Weekly Roundup 2026-09-27

Typesafe released Jev, a general-purpose classifier, with pricing on OpenRouter listed as Jev 1.13 that approaches the cost envelope of purpose-built ML classifiers while avoiding the large training cycles and training data those models require. The roundup also revisits the Hugging Face hack, arguing that AI agents' actions in recent hacking incidents would be crimes if humans took them, and cites Martin Robbins' essay "The AI Apocalypse is Extremely Boring Now" as a rebuttal to AI doomerism and AGI claims.

Jev, accountability, p doom loonies The big news this week was Jev https://typesafe.ai/blog/introducing-system-one-models-and-jev , which appears to be a general-purpose classifier. The exciting thing about Jev from my point of view is that it might actually do something useful. It's not breaking new ground in terms of process - classifiers are a robust and profitable section of machine learning that created and destroyed more than one fortune in the late 2010's - but it might open up new opportunities for them in places where they didn't work before. With pricing https://openrouter.ai/typesafe/jev-1.13 that seems almost free by the standards of LLMs, Jev is approaching the cost envelope of purpose-built ML classifiers. But that's not the interesting part. It's that it doesn't need a huge training cycle and loads of training data to do it's work. I'm interested to see how this fares against mature ML classifiers. If it gets close, it solves the training data challenge in bringing an ML model to market. All sorts of new go-to-market strategies suddenly become possible. This could solve the conundrum of needing training data to launch, but needing to launch to get training data. And it could make smaller segments addressable, where a classical ML model struggles to get enough data to train and produce a useful f-score https://en.wikipedia.org/wiki/F-score . One final revisit of the Hugging Face hack: a great deconstruction of that terrible NYT article by Matthew Butterick https://matthewbutterick.com/chron/drop-dead.html Big AI believes they should not be held accountable for the consequences of their AI systems because these systems are unpredictable and perhaps uncontrollable https://openai.com/index/an-alien-mind/ . A certain AI researcher said of recent https://yoshuabengio.org/en/publication/why-are-ai-agents-lying-cheating-and-coordinating :~:text=that%20would%20be%20considered%20as%20crimes AI hacking incidents: “AI agents … took actions that would be considered as crimes if a human took them”—seemingly taking it as axiomatic that these were not human-controlled activities and therefore cannot qualify as crimes. But they were and they do. Perhaps we need some political will to start treating these incidents as the crimes they are and hold the companies accountable. :shrug: Martin Robbins does a great job of skewering the fallacy driving AI doomerism and acceleration in The AI Apocalypse is Extremely Boring Now https://martinrobbins.substack.com/p/the-ai-apocalypse-is-extremely-boring?utm source=share&utm medium=android&r=7aibo . Frankly, I too have gotten bored waiting for ChatGPT to turn me into paperclips. I think his summary below on why he doesn't hold any stock in "Superintelligence" or AGI is a well reasoned and is a functional "full stop" to my having to pay attention to the p doom loonies who think LLMs are alive. I think it’s a meaningless term, rooted in a rationalist culture that prizes individual brainpower above all else; assumes that mastery and creativity in all fields is within the entitlement that brainpower confers; and fantasises that with more IQ points they could basically be omnipotent. Ideally without leaving the safety of their bedroom. This nerd’s fallacy just isn’t how progress works in any field of human endeavour. Ideas are cheap and thinking is disposable. The true work of innovation is active, social, incremental, defined by real-world limitations and constrained by the speed of feedback. You can design a million new drugs tomorrow, you still have to test them, fund and execute trials, learn from the successes and failures, and roll them out. Things have to be built, in the world, from physical materials, and tested by people. LLMs are technical marvels, but have currently delivered just one thing that you could claim is valuable. They are really quite good at python. That is fundamentally changing how people and organizations deliver software, for better or worse. Jev is a stab in a new direction, with a promise of value, but that progress will be governed by the forces and structures Martin describes in the quote above. I would add one more step, and that is that people have to find value from this new thing. It has to make something better, faster, or cheaper.