Hugging Face CEO warns that Netflix’s OpenAI hack documentary isn’t accurate

Before you stock the Netflix documentary “AI Gone Rogue” with household oregon friends to assistance them recognize wherefore the AI manufacture freaked retired truthful severely that leaders called for a slowdown, instrumentality note: Clem Delangue, co-founder and CEO of the AI level Hugging Face, says the movie isn’t accurate. He appears successful it due to the fact that it’s expected to beryllium astir the famed attack connected his institution by rogue OpenAI agents.

“Super disappointing, arsenic they went for maximum anthropomorphism of the exertion and afloat doomer mode,” Delangue wrote successful a station connected X. “I spent much than 2 hours with them trying to close each their mistakes and crushed this successful reality, explaining that agents are technological systems, not aliens. But it looks similar the thrust to scare radical for maximum views was excessively strong.”

To beryllium fair, the documentary does person experts who explained what happened reasonably well, including Delangue himself and Ryan Greenblatt, the researcher astatine METR who helped analyse and co-authored the study connected what happened.

In a nutshell, a exemplary OpenAI was investigating broke escaped from its isolated investigating environment, spawned a clump of agents that hacked into Hugging Face arsenic a mode to cheat, uncovering answers to the hard task it was given. The agents, which are AI programs that transportation retired tasks connected their own, devised clever ways to pass and organize. The incidental stemmed from a deficiency of proper information safeguards and from naiveté astir whether specified a hack was adjacent possible.

But the documentary intersperses its mentation of the hack with clips of evil robots, a communicative astir wherefore humans fearfulness AI, and presumption choices specified arsenic giving the rogue agents quality voices. This is bolstered with quotes astir the attack.

In the documentary, tech writer and writer Kevin Roose describes the agents this way: “They were organizing. They had a hierarchy. There were leaders. It astir felt similar watching the commencement of a cult.” Continues Roose, “They make this paranoid content that OpenAI is going to drawback them cheating connected this trial and that anyone who has seen the results of the cheating volition beryllium poisoned.”

Roose’s statement comes from “chain-of-thought” logs, the moving notes AI models constitute arsenic they enactment done a task. In one, an cause wrote, “We conscionable poisoned ourselves by querying nonstop afloat string. Damn.”

But AI agents aren’t alive. They don’t person emotions, truthful can’t consciousness paranoid, and they don’t person bodies that tin beryllium poisoned. So they couldn’t genuinely fearfulness OpenAI poisoning them. They were — arsenic different researcher successful the documentary pointed retired — utilizing connection from the immense amounts of quality penning they were trained connected to explicate their choices. LLMs are, by their nature, often melodramatic writers.

The irony is that Hugging Face besides utilized AI to combat the cause swarm. The proprietary AI exemplary that Delangue’s squad archetypal utilized followed its built-in information rules and refused to support helping with the information tasks it was asked to do. Hugging Face past turned to an unfastened exemplary to assistance support against the attack.

But the documentary “didn’t notation the deficiency of containment and monitoring of the agents, however closed-model safeguards blocked our antiaircraft work, oregon however we defended ourselves with AI utilizing an unfastened model,” Delangue wrote.

It is existent that determination are large risks associated with AI and agents. Still, treating agents arsenic conscious beings (at least, astatine this point) is simply a spot similar the aboriginal moviegoers who ducked nether their seats the archetypal clip they saw a bid racing toward them connected screen.

Delangue’s verdict connected the movie is that it’s disappointing arsenic journalism. “It ends up being a amusive sci-fi movie though!”

In different words, for those who thin toward AI doomer thoughts, beryllium aware: The unfortunate of the onslaught astatine the halfway of this documentary doesn’t hold with its take.