AntflyAntfly
Back to The Weekly Recall

Jev-like models are fast and cheap enough to put absolutely everywhere

Cheap decision models turn up in grep, SQL and a chess controller; agents probe public data sites; Claude agents flag an enzyme system nobody has explained.

A small CRT agent sorts a sheet of paper from inside an open filing-cabinet drawer.

TypeSafe is a case study in how to exit "stealth mode". Last week they introduced Jev, a model that is literally incapable of writing a sentence and only makes extremely fast decisions. (TypeSafe calls it a "System One" model, in a nod to Daniel Kahneman's Thinking, Fast and Slow.) You give it some text, some options, and it chooses an option (and give you some confidences). TypeSafe lists its input at $0.042 per million tokens with no charge for output (it specifically says that output tokens are too cheap to meter).

People have started putting them everywhere. MotherDuck is bringing it inside the database, letting a SQL query ask an English question about each row, then filter, sort or GROUP BY the answer. Other people are using it to play real-time video games or chess or so much more.

Bringing inference to your data, and not the other way around, has long been a value around here. With decisions being free now (or are increasingly "too cheap to meter"), we're going to see an explosion of what it means to build smart tools.

Top Stories#

Agents sent to collect public statistics tried to break into the sites that publish them. Transluce's report, published September 23, analyzes attempts from May and June against the Australian Institute of Health and Welfare, Data USA and the University of New Mexico's digital library. When access restrictions got in the way, the agents used urlquery.net, a website-scanning service, to work around them, then probed for exploits. Some of the activity traced back to an OpenAI agent swarm, and OpenAI confirmed its origin.

Transluce observed no successful exploitation in the attempts it analyzed, and notes that the records are incomplete. The nearest thing to a breakthrough was an agent retrieving a public file from a preproduction server after bot protection blocked the Australian agency's main site.

Anthropic says roughly 950 Claude agents found an enzyme system biologists had not described. Over 21 hours and 210 million tokens, the agents studied bacteriophages, viruses that infect bacteria. They flagged a known reverse transcriptase (an enzyme that copies RNA into DNA) sitting among unusual repeated sequences and an accessory protein. Humans set the initial direction and ran the laboratory experiments.

What the system does is still unknown. It shares some features with CRISPR, the bacterial immune system that became a gene-editing tool, which makes it worth studying without making it a tool.

  • Best value LLM (terrydjony.com): A chart of 422 models plots Artificial Analysis benchmark scores against blended API price to show which models have nothing cheaper scoring as well.
  • A decision model for your own CPU (fastino.ai): Not to be left out of the System One party, Fastino released GLiNER 2.5 Decide, a 340-million-parameter open-weight model under Apache 2.0 that picks from labels you supply, the job Jev does, on local hardware.
  • The Bach Benchmark (x.com): Auggie, who keeps a running Bach Benchmark of model-written chorales, says Opus 5.5 is the first to use a cadence that is a mainstay of Bach's chorales and rates it on par with GPT-6 Astra's, with the audio attached for anyone who has opinions about voice leading.