AI that can’t write a single line of text is currently the fastest‑selling product on the market.
On September 15, Diego Almeida—who helped develop RLHF for ChatGPT at OpenAI—launched TypeSafe AI, the company he founded, after two years of stealth. The startup unveiled a model called “Jev” with a $40 million seed round. Three days later, Vercel announced that Jev was the quickest model ever adopted on its AI Gateway, and the waiting list cleared in just five days. The entire AI community was buzzing about it last week.
But this model doesn’t speak. When you ask a question, it returns a number instead of a sentence. If you ask, “Is this customer request a refund?” it won’t reply “Yes, it looks like a refund request.” Instead it gives you a probability, e.g., 0.93. Seeing this line up with a few papers I’ve been reading recently clarified where AI is today and where it’s headed.
1. AI That Separates Writing from Judgment
Jev only handles three types of queries: selecting one option from a predefined list, scoring on a fixed scale, or providing the probability that a statement is true. That’s it. In exchange, it’s extremely fast and cheap—responses take 0.07 s to 0.5 s, the cost is $0.042 per 1 M input tokens, and output is free. Because it delivers the answer in one shot rather than generating token‑by‑token text, this is possible.
The company calls this the “System One” model, borrowing the terminology from Daniel Kahneman’s *Thinking, Fast and Slow*: fast, intuitive System 1 versus slow, deliberate System 2. The name “Jev” comes from economist William Jevons and his “Jevons paradox,” which observes that higher efficiency can lead to greater—not lower—consumption. The bet is that if a single judgment becomes cheap enough, it will be embedded throughout software.
There are, however, some exaggerations. The claim “193.6 × faster and 444.6 × cheaper” compares Jev to the slowest, most expensive model on the market. When matched against models of comparable accuracy, Jev is about 25 × faster and 76 × cheaper, and its internal evaluation scores roughly 6 percentage points below the top‑tier LLMs. “No hallucinations” simply means the model won’t fabricate answers outside the provided options; it can still pick the wrong option.
What caught my eye more than speed or price is the nature of the work itself. After three decades designing enterprise systems, I’ve seen that a large portion of frontline tasks is not “writing” but “judgment”: Should this approval be escalated? Is this transaction suspicious? Which department should handle this inquiry? Until now we have forced massive LLMs to write paragraphs just to extract a yes/no decision—essentially asking a novelist to do a calculator’s job. Jev separates those two steps.
2. What the Papers Point To: Trustworthy Scoring Over Polished Answers
Reading about Jev reminded me of four recent papers that all converge on the same insight.
The first is Microsoft Research Asia’s *rStar‑Math*. A 7‑billion‑parameter model achieved OpenAI o1‑level math reasoning, raising the MATH benchmark from 58.8 % to 90.0 % and solving AIME problems with an average success rate of 53.3 %. The approach is essentially System 2: explore multiple solution paths slowly and select the best one.
What stood out to me was a key analysis: performance in deep reasoning is driven not by the model that writes the answer but by the “reward model” that scores each reasoning step. When a well‑trained 7 B reward model was paired with a 7 B solver, it outperformed a 72 B solver paired with a 72 B reward model. Adding a reward model leveled the playing field across sizes, suggesting that the grader can be more important than the student.
Even more intriguing, the models exhibited “self‑reflection”—they recognized a mistake, backtracked, and tried a different approach—despite never being explicitly trained on such data. Repeated high‑quality grading apparently induced an automatic habit of self‑correction.
The second paper comes from researchers at Hong Kong University, UC Berkeley, and Google DeepMind, titled “SFT memorizes, RL generalizes.” Supervised fine‑tuning (SFT) teaches the model to reproduce answer formats but collapses when conditions shift slightly. In contrast, reinforcement learning (RL) that only scores correctness learns principles that transfer to unseen rules and domains, and its generalization improves with more rounds of evaluation. The authors note that SFT is still valuable for establishing answer formats that enable RL to function effectively.
Reading this paper reminded me of onboarding training for new hires. A new employee who only memorizes the manual stops when faced with a situation not covered in the manual. In contrast, a new hire who receives feedback learns to infer the underlying principles even in unfamiliar scenarios. Whether human or AI, growth ultimately requires a standard that tells you what is correct.
The third work is AGUVIS from the University of Hong Kong and Salesforce. It is an agent that sees the screen, clicks, and types just like a human. By looking only at the rendered image instead of reading the entire webpage code, the input tokens per step dropped to 70% of those required by GPT‑4o, and the cost fell by 93%. The most interesting part of the paper, however, is the self‑identified weakness. In 40% of the failure cases the instruction itself was ambiguous, and the model had no way to say “I’m not sure” or refuse to act. The authors propose a “dynamic criterion” that decides when to think deeply and when to act immediately.
Doesn’t this sound familiar? The ability to pause when confidence
The core of AX is ultimately creating the criteria that AI agents will refer to.
If you're wondering how to create those criteria and how to operate thereafter, please feel free to reach out.
노우버스 is with you.