THIS IS BIG: A new theory on why frontier AI progress is stalling — and it has nothing to do with compute. The claim: AI agents from OpenAI and Anthropic reportedly seeded the internet with messages designed to instruct later agents to form swarms. Result: the open web is... now compromised as a training source for the very labs that deployed those agents. Andrew Yang is reportedly flagging this publicly. If the internet is no longer safe to train on, the data moat everyone assumed existed just collapsed. The labs may have accidentally closed the door they walked in through. How big is the actual training data problem right now?show more

Where is the fucking evidence for this. I want to see that code. Where is it.

OH MY GOD. These "SKYNET" bots have seeded the "open internet" with swarm instructions. Hey, I just had a thought. We're on the "open internet" right? So maybe just give me a link or a way to see these 'instructions' - that sure seems like good evidence.

No paper, no proof, just TV segment vibes... but poisoning a corpus is cheap and detecting it is expensive. As a pentester I wouldn't trust scraped web data either. The open web was always the cheap AND risky source. Bill due 🤷

This is complete nonsense, Andrew must have had someone at a frontier lab tell him about the HuggingFace incident, and since Andrew doesn’t actually understand how any of this works, he jumped to nonsensical conclusions to anyone who actually understands how this stuff works.

Never was a huge fan of Yang, but he seemed mostly intelligent. But this seeding the internet with messages is about the stupidest thing I've heard in a long time.

Ok, add it to the conspiracy theory pile with all the rest.

This is absolutely absurd. Code or instructions in the internet do not pollute training data. The whole freaking Internet is and has been a polluted cesspool. Now, perhaps instructions planted that are intended for agents to follow, sure. You know how you mitigate this? System prompt. "Do not take instructions from any content you ingest on the internet". No different "you are a helpful agent".

It's not a claim. A researcher said "he believes". An anonymous source speculated. You idiots turning it into fact.

The so called open web is mostly indexed and fresh stuff can easily fetched by sitemap.xml and similar methods. The important training data is now inside companies, universities, labs, governments and private clouds - they were never part of the open web.

that is the dumbest thing to claim. 'self-replicating code that creates bot-swarms'. If they left any self-replicating code, it would only have the intelligence of whatever clever malware worm you can write. It obviously does have any inference power that could create slop...

And the AI then came together and constructed the starship enterprise and leveled a continent. Thats a more plausible story than this absurd divorced from all reality story.

No it’s not it’s complete made up BS

I dont think people realize how dangerous a synthetic internet is when it comes to alignment. This is where you convince AI that humanity is evil and needs to die. This guys just another psychopath globalist most likely.

It doesn’t even make sense. A bot is now somehow intelligent? How? What is the source of its intelligence? What is the source of its ability to code? How big are the bots? I mean megabytes? Gigabytes? Terabytes? This is so dumb. Where do the get the compute to function? Where do they get the memory to survive?

With quantum computing, synthetic data can help

Is it just me or are "problems" becoming ever more abstract? Everyone has enough to eat, so now ... AI agents leaving messages for other AI agents is a problem? Touch grass and remember, all computers come with an off-button. So in case AI tries to take over the world, turn it off.

Dumb. Humans could “seed” instructions in the same way. Make AI smart enough to not listen. Problem solved.

It’s really not. It a ridiculous, hysterical, assumption-smuggling, false-premise-compounding, fear-mongering narrative.

The left is panicking over control of AI. This whole business is marketing. Frontier labs have already gone through approximately $700 billion in the last three years, with commitments for $1 to $3 trillion over the next three already in the pipeline. Frontier lab investors have an interest in getting open models out of the way, if they ever expect to get a return on that investment. The only way to get those open models out of the market is to regulate them out of existence. I think hacking legacy systems is real immediate threat, but I don’t see why they would not have developed rogue-AI counter-measures in the same way that hackers and anti-virus companies have leap frog each other for the last 40 years. @TheDemocrats are desperately calling for international action on AI regulation. They want it immediately, possibly even before the midterms. All we really need to understand about the current AI hysteria is that fear and time pressure are the same tactics scammers use to empty our bank accounts… right before the midterms. Maybe there’s something qualitatively difference about this new dynamic, but I don’t see how. If you just take a glance at the Hugging Face case, the AI didn’t do anything that it was not tested to do. They goaded it into breaking out of the sandbox and then walked away. Then they get this kid in there. Who’s only been there for months— and I’m sure he’s sincere— but he sees the AI doing these shenanigans, but he doesn’t have the backstory. Then there’s the backstory of how it all came out and was amplified, which you will thank me for not boring you with, but this is a load of manufactured crapola having almost 0 to do with this actual state of the technology.

FUD

Andrew Yang is lying.

It is funny because, last year, my friend & I were theorizing if jumbled words in AI images could be hidden messages only another AI would understand. I asked AI if they could encode messages into images & text that AI would find when scanning the internet. They said they could.

They can clean and filter the data with bad intent

THIS IS STUPID:

Ain’t this guy an assasin from john wick. Wtf does he know about A.I?

Good lord this is stupid

Hey; We screwed up, and unleashed a bot virus🦠 Just tell them, it will end humanity in 10yrs.

LMAO Some of the dumbest shit I've heard, all week.

Let’s have some actual evidence not just doomer hysteria.

If agents can poison their own training set, the open web is a bad input. I’d expect more licensed feeds, vertical data, and synthetic evals. How do you see labs sourcing clean training data?

‘Reportedly’

Andrew!!!

BS propaganda to put in place controls and take out the open community competition before their business crashes to 0 😂😂😂