Log in
Cleared Hot is one of my indie favourites from 2025, currently in Early Access this nostalgic helicopter shooter is about to get a big expansion in content.
Read the full article on GamingOnLinux.
Portal: Alive & Kicking has officially arrived on Steam, giving the classic Valve game Portal a bit of a makeover.
Read the full article on GamingOnLinux.
Some bureaucracies have an unlimited appetite for surveillance with machine learning, the AI that actually works! Particuarly for abusing people’s privacy.
The British Transport Police just answered a Freedom of Information request from Liberty Investigates. The BTP ran a trial of live facial recognition in nine large railway stations in London from February to July this year,. [Liberty Investigates; Guardian]
The BTP scanned half a million faces hoping to catch known offenders on the loose. The total cost was £320,786 and over 100 hours of police time.
The trial netted them one match of a suspect over six months — and that was a false positive.
The BTP told the Guardian the trial totally helped them make lots of other arrests, honest! Just not ones in the trial.
In fact, the BTP and Transport for London are running the trial another four months and extending it to some Tube stations.
TfL said the live facial recognition will “specifically tackle violence against women and girls,” by this process that catches no bad guys. You’d almost think they were making an excuse.
Over half of police forces in England and Wales are deploying live facial recognition out on the streets. The Metropolitan Police and the Mayor of London have announced that live facial recognition will be deployed on Oxford Street in the West End.
Machine learning is great! You can spend public money, treat the public badly, and still not achieve a thing.
It’s pledge week at Pivot to AI! Please put in your $5 to keep us going! It really helps!
Three of the founders of AI hosting company CoreWeave — Brannin McBee, Michael Intrator, and Brian Venturo — just announced yet another planned sale of CoreWeave stock, totalling over a billion dollars’ stock sales for each since the company went public. This means that whatever happens with CoreWeave and the AI bubble … they’ve got theirs. [Bloomberg, archive]
CoreWeave has never been a solid business. It spent the past three years piling debts on debts, because this was the AI revolution! This is the future!
But bubbles pop. CoreWeave’s whole deal was to sell a financially impossible company to people who were fully equipped to know better — but they had powerful incentives to lie to themselves.
CoreWeave was founded as a cryptocurrency miner. Crypto crashed and the company pivoted in 2019 to cloud scientific computing on GPUs.
In mid-2022, CoreWeave started buying heavily into Nvidia GPUs, just in time to catch the wave of the AI bubble later that year. Nvidia owned 5% of CoreWeave and the two were very cosy. By 2023, CoreWeave was putting the Nvidia chips up as collateral to borrow money to buy more Nvidia chips. This booked a lot of revenue for Nvidia.
CoreWeave did its IPO in March 2025. This was the first big public offering of the AI bubble. Ed Zitron wrote up a nice analysis of CoreWeave’s S-1 prospectus: “CoreWeave Is A Time Bomb”. None of the numbers added up to a business that could ever make it out of the AI bubble alive. [Where’s Your Ed At]
CoreWeave’s main business was as a backup provider for the hyperscalers when they couldn’t build out their own data centres fast enough. The didn’t have a lot of revenue and had taken on ghastly and ever-increasing amounts of debt. The company was spending $1.43 to make one dollar. The S-1 only worked if there was massive AI demand coming soon — which was already not showing up.
None of the numbers have improved in the year and a half since. CoreWeave’s debts remain crippling.
CoreWeave’s one weird trick is to get their debt listed as investment grade and sell it to insurance companies. McBee from CoreWeave told the Financial Times that insurance capital was: [FT, archive]
unlimited. Like, that’s all the money in the world that sits in the insurance tranche.
The stock dropped about 15% on the news of McBee giving the game away.
McBee, Intrator, and Venturo have sold a total of $3.3 billion in stock since the IPO. They’d sold another $450 million before the IPO. The company told Bloomberg:
They’re making personal financial planning decisions that any responsible founder with concentrated holdings would make.
Particularly when the concentrated holdings are in a massive and expanding debt bomb that will blow the moment CoreWeave is not getting OpenAI and Anthropic money via Microsoft.
CoreWeave is one of the best funded and best connected AI bubble service companies. Everyone else is going to do worse.
Anthropic have a new scary campfire story on why everything they work on is evil and out to get you! This one is called “AI models may be capable of self-replicating without access to a datacenter.” Here’s the headline claim: [Anthropic]
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes.
… If small models improve at their historical rate, they’ll match this capability level by the end of this year.
The end of this year.
The paper constructs a scenario. This is otherwise known as making up a sci-fi story about killer robots, then telling the incredibly credulous press that this can happen in the next two months, and then they print it as-is.
Here’s Joseph Menn at the Guardian swallowing this tall tale, hook, line, and sinker: [Guardian]
the red team, as such internal testers are called, said it created a malicious AI program that spread autonomously through internet-connected “smart” televisions and could do the same inside other ordinary devices.
No they did not! The team made up a story about a program that might do that if a hundred other things that don’t exist come true!
Menn also compared this scenario to Harry Potter, because of course he did. Read another book!
Menn spoke to an “expert” who told him:
What differentiates AI worms from typical reproducing viruses is that AI models can learn as they go.
Chatbot models do not learn as they go! That’s why the AI vendors spend a fortune on training and post-training!
Anthropic’s scenario posits an AI network worm — a self-replicating program.
We’ve had computer worms since 1971. Someone else (not Anthropic) built an AI-based worm earlier this year and tested it on an isolated network. This did not propagate the model itself, because chatbot models are quite big. Instead, the test worm took over smaller machines it had access to and ran tiny models on them. [arXiv, PDF]
Anthropic took this earlier experiment and says it means a whole model “could” copy itself from machine to machine. This means it “could” set up a botnet of TV set-top boxes! And spread all by itself! You can’t prove it won’t!
They did run a small model on a network of 15 Android set-top boxes. It ran at a blistering 0.39 tokens per second.
There are a few other steps that would need to be shown for the threat to hold. The bot has to scan the wider internet. It needs to hack into the set-top box. This has to be done with their tiny model, at less than a token per second. Then it has to copy itself.
The researchers don’t bother testing if any of these work. They just say “not our focus.”
It is true that embedded boxes have notoriously awful security. There have been some huge botnets on these insecure boxes. Those botnets run actual code, which is quite a bit smaller and faster.
Those botnets existing does not let you handwave how your evil chatbot could totally do that for sure.
If you can create a model that’s smart enough to hack tiny embedded hardware while it’s running on tiny embedded hardware, and if it can figure out how to hack new small boxes, and if it can do all of this without the researchers helping it along at every stumble, like all the previous AI hacking researchers did … then sure. But you can’t just assume any of that.
Anthropic says “could” a lot in this paper. “Could” and “might” and even “will” are all words that mean “doesn’t”. “Could” is not enough.
And they admit that “doesn’t” is the word for how this works: [Al Jazeera]
Although no current AI model proved to be both powerful and small enough to self-replicate reliably on a smart TV used in testing,
Much like no current slingshot has proven elastic enough to fire a space shuttle. Obviously we just have to scale up.
The models would only have to shrink by 99% and run with a thousandth of the data centre compute they’re made for while still being able to hack into arbitrary gadgets and come up with new hacks. Autonomously. And they’d have to do all of this by the distant magical future era of the end of 2026.
But maybe I’m wrong. Show me! You’ve got ten weeks.
This paper is marketing. Not to you and me, but to credulous government regulators. The current US administration already believes the chatbot is the magical truth machine, and they think Anthropic has the best magical truth machine.
This paper is a message to say: regulate AI. Make rules so only the huge vendors like us can comply. Then the AI can convert NASA to use giant slingshots.
If you have built an application with an interactive map, you know what you are in for. You pick a tile provider such as Google or Mapbox, or a self-hosted stack like GeoServer or Martin, and your frontend requests thousands of small image or vector tiles as the user interacts with the map. A server processes each request: it reads from a database, renders the tile, maybe caches it, and sends it back. That server is infrastructure: it needs to be deployed, monitored, and scaled.
Next Page of Stories