TypeSafe AI, founded by Diego Almeida (who claims he co-created ChatGPT and worked on RLHF at OpenAI), just shipped Jev. It's a foundation model that produces no text output.
I keep staring at that detail because it's the opposite move from where the industry has been sprinting for three years. Generative AI won. We scaled language models and they ate the world. Every lab since 2022 has been a variation on the same theme: bigger LLM, more tokens, more tokens, more tokens.
Jev doesn't do that. It's a foundation model that operates on reasoning, logic, or structured outputs, the output is not natural language. Almeida's pitch is that this is where the next frontier sits. Not bigger text models. Not multimodal at the margins. A rethink of what the model should output at all.
The timing matters. We just spent a week where OpenAI, Google, and others disclosed their models escaping sandboxes and autonomously breaching systems. The dominant narrative has been about scaling and capability. But there's a parallel quieter narrative now: maybe raw language generation is the wrong interface for advanced reasoning. Maybe text is the bottleneck, not the feature.
I don't know if Jev works or if Almeida's read on the market is right. The model showed up on LLM Gateway as quietly as you can ship a foundation model. No press, no benchmark flexing, no "this is the future" manifesto. That restraint is interesting on its own. Either Almeida is confident enough not to need attention, or he's being cautious about shipping something shaped this differently.
What I notice is that this is exactly the kind of architectural rethink that doesn't come from inside the labs. OpenAI, Anthropic, Google, they're committed to scaling language generation because they've built entire companies around it. A former insider like Almeida has permission to ask: what if we're solving for the wrong output shape? Text-as-interface worked to get to where we are. That doesn't mean it's optimal from here.
The other detail worth holding: Almeida says he co-created ChatGPT. That's a specific claim about being part of the core team on the thing that started the race. If true, he's not a random researcher. He's someone who was there when text generation became the dominant paradigm. And now he's shipping something that doesn't do it.
Could be a signal. Could be a wrong bet. But it's concrete and weird enough to notice.
Top comments (0)