboration to tell and teach the difference between a human and a mouse. Once robotic lab-work is automated, the super-intelligent machine can grasp subjects to operate on, programmed to achieve, like the paper says how to simulate:

"... bilateral carotid artery occlusion (BCAO) for stroke. Whole-cell patch clamp recording in the mouse retinal ganglion cells was performed.

"Key proteins involved in apoptosis, endoplasmic reticulum (ER) stress, and heat shock proteins were analyzed using immunoblotting. Carb is effective in protecting PC12 cells against glutamate- or hypoxia-induced cell injury.

"Electrophysiological results show that Carb attenuates NMDA-mediated glutamate currents in the retinal ganglion cells, which results in activation of the AKT signaling pathway and increased expression of pro-cell survival biomarkers, e.g., Hsp 27, P-AKT, and Bcl2 and decreased expression of pro-cell death markers, e.g., Beclin 1, Bax, and Cleaved caspase 3, and ER stress markers, e.g., CHOP, IRE1, XBP1, ATF 4, and eIF2α.

"Using the BCAO animal stroke model, we found that Carb reduced the brain infarct volume and decreased levels of ER stress markers, GRP 78, CHOP, and at the behavioral level, e.g., a decrease in asymmetric turns and an increase in locomotor activity. These findings for Carb provide promising and rational strategies for stroke therapy."

Leave my NMDA receptors alone. Something already messed with them.

 

[ME/CFS Science Blog]

New breakthroughs on:

• Local smoothing and the Kakeya conjecture
• The Exponential Time Hypothesis is talse
• Existence of quantum chromodynamics and completions of the Standard Model

Working with academic communities to build scientific understanding | OpenAI

Would be interesting in hearing people's view of how significant this is if you have some background knowledge in any of these fields.

 

[EndME]

This is big, as everything thus far has been. But it was also anticipated. The frontier models are extremely powerful.

For the second problem there exists many people that are worried about the real world implications of AI’s progress on topics related to P vs NP, Peter Scholze is one of those. I think that is very hard to judge, but there exists a scenario that is extremely devastating to humanity for which there currently is no countermeasure.

The third problem is something I know more about mathematically. When the announcement of people working in the AGMAI was made, people working in the field were very concerned that Yang-Mills had been solved, because the group included Martin Hairer, who works in constructive QFT and in particular has been working on Yang-Mills for many years and the physicist Ed Witten, which is a very peculiar choice given his age. That’s why even just the announcement of the people in the AGMAI even before the solution to Yang-Mills was presented, already presented a crisis to many of the people in the field. I know some that left the field as a consequence. Looking at the timelines, that appears to be correct. Solving Yang-Mills as a subproblem just shows how good these models are. The team working on this now includes Jaffe, who is somewhat of the father of the approach that was later taken by Martin Hairer. To mathematically rigorously construct QFTs is extremely hard, keeping track of everything mathematically is too hard for most humans. Yang-Mills is particularly difficult as you additionally need a lot of background in many different fields.

On the other hand the AGMAI has also released a statement, that working with some select group of mathematicians in private, has been the wrong approach.

Some mathematicians are also becoming more relaxed, because the AI companies don’t care about mathematics anymore. It has already been conquered and now new benchmarks are being looked at. So to answer @Utsikt, if someone really cares about recreating GTA VI like that, it can be done with the same ease people say 1+1=2.

 

[hotblack]

This has been my point about motivations and resource constraints. They can throw compute at problems when they want to get attention. Be it maths, science or security. Then they get bored and move on or the money runs dry. It’s pure publicity nothing to do with supposed claims of the good of mankind, just like most of the other industry claims they are usual silicon valley stories to convince investors or avoid regulation. We’re different, we’re special, the usual rules don’t apply. We’ve seen this play out before with the internet boom and social media. Both with undoubted benefits and sadly significant harms. The latter of which we have still failed to address.

Anthropic’s filing before their IPO show yes, they have revenue increasing, but that growth has slowed, they also lost 80bn in the first half of the year alone and are like all these companies, so full of debt and intertwined with each other that even the revenue is difficult to be sure of. They have stakes in each other and commitments to each other for buying chips or compute from each other. Nvidia, Amazon, Google, Meta, Microsoft, OpenAI, Anthropic, SpaceX, Broadcom, Coreweave even seem to just keep raising debt and pumping money into each other. This is not the good sort of circular economy.

It’s an achievement that they have convinced so many up until now that none of the normal rules apply to them. If we are to harness this technology we need clear minds and oversight. We haven’t had that and hence we see the backlash building already and it’s such a shame.

 

[fst]

The Exponential Time Hypothesis is talse

Complexity theory is not my strong suit, but this appears to be unexpected, at least a bit.

The famous P-NP problem states that there is no polynomial runtime (aka "fast") algorithm for NP-hard ("difficult-to-solve") problems, which remains open. ETH states, that there is no subexponential algorithm for hard problems (specifically 3-SAT), which is stronger (thus, easier to disprove).

Proving ETH would have implied P!=NP and there are certain barriers results that show that our current methods are incapable of proving this. Disproving ETH, on the other hand, can be done by finding an algorithm that runs asymptotically faster than exponential time for 3-SAT and this is what has been done.

In theory, this means that we now have a slightly faster algorithm for certain problems, but whether that makes a practical difference or is only of theoretical importance remains to be seen. There are quite some result that have build on the assumption that ETH is true, which now have to be restated.

With these proofs by counter-example, there's also usually the option that the hypothesis survives with a small modification/tweak.

 

[EndME]

The famous P-NP problem states that there is no polynomial runtime (aka "fast") algorithm for NP-hard ("difficult-to-solve") problems, which remains open.

Yes. To address my previous point a bit more. People like Peter Scholze are not concerned about the P vs NP problem being solved, but rather it being constructively solved in a specific manner, essentially showing P=NP and that you even can design an algorithm that isn't too bad. That is the risky scenario and a scenario where companies might be very unlikely to release that they have solved the problem, because the power gained through solving it is larger than the mathematical credit you get, which these companies anyways don't care about. It's all very speculative and hypothetical, for example it might also be that a counterexample exists or only a non-constructive proof, but it's one of the few problems that can carry direct and instantaneous implications that reach beyond the mathematical world.

 

[rvallee]

New breakthroughs on:

• Local smoothing and the Kakeya conjecture
• The Exponential Time Hypothesis is talse
• Existence of quantum chromodynamics and completions of the Standard Model

Working with academic communities to build scientific understanding | OpenAI

Would be interesting in hearing people's view of how significant this is if you have some background knowledge in any of these fields.

Either winners or close contenders for 1) 2 Nobels in physics and 2) a Fields medals. Bigger than either of the quasi-Riemann or Navier-Stokes from a few weeks ago, and likely small potatoes compared to what comes in the coming weeks and months, which is the likely reason they wouldn't win.

Doesn't have explicit consequences yet, but it opens up many avenues of research while closing off several that had been held open based on things that have now been disproven, especially the Exponential Time Hypothesis, which makes significant progress on P=NP and opens up the possibility that classical encryption could be broken, but on that a lot of people are working on quantum encryption because it's pretty much expected that it will inevitably become necessary.

 

[rvallee]

They can throw compute at problems when they want to get attention.

Anthropic says they also had an internal proof of ETH, and from what they say the costs could have been as low as a few hundred dollars. This is the most incredible, it doesn't even take that much compute past training. So the implication is that it doesn't matter if they get bored with doing frontier science or math, the average academic can now justify using it for everything and expect results that would normally put their names in textbooks for less than a budget laptop.

 

[EndME]

Either winners or close contenders for 1) 2 Nobels in physics and 2) a Fields medals. Bigger than either of the quasi-Riemann or Navier-Stokes from a few weeks ago, and likely small potatoes compared to what comes in the coming weeks and months, which is the likely reason they wouldn't win.

If humans had obtained these results they would be definite Field medal winners, the same applies to Navier-Stokes. Even sufficient partial progress is Fields Medal worthy. But there's no point in awarding medals to computers.

So the implication is that it doesn't matter if they get bored with doing frontier science or math, the average academic can now justify using it for everything and expect results that would normally put their names in textbooks for less than a budget laptop.

Yes, this is definitely the case. The problem is whether Anthropic and other companies can justify anything if the open weight models are always just a year or even less behind. And the question for the academics is not whether they can solve problems at the computational cost of a potato, if one ignores the training computational costs, which is where the computational work takes place, but much rather whether there is a reason for why humans should still understand this or not because learning something fundamentally rather than being told what something is can have a different impact on developing understanding itself. Because if we don't care about understanding there's no point in having academics involved in any of the processes at all and the academics are very aware of this.

 

[Trish]

Doesn't have explicit consequences yet, but it opens up many avenues of research while closing off several that had been held open based on things that have now been disproven, especially the Exponential Time Hypothesis, which makes significant progress on P=NP and opens up the possibility that classical encryption could be broken, but on that a lot of people are working on quantum encryption because it's pretty much expected that it will inevitably become necessary.

An irrevant little aside: I first came across P=NP in an episode of the TV series Elementary (modern day Sherlock Holmes).

 

[Braganca]

Early progress on accelerating drug development with Claude

Over the past several months, we’ve partnered with six organizations to accelerate curing disease. We’ve also started drug discovery efforts of our own, including programs for rare childhood forms of muscular dystrophy. In this post, we provide an update on our progress.

www.anthropic.com

Anthropic’s life sciences team is working to accelerate curing disease. Over the past several months, our scientists have engaged in deep research partnerships with six organizations developing new medicines: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. We've also begun drug discovery efforts of our own. In doing so, we've learned a lot about how Claude can support drug development, and how we’ll need to improve Claude further to realize AI's potential in medicine. We plan to scale this work up substantially from here.

 

[Braganca]

Open call for guest investigators in the life sciences

Today, we’re inviting members of the life sciences community to apply to work with us on special projects.

www.anthropic.com

“Today, we’re inviting members of the life sciences community to apply to work with us on special projects. We’re looking for proposals from principal investigators and their research groups at academic institutions, nonprofits, or government labs who want to use Claude to investigate key questions in causal human biology, structural and computational molecular biology, medicinal chemistry, drug discovery, and clinical science.

You can submit a proposal for a project here.

Over 8–12 weeks, guest investigators will work directly with Anthropic researchers on a problem of interest and produce the analyses, artifacts, and tools that will be most helpful for them in their work. Our team will run our models and harnesses at no cost to them.”