AI Labs Want to Slow Down. China Isn't.

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: AI Labs Want to Slow Down. China Isn’t. · Channel: Fahd Mirza · 08:08 · Watch on YouTube

Topics: AI governance and regulation, AI models, AI safety

Summary

The video examines a call by AI leaders to pace frontier development, following researcher Jacob Coxen’s resignation and a warning from Dario Amodei about recursive self-improvement. Sam Altman said OpenAI would match the commitment, and Elon Musk publicly agreed, despite their past disagreements. The narrator contrasts this alignment with the rapid release of Chinese models and asks whether calls to slow down reflect safety concerns or anxiety about a narrowing lead. He also suggests that researchers close to model development may be among the first to feel genuine fear, while much of the public remains unaware of AI’s capabilities.

Key points

Jacob Coxen resigned after three years of pre-training research at OpenAI and Anthropic, saying neither company was acting responsibly.
Dario Amodei warned that AI systems could help build their successors faster than people can understand what is happening.
Amodei proposed three steps to slow the process without stopping it, including an immediate commitment from Anthropic.
Sam Altman said OpenAI agreed with pacing and would match the commitment; Elon Musk also publicly agreed with Amodei.
The video contrasts the leaders’ recent alignment with their inability to sign the same safety letter in 2023.
The narrator says 1,386 employees across frontier labs had already supported the option to slow down, while noting that not everyone was impressed.
Chinese AI projects including DeepSeek, Qwen, GLM, and MiniMax continued releasing models as the quality gap with American frontier models narrowed.
The video questions whether calls for pacing are driven by alignment concerns, concern about competition, or both.
The narrator suggests researchers working closely with these systems may be especially likely to experience fear about their development.

Chapters

Start	Chapter	Summary
00:01	A resignation and a warning about AI	Jacob Coxen’s resignation leads into Dario Amodei’s warning about recursive self-improvement and his proposed steps to pace AI development.
01:38	AI leaders align on pacing	Sam Altman says OpenAI will match the commitment, and Elon Musk agrees with Amodei despite their past disagreements.
03:18	A broader safety effort meets a challenge	The narrator describes support for slowing AI from employees at frontier labs, then contrasts it with a Chinese AI leader’s response.
04:27	Chinese models keep shipping	The video points to continued Chinese model releases and a shrinking quality gap, raising questions about the motives behind calls to pace development.
06:02	Fear, proximity, and public awareness	The narrator reflects on why researchers close to model development may feel fear first and notes that many people remain unaware of AI’s capabilities.

Quotes

Neither company is acting responsibly. — 00:01

He called it recursive self-improvement — 01:05

Dario is right. — 02:10

Stop pretending you need anyone’s permission. — 04:27

the story ends not with an answer with a question — 07:38

Concepts and entities

recursive self-improvement (RSI), pacing, frontier labs, open-weight models, Anthropic, OpenAI, DeepSeek, Qwen, GLM, MiniMax, Dario Amodei, Sam Altman, Elon Musk, Junyang Lin

Full transcript (original language)

00:01 Four sentences. That’s all it took. I resigned from Enthropic today. I spent the last 3 years doing pre-training research at both OpenAI and Enthropic. Neither company is acting responsibly. A researcher named Jacob Coxen posted that on a Wednesday morning and inside a few days 170 million people had read it. 4 days later, Dario Amodell posted a link of

00:33 his own 9,000 words, a title that reads like a warning. We must pace the frontier that also accompanied by his tweet. Amodai has spent 12 years telling the world AI could cure the disease that took his own father and he still believes that. But this time he said something plainer than he ever had. The machines are starting to help build the next generation of themselves faster than anyone predicted and left alone

01:05 that could outrun anyone’s ability to understand what’s its doing. He called it recursive self-improvement and he laid out three steps to slow it down without stopping it. The first one he said enthropic would do immediately regardless of what anyone else chose to do. And this is his full tweet. I will also drop the link to this blog post in video description. This is Fahad Midsa and I welcome you to the channel. I recently also have started this free

01:38 weekly AI newsletter which you can subscribe at fadmmitza.substack.com. Well, the story which started with Jacob’s resignation and which maybe likely prompted Dario to come up with this tweet didn’t end there. Then we saw this tweet within the hour. Sam Alman replied he agreed with Dario completely pacing. He said had already been a live conversation inside OpenAI for weeks.

02:10 They would match the commitment. Right after that, Elon Musk, a man who has spent years at war with both of them, and to be honest, do it in a very bad taste every day, posted three words under the same essay, Dario is right. And then this thing goes on and on and on. You can see that if you go back to 2023 and none of this makes sense. That year these three men could not agree to sign the same safety letter. Musk signed one

02:45 in March. Open AI wouldn’t touch it. Altman and Amodai signed a different one in May without Musk. Two letters, one year. Never once all three names in the same place. Now on a random weekend, all three lined up within hours of each other. So what is happening? Well, I believe lot of things are happening. But before that have a read through this. And Coxen wasn’t even where it started. I would I must say I mean media is making it like that. But the thing is

03:18 that back in July 1386 employees ac across the frontier labs had already put their names on one page. Not outsiders not activists. the chief scientist of OpenAI of Enthropic of Google Deep Mind of Meta um Ilia Jared Kaplan Shane Lag people who compete with each other every day for the same researcher the same chips the same headlines agreeing on one sentence the world may need the option to slow this down and no single company can buy that time alone but here comes

03:54 the real twist not everyone Everyone was impressed. A man who was until this year the face of China’s biggest open source AI project saw the whole thing unfold. And just 29 minutes ago. Um he came up with this tweet. And by the way he uh Jen Jianglin who built Quen into a model family downloaded hundreds of millions of time wrote one line that when we try to accelerate you tell us to slow down then sharper.

04:27 You can do whatever you want. Stop pretending you need anyone’s permission. So this is so so um I would say insightful and maybe the crux of the whole story. And the reason why I say that that while all of this was happening on American timeline, something else was happening on Chinese ones. Deepseek kept shipping. Quen kept shipping. GLM, Miniax, and end group models, one release after another, arriving faster than most people outside

05:00 the industry could track them. And we have covered them all on the channel, including the closed labs like OpenAI Enthropic. All of the models we have covered. Sanctions had cut off the newest ship. We all know that export controls had tried to choke the pipeline in the Chinese lab at the hardware level. None of it slowed the releases down and the gap in quality that used to separate these models from the American frontier kept shrinking month after month benchmark over benchmark. So when the

05:31 same week produces both an unpreced un unprecedented safety alliance between anthropic and open AI and Elon Musk and a fresh wave of competitive Chinese openweight models landing right on top of it. The question writes itself is this really about buying time for alignment research or is it about the fact that for the first time the lead um you know these labs have taken for granted for years is actually closing and pacing is a much easier thing to ask

06:02 for once you are worried about being caught. So look, let me know what do you think. But I just want to finish this video by saying few lines on these sort of resignations that beyond this all conspiracy theory and tactic and all this. It might be simpler than that. Maybe when you spend 3 years inside the room where these models are actually built, watching numbers move faster than your own models of the world can keep up with, something in you starts to bend. Jacob Coxson was

06:35 not a policy person or an outsider with an agenda. I mean, I believe so. Um, he was pre-training staff, close enough to the work to watch change week to week. And that closeness may be exactly what broke him. The people building this technology are also the people most exposed to it and the ones most likely to be the first to feel real fear, whether or not the rest of us ever see what they saw. >> [clears throat] >> And I would tell you from my humble experience, I’m not a researcher or

07:05 machine learning PhD, but I have been covering these models from day one more than 4 years I would say when the first chat or something like that launched. Even now when I go out and talk with people here in Sydney or elsewhere online, you won’t believe that most of the people are still unaware of what you know AI can do. And that’s the truth. So we are living in our bubble of models tools every day but there is a world outside of it and that is where the

07:38 story ends not with an answer with a question about who among us is close enough to be right and who is simply too close to still be thinking straight and that is why this channel exist where we try to bring these things in open so that we could discuss it together and then we could um ensure that this technology would be used for the betterment of everyone. That’s it. Let me know what do you think in the comments. Um very very keen to know your

08:08 thoughts. Also again please subscribe to this newsletter as that helps and if you want to help out the channel please become a member. Thank you for all the