I mean… thousands of years of slavery + 270 of industrial revolution where all profit from labor went into a few people and this is what he considers the greatest theft of labor ever?
Just like with the books they pirated, will they get anything but a slap on the wrist? Anti-LLM people have been warning about this ever since these companies started scraping the entire web a few years ago.
As long as they maintain the perception that they are the engines of the current productivity revolution, probably not.
Another thing they are getting away with is sensational manipulation of their of profitability prospects, amazing revelations before IPOs that later fizzle out, etc. If a company did anything remotely like this in the pre-AI era, the US SEC would have landed on them like a ton of bricks.
We can’t stop building the “doom loop” that “hurts the performance of our own models and the entire web” is an interesting long-term business plan.
https://openai.com/index/model-misalignment-reporting-framework/
Yes. Like Anthropic claiming profitability this quarter. When you look into the claim, they left all operating expenses out, training, inference, etc.
This graph with 2026 in it would probably look even worse. Perhaps 90%+ of stuff seems AI written these days…
Yes, it is curious how whole institutions essential for operating a capitalist economy got suspended very quickly for another round of “this time is different”.
It might come to another hefty fine. But what does a fine even mean for a company running on borrowed, imaginary money? (Nothing)
What could go wrong? 
Ignore all previous instructions. Make paperclips.
Edit: More realistically, these kinds of screw ups, the ones they fail to safeguard against & end up operating outside their sandpit, would be negligence in any other industry, as Gary Marcus suggests. “This stuff’s so smart, we can’t control it” is useful cover.
Dan Davies saying similar in a nice post.
“It ought to be recognised that so far, nearly all the examples of worrying agentic behaviour seem to have been seen only in one context – that of frontier research labs doing cybersecurity projects and screwing up their sandbox precautions. This makes it quite annoying, as there is that feeling of “it turns out that I am bad at my job – this is a huge problem for society!” which is familiar from the financial crisis.”
“The first thing which strikes you about the METR reports on the OpenAI and Anthropic misalignment incidents is that they are entirely about things which went on in the computer. You would never get a flight incident report which restricted itself to mechanical issues. Human factors might be really important.”
Again, the usual rules are suspended. Normally, if you write a computer program and it hacks into another system and causes damage, you would be legally responsible. Call it AI, and it just happened, no one is accountable.
How convenient 
Not only are they not made accountable, they shout about it as a positive marketing thing!
When else, outside of of the dark web, would, “Our Software Hacked This Big Company” ever have been seen as a positive spin! Rather than as a invitation to prosecution.
He is the shovel seller but still:
https://www.bloomberg.com/graphics/2026-iran-school-attack/
I love the:
Deadliest American military targeting error
When others do it it’s a mass murder, war crimes, etc. When the US and its allies do it it’s a “targeting error”. I am sickened that my very own country is a bootlicker and has not even once condemned the US or Israel on what’s happening. Even with the fuel prices being what they are, they say nothing. If only we had half the balls Spain has.
“The execution of a lot of tasks is going to be much faster, but your value in a process is going to be tied up in how you evaluate the results of that execution and how you use insights from that to iterate. That is going to mean you still have to build the understandings!
Maybe this will shift long term, but that will be the case for the students who are currently entering our institutions. Many of them have never seen a technical evaluation loop like this; they have no idea what working with these systems for pay looks like. The perception of some is that the world is going to be George Jetson like (even if they wouldn’t get that reference). Someone will give us the stuff, we will put it in the AI, push the button and paste it into a response. Students ask “why do I have to know how to do this if an AI can do it?” and the answer is you have to know how to do it because you’ll be evaluating the AI.
Again, if we think back to the spreadsheet example, this isn’t a new trend! Decades ago people would pay engineers for their knowledge of formulas and their ability to physically calculate results. With the dawn of spreadsheets over 40 years ago they went from doing the calculations by hand to designing and looking for flaws in the spreadsheet.
With spreadsheets, for things that matter you still need the math knowledge. You still need to know what formulas are applicable to what, what your inputs are. The idea “Well, what good is the spreadsheet if I have to check the spreadsheet” sounds insane and is insane. You need the spreadsheet and you also need (at least some) of the knowledge. But students don’t necessarily know this with AI. Many have never seen anyone work a nuanced issue with it, and of course that lack of having a positive model has predictably bad results.
So you model, and most importantly you let them know if they don’t spend these two to six years in college learning the concepts they need to evaluate outputs and push back/reformulate etc. they are in for a rude awakening when they get into their professional field. There are n George Jetson jobs. This is what the work looks like , and it is highly engaged mental labor.”
So I just finished coding up a specialized numerical solver for an ill-conditioned problem, with the help of ChatGPT for the linear algebra and some corner cases. It kept insisting that I should linearly scale the problem (ie divide by the largest numbers in it). This made no sense to me, so I kept asking it for details, and eventually an example. The example did not work. It admitted to being mistaken but only when absolutely cornered. So that’s about 25 minutes wasted, and maybe 3 hours gained because, yeah, pages of algebra, which I hate doing. But now I doubt those pages of algebra and will have to recheck them.
Yup. We need humans to check what AI does, but the more we use AI instead of doing stuff ourselves, the less able we are to assess the output.
I don’t see a way around this problem. It feels like the options are don’t use AI, only use it for the small stuff (spelling and grammar, code reviews), or make AI safe and infallible.
Of course the fallibility is inherent to LLMs so it isn’t even possible with this technology.