AI megathread

It’s very far from human intelligence if this is the benchmark for agi. With that said I believe it’s only a matter of time before there will be world models that can recursively self-improve from continuous input. It’s certainly not possible to disprove we can build such thing.

My guess is we won't have AGI because humans intelligence isn't generally applicable.

Even mathematics and music, which seem to have some connection, doesn't by necessity produce great mathematicians who are also great musicians or vice-versa.

I think we'll have an improved version of what we have now - a general search function that can go to the necessary programs and produce output. 

For consciousness I think that we'll need to see if structure is necessary. I do think we will have synthetic (android) life but it won't be mere Turing Machines in structure. My guess is we'll need deep structures akin to what we have in our own brains and possibly the larger ecosystem of our bodies.

What I find remarkable about current LLMs is not that they can search for information or call tools. What's surprising is that a system trained only to predict the next token appears to acquire abstractions, world knowledge, planning abilities, coding skills, mathematical heuristics and many other capabilities that were never explicitly programmed into it.
Whether this ultimately scales to something resembling human-level intelligence remains an open question. But reducing today's models to "search functions" seems to miss the most interesting phenomenona - complex cognitive behaviour emerging from a surprisingly simple learning objective.

Nobody can explains why this happens - it just happens in these models.

The emergence of advanced cognitive abilities in purely physical systems should at least make us cautious about claiming that intelligence requires anything non-physical. Every year now seems to shrink the set of mental capabilities that appear to demand supernatural explanations.

The emergence of advanced cognitive abilities in purely physical systems should at least make us cautious about claiming that intelligence requires anything non-physical. Every year now seems to shrink the set of mental capabilities that appear to demand supernatural explanations.

If you define the "physical" as something that has no mental character, then it simply follows that thoughts are non-physical.

If anything LLMs would suggest Platonism, Panpsychism or Idealism are more plausible if they truly provide all the bonus features you claim they do.

It seems mental concepts are woven into the fabric of reality, recalling Wigner's question about the uncanny efficacy of maths.

If anything LLMs would suggest Platonism, Panpsychism or Idealism are more plausible if they truly provide all the bonus features you claim they do.

I actually don’t believe LLMs have any mental character.

An LLM is a very large collection of numbers. During inference, it performs a long sequence of matrix multiplications on numerical representations of text and produces the next token. That is all it does.

What unsettles me is how quickly their capabilities are evolving. Since the start of this thread, every three months seems to have brought another leap forward with increasingly impressive models. Yesterday, Claude Fable 5 was released and, judging from the early reactions, it appears to be another significant step forward.

While I remain cautious about whether these systems can truly invent something fundamentally new - such as proving a previously unproven theorem - I have to admit that I am less certain than I used to be.

I actually don’t believe LLMs have any mental character.

An LLM is a very large collection of numbers. During inference, it performs a long sequence of matrix multiplications on numerical representations of text and produces the next token. That is all it does.

What unsettles me is how quickly their capabilities are evolving. Since the start of this thread, every three months seems to have brought another leap forward with increasingly impressive models. Yesterday, Claude Fable 5 was released and, judging from the early reactions, it appears to be another significant step forward.

While I remain cautious about whether these systems can truly invent something fundamentally new - such as proving a previously unproven theorem - I have to admit that I am less certain than I used to be.

I'm still unconvinced any program running on a Turing Machine can become conscious, but if the fundamental nature of the world has mental aspect th[e]n it isn't that surprising patterns can emulate intelligence.

What I'm wary of is this idea that LLMs merely are word predictors, as if there isn't a lot of background human work to make them apply to multiple domain spaces.

I think we'd need a deep dive into *how* they work, made understandable to the public, before we decide exactly what level of intelligence they have.

I think we'd need a deep dive into *how* they work, made understandable to the public, before we decide exactly what level of intelligence they have.

Indeed. Right now, many believe that LLMs are, basically, "magic", and there lies all of my misgivings and worries ~ so many resources are being poured into it blindly. The LLM companies prey on this to make their billions from gullible fools who don't know how it works. The ignorance is part of what makes it work so well.

Which is why instead of replying to Laird, I thought that the above video would do a better job. Because unmasking the hype to get an understanding is crucial.

A court in 🇩🇪Munich ruled yesterday that Google was directly liable for incorrect answers generated by its "AI overview" feature in Gemini.

The legal dispute had centered on whether the service should receive the same legal treatment as conventional search results. Google's overview feature had erroneously linked the 2 plaintiff companies to dubious business practices, subscription traps and fraudulent schemes. It had linked the plaintiffs with information about other, genuinely shady companies and invented connections that did not exist.

The court ruled that the AI summary does not merely display or link to search results but constitutes “distinct content” with “independently comprehensible” statements, different from merely listing 3rd party links. The existing German jurisprudence protects search engine operators from direct liability regarding the simple listing of 3rd party content.

The court rejected Google’s defence that "AI-generated information should not be trusted blindly."

Google was ordered to “stop spreading false claims” and cover 80% of legal fees. Google will appeal.

https://x.com/Dovydas44444/status/2065732222510928213/

https://x.com/Dovydas44444/status/2065732222510928213/

I don’t think it’s new in 2026 that AI hallucinates especially when output is based on a single-step inference of the input as with google search.

The massive development is when models iteratively optimizes the outcome by generating chain-of-thoughts. Even though it’s a heuristic approach it’s tremendously successful in domains like mathematics and software development.

A great example is this recent article in nytimes https://www.nytimes.com/2026/06/08/scien...cians.html

Earlier this year, Sidharth Hariharan, a graduate student in mathematics at Carnegie Mellon University, received news that sent him rushing into his adviser’s office with tears in his eyes.

He had just gotten an email from Maryna Viazovska, a professor at the École Polytechnique Fédérale de Lausanne in Switzerland and a 2022 winner of the Fields Medal, the highest honor in mathematics.

For more than two years, she and Mr. Hariharan had been leading a team of six mathematicians in an effort to break down one of Dr. Viazovska’s most celebrated proofs into distinct logical steps, a task known as formalization.

But hours earlier, Dr. Viazovska had received a tip-off from a colleague: They’d been scooped.

Or, as she soon became fond of putting it, they’d “gotten Gaussed.”
Gauss is an artificial intelligence system built by Math, Inc., a California start-up. It had taken the team’s road map for formalizing Dr. Viazovska’s result — a solution to the densest possible arrangement of eight-dimensional spheres, popularly known as the sphere-packing problem — and completed it in just five days.

I don’t think it’s new in 2026 that AI hallucinates especially when output is based on a single-step inference of the input as with google search.

LLMs don't "hallucinate" ~ LLMs algorithmically pattern-match what the next most probabilistic token should be based on training data and what exists in the context window. There are no "errors", no "delusions", nothing but tokens.

The massive development is when models iteratively optimizes the outcome by generating chain-of-thoughts. Even though it’s a heuristic approach it’s tremendously successful in domains like mathematics and software development.

There is no such thing as "chain of thought" for an LLM ~ there is a context window which just adds to what the algorithm has to work with for predicting the next most probable token in the existing output chain.

LLMs only appear to be "thinking" because we are deceived by patterns we unconsciously associate with intelligence. Very simply, LLMs are designed to mimic human language patterns.

LLMs are absolutely abysmal at mathematics ~ LLMs only appear to be "good" because they have been trained on the data in question, so it is easy to just cheat and pattern-match. LLMs are also abysmal at software development ~ again, any apparent "good" results are because of pattern-matching against training data, and even then, the results are anything but deterministic. If you start with a fresh context each time, and ask only a single query, you will not get a deterministic answer, which makes LLMs extremely bad for something as deterministic as computer programming. CPUs need explicit instructions as they are deterministic ~ or as damn close as we can get with all the fault tolerances built into computer hardware to deal with random fluctuations in electrical current.

Valmar, I'll be honest about why I rarely engage with your posts: it's consistently clear that you're arguing about fields you haven't studied, and it shows. Not in a subtle way. This post is a good example.

Claim: "LLMs are abysmal at mathematics."

Clearly false. One example (note how I link to references when making strong claims)

The tipping point came in the summer of 2025. That July, several artificial intelligence models solved five out of six problems at the International Mathematical Olympiad, an annual challenge for some of the world's best high school students. But while mathematicians were shocked — few had expected the programs to get that good that quickly — the impressive results didn't necessarily mean that AI would make important strides in research math. After all, Olympiad problems are challenging puzzles with known answers, not open questions.
Nevertheless, the results made people pay attention. Mathematicians who had dismissed AI models as too error-prone to be useful started playing around with them. Those early adopters found, to their surprise, not only that the models were good at puzzles, but that they could help break genuinely new ground. Soon, mathematicians were using AI to discover and prove new results, accomplishing in a day what would have once taken them weeks or months. "2025 was the year when AI really started being useful for many different tasks," said Terence Tao (opens a new tab), a prominent mathematician at the University of California, Los Angeles.

https://www.quantamagazine.org/the-ai-re...-20260413/

The International Olympics in Mathematics isn't for the average high school student. You have to qualify and be extremely outstanding on a national level to qualify. The problem set are "new problems" not explicitly defined before so all your pattern-match BS does not apply here.

Claim: There is no such thing as "chain of thought" for an LLM

"Chain of thought" is a well known llm terminology defined in this paper https://arxiv.org/abs/2201.11903

—————————————————
The claim about software engineering is also demonstrable off, as the adoption of the coding assist tools is so abundant that Microsoft in June effectively increased the cost of using Github Copilot with what for many amounts upward to a 10x increase. Do you think companies would pay those money without benefit?


Even your comment about random current fluctuations in semiconductors is factually wrong. Logic gates operate on discrete voltage thresholds precisely to be immune to minor fluctuations. This is undergraduate-level electronics, I'll leave the rest for your own research.

I don't say this to humiliate you. I say it because a debate forum functions on the assumption that people have done at least basic due diligence before asserting things as fact. When that's absent, it wastes everyone's time - including yours.