Cameron Thacker

PhD | Co-founder @playmythical | 0-to-1 builder | recovering physicist | AI since ML | explore gt exploit | type II fun

[Cameron Thacker @CameronMThacker]

But I just don't see how it will retain enough specific information to even let the agent know how to backtrack to the level of detail it needs. Like you said, I think the summarization will drop the exact thing that provides relevant context. So even though the source is there, the agent can't find it

[Cameron Thacker @CameronMThacker]

@iso_morph Well one thing that maybe helps here is that nothing is ever really pruned. You can always follow the summarizations back to the source which is some append only memory log.

[Cameron Thacker @CameronMThacker]

@iso_morph I have played around with it, just to understand it, but I have not actually used it in an agentic harness or anything. While I think it's a clever idea, I'm a bit skeptical that it will work over longer time horizons for the reasons you listed above.

[Cameron Thacker @CameronMThacker]

@iso_morph Opt-memory by Victor Taelin seems to have some promise. Save all memories then condense into a binary tree of summaries. The idea being that if an agent needs to, it can always go 1 layer deeper for more fine grained understanding.

[Cameron Thacker @CameronMThacker]

@xeophon We have all been there. Luckily we have agents now so I don't have to devote any brainpower to that stupid problem any longer.

[Cameron Thacker @CameronMThacker]

@badlogicgames My agent earlier today told me that there are changes on main that are going to need a migration and aren’t backwards compatible.

You are really making me choose between stability and shiny new features and you know damn well which one I’m going to choose.

[Cameron Thacker @CameronMThacker]

@thdxr No shade meant here but isn’t that kind of what you are doing though? Working on a custom agent setup?

I think people focus on that because it’s the first new problem to solve with agents. Creating great products is still hard.

[Cameron Thacker @CameronMThacker]

@thorstenball How do you like buildkite compared to just doing the testing and building inside GH actions?

[Cameron Thacker @CameronMThacker]

@Porsche Beautiful car, but you need to hold on this for a bit.

[Cameron Thacker @CameronMThacker]

@_chenglou Okay I know this is a “fun” project … but I think you might have landed on the start of something really cool

[Cameron Thacker @CameronMThacker]

@mitsuhiko Decisions imply more intelligence than classification model or something I guess.

[Cameron Thacker @CameronMThacker]

@thdxr My gut tells me this is sort of temporary while we figure out what the new things to specialize in will be. Definitely moving up the stack.

[Cameron Thacker @CameronMThacker]

I've always thought about unit tests as catching regressions, not really type errors. The problem is, agents don't seem to know if the test is right or the code is right and it just fixes one.

The problem with the blackbox or hidden test suite is that it has to be maintained separately - and that is difficult! Agents are great at finding answers if anything leaks. I don't think just having these in a separate repo will work that well in practice.

[Cameron Thacker @CameronMThacker]

I find myself always drawn in to problems like this. It might have been initially shaped by the limited resources I've had, but I just see so much value in being able to run on device everywhere. I want background agents working 24/7 cheaply. So many use-cases don't require frontier intelligence.

[Cameron Thacker @CameronMThacker]

@VictorTaelin It’s a great idea and I can see this improving and evolving as well. Not sure if it was you, but someone said this has really only be effective with the newer models. Do you find that to be true?

[Cameron Thacker @CameronMThacker]

Is this for bots? Why do posts like this end up with so much engagement? Are 35k real people commenting on this???

[Cameron Thacker @CameronMThacker]

I also live in LA and one of my favorite things is that our friend group has really diverse jobs. While it would be fun to have more close friends who really know ML/AI better, I also really enjoy hanging out with people who all do different things. I’ve never lived in the bay but it must get tiring.

[Cameron Thacker @CameronMThacker]

@yacinelearning If you can do it efficiently enough then you get a nice tradeoff between compute and memory. I'm sure you know, but backprop can require a ton of extra memory, so if the tradeoff isn't too bad its interesting at least for some type of pretraining step.

[Cameron Thacker @CameronMThacker]

@willcb Goodhart strikes again! You don't just want "correct eventually".