Yes, AI models actually understand things and are not stochastic parrots. David Shapiro does this video (of AI-generated animations).

On October 15th, a tech startup CEO named Gregory Curtzer posted, allegedly, on X:

"An LLM doesn't understand anything. It is a statistical model that spews out tokens based on probabilities calculated through massive matrix operations and context, not an understanding of the subject matter."

I like Scott Aaronson's response to this sort of thing, which is whenever someone says AI is "just a ...", he asks, "What are you 'just a ...'?" An AI is "just a" probability calculator, and you're "just a" blob of neurons and synapses carrying out chemical reactions.

This video explores a rather ingenious way of testing the "actually understand" hypothesis:

A model was trained only on the moves of the game Othello (a variant of Reversi). Then the model was cracked open, and inside the researchers found an internal picture of the board. The model was never given a picture of the board, only the moves. It figured out on its own that it needed to create a picture of the board and then created it.

The proof that the picture was real and what the model was really using was when researchers changed that picture, the model changed its moves.

The implication is when people ask, does a large language model actually understand whatever subject it is talking about with you, or is it "just" statistically predicting next token probabilities. While statistically predicting next token probabilities is part of what it does, it actually does have actual understanding of what it's talking about. If it's writing code for you, yes, it has actual understanding of writing code.

Obviously, in order to do this, the model has to have enough capacity. The Othello playing model was small, but today's largest coding models have more than 10 trillion parameters. That leaves room for a lot of understanding what their input and output tokens mean.

The Othello research was actually done in 2023 -- it's the animation to visualize it that was only done just now (and in order to argue with a post on X). See below for links on that.

Does AI Actually Understand? YES! - David Shapiro

#solidstatelife #ai #genai #llms #reversi #othello

Not self-improvement, but models helping each other out.

Researchers from the Massachusetts Institute of Technology (MIT) Computer Science & Artificial Intelligence Laboratory (CSAIL) reported today that they used OpenAI’s GPT-6 Astra language model to raise a small, open source Qwen model’s math score roughly sixfold by directly editing its weights—the numerical parameters that govern its behavior.

In the strongest language-model example, Astra’s intervention raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% while touching approximately 0.016% of its weights.

The improvement came primarily from correcting repetitive reasoning—a distinction that shaped both the technical findings and the debate over their implications. Before the intervention, Qwen frequently repeated text until it exhausted the permitted output length. The edit reduced that obstruction and allowed the model to complete more answers using abilities it had already acquired. The experiment therefore demonstrated a form of model repair. It did not establish that Astra had inserted six times as much mathematical knowledge into Qwen.

#AI

venturebeat.com/technology/lik…

A nice overview of the project to decipher the Hercalaneum scrolls. No new announcements, but the work continues. Then there's the second project, which I hadn't heard of before :

Enter the RePAIR project – short for Reconstructing the Past: Artificial Intelligence and Robotics meet Cultural Heritage. Its mission is to use cutting-edge AI and robotics to digitally reassemble frescoes, including the House of the Painters at Work, before handing the final reconstruction process over to a robot that will physically piece together the fragments.

“Our dream one day is that archaeologists can find a bunch of fragments of artwork while excavating, lock them in a room, and then come back the next day to find them reassembled,” says Prof Marcello Pelillo, a professor of computer science at Ca’ Foscari University of Venice, in Italy, and a key figure in the project.

The fresco pieces have been scanned using hyperspectral imaging, which captures data from a range of wavelengths beyond what the human eye can see. This allows the AI to detect degraded pigments, revealing details lost over time.

Once the digital reconstruction has taken shape, the robot steps in. Using its mechanical arms and hands, it translates the AI’s digital solution onto the real puzzle, physically moving the fragments into place.

Importantly, human experts remain a crucial part of the process. The AI operates within a ‘human-in-the-loop’ system, meaning archaeologists provide feedback and refinements, guiding the algorithm toward a final, accurate solution.

#Archaeology
#History
#AI

sciencefocus.com/future-techno…

Humans have been very bad at AI forecasting, even those who are normally good at predictions.

Why did the best forecasters do so badly? I’d speculate it’s because the best way to predict short-term geopolitical events is to assume things will stay the same, or at least remain on trend, while AI is highly trend-breaking.

An incorrect implication of all this is that “expert views are useless”. Experts in AI were better at forecasting progress than the general public. And the main way experts have been wrong is not expecting progress to be even faster. In other words, they’ve been underhyping AI, rather than hyping it.

The AI company leaders, the people most accused of hype, are among the few who’ve erred optimistic. We don’t have many quantified predictions, but the ones that have resolved mostly came true, just sometimes later than suggested. Here are the ones I could find:

• Jan 2025: Altman said agents would join the workforce in 2025. They trickled in but took off in early 2026.
• Jun 2025: Altman said 2026 will see AI that can have novel insights (correct)
• Oct 2025: Altman said an intern-level automated AI researcher would arrive by Sept 2026. OpenAI says it landed Sept 2026 (self-graded, but seems close)
• Mar 2025: Dario said AI would write 90% of code in 3–6 months. Anthropic’s own codebase hit 80% after 14 months (too optimistic).

At the broadest level, people at the companies have consistently said that scaling isn’t hitting a wall, that more powerful capabilities will be here soon, and they’ve been right. Today, all three leaders seem to be saying pretty broad beyond-human AGI will be here in 1-3 years.

Rather than dismissing what the companies tell us as “hype”, talking about stochastic parrots, or continually pointing out “fundamental” limitations (which get overcome months later), the best AI forecasters take what the companies say seriously, but tend to assume things will take just a bit longer.

Though they seem very bad at predicting when it will become profitable.

#AI

benjamintodd.substack.com/p/al…

Riiight.

The man, 45-year-old David Durman, identifies as a humanist and told Religion News Service that he felt being asked to use AI tools like ChatGPT and Microsoft Copilot, as the University of Kentucky reportedly encouraged its employees to utilize, undermined his belief in the importance of the human parts of life and ran counter to his desire to see human flourishing.

“The process and work involved in thinking through situations, through my work, through my ethics, through the way I handle other people, is a key component in being able to develop myself as a moral person,” Durman told Religion News Service. “And offloading that labor, regardless of whether it’s taxing or not, is undercutting my ability to be a moral actor.”

In the letter he submitted to the University of Kentucky’s Office of Equal Opportunity as part of the accommodation request, Durman stated, “Substituting automated machine processing for human presence, authentic listening, and personal responsibility in my work directly conflicts with my obligation to uphold human dignity and flourishing as primary.”

What a complete loon.

This might sort-of work if he'd gone for "it's degrading my ability to think clearly and critically and therefore engage with people responsibly". It would at least be somewhat defensible in the case of preventing addiction, if you really feel you lack the self-control not to use AI for literally everything. As it is, he sounds like he thinks that anything that makes his life easier undercuts his moral worth for some reason. Wouldn't it actually be more moral to use AI as much as possible to do the drudge tasks that prevent you from doing the human side of things ? Free you up from editing spreadsheets so you can concentrate on the things that more directly matter ?

I think the religious angle is pretty dumb. If your job requires you to use specific tools, then that's what you've got to do. You don't get a religious exemption from using a tractor because you object to the internal combustion engine, or at least you shouldn't. At the same time, employers shouldn't require you to use any particular tool provided you can demonstrate equal performance using alternatives.

#AI
#Bullshit

gizmodo.com/man-who-identifies…

So far I can tell, the real-world implications of this are "sod all" on the direct front. Which is fine with me - God knows I'm in no position to judge.

In terms of demonstrating what AI can do, it's bit of a mixed bag. For raw capabilities this appears impressive :

For decades, Sarnak has argued that mathematicians should approach the Riemann hypothesis by considering the zeta function as just one member of a broader family of functions, and then proving statements about that family. “But it was not clear exactly how to do so, or which family to consider,” said Michael Rubinstein(opens a new tab) of the University of Waterloo.

As it turns out, Sarnak had the right intuition. But the family that OpenAI’s model identified had previously been overlooked. Finding the right family required obscure mathematical objects called metaplectic forms, which Dunn has spent the past five years studying. “Nobody expected this,” he said. Metaplectic forms had been somewhat uncommon in number theory until now, he added, “so I hope this beautiful and esoteric theory gets the attention it deserves.”

In retrospect, Radziwiłł said, mathematicians who study the branch of number theory that all this work belongs to were mistaken in thinking that their tools weren’t strong enough to resolve the quasi-Riemann hypothesis. They’d been convinced that proving it would require techniques from another branch of number theory entirely. The AI model’s “belief” that their tools might be enough was a crucial perspective. “That was a pretty bold belief to have!” Radziwiłł said.

Whereas in terms of explanatory power, it's rubbish :

As with many recent AI-generated results, both papers “are horribly written (in a way that would lead to immediate rejection at any mathematical journal),” Radziwiłł said in an email. “They fall short of the standard of exposition and clarity in our area by a wide margin.”

In my experience, GPT 6 is great at writing code to do the things I want to do. It's excellent at immediately finding gaps and problematic interconnections i wouldn't find for months, or necessarily even care about. It's far more likely to fix edge cases than I would bother with. It even does a good job of documenting its plans and explaining what it did afterwards.

Where it does considerably less well is in designing GUIs, which remains (as in GPT 5) passable for testing but completely inadequate for real use. And in terms of writing user guides, it's a miserable failure. It confuses development documentation with user instructions, and I cannot make it understand that intended users aren't computer scientists. One example of extreme and pointless pedantry : it wanted to correct "approximately two" to 2.154. This is after giving it custom instructions explaining the intended audience and project scope.

#AI
#Mathematics

quantamagazine.org/ai-proved-t…

Water, water, every where,
And all the cores did think;
Water, water, every where,
And many drops to drink.

The structure of a MOF looks somewhat “like a jungle gym on a playground,” says Seth Cohen, the dean of the School of Physical Sciences at UC Irvine. MOFs are essentially “molecular sponges—they’re a solid material that’s very, very good at absorbing other molecules,” says Cohen, who is not involved with Atoco. Like a sponge, the materials can also be wrung out using methods including heat to release whatever they’re holding onto.

These jungle-gym-like frameworks create open spaces on the microscopic level that can hold an unbelievable amount of material, and small tweaks can yield even more interior space. “The footage of a gram of MOF could cover two football fields,” Yaghi explains. “They can store hydrogen in there to make clean energy, or you could store CO2 in there to make clean air. Or, in this case, you can store water.”

“The scientists reach a level of precision [where] we really know even the angle of the atom,” in a given material, Taha says.

In Atoco’s technology, air passes through the MOF, which has been specially tailored to attract water molecules. Those molecules adhere to the pores. But the material feels completely dry to the touch even when saturated, and squeezing it doesn’t release a single drop. Instead, the company relies on heat to release the captured molecules.

There are other means of harvesting water. Cohen notes that silica, a popular desiccant, can also pull water out of air, but it doesn’t hold as much moisture and requires you to “heat the heck out of it to get the water back out,” he says. The MOFs in the device can harvest and release water with heat as low as 100 degrees Fahrenheit (38 C), Taha says.

#Science
#Technology
#AI

archive.li/bvQ6A

Worth a read. Most of the paper is appendices, which I skipped.

We present a suite of 50 unpublished research-level theoretical physics problems with verifiable answers. Without AI assistance, a typical problem in our collection could be solved by a world expert in about a day or less but would take considerably longer for experts in adjacent specialties. We test the ability of off-the-shelf AI agents to tackle these problems autonomously, without human hints. When given access to tools available in a realistic setting, including the ability to search the literature and write scripts for numerical and symbolic work, the best-performing agent, GPT-6-Astra harnessed in Codex, obtains the final answers with a near-perfect accuracy of 98% at an average cost under USD 3 per problem. On average, agents show only small changes in accuracy when denied web search and access to background research papers, suggesting significant memorization of the literature. These results suggest that state-of-the-art agents can accurately perform concrete calculations required across theoretical physics research. We comment on the broader implications of these findings and discuss possible explanations for the current lack of agent-led breakthroughs in theoretical physics despite their outstanding problem-solving abilities.

That last point they take more literally than I was expecting. AI isn't leading breakthroughs, they say, because it's not very good at deciding what problems to solve. They don't really discuss this though, they just assert it.

It was interesting to get the perspective of other institutions on my recent trip. Some are much more AI-reluctant than my local Czech group. But pretty much nobody is skeptical that AI isn't good enough any more, or concerned that it makes too many mistakes. Yes, of course, it still does make mistakes, but it feels now as if we've reached the point where the mistake rate is acceptably low. Instead, what people are worried about is that the problems will arise from AI being too good.

#AI
#Science

arxiv.org/abs/2610.16733