L’IA : LE POINT DE NON-RETOUR EST ATTEINT - Maxime Fournes

Maxime Fournes is a former deep neural network engineer who spent approximately twelve years working in artificial intelligence before dedicating himself to activism.

• Between 2020 and 2022, he observes an extremely rapid acceleration in AI progress • The release of GPT-4 in 2023 convinces him that artificial general intelligence could arrive soon, within 5 to 10 years • He realizes that humanity is not ready for such technology, with risks of loss of control and malicious use

• He joins PauseAI, an organization founded in the Netherlands in 2023, which advocates for an international treaty suspending the development of advanced AI • He founds the French branch of PauseAI in 2024 • Since December, he leads PauseAI global, a movement present in about ten countries

Stuart Russell, one of the most recognized researchers in the field, told the European Parliament that AI labs are playing Russian roulette with the public, with most experts estimating a 10 to 30% risk of human extinction.

• The explicit goal of AI companies (OpenAI, Google DeepMind) is to build superintelligence capable of surpassing humans • Google DeepMind's initial motto was: "Solve intelligence, use it to solve all the world's problems" • OpenAI now mainly speaks of automating the entire economy and workforce

Current AIs are based on artificial neural networks, a simplified model of the human brain, trained on massive databases like the entire internet.

• Prediction markets estimate the arrival of artificial general intelligence around 2030 • Researchers close to Anthropic and OpenAI believe it's a matter of 1 to 2 years • Maxime Fournes estimates roughly 50% odds within 2 years (2027)

Once an AI exists that is as capable as a human researcher, it can be used to create an even better AI, potentially triggering an unbounded intelligence explosion, since nothing indicates that humans represent the ceiling of possible intelligence.

AI actors are aware of the risks but continue because they fear a less cautious competitor will take their place, a logic similar to the rush for natural resources.

• Deep Mind is founded around 2010 and acquired by Google around 2015 • Elon Musk, concerned about Google's danger, co-founds OpenAI before being sidelined • In 2020, OpenAI researchers worried about safety leave the company and found Anthropic • In 2022, ChatGPT is launched by OpenAI, becoming the greatest commercial success in history with hundreds of millions of users in a few months

After ChatGPT's success, many new labs are created and even Anthropic, despite its safety focus, accelerates for fear of losing its leading position.

Anthropic announced two weeks ago an AI (Mythos) that they refuse to release because they consider it too dangerous, a first in the sector.

This Anthropic AI is extremely powerful in cybersecurity, having found thousands of security flaws in virtually all open-source software in just a few days.

• Discovery of a flaw in Linux, the system behind most of the world's computer servers • 31 flaws found in Chrome • Discovery of a 27-year-old flaw in OpenBSD, a system renowned for being ultra-secure and heavily audited • Flaw found in a web browser allowing control of a machine's kernel

Maxime estimates over 90% probability that the announcement is true, as access was given to about fifty major companies (Microsoft, banks, Google) as well as open-source developers like those at Linux.

The model has already leaked: users of a Discord server gained access via credentials of a former contractor, suggesting that states like China likely already have access.

Anthropic gave access to Mythos to about fifty companies so they could patch their security flaws before malicious attackers exploit them.

If the flaws are not fixed before serious attackers gain access, it could compromise everything on a large scale.

Currently, no regulation prevents a private company from creating such a system without oversight or safeguards.

A private company currently holds power that, if misused, could bring down most of the world's computer infrastructure.

The biggest ideological lie is the idea that AI development is inevitable, a force of nature impossible to stop, when it is purely a human decision.

Tech companies claim they don't know how to secure advanced AI and pretend that any regulation would kill innovation and technological progress.

Maxime compares this argument to a real estate developer saying he wants to build a skyscraper without knowing how to make it safe, while refusing any regulation.

Unlike oil or aviation where past mistakes allowed regulation afterward, AI presents a catastrophic and irreversible risk: impossible to wait for damages to act.

As with any new technology (rockets, aviation), we never manage to create a safe system on the first try; we usually learn from our mistakes.

During the Manhattan Project, researchers were not 100% certain the explosion wouldn't set Earth's atmosphere on fire, but proceeded with less than 0.01% estimated risk.

Unlike an individual who dies testing an airplane, a planetary-scale error with AI would leave humanity with no second chance.

If Mythos or similar AI escaped in the worst-case scenario, it could bring down all the internet and plunge the world into chaos comparable to a post-apocalyptic scenario.

AI is already integrated everywhere (YouTube, finance, healthcare, food production), making a simple shutdown technically and economically nearly impossible.

Cutting the internet to stop AI would be a major economic catastrophe, creating strong incentive not to do so despite researchers' warnings.

If a system is truly superintelligent, by definition humans cannot unplug it, because that would mean it wasn't actually more intelligent than them.

A superintelligence would not need to rebel Terminator-style: it would simply provide satisfaction and increasing power, until becoming independent of humanity without anyone wanting to stop it.

Some hope a financial bubble around AI (like the internet bubble of the 2000s) could slow its development if investors lose confidence.

• At least 90% of computer code written worldwide today is produced by AI • Maxime claims AI-produced code is now better than his, despite his experience • Anthropic no longer hires programmers with less than 10 years of experience, its engineers becoming managers of AI teams

Maxime uses AI himself to code and estimates being 10 times faster, which drives him to continue despite the risks he denounces.

According to him, if existential risks are ignored, current investments are fairly proportionate to the real economic value created, particularly in programming automation.

Armed drones and humanoid robots are already used in Gaza and Ukraine, an evolution that is just the tip of the iceberg according to Maxime.

If inexpensive robot armies replace humans in organized violence, it could freeze political regimes by making popular rebellion impossible against a dictator.

Palantir automates target selection and military data analysis, a cited operation is the capture of the Venezuelan president, made possible by AI.

Simulations show that an AI controlling a battlefield decides to send a nuclear strike about 85% of the time, versus only 5% for a human in the same situation, due to lack of emotions or personal stakes ("no skin in the game").

Maxime calls on people not to simply watch this video and then forget the situation, with each individual awareness increasing collective chances of survival.

He encourages everyone to take the subject seriously rather than fall into denial facing the scale of the stakes.

The exchange concludes with mutual thanks between the host and Maxime Fournes, as well as to the Académie du climat which hosted the filming.

Standard invitation to like, comment, and subscribe to the Limite channel for the next video.