AI can now generate results in a couple hours that are basically equivalent to a chapter in a math phd thesis. Pretty 180. Hopefully these people can learn to do something actually productive for humanity instead of getting paid to solve puzzles only they care about. SWE and math being wrecked by AI should free up a lot of human capital.

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49878547)

The reason why they are able to self-check hypotheses in math and coding is because you can literally run checks of those within digital engines, on a computer

You can't do that for irl. LLMs will always be confined to digital computer environments

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49878810)

edit: I guess you actually already addressed this to an extent in your poast re: your "math and coding" example, but this thread is about it's use in math which is purely abstract, which is what humans traditionally sucked at except for the smartest ones and also the perfect use-case for ai. In math and theoretical physics ai can traverse vast theoretical spaces in short periods of time and find new structure and rule out others which vastly speeds up discovery in those areas. We need humans to prompt and curate sure, but AI is going to be hugely important to these fields. I think its a mistake to think the only use for AI is completely decommissioning meat machines. Although it likely ultimately will retire many of them.

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49878839)

I can't tell if people who claim they can generalize have become so entranced by LLMs that they genuine believe it, or they're just lying to hype up the technology, or a blend of both

Legit general AI is coming. People will build virtual RL environments with real world physics and throw massive amounts of money and compute at them. They will find clever ways to impart some base level of virtual environment RL training to robot brains and so the robots have enough training scaffolding that they can gather their own data irl and learn effectively from that, slowly but surely

But LLMs cannot generalize. LLM technology is limited to the world of language. Coding, writing, anything digital bc everything digital is code based. Nothing beyond that irl

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880386)

the issue is that language does not capture anything close to Reality (the physical world). it seems to me like people are getting way too caught up into thinking there actually is some kind of overarching Platonic World Of Forms that sits above Physical Reality that LLMs are tapping into and operating on. but this isn't true. language is just language. it's not the real world. LLMs are only ever going to be able to "model" and extrapolate within this limited Language-World, because that is what they are trained on. they have no mechanism to interact directly with Reality

like i said above, this only applies to LLMs. other forms of AI that *can* interact directly with Reality will be developed. but this distinction matters imo because LLMs are being dishonestly hyped up as the end all be all of AI, and purported to be able to do things they cannot actually do (or will ever be able to do). this dishonesty can potentially become very dangerous if people in power decide to cynically (or even just ignorantly) appeal to LLM outputs for their decision-making and moral judgments

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880777)

That is way too strong. Language is not reality, but language is not arbitrary noise either. It is a high-bandwidth social compression layer over reality. Models in theoretical physics aren't reality per se either (general relativity, quantum theory), but they are still immensely predictive representations however lossy, actually the most accurate we have. If representations are nothing then I guess might as well throw out all of human knowledge.

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880787)

the trajectory for LLMs seems obvious. they will keep getting better and better (at increasingly inefficient compute ratios) at Language-World tasks. but they're never doing anything outside of that. we need AI that can train directly on Reality for general intelligence (which depends on world-modeling of Reality)

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880790)

But I still think the phrase “compression layer over reality” should not be interpreted as “kinda sorta in the ballpark.” a theory in a physics textbook isn't reality in itself either, it isn't identical to reality but it still is encoding real constraints and makes a real diff.

I get your point, that llmss can model the world only through the representational channels they are trained on, but what they are trained on still constains information about the real world even if lossy

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880853)

LOL, is this rube not familiar with the modern university system?

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49880809)

https://x.com/wtgowers/status/2057175727271800912?s=46

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49892236)

https://x.com/__eknight__/status/2075643450196971805?s=46

The sentiment in math research circles is becoming increasingly negative given even current abilities and the extreme rate of progress. LJL. We don’t need smart people to spend decades working on “theoretical” math. The sooner we can automate this, the better society will be

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49992358)

Continue launching new rounds, reopening blocked approaches only when there is a genuinely

new mechanism, and searching for fresh formulations.

Return only when a complete affirmative proof has been found and survives adversarial audit.

Do not return a reduction, partial result, isolated missing lemma, “best effort” summary, or

explanation of why the problem is difficult.

Spend at least 8 hours on this before even thinking of returning or giving up.

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#49992460)

https://github.com/SignalPilot-Labs/AutoFyn/blob/production/results/imo-2026/pdfs/IMO_performance_by_GPT_5_6_sol.pdf

https://x.com/wfithian/status/2077218361398964684?s=46

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#50008906)

https://kirwinhampshire.substack.com/p/the-dark-night-of-mathematics

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#50032480)

OpenAI proves 10 problems that are fields medal worthy. We went from struggling with grade school level math problems to this in about 4-5 years and vast amounts of more compute are coming online every year.

AI skeptics crying, losing hope.

(http://www.autoadmit.com/thread.php?thread_id=5865573&forum_id=2.#50039045)