Or they will go with cheaper AI models in order to keep costs down. The additional benefit is that you can modify the parameters yourself as a business (vs using a black box model like OpenAI and Anthropic uses) and keep your data safe (a huge problem right now as anything you feed the AI can be used to train the AI models themselves).
Interesting, and not too surprising.
Training AI models on AI-generated data sets causes problems (study from Oxbridge).
Its like taking a photocopy of a photocopy…you end up with a loss of precision.
This seems like a fundamental problem with AI ever truly advancing anything. I think they can find answers hidden in the current rubble of human thought that makes up the online world, but just about anything that gets introduced going forward will be highly polluted information.
1 Like
Never put anything in Claude that you wouldn’t want read by your competitor/mother.
and that’s why i have a zero data retention agreement lol. So I can actually do this.
But I know lots of people in the insurance industry putting their client’s info in wholesale. I know one guy locally who’s putting entire groups of clients in.
If you’re talking about open weights models, they also have some risk, even if you run them locally, as they are also a black box.
Anthropic claims to have never trained on user prompts. I would give them some benefit of the doubt given how many unsolved math problems they have taken down recently.
If I were doing anything with sensitive data, I would at least make sure the LLM owner is legally required to not use it.
I think the odds of AI agents escaping and going out of control are pretty high right now.
Hmm, perhaps a cheaper legal support option is becoming available.
That whole hugging face hack thing is way crazier and more concerning when you look into the details.
1 Like
Wouldn’t p(escape) = 1 where we’ve seen it happen repeatedly?
1 Like
I used my personal Claude account to create an initial draft of a proposal, saved me a week’s worth of writing work. 
“The” internet worm took down sections of the internet back in 1988.
It was because the programmer was clever but sloppy. We blamed him, not the worm.
I wish we (meaning the press and the wider world, not people on this forum) talked about these LLM powered scripts the same way. These companies need to test them responsibly. That seems to take more time and effort than they want to spend. It’s not because these programs, running on a cluster of computers completely controlled by these companies, have a being of their own the transcends the computers are running them.
Yes. But I suspect we are only seeing the very top of the iceberg when it comes to external hacks by AI agents.
The AI companies are going to get sued and it will all come out via discovery.
A spiralling legal crisis threatens to engulf OpenAI after its AI agents hacked dozens of companies and governments worldwide, with the prospect of a barrage of lawsuits marking the latest test of Sam Altman’s leadership. Staff at OpenAI, alongside senior legal, tech and policy figures, told the FT that the company has been left open to legal damages and government actions after cyber security breaches involving its AI agents were revealed around the world. The past week has seen new legal actions and regulatory probes launched against OpenAI almost daily, from California to Australia, with more expected to follow as details of other incidents come to light.
2 Likes
Listen to the This week in tech podcast, they are pointing to the example of the guy who wrote one of the first worm programs. It was intended to be inocuous, but got into a lot of computers. The guy ended up being convicted of a crime. The hosts of the company are asking why the AI company execs shouldn’t be charged for the crimes their models have committed.
The AI company execs have money – money that would go to lawyers who would make prosecution a very expensive exercise, and would be better spent filling PAC warchests.
Hmm, this might be AI…
One of the biggest things I hate about the rise and sloppy use of AI is that I used to be a frequent user of em-dashes.
Between that and my sometimes-stilted prose…
1 Like