←
Kyle Corbitt3,237 posts
Kyle Corbitt avatar

Kyle Corbitt

@corbtt

Today is a great day to grow

◷ Joined September 2012
Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 17

In the not-too-distant future any PR that doesn't have the full traces that generated it attached will be auto-rejected

i believe this is true. if a rogue employee says "add a backdoor that can never, ever be detected" it'll be very hard to detect. unless you have the intent.

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 10

Contributing to my 2yo's 529 plan like I'm Jeremiah buying his cousin's field

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 7

I also have it use sol subagents or just use sol directly myself for CUA or search-heavy tasks which OpenAI has a significant edge in

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 7

I have a solution to opus 5.5 over-refusals I'm happy with! 1. Use bb 2. opus 5.5 is main agent 3. when it refuses (eg. pasting API keys between systems), tell it to have sol 6.1 do it 4. Opus is fine responsibility-laundering thru sol, sol is fine doing the dirty work!

Post attachment

Back to 6.1-sol for most life automation tasks since opus 5.5 keeps getting hobbled by guardrails like "I can't copy/paste your known traveler number into the airline booking form since it's a government ID"

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 4

Wait, like, this is a huge training alpha drop. OpenAI claims that you can attribute specific behaviors in a downstream, post-RL model to a specific SFT dataset by just subtracting the SFT weight deltas from the final model? Was this a known technique?

We also identified a SFT dataset which massively increases rates of inventing permission restrictions or justifying inaction with safety. When we subtract the weight update from this SFT from the final RL checkpoint the performance gap goes away. SFTing with this dataset on other models and then doing RL did not cause underperformance however.

Post attachment
Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 4

One remaining gap in AI-as-employee tooling is "monitoring the situation". Agents can work autonomously until they're blocked waiting on a customer responding with some data, a days-long training run to finish, etc. I have to tell them to fetch the update. Missing primitive?

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Nov 2

Sorry you're telling me deepmind's first 1GW datacenter, going live in late 2027, right on track for the singularity, is named "Goodnight"? our timeline has the laziest screenwriters...

In late 2027 Google will have the largest datacenter: Goodnight - Claude, Texas either Google trains AGI there at the end of 2027 or Anthropic already took off and just buys the place

Post attachment
Kyle Corbitt avatar
Kyle Corbitt@corbtt·Oct 28

Building with coding agents and building as a middle manager are wildly different experiences and anyone who says they're the same is bad at one or both

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Oct 28

Ironically AI's most superhuman ability may be its ability to forget. With the proper setup you can trust it with secrets you'd never tell a living soul

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Oct 22

Maybe you can survive the small waves by scrunching your eyes and plugging your nose. But when the Big Wave comes you just gotta stand up and surf

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Oct 22

Back to 6.1-sol for most life automation tasks since opus 5.5 keeps getting hobbled by guardrails like "I can't copy/paste your known traveler number into the airline booking form since it's a government ID"

Kyle Corbitt avatar
Kyle Corbitt@corbtt·Oct 22

3 weeks later and I bet if I re-did this the answer mix would look different! My personal estimate of a temporary internet blackout (to control an AI worm, or massive hacking of ~everything simultaneously by a motivated bad actor) is probably 30%.

The USA has never had a nationwide Internet blackout. What are the chances we have one in 2027?