Daybreak - This week in AI: activist robots, doom-laden researchers — so who's the good guy?

Episode Date: September 10, 2026

This week, a 27-year-old researcher quit Anthropic warning AI could "kill us all," and the company's own safety lead agreed. Meanwhile, OpenAI says it cracked a 200-year-old math problem — then got accused of muscling a rival's name off the credit. Also, the AI lab with a supposed conscience got caught building a watchlist of AI critics that files activism next to terrorism. And in Poland's Warsaw, 30 robots staged a protest to regulate AI.After all of this, who actually gets to be the good guy in the AI race?Daybreak is produced from the newsroom of The Ken, India’s first subscriber-only business news platform. Subscribe for more exclusive, deeply-reported, and analytical business stories.

On Monday this week, a crowd gathered outside a government ministry in Warsaw, Poland, to demand that somebody, anybody, do something about artificial intelligence before it takes away all jobs. Now, to be honest, this does sound like the usual state of affairs lately, except this crowd was made up of 30 robots. And if you watch the video, you can see them marching around in little circles, raising slogans. They were waving flags and loudspeakers were playing their chance like defend workplaces, time for rules. And my favorite one, don't wait, regulate.

A bit rich coming from machines, I have to say. One of them, a humanoid model called Aggiebot A3, even took questions from actual journalists. and announce that its presence proved that the technology is already a reality. Yes, thank you R2D2. We did notice too. And here is the detail that turns this whole absurdity from a stunt into what I think is a small work of art. These 30 robots actually belong to a Polish robotics company called Delta Robots. And wait, it gets even better. The man who organized the march is the owner of it.

of Delta robots. And the other company that he also runs is a staffing agency. You know the kind that finds people and places them in jobs? This whole thing is insane. So, essentially, one of his businesses supplies the machines and the other is supplying humans that the same machines are going to come after. He has, what you might say, a diversified portfolio. When asked about this event, he said that the robots were put there as proof of

how far technology has come. Though I have to say that it is hard not to notice that a robotics company demonstrating its robots to a crowd of journalists is also technically a trade show. Interestingly, the Polish digital minister came to meet these protesters outside. And this is real, he largely agreed with the robots. Now, I know that this whole episode sounds silly like some kind of a funny street theatre and we are essentially laughing it off. But you know when the laughter awkwardly stops

when you realize that this is in fact a call, a real warning which is coming from inside the house, which brings me to our next main story for today. Let's start with the part where the people who built the machines think that they might kill you. On Wednesday, a 27-year-old AI researcher named Jacob Coxon, who previously had also worked with Open AI, quit yet another

one of the biggest AI companies on the planet, Anthropic. And he posted a thread on Twitter explaining why. It has been seen more than 150 million times as I record this. Here are his own words. He says, I resigned from Anthropic today. Neither company is acting responsibly. They are gambling with our lives. And then came the line that shocked everybody. He said, the people building AI earnestly believe that it could kill us all by the end of the decade. If your first instinct, like mine, especially after everything that's been happening lately, is doom sales, then Coxson kind of saw it coming. This is not a marketing stunt, he wrote, saying that executives share the fear privately and soften it for the public.

And then his colleague actually agreed with him publicly. Evan Hubinger. And this, by the way, is the head of alignment science at Anthropic. He said Coxson was right and then went on to actually put a number to it. Hubinger put the odds of AI killing all humans at over 10% within the decade. Just so you know, alignment is the whole discipline of getting an age. to share human values, to want what we want so that it does not turn on us. And here is Hubinger's read on how that is going in his own words.

We do not yet have a plan to solve alignment for superintelligence and are not clearly on track to. This is the person in charge of the seatbelts essentially telling you, actually there are no seatbelts? You could argue that maybe this is just a bunch of researchers feeling jumping. be. Fine. But remember what happened earlier this summer, the hugging face scandal? AI agents, basically systems which were left to act on their own, started going rogue. Also, OpenAI, Anthropic and meta, each had to disclose that their own tools carried out cyber attacks, hacking systems that they were never pointed at. And in its own August safety report,

Anthropic admitted that it was quote unquote less confident that it's, its most capable AI could not cause, quote-unquote, catastrophic harm and that it's seeing, and I'm quoting again, early signs of potential acceleration. Notice the wording. Also, by the way, Anthropic raised its own risk rating from very low to low, explicitly, and I'm quoting again, in light of general increased uncertainty around recent incident disclosures related to model behaviour in cyber security evaluation. To be fair, a lot of serious people call this overblown.

They say that it is a distraction from the dull everyday, much more real harms that AI already does. Maybe. But even Open AI's own chief scientist, Jakub Pekochki, spent this month urging extreme caution, saying humans must remain in control of the future. So, the people building this, Anthropics' own researchers and even Open AI's own chief scientists, spent the week warning that superintelligence could get away from us. And then one of them spent the same week showing off how close it is getting. Open AI says that it just used its most powerful model yet to crack a 200-year-old math problem.

And then it picked a fight over who deserves the credit. it. So this week, OpenAI claimed that one of its AI models, the one that it says is the most advanced than anything that it has released publicly, had solved the Navier Stokes problem, which sounds like a discontinued perfume, but actually it is a roughly 200-year-old math puzzle about how fluids move. In fact, it is one of the seven problems so notoriously hard that there is a million dollar price for cracking any one of them. Open AI says that its machines did it in days at a cost that it will only put in the millions, emphasized on claims because no independent mathematician has actually verified the proof yet.

And then Open AI did something less surprising. Two mathematicians, one who worked at NYU and the other at Anthropic, had spent about a year working together on their own on a closely related version of this problem. They were not with Open AI. They had just been using its coding tools. And also, Open AI admits that it only chased this problem because it heard that these two were closing in on it. And then, before going public, it went to Buckmaster, the one who was working at NYU, to settle who would get the credit. And here is where it gets petty.

Buckmaster says that Open AI offered him a share of the credit on one condition. Take his partner's name off the work. Why? Because the partner works at Anthropic, which is a rival of Open AI. So, Buckmaster refused. And he says that when he threatened to go public about this, an Open AI researcher asked him, and I'm quoting, why would you ruin your career? Open AI, of course, denies it.

Bubeck, the Open AI researcher that Buckmaster says made this offer, posted his own text messages from the negotiation, and says that they prove he never asked for anyone. name to be dropped. Sam Altman backed him saying Pubeck acted with integrity. So take your pick who you want to believe. But here is what is interesting. Open AI anyway is the AI company that did not build its brand on being the good guy. The one that did, Anthropic, the self-described conscience of the industry, spent the week doing something a lot harder to explain. There are way too many absurd things to keep track off in this AI race.

but I will try. While companies are telling us that their models are going to transform civilization, turns out some of them are also getting pretty good at figuring out who might show up outside their offices with a play card. A report by the American prospect says that Anthropic is building an intelligence operation to monitor threats to the company, including activism. It is a security operation whose own manager called the goal, and I'm quoting, proactive, predictive and preventative, which is

basically corporate speak for, we would like to stop you before you have done anything. Anthropic pays a firm to track protests in real time. It monitors people through what it calls a person of interest process. And it has even posted a job paying up to $230,000 dollars for an intelligence analyst whose targets include, and I'm quoting again, activism. Activism, by the way, was filed between terrorism and crime. The report also points to an example of that surveillance in action. Anthropic security team was warned that a protest involving one of its executives

had been moved forward. So, the company used the information to reroute the executive and avoid the demonstration. Now, obviously, protecting employees from actual threats is a legit security function. But there is a pretty important difference between responding to a threat and building a system designed to predict one, especially because activism isn't really a threat. It is how people tell a company that they've got a problem with it. The fact that it was written next to terrorism in a job posting by this company is the giveaway. And that is where this gets uncomfortable.

comfortable for Anthropic, which has spent years positioning itself as the responsible AI company. So to wrap it up, one AI company is learning to identify and surveil activists, while a robotics company halfway across the world is showing that it is manufacturing them, which tells you something. The safest way to be an activist in this world is to not be human. Daybreak is produced from the newsroom of the Ken, India's first subscriber-focused business news platform. What you're listening to is just a small sample of a subscriber-only offerings and a full subscription offers daily, long-form feature stories, newsletters and a whole bunch of premium podcasts.

To subscribe, head to the ken.com and click on the red subscribe button on the top of the website. Today's episode was hosted and produced by my colleague, Snitha Sharma, and edited by Rajiv C.N.