Latent (YC F26)

Know when your model is about to be wrong. Latent scores every answer from inside your open-weight model. Backed by @ycombinator

[Kunal Ranjan @kunalranjan23]

@danellisona Billions go into making models smarter, and a tiny fraction goes into making them verifiable. If I'm worried about anything, it's that ratio. More bets should go to teams like @runlatentai, catching hallucinations by reading model internals.

[Vedant Gaur @thevedantgaur]

In SF for Tech Week + COLM next week and would love to grab coffee with anyone building with LLMs. Especially if your model has ever been confidently wrong in front of a customer. DMs open!

[Vedant Gaur @thevedantgaur]

Some news: I'm in YC F26, building Latent.

I've done interpretability research for a while, and one thing keeps showing up: models often know they're about to be wrong before they say it.

Latent turns that into a score on every answer your open-weight model gives. Demo below 👇