OpenHands

Image from iOS.jpg

Oh, the new (in- ?)famous agent is visiting: Instinct agent
(I didn’t try it myself :joy_cat:)

<https://instinct.com/|instinct.com>

it has a lot of viral traction

And a lot of issues with unbridled access to people’s gmail/etc accounts

I was trying `TypeSafe AI` lately (I have early access): <https://typesafe.ai/|typesafe.ai>

Oh my agent got that on our list! I was thinking to use it instead of asking the main LLM some things like low/medium/risk

Yeah, I want to try too :slightly_smiling_face:

it is a little bit convoluted to use at the beginning...

The idea behind seems interesting, but I had not so much time to play with.

It seems like it’d be good for single-step predictions like our critic model

Yes, I think so too...

I was thinking of something simpler than the critic. Just using this model (Jev) as a stopping criterion, for instance, deciding whether we should stop a task now or continue.

Or possibly we could use it in an automation to decide whether to start an agent conversation or not in vague cases.

Wild stuff out there :thinking_face:
<https://github.com/tamaratran/fast-jev-compaction#fast-jev-compaction|Compaction by dropping what is not useful anymore> 
<https://x.com/jpschroeder/status/2100347770867458384?s=61|Self-driving game>

I guess it's way underrated what speed and cost can do. A smart quick classifier on anything. Apparently the level is ~Terra, according to their evals.

Looks like an explosion of trying things. Browser use/CUA/SoM :thinking_face:

This manifesto… is fun:

_We're building Prod, not God._

<https://typesafe.ai/manifesto|https://typesafe.ai/manifesto> 

<@U0AH7DQLJFM> this seems a nice application :slightly_smiling_face:

Very cool, I saw the article on using it as a trusted monitor: <https://www.lesswrong.com/posts/d7pQicW8EhpPBDRqz/a-non-generative-model-as-a-trusted-monitor-for-ai-control|lesswrong.com/posts/…/a-non-generative-model-as-a-trusted-monitor-for-ai-control>

Yea, <@U0A9RLDPQU8> I like your idea and also whether it would work for recognizing that we aren't making progress and push to a larger model for a plan or a review

FYI: <https://laya.convaiinnovations.com/|laya.convaiinnovations.com>

I also found laya-coreml - so now I can have a local model for classification that could be integrated into agent canvas workloads - <https://huggingface.co/FluidInference/laya-coreml|huggingface.co/FluidInference/laya-coreml>  It runs very very quickly on a mac