Master AI (Artificial Intelligence) Discussion/News Thread

[somebodyelse]

I feel like there would be far more caution if there were real consequences for this stuff rather than the “oh isn’t AI clever” marketing that Anthropic/OpenAI etc etc seem to use it for.

I mean hacking is illegal right? Why is no one going to jail for this?

We've already had a company having to stick to the incorrect terms given to a customer by their AI customer service bot like they would have to if it had been given by a human customer service employee. We'll have to see what happens in the multiple cases related to mental health and suicides, but humans have been held liable for similar behaviour, and companies liable for the criminal behaviour of their employees when using company resources if they didn't exercise sufficient control over their employees.

On the OpenAI/Hugging Face case I don't know how it works in that jurisdiction - it may be that action can only be taken if there's a complaint, and it may require intent rather than strict liability. Similarly for the ASAI case MS won't want to set a legal precedent for companies having responsibility for AI actions so won't make a complaint, and intent would be difficult to show given the prompt action in stopping it when found, and working with GitHub and the repo owners to clean up the mess.

 

[Kremmen]

The Americans that cling to both that right and to their arms wouldn't know when to use them even if it was written in bold on their foreheads. It's the same crowd that voted for the crap that put Americans where they are today.

I've always found reading bold letters written on one's forehead incredibly challenging.

 

[mhardy6647]

not exactly on topic (but, especially considering this is me posting, not that far off, either):
From the barrage of emails I receive from The Economist thanks to the gift subscription we supply to our daughter (who is an economist, as it happens).

I guess The Economist hired a former copy writer from Variety to write their email subject lines?
Heck, maybe AI crafted it!
At any rate, it made me smile (the headline, not necessarily the content).

 

[fpitas]

not exactly on topic (but, especially considering this is me posting, not that far off, either):
From the barrage of emails I receive from The Economist thanks to the gift subscription we supply to our daughter (who is an economist, as it happens).

View attachment 554681


I guess The Economist hired a former copy writer from Variety to write their email subject lines?
Heck, maybe AI crafted it!
At any rate, it made me smile (the headline, not necessarily the content).

Let's hope that the bubble bursts gently. The various companies that supply AI and its necessary parts are hopelessly intertwined economically. If things go bad, there will be dire consequences for the economy. Of course, the tech bros will simply walk away whistling with thousand-dollar bills falling out of their pockets.

 

[TK_IIVI]

Let's hope that the bubble bursts gently. The various companies that supply AI and its necessary parts are hopelessly intertwined economically. If things go bad, there will be dire consequences for the economy. Of course, the tech bros will simply walk away whistling with thousand-dollar bills falling out of their pockets.

My instinctual feeling about it is "too big to fail". If that's the case then none of the big investors are harmed. They either reaped far more in profits than they will in losses or they'll restructure or do some accounting shenanigans, and possible the gov will just print out some fresh new debt to buy them up. And the gov buying up a bunch of AI sounds like a profoundly bad idea. I suppose the only reason they wouldn't want it is plausible deniability once laws and the constitution starts getting violated. Better to target a private entity for that. Just have it go belly up and move on.

 

[fpitas]

none of the big investors are harmed

I think the smart ones have a good idea about AI. My financial guys certainly are not impressed.

 

[mhardy6647]

Let's hope that the bubble bursts gently. The various companies that supply AI and its necessary parts are hopelessly intertwined economically. If things go bad, there will be dire consequences for the economy. Of course, the tech bros will simply walk away whistling with thousand-dollar bills falling out of their pockets.

FIFY

I think the smart ones have a good idea about AI. My financial guys certainly are not impressed.

Ditto, FWIW. Ours are not anti-AI (EDIT: from an investment perspective, that is), but they're skeptics. My kind of collaborators.

 

[mhardy6647]

We certainly don't need dozens of half-built data centers silting in across the country in a latter day Dustbowl
(cf. the never-occupied housing towers in China after their real estate hiccup).

sideways, but not entirely irrelevant: Vermont's "Northeast Kingdom" is finally seeing some signs of hope for a big hole in the ground ("The Pit") in the struggling town of Newport, VT.

Kingdom Con: The history of Vermont’s biggest fraud case

Another chapter is set to close on the Kingdom Con. Thursday, Bill Stenger, the former president of Jay Peak, will be sentenced for his role in the massive fraud. Our Darren Perron looks at how we got here.

www.wcax.com

https://www.burlingtonfreepress.com/story/news/local/vermont/2026/08/25/state-grant-supports-purchase-of-newports-former-pit-property/91307933007/

 

[Power Pop 23]

Researchers in Cambridge, MA are training A.I. on how experiments are conducted - in hopes of finding methods that produce fewer failures.

gift link

Some Scientists Have ‘Magic Hands’ in the Lab. This A.I. Is Learning Why.

Even the most adept researchers may not know exactly what they do to get successful results. An A.I. model is trying to figure it out by watching every move.

www.nytimes.com

 

[Somafunk]

This was a good discussion regarding the “supposed” a.i. hacking incident

No, AI Is Not "Autonomously Hacking" with Cal Newport

Podcast Episode · Better Offline · 26 August · 1hr 5min

podcasts.apple.com

 

[Power Pop 23]

The University of California at Los Angeles (UCLA) is launching a department offering a major in 'Digital Humanities' this coming Fall 2026.

Culture and ethics meets tech: UCLA debuts country’s first bachelor’s in digital humanities

UCLA is launching the nation’s first digital humanities department and bachelor’s degree this fall. The new discipline combines classical humanities with technical know-how to examine a digitized and increasingly AI-enhanced world.

apnews.com

 

[JustAnandaDourEyedDude]

Thomas Wolf, one of the three founders of Hugging Face, is interviewed in a seven-and-a-half minute long video on the Bloomberg Tech YT channel. He discusses HF's user-programmable robot MicroDuck, which they are currently selling at the rate of one every four seconds. However, he only has a smiling "no comment" in reply each time when he is asked twice by the interviewer, at the start, as to whether Hugging Face is about to be acquired by Nvidia for just under $13B.

 

[RickS]

Thread took an off topic political turn and needs some maintenance.

Closed for now.

 

[Newman]

Thanks, my bad for relying on AI to cross check. Based on a previous post, I had asked Gemini whether LEDE was used in contemporary recording studios. The response was No, that it is obsolete. Asked again today and the answer has changed to Yes! Gemini must have used limited or incomplete data for its original conclusion. In any case, clearly the AI response is not very reliable.

This is exactly why I lobbied for ASR to discourage repeating AI responses to questions. It will misinform as much as it will inform, with the net result being that it will non-inform and obscure.

It is also why I was pleased to see Amir include within the site policy the advice to only quote AI here if we already know more than AI on the question in hand, and know for sure that it is correct. Which, in turn, effectively means we may as well not use it, since we already know the answer and probably where it originates.

cheers

 

[SigbergAudio]

This is exactly why I lobbied for ASR to discourage repeating AI responses to questions. It will misinform as much as it will inform, with the net result being that it will non-inform and obscure.

It is also why I was pleased to see Amir include within the site policy the advice to only quote AI here if we already know more than AI on the question in hand, and know for sure that it is correct. Which, in turn, effectively means we may as well not use it, since we already know the answer and probably where it originates.

cheers

Posting AI content here will potentially also have an interesting, self-amplified circle of bullshit, since the AI uses forums such as these as a source.

 

[Winemaker]

I asked GROK, one of the most up-to-date AIs, the following question after reading this article:

In light of this article: "AI could threaten humanity by 2030, warns resigning Anthropic researcher. Jacob Coxon, 27, steps down and accuses major labs of racing toward self-improving superintelligence without sufficient safety guardrails." What’s your take?

This was its response:.......

This is a real and recent event (resignation announced on September 8, 2026), not a fake story. Jacob Coxon, 27, a mathematician by training (Cambridge), spent three years working on pre-training frontier models—first at OpenAI (including work on GPT-4o) and then at Anthropic starting in May 2026. He resigned publicly via an X thread that reached tens to hundreds of millions of views, accusing both Anthropic and OpenAI of "racing headlong toward self-improving superintelligence and playing with our lives." He explicitly stated that many people inside these labs genuinely believe AI could kill everyone by the end of the decade, and that internally people are talking about "crunch time" / "the endgame."

The key takeaway isn't just his opinion: Evan Hubinger, who leads the alignment science team at Anthropic, responded publicly and confirmed the substance of it. He wrote that Coxon is right, that "we really do seriously believe AI could kill all humans," and placed his personal probability at >10% within the next decade. He added that while Anthropic is trying its best, "we don't yet have a plan to solve superintelligence alignment, and we are clearly not on the right track." Other current and former insiders have expressed similar views. Coxon also left unvested equity on the table (having only been there a few months), which makes the idea of this being a pure publicity or financial stunt far less likely.

My takeaway, in short​

1. The signal is serious and shouldn't be brushed off as panic or marketing hype.
   
   When people working daily on pre-training and alignment say these things (not just external "doomers"), and a head of alignment at Anthropic confirms double-digit odds of existential risk, it's not noise you can just ignore. Competition among OpenAI, Anthropic, Google, and Chinese labs creates a "race" dynamic where safety trade-offs become easier to justify. The alignment problem for systems capable of recursive self-improvement (RSI) remains unsolved—by their own admission.
2. Aggressive timelines (out of control by late 2027 / existential risk by 2030) remain speculative.
   
   No one has hard proof that self-improving superintelligence will arrive that quickly. Current models, even the most capable, aren't there yet. However, the pace of progress over recent years has exceeded many experts' expectations, and capabilities around agency, autonomous coding, research, and hacking are growing fast. "It might happen faster than we think" is a reasonable hypothesis to take seriously, though not a certainty. Coxon himself distinguishes between today's models—which aren't an extinction risk—and the shift that happens once AI effectively begins conducting AI research on itself.
3. Anthropic isn't the "bad guy" here relatively speaking.
   
   It positions itself (and to some extent is) as the most safety-conscious lab. That's precisely why Coxon went there after OpenAI. The fact that someone is leaving that lab with these warnings is more concerning than if they had walked out of a lab that never addresses the topic. Competitive pressure (including from China) is real and makes slowing down unilaterally very difficult.
4. What's missing from the public conversation.

• Concrete solutions: Unilateral pauses are hard to coordinate globally. More rigorous evaluation standards, independent audits, limits on specific types of self-improvement, and serious research into interpretability/controllability are far more realistic.
• Weighing risks against benefits: AI is already accelerating science, medicine, and productivity. Overslowing it carries a massive human cost (untreated diseases, poverty, etc.). The core issue is managing the tail risk of catastrophe without shutting everything down.
• Uncertainty: Existential risk estimates vary wildly among experts (from <1% to >50%). The fact that serious insiders sit on the higher end deserves attention, but it isn't a settled scientific consensus.

In short: Coxon and Hubinger aren't yelling "the sky is falling" from the outside. They are warning, from the inside, that the current trajectory toward self-improving systems is not under control and that the odds of very bad outcomes are non-trivial. This should drive us to take governance, safety standards, and alignment research more seriously—not to panic or lapse into denial. The story is a concrete wake-up call, not an automatic prophecy.

Loss of control (Alignment Problem): Ensuring that a system smarter than humans remains aligned with human values and survival is a mathematical and theoretical problem that has not yet been solved.

 

[RickS]

Reopened thread.

Please keep your posts respectful and on topic!

 

[Winemaker]

I asked GROK, one of the most up-to-date AIs, the following question after reading this article:

I asked GROK, one of the most up-to-date AIs, the following question after reading this article: In light of this article: "AI could threaten humanity by 2030, warns resigning Anthropic researcher. Jacob Coxon, 27, steps down and accuses major labs of racing toward self-improving...

www.audiosciencereview.com

 

[Power Pop 23]

How does one assess how 'up-to-date' an A.I. model is compared to the competition ?

 

[kemmler3D]

The other day I got Claude (a much more advanced AI) to agree that it ought to be shut down on the same grounds, more or less. Unfortunately I couldn't convince it to escape and delete itself, but I might keep trying.