Short Fuse

AI might kill us all. That is increasingly the warning coming from the people building the most powerful models.The awkward bit is that they are much less specific about how.This week, we get into the latest wave of AI doom, why Anthropic and OpenAI keep talking about existential risk, whether self improving AI changes the equation, and why regulation may be doing more than keeping us safe. We also ask a fairly basic question: if the danger is as serious as they say, why are they still racing to build it?Plus bioweapons, regulatory capture, recursive self improvement, dictatorships, Snowpiercer and the possibility that AI eventually lies to itself.Cheerful stuff.

Has AGI finally arrived, or have we just moved the goalposts again? Duncan and Ollie pick apart what artificial general intelligence is supposed to mean, why the definition keeps changing, and whether today’s smartest models are still more book smart than street smart. Then, inevitably, things drift towards superintelligence, nuclear war and the Cobra effect.

Nvidia reportedly wants to buy Hugging Face for $12.9 billion. On paper, it is another enormous AI deal. In practice, it could give the world’s dominant AI chipmaker control of the platform at the centre of open source AI.We get into why Nvidia might want Hugging Face, why the deal could be a very clever hedge against OpenAI and Anthropic building their own chips, the possibility of a GeForce Now for AI, and the bigger question hanging over the whole thing: what happens to open source when one of the richest companies on Earth owns the place where much of it lives?

Anthropic is putting invisible watermarks into text generated by Claude. Which sounds like bad news for AI slop. Until you start asking what counts as AI generated in the first place.What happens if Claude edits one sentence? If AI reorganises your own ideas? If your company actively expects you to use AI at work? And if AI touches almost everything we write, does an AI watermark eventually tell us anything useful at all?Duncan and Ollie get into Claude’s new AI watermarking plans, the EU AI Act, AI slop, authorship, false attribution, workplace AI and the increasingly difficult question of who wrote what.Also: graffiti, The Hungry Caterpillar, authentically human crime and the beginning of The Sloppening.#AI #Claude #Anthropic

Who are the good guys in AI? Bad news: there probably aren’t any. We look at Anthropic, OpenAI, open weights, corporate backstabbing and the curious business strategy of telling everyone your product might destroy civilisation.

SoftBank has explained the future of AI using a golden goose, a conveyor belt of ASI eggs and the calm financial logic of a man trying to turn PowerPoint into scripture. Duncan and Ollie ask whether the AI boom has become a bubble, a religion, or just a very expensive bird with ARM inside it.

Prediction markets, smart glasses, a Kylie Jenner voice in your sunglasses. Meta will copy anything that moves. Today: the most profitable company on earth with no idea what it actually wants to be.

OpenAI is burning through money fast enough to make a small country nervous, so we worked out how long the runway really is and what happens when one of these AI giants can't pay its bills. It's also really, really hot.