When instructions and criteria ask for different things, accuracy drops; treat criteria as an extension of the instruction.
The clearest example in the docs is a Noul whose true maps to no and false to yes, which performs worse than the natural phrasing. Clear, human-readable alignment between instruction and criteria is what the model was trained on.
The guardrail is to align the two with precise language and to prefer question shapes an average person could read and answer consistently.
Definitions that connect to this one.
Posts from builders shipping with Jev right now.
Computer use at 155× cheaper than Opus 5
i built computer use using @typesafeai ! it is 155x cheaper than opus 5, ~20x faster, and generalizes across OS's more on how it works in the vid & thread below:
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
Foreman keeps coding agents on task
I just open sourced Foreman: a software factory foreman built with @typesafeai's Jev. Coding agents work the factory floor. Foreman watches them, continuously assessing progress, completeness, tests, drift, and verification, and intervenes when needed. GitHub: Show more
Unclutter: an ad and slop blocker that runs on Jev
introducing Unclutter: a smart ad + slop blocker with Jev 🤓 it auto cleans up pages from slop elements: ⬖ ads ⬖ cookie banners ⬖ upsells ⬖ bs dialogs BYOK. open source + free, download below 👇
An open-source BS meter for debates and investor calls
🚨 Open Source Jev BS meter you can use this to analyze any debate / investor call / interview / sales pitch / podcast video fact check live , for example this dario interview cost 60 Jev calls / 111K tokens / $0.0047 github.com/ChetasLua/jevm…
🚨 I gave the Trump vs Kamala debate a live BS meter using Jev every sentence, both candidates, 5 yes/no questions each 1,191 Jev calls / 1.18M tokens / 415 ms median total cost : $0.0497 same questions for both, clips picked by one fixed rule, not a fact-check
A Jev-shaped model on Cerebras and Qwen
Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost. TypeSafe was way cheaper, and did beat Qwen on performance. Closest we can get using LLMs I think. Source: github.com/iammrduncan/ty…
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
Jev benchmarked on two public safety corpora
1/ Benchmarked TypeSafe's Jev on two public safety corpora. It doesn't generate text, it returns calibrated probabilities you threshold in code. 96.5% on prompt injection, all 662 messages in deepset/prompt-injections. No tuning. 325ms p50.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
