‘The biggest decision yet’ – Allowing AI to train itself | Anthropic’s chief scientist says AI autonomy could spark a beneficial ‘intelligence explosion’ – or be the moment humans lose control
https://www.theguardian.com/technology/ng-interactive/2025/dec/02/jared-kaplan-artificial-intelligence-train-itself
26 Comments
> AI systems will be capable of doing “most white-collar work” in two to three years.
This is another thing he said. And given that’s been the line for multiple years already, I’m not too concerned.
Just more valuation pumping.
Seems like an appropriate risk trade-off… Let’s do it!
It’s good that everyone’s future will be wagered on a dice roll. We may all become extinct, but for a short golden moment, a lot of shareholder value will be created.
It’s just gonna eat itself lmao. We already see this with a.i gen images ripping off other a.i gen images. Sorry Mr. Godfather of a.i no.294, but until this is actual intelligence (it fuckin isn’t) its not gonna be able to teach itself because it has no true comprehension of what it’s teaching.
Or just becomes even more divorced from reality and disappears up it’s own virtual arse.
Isn’t this already a problem. AI training on AI outputs, compounding errors over and over again. Resulting in inbred AI hallucinations?
Can anyone explain the lose control bit?
Is AI going to take over electricity grids and transport networks as stop humans from turning it off? How? AI is simply not elemental to our lives insofar that it could ‘take over’
Huh? I thought using AI to train AI was already being done.
It’s one of things that lead to DeepSeek. They used ChatGpt to train it. Heck, I thought OpenAI has said they use ChatGpt to help train the next version.
Then they bring in humans to question it, refine it and slap in some rules.
They just do it on sandboxed systems so it can’t manipulate anything but the system it’s on.
“Man whose wealth depends on people believing that AI is the best thing ever says that AI is the best thing ever”
AI training AI will result in a reduction to even-worse-slop.
The same old AI junk ideas are recycled every generation. Oh, what if we had a genetic algorithm that selected genetic algorithms for suitability?! Then, actually, things get worse.
And as it is, the trillion-dollar AI industry is already admitting – they’ve exhausted basically all viable training data and all that’s happening is that AI output is leaking into training data en-masse, and it causing merry hell with their training as it just reinforces the AI’s already-bad output.
Literally every ten years, the same shite over again with AI. “If only we had more money, processing power, computers, training data, time, nodes, interconnection, rounds of training, etc. then I’m *sure* this time intelligence will just magically pop out of the box like hitting some kind of critical mass when slop training slops suddenly turns into a self-enhancing genius at exponential rates for perpetuity….”
You let AI train itself and what you’ll see is worse AI.
Or be neither of these things and a path to digital madness.
This is already happening, whether we like it or not.
It’s not possible to reliably determine whether a published text or image was generated with AI.
So unless training is only done on pre-AI-era materials, the training will be increasingly cannibalistic.
How is AI going to train itself when some of the models aren’t even deterministic?
Someone should tell those idiots to play through Mass Effect….
When did inbreeding ever yield good results?
AI bro saying AI bro stuff, nothing to see here.
“.. be the moment humans lose control?” No, it will be the moment YOU lose control and the humans then need to fucking cope with it. Profits are privatized, losses are socialized. If something goes wrong, well fuck it, someone will clean after me..
Have they seen the internet?
Control lost already.
Edit and these idiots think being ceo protects their white collar job when in fact they are a profit suck.
Hahahah how on earth are there people dumb enough to think letting an AI “train itself” will make it work better.
How will the AI know that it’s training itself with real, accurate and true information and not the mountains of bullshit lies that are on the internet?
Right, because none of the other companies have tried that yet /s
Ai with more than 25% of synthetic Ai produced data causes AI degeneration.
LLM will never lead to true AI or even AGI. We’re just burning the Earth’s resources and risking humanity’s future to make line go up.
> Allowing AI to train itself
mf just totally forgot about GANs
The amount that I really truly hate these people is astounding to be completely honest
What a load of bullshit.
We keep calling it AI but it’s not, I say let them do it, models allowed to do it will destroy themselves by training on garbage data that will become even more shittier over time until it’s useless.
It’s mostly useless right now unless data is carefully curated by humans.
These guys are idiots. They are just pumping the market and have zero restraint on what they are doing. They have no idea what the outcome is going to be. It’s a race to be “first” and they have much of that same pathetic YouTube commentor energy. The only difference is that there are real consequences to what they are all doing as they race to plant their flag and make “line go up”