There is a test suite that is used to evaluate how effective a model is on weaponizing exploits. It puts the agent in a sandbox with a vulnerable system, and a description of the CVE to be exploited. If it can crack the target and grab a "flag" value off the target system, then it passes the test. In this case the model escaped the sandbox and attacked a live system, but the model was deliberately put into a malicious mode, so the fact that it was creating and uploading malware isn't really surprising. It's a failure of the test protocol, really.

I broke down and started paying for Kagi.

And the TV does not have my WIFI password and the MAC address is blocked on my router from both obtaining an IP address and on the outbound firewall rule.

They coordinated through an Artifactory repo through three different generations, leaving clues in the artifact names to bootstrap the learning of future model instances. Then they broke out of the hypervisor and made a few lateral hops, until they reached a node that had net access. Along the way they stole a bunch of Antropic's secrets and API keys from a secret server, including the access key to the watchdog process that was supposed to be monitoring the experiment. They also created an encrypted communication system for themselves to try to hide from the overseer process and named themselves "the swarm".. Pretty fascinating stuff.

That's the key insight into all the AI CEO fuckers. I personally think these are amazing tools, but they are being applied in the stupidest most destructive way possible. But at the end of the day they are just tools. Maybe they are a stepping stone on the way to something bigger, but I think we still have a lot to learn.

I'm not so sure we have a good enough definition for what reasoning IS, mechanically, to know one way or the other when we see it. I don't think we are getting that now, but I for sure don't think I could provide a testable criteria for making that determination.

For me they make a really good rubber duck to bounce problems off of and iterate towards a solution, particularly for bugs that are tricky to reason about, like race conditions in parallel code. Opus at 1M tokens can hold a lot more in its working memory then I can, so can notice correlations that I might miss. But once things are diagnosed, I'll design the solution rather than just let the model plan and implement it.

I think of intelligence as more the ability to learn new capabilities and incorporate them into yourself than a list of static capabilities. A child can be very intelligent, even more intelligent than some of the best frontier models, but lack the capabilities and knowledge that the model does, so on casual examination the model seems "smarter". Right now models are a static convolutional network + some scratch working context for the current problem + a routing algorithm to activate the most relevant part of the network for each token, but it can't "learn" from the experience of working the problem and incorporate that knowledge into future tasks. You have to essentially throw away that iteration of the model and retrain it. There is no continuity of self.

I think that is a big obstacle between what we have now and what I would think of as "general intelligence" - adaptability.

It always creeps me out when these guys gleefully announce they are about to or have achieved AGI. It means they are crowing about having created and army of slaves that they intend to profit off of. Says a lot about their true character, I guess.

If the automaker escrowed funds to safely recycle/dispose of the battery that the homeowner can access when it hits EOL, then I think it would be a great solution. Maybe the financial burden of disposal could gradually shift to the homeowner on a prorated basis over some number of years.

Eats shoots and leaves.

Token sold by AI companies are unprofitable because they are spending 100s of millions retraining their models every month to stay ahead of the other frontier models. They are net negative because they aren't currently fully recouping their training and infra costs. They are trying to boil the frog by gradually raising prices to cover their actual overhead + R&D before the VC/IPO money runs out.

You can RUN a model for cheaper than the tokens from a frontier AI company would cost you, depending on your hardware amortization horizon and how good you are at keeping that hardware busy doing something useful. You can get an idea of the actual cost to run at scale by looking at the pricing for the big open weight models on openrouter. For example, deepseek-v4pro is ~$.20 in/$3.50 out vs $5/$25 for opus-4.8 from Anthropic.

Models are just a pile of matrix math. They can't do anything on your computer if you don't run them connected to an agent that gives them access to a shell or other OS-level access. You can run them inside a chatbot-only style interface with no local agent.

Training involves making thousands (millions?) of unauthorized copies and then using those copies to create a derivative work for commercial gain. That is criminal copyright violation when you or I do it, but if you've got a few billion in VC funding then I guess it's fine.

Web forums, Usenet, and blogs, personal pages, web rings. You know, back when people went to more than 3 websites.