It's not every day the show gets to cover something that isn't another LLM. TypeSafe emerges from two years of stealth with Jev, a 'judgment model' that answers narrow questions with calibrated probabilities instead of prose — a cheap, fast decision layer that could get embedded everywhere in business software. Plus: Zuckerberg's rebuttal to the slowdown discourse, and Bernie Sanders and Steve Bannon share a stage against the oligarchs.
Judgment models could be the cheap decision layer business software has been missing.
TypeSafe's Jev doesn't generate text at all — it answers narrowly defined questions with epistemically honest probabilities, and claims to do it 20–200× faster and 40–400× cheaper than LLMs. That makes it useless for chat and potentially perfect for the thing so much office work actually consists of: reading something and deciding what happens next. If the category holds up, the pattern to watch is LLM proposes, judgment model decides, code executes — with judgment cheap enough to check everything, every time.
TypeSafe emerged from two years in stealth with Jev, trained via a new method called reinforcement learning for calibrated decisions — claiming 20–200× faster and 40–400× cheaper than LLMs with output tokens free. The founder, who says he co-invented ChatGPT, frames it as the answer to why superhuman chat models haven't led to AGI: "the shortest path to AI-based economic revolution."
Where existing LLMs optimize for chat responses human raters prefer, TypeSafe's new "System 1" models optimize for calibrated decisions — answers to narrowly defined questions delivered as honest probabilities, categories, or scores, built on a new architecture and a parallel sampler for maximum efficiency.
Every's Mike Taylor: ask "does this customer sound angry?" and Jev returns 0.9 — a 90% probability — which software can act on directly, escalating or deprioritizing the ticket. A chatbot's flowery "You're absolutely right, this customer does sound very angry" would crash the same program, which expected a number between zero and one, not an essay.
Jev cannot generate text, so it's no substitute for LLMs in general. It replaces the category of work where LLMs have been square-peg-mashed into a round hole: a class of decision-making suited neither to dumb, unintelligent code nor to slow, expensive LLMs — fast, cheap AI decisions embedded in software.
Does this message need a response? Which department handles it? Is this a bug or a feature request? These are judgments about meaning that resist fixed rules. TypeSafe's docs explicitly recommend breaking complex decisions into small questions and combining the results in code — routing tickets in support, sorting inbound leads in sales, flagging copy against style rules in marketing.
YC founder Nathan Fleury's framing of where this fits alongside generative models. In a support workflow, judgment models score whether a message describes a product problem, indicates repeated failed support, and sits near a commercially significant deadline — then a generative model drafts the reply, and Jev checks that draft against narrow criteria before it goes out.
When a check adds noticeable delay or expense, teams run it only on selected cases or at the end of a task. When it's fast and nearly free, it can run on every incoming request, after each draft revision, across many candidate documents, and before an agent takes a consequential step.
Mike Taylor fed Jev 37 documents and asked 21 AI-tell questions of each — does the text repeat an idea without adding evidence? force a symmetrical both-sides argument? — and got 777 judgments back in under 0.7 seconds for an estimated quarter of a cent. That's fast and cheap enough, he notes, to AI-check everything everyone at your company has ever written.
One framing from the reaction: most software is ultimately a giant tree of if-this-route-here, if-this-escalate, if-this-ask-a-human — and Jev asks what happens when those if-statements can parse messy human context. Early fits people see: fraud and risk, support routing, moderation, QA automation, lead scoring, compliance, and agent orchestration.
Matt Stockton's read: lots of business problems are really classification or regression problems, solved today with people and process — or, worse, with LLMs that are the wrong tool for the job. Classic ML demands labeling data, training, and hosting; Jev offers LLM-style API ergonomics for pre-LLM techniques.
A personal agent gets far by learning one person's preferences; a team agent has to judge who owns this, who's waiting, and whose approval is required. A salesperson's "we should be able to support that integration before your renewal" implicates sales, engineering, product, and customer success — and a judgment model can flag it as a possible cross-team commitment and request the missing decisions.
Judgment models cannot do the work generative AIs currently do, so they'll have to live inside the more complex model stacks the show has been discussing for months. The flip side: because judgment intelligence is this inexpensive, it can be integrated extraordinarily deeply into the automated systems everyone is building.
It's day one for a concept a very competent team spent two years building, and it still has to prove out in practice. But NLW's read after digging in: this has the feel of something both important and obvious — the type of thing that, once it exists, we'll be surprised we didn't have for so long.
After staying quiet over the weekend, Zuckerberg argued every lab has "the responsibility and incentive to move at the pace required to train its models safely." His case rests on two ideas: people won't use agents misaligned with them, so labs have a natural incentive toward alignment — and labs face significant liability if their models cause harm.
Zuckerberg says Meta delayed Muse by several months to work on safety — "we didn't call for everyone else to do this before we would. We just did it" — and has committed the significant majority of its compute to serving people rather than racing toward recursive self-improvement, a commitment he argues other labs can match.
One camp — Matthew Berman, Bill Ackman — cheered the framing of safety as an economically incentivized feature and "the proper approach to AI development." The other argued Zuckerberg simply lacks the trust or standing to make the argument regardless of its merits, with Kevin Roose's sarcastic line about Zuckerberg being the person best suited to protect us from powerful new technology capturing the mood.
At the Future of Life Institute's Pro-Human Assembly in Washington, Bernie Sanders and Steve Bannon delivered near-identical messages in the strangest alliance of this political cycle. The event was less about existential risk than about the class struggle AI has come to represent — Bannon: "we can never trust what an oligarch says."
— Senator Bernie Sanders, at the Pro-Human Assembly in Washington. Sanders' framing, if the lives of every man, woman, and child are going to be fundamentally changed by this technology. Bannon called it "a hinge in history" that has to be handled correctly — starting with never trusting what an oligarch says.
Both movements trace to the populist anger of 2016 — Bernie's near-miss nomination run and Bannon's architecture of the first Trump administration — rooted in post-GFC fury at unaccountable institutions. NLW isn't dismissing the merits of the common ground; he's noting these stories were always more intertwined than they appear, and this issue has real power to scramble political alliances.
Echoing Jasmine Sun's reporting on data-center opposition, speakers framed the risk not as cybersecurity or bioterrorism but as a future forced on communities with little input. Texas Democrat Greg Casar, sponsoring Sanders' superintelligence bill: "The answer is simple. We ban AI systems that are too powerful for humans to control."
NLW contends we've moved into a phase of negotiating a relationship where citizens and governments have a stake — which means more specificity and, dare he say, nuance. Exhibit A: Glenn Beck signing the Pro-Human pledge alongside people he disagrees with on the specifics of data centers.
— Jensen Huang, at Salesforce's Dreamforce. The safety debate followed everyone to Salesforce's Dreamforce: Dario Amodei pitched standards the industry could organize around, Altman said he's very confident the industry can do this safely, and Benioff said every company has a responsibility to uphold ethical standards.
Salesforce's first in-house model in quite some time is a fine-tune of an NVIDIA model designed to handle sales management within the CRM — reinforcing the role open source plays in the enterprise by enabling narrow vertical models.
The new umbrella term covers connectors that let third-party agents access Salesforce data — a platform-agnostic commitment to letting any agent become the interface.
