Diogo Almeida Reveals the OpenAI Memo That Sparked TypeSafe AI and Jev
TypeSafe AI founder Diogo Almeida shared how an internal OpenAI memo titled “We Could Have Had AGI Last Year” led him to spend two years in stealth building Jev.
Two months after TypeSafe AI emerged from stealth with Jev, founder and CEO Diogo Almeida published a candid origin story on LinkedIn tracing the model back to an internal document he wrote while working at OpenAI: “We Could Have Had AGI Last Year.”
In that memo, Almeida started from OpenAI’s own working definition of AGI—AI performing most of the world’s economically valuable work—and pointed out an awkward gap. Even GPT-3.5-era models had already passed median human performance on many tasks, yet AI was still automating less than 1% of economically valuable work. By volume, most of that work was repetitive, unglamorous software work such as data entry and classification that models should have been handling at scale.
“Making models reliable turned out to be MUCH harder than making them expert people-pleasers.”
Diogo Almeida, Founder & CEO, TypeSafe AI
Why people-pleasing models stalled automation
Almeida’s diagnosis was that nearly all post-training investment had gone into optimizing models to please human readers in chat rather than building machine-native models that software can depend on. His pitch inside the memo was straightforward: dedicating an RLHF-scale effort for one year would be enough to create a new class of models designed directly for software interfaces.
The original OpenAI pitch
- Target repetitive, high-volume software judgment work rather than prose chat.
- Dedicate a one-year RLHF-scale effort to machine-native reliability.
- Optimize for calibrated probabilities software can branch on directly.
What happened next
- Sam Altman urged Almeida to drop his existing work and build it.
- A planned one-week prototype grew into a two-year stealth research effort.
- Jev launched Sep 15, 2026 and triggered an industry-wide System One wave.
When Almeida showed the memo to OpenAI CEO Sam Altman, Altman told him to drop his current work and pursue the idea instead. Almeida joked that he already had a job and set the proposal aside for a while. When he eventually began experimenting on his own, he expected a prototype to take a week; it took far longer, but early signs of life convinced him to leave OpenAI and commit full time to TypeSafe AI.
Two years in stealth, two whirlwind months in public
After two years building in stealth with co-founders Sasha Sheng and Erik Gafni, Jev launched into early access in mid-September 2026 and quickly triggered an industry-wide rush toward non-generative decision models. Developers responding to Almeida’s post echoed his central point: in production analytics and software pipelines, consistent, verifiable behavior matters more than conversational polish.
In analytics work, the failure I see most is an output that stays consistent while the metric underneath it has quietly changed definition between teams, so it is wrong and nobody notices.
Yogesh Arora, commenting on Diogo Almeida’s LinkedIn post



Developer reactions & follow-up context from the LinkedIn thread
Commenters across ML engineering and infrastructure teams—including Rohit Gajaraj, Aaron D. H., and Pahul H.—highlighted that static public leaderboards decay quickly once production definitions shift. Several noted that Jev revived classical ML calibration discipline inside modern transformer stacks.