OpenTube

You’re Probably Thinking About AI Risk the Wrong Way

Prof Gio | Giordano Scarciotti

9,637 views

Could artificial intelligence actually become impossible for humans to control? And why do some AI researchers take existential risk seriously when ChatGPT still makes obvious mistakes? In this video, I look at what has actually changed in AI: the rise of AI agents, tool use, autonomous action, large-scale collaboration, and the growing ability of systems to search for solutions in ways humans did not explicitly specify. Using examples including the Hugging Face cyber incident, AI-assisted mathematics (Navier-Stokes Millennium Prize Problem), alignment failures and autonomous-replication tests, I separate capabilities that already exist from the larger steps that would still be required for genuine loss of control.

I also look at what AI researchers themselves believe, why there is so little consensus on existential risk, and why even a world in which the most extreme AI-risk scenarios are wrong could still be profoundly disruptive. AI is already beginning to affect parts of the labour market, especially white-collar work and junior hiring, while at the same time offering the possibility of major scientific and economic gains. So is AI risk mostly hype, a serious warning, or simply something we still do not understand well enough?

👇 If you are finding this breakdown helpful, leaving a comment below is one of the best ways to support the video.

🔔 To help scale up the channel's operations and bring you more high-level science and engineering outreach, consider supporting my work on Patreon:
Patreon: ProfGio

☕ If a monthly commitment isn't for you, but you still want to help support the channel's growth, I have also set up a Buy Me a Coffee page:
https://buymeacoffee.com/profgio

🔔 Subscribe for more deep dives into AI, engineering, control systems, energy technologies and the science behind the technologies reshaping our world.

📚 References / Further Reading
Some of the main sources discussed in the video include:
International AI Safety Report 2026 — https://internationalaisafetyreport.o...
METR on the OpenAI / Hugging Face incident — https://metr.org/blog/2026-08-26-open...
OpenAI on the Hugging Face incident — https://openai.com/index/hugging-face...
OpenAI model-misalignment disclosures — https://openai.com/index/model-misali...
Hugging Face technical reconstruction of the intrusion — https://huggingface.co/blog/agent-int...
UK AI Security Institute — RepliBench — https://www.aisi.gov.uk/research/repl...
Grace et al. — Thousands of AI Authors on the Future of AI — https://arxiv.org/abs/2401.02843
O'Donovan et al. — What are AI researchers worried about? — https://arxiv.org/abs/2603.06223
Stanford Digital Economy Lab — Canaries in the Coal Mine? — https://digitaleconomy.stanford.edu/p...
Federal Reserve Bank of Dallas — AI and job postings — https://www.dallasfed.org/research/ec...
OpenAI — Navier–Stokes Millennium Prize Problem — https://openai.com/index/navier-stoke...
Clay Mathematics Institute — Navier–Stokes announcement — https://www.claymath.org/news/navier-...

⏱️ Chapters
00:00 Is AI Going to Kill Us All?
01:33 From Chatbots to AI Agents
05:53 The Hugging Face Cyberattack
09:10 The AI Alignment Problem
11:53 AI Takes on Navier–Stokes
14:55 AI Helping Build the Next AI
16:08 What Would Loss of Control Actually Require?
19:33 What Do AI Researchers Actually Believe?
21:50 Why Are AI Companies Warning Us?
23:17 Can Governments Regulate AI?
24:38 Even If the Extinction Risk Is Wrong
27:26 So… How Worried Should We Be?

🏷️ Tags
#ArtificialIntelligence #AI #AIRisk #AISafety #AIAlignment #AIExistentialRisk #AIAgents #Huggingface #NavierStokes #AgenticAI #ChatGPT #AGI #AIControl #MachineLearning #OpenAI #Anthropic #FutureOfAI #AIJobs #Automation #FutureOfWork #Cybersecurity #ArtificialGeneralIntelligence #ProfGio

20 Comments

@Prof_Gio Creator 3 days ago Pinned

What AI topic would you like me to cover next? I’m also happy to bring colleagues on board for specific topics. Leave a suggestion below and I’ll add it to my list. I may not reply to all of them, but I read them and they genuinely shape future videos.

This is an independent channel, and making these deep-dive videos takes a lot of time and resources, especially research, editing, software and licenses. If you value this kind of evidence-based science and engineering content and want to support the channel:

🧡 Patreon (monthly): https://www.patreon.com/cw/ProfGio
☕ Buy Me a Coffee (one-off): https://buymeacoffee.com/profgio

Thank you for watching and for being part of the community ☺

@Beniguitar94 2 days ago

Very high quality and nuanced content - thank you!

@statsmad2812 2 days ago

Considering the AI progress in its entirety, it is tempting to think about it through an evolutionary lens. If the progress of AI, which involves all kinds of human interactions (programming, regulating, lobbying pro / against, etc), does follow a chaotic, stochastic process like Darwinian evolution, Prof Gio's concluding remark seems the most apt: we really don't know how it may end up. If AI evolution is really 'chaotic', it could end up really 'anywhere'.

Can we put some boundaries to totally, 100% random evolution? Probably, but the results (like the Hugging Face incident suggest) may not necessarily make the system evolve in the directions we expect.

Food for thought, and possibly another video.

@GooDgRaFT 2 days ago

Thank you, Professor. Your video covers the topic very well.

As a mathematics PhD student, I'm really worried about the current path of AI.
But I'm also very excited about it, just as you illustrated at the end of the video. We don't know how far it could go, or where we will be five years from now.

Who knows, maybe one day we will be forced to attack data centers to stop an AI. And even then, would it be enough?..

I like this phrase: “Humans don't understand the exponent.”

@CliffNickerson 2 days ago

The issue is humans using AI to do stuff it shouldn't do. Nearly everybody trusts ChatGPT's answers. A small business owner I know fired his IT pro, hired an intern and gave him Claude Code.
If we give AI permission to do stuff, it'll generally do it.

@1988Tull 2 days ago

Interesting video - thanks!

@gogutier1 2 days ago

Great video as always.

@themixx8228 13 hours ago

Excellently researched video!

I will say that as a PhD student in AI research myself (in the realm of video rather than LLMs, though my work often bridges the gap between the two), I am very wary of giving any agent unmitigated access to my computer. I have caught it multiple times trying to look at files it wasn't supposed to.

In my opinion, the biggest threat to humanity is the extreme energy and water usage from data centers (worsening the effects of climate change significantly) which coincide with the installation of millions of flock cameras (at least in the USA). It seems quite obvious that one could use the data generated from these cameras to train an AI that could simultaneously identify and track every single person in America. The solution to these problems is relatively simple: a moratorium on new datacenter construction, better monitoring of the "internal models" these companies are producing, and a pause on AI research from private companies that will give us time to create meaningful legislation to combat AI's negative effects. Fortunately, these solutions also help protect us from AI systems that could end humanity.

@isoscelestriangle 1 day ago (edited)

There isn't much difference between a few years ago and now. An "agent" is just an LLM running in a loop. That's it. Yes, that makes it more useful, but it's still just the same kind of LLM as before, it can just run multiple times now to improve its answer and abilities. This can patch over hallucinations, but mostly in tasks that can be quantitatively tested, like code.

@silvio5266 2 days ago

Great video, it’s hard to find a video so balanced on this topic today. Also very complete, one thing you didn’t discuss that is also relevant is the impact of this revolution on humans’ mental health: we are seeing that basically all we can do can be done much better and faster by a machine, a lot of us are losing what made them feel useful, or important: being good at something that only few people can do. Now anyone is a prompt away from that. I think this is not something to ignore and it will concern more and more people as this technology evolve

@JDon8110 2 days ago

Seriously you are one of the most capable minds on the internet. Your clear and thoughtful approach is so appreciated. Thanks ♥️

@Eddison33 27 minutes ago

I think there's a good reason why there are almost no AI researchers (who don't work for OpenAI and Anthropic) who are fixated on existential risks, AI takeover, or extinction scenarios. The hype about doomsday AI, often spurred by OpenAI itself, distracts from the actual problems we're facing - the chips crisis, the energy demands, unethical training on the work of thousands of artists, deepfakes, pollution, questionable pictures of minors generated by AI, the Internet being filled with AI slop... As always, I'm more worried not about AI itself but in whose hands (greedy for-profit companies registered in the US) it is concentrated, and how little democratic oversight they experience. Of course, they want to regulate themselves to avoid some catastrophic doomsday scenario - anything that doesn't make them accountable to democratic institutions for the damage they're doing right now.

@David-nq4qi 2 days ago (edited)

The question I want answered is why people say AI is going to make humans extinct.I can see how it could put some people out of work, and I can see how it might mess some important things up, but how is it going to actually attack people? Is it going to take over factories, build robots, and send them to my home to kill me? Or what? Whether it is real or imagined, what exactly is this worst-case scenario that will result in extinction? Nobody is clarifying that, and you might be the best person to do it.

@Biggest-hz7ng 2 days ago

How will we know whose AI is going rogue? How do we assign blame and culpability to the developing company/nation state? I see cyber security being the likeliest area of damage from AI in the short term. How useful would the internet be if AI worms prowled it? Can companies, individuals and countries turn off the internet? How long would it take to 'fix'? It seems likely that there will be periods where we will see degradation in the usefulness of the internet and reductions in the numbers prepared to expose themselves to it.

@casualredditor9918 2 hours ago

Finally actually nuanced content with knowledge in the field

Loved your explanation of RLVR etc!!

@xidxip 1 day ago

If we view the actual stochastic parrot AI as an existential threat only through its zero intellectual existence, we are safe; it never, ever will be intelligent. But if we look ahead at electricity consumption, all kinds of pollution, water use, and so on... Well... if we look at those aspects... we're dead.

@charlesmanning3454 2 days ago

It's very unlikely A.I. will go rogue, decide to take over the power grid and cause mass havoc. It's much more likely that power grid operators will use A.I. to control the power grid, fail to monitor it and it will malfunction and cause mass havoc.

@whatabouttheforests4631 7 hours ago

Im not a mathematician but have watched several videos on the navier-stokes solution.

The thing that fascinates me is what type of intelligence is used to solve these problems.

Are these llm/cot/agent swarm systems the equivalent of an einstein+++ intelligence with an iq of 500 or whatever or is it more a case that they spawn 100 einstein level intelligences to work together for the equivalent of 100s of years.

If the latter then this is far less interesting imho. Openai spent the equivalent of 10-40 million dollars to solve the problem.

And from my limited knowledge solved it by trying huge numbers of approaches that mathematicians could, if there were enough of them with infinite patience, also do.

@tonig2757 1 day ago

I think the growing field of local AI is also going to throw a monkey wrench into those predictions.
Like, the evil AI that would hack our power grids, might spawn in someone's garage.

@Me__Myself__and__I 2 days ago

"We do not have evidence for the complete chain from today's systems to an AI that humanity can no longer control."
THIS is the problem. Once that evidence exists, its far too late and we are done for. If we wait until its a certainly then we are certainly extinct It is in fact pretty darn obvious where things are headed and that those gaps WILL be closed, we just don't know when or how. By the time we do, its too late. If you wait to slam the breaks until AFTER the car goes off the cliff guess what, they are useless at that point!