OfficialJev Media & Talks
AI: too good to be true, too bad to be useful | TypeSafe AI
A TypeSafe media piece arguing that RLHF-trained language models please humans and assist rather than make reliable autonomous decisions, and asking what class of model comes next.
- Format
- Video
- Published by
- TypeSafe AI
- Last verified
- Oct 2026
Works forJev for Classification
Used inJev for Software & DevTools
Highlights
- Filed as media rather than a written essay: the page is the embedded video, with no accompanying transcript.
- Its premise is that RLHF optimizes models to please humans, which is not the same as reliable automation.
- It sets up the problem TypeSafe answers with System One models and RLCD in its launch material.
Watch out
The page carries only the embedded video and no body text, so this summary rests on the blog index blurb and the title rather than a quotable argument.
Questions about this entry
Short answers drawn from the linked project or page.
No. The page is essentially the embedded video with the standard site chrome; unlike the Bitterest Lesson or the benchmarks essay, it carries no body text to quote.
That RLHF-trained language models are optimized to please humans and assist, and that pleasing a human is not the same as making a reliable autonomous decision inside software.
The blog index only asks the question in its blurb. TypeSafe's written answer appears elsewhere: System One models trained with Reinforcement Learning for Calibrated Decisions return typed decisions rather than text.
Readers weighing why capable chat models still need humans in the loop, and what would change if a model were built for machine-to-machine decisions first.
It states the problem the launch post answers: chat models require human oversight, so TypeSafe built a model whose output space is typed decisions and calibrated probabilities instead of prose.
Media, alongside the Founders You Should Know interview, kept separate from the Thoughts and Company News posts such as the Bitterest Lesson and the Series A announcement.
There are no published figures on the page itself. Benchmark and case-study numbers live in the separate evals, antibenchmaxxing, and Jack & Jill posts.
More Jev resources like this
A TypeSafe media profile of co-founder and CEO Diogo Almeida, framed around why some AI applications over-promise while capable models can still under-deliver in practice.
Jev Media & TalksVideo#official#video#interview
Official
AIEdTalks part 4 on the batching trick: a loop that resends a long document for every question, versus one call that asks all thirteen.
Jev Media & TalksVideo#community#video#cookbook
Community
Max Jesch, on IBM Developer, walks Jev as a System One model that returns probabilities for classification instead of generating text.
Jev Media & TalksVideo#community#video#explainer
Community
Martin Keen, on IBM Technology, explains Jev as a System One model whose output is a decision rather than generated text.
Jev Media & TalksVideo#community#video#explainer
Community
AiNow dates the Jev release to September 2026 and walks live demos of a model that returns a decision instead of the next token.
Jev Media & TalksVideo#community#video#demos
Community
Fireship on Diogo Almeida's System One launch claims: no chat output, and the headline 200× speed and 400× cost multiples.
Jev Media & TalksVideo#community#video#explainer
Community
From the Jev community
Posts from builders shipping with Jev right now.