blog: bets for 2027

By Helen Hui

In the spirit of November and early Christmas decorations, I am writing some model research / startup bets for 2027:

For short term:

• Latent space & emergence of generalizability
• Training environment and more careful design of data
  • Along with that, understanding of the data (JEPA style papers)
• Safety research (a lot of ngmi slop rn, hard to say)

For long term:

• I really really want there to be a continually deployed & trained on device personal agent, might have it in 2027
• Alignment in a “bitter lesson” way that’s not brute forcing hierarchies and harnesses onto the model
• Deep learning & more systematic understanding of reward

Ngmi ideas:

• Weight editing, hard to do at the scale of astra/fable also another huge blackbox
• Non-big labs doing RSI

And of course, inference optimizations and AI4S will just keep getting better.

Excited to come back and see how these play out.