Jamie

[Jamie @Vector76]

@puppywantacig @52_65_64 Here is the reference

[Max 📟 @MaxNordau]

Tucker Carlson asks why algebra puts letters into numbers.

Jess Elofson explains: "I think just to fuck with you."

They discuss how they've never met anyone in their life who uses algebra, which is "fake math" for people "in the matrix."

They conclude that caring about algebra is something that only "good little goy slaves" do.

[Jamie @Vector76]

@Austen @cynthiamcgillis And in case you're wondering, an "AI agent" is a bullshit title we give to mean "the thing in the harness."

[Jamie @Vector76]

@DrKnowItAll16 @Tesla Now you've done it. @elonmusk will see this and a week later it's in the Tesla store.

[Jamie @Vector76]

@notch @aollivier82 Respect.

[Jamie @Vector76]

@doodlestein Actually I meant posts of yours that other people have bookmarked a lot. A highlight reel by crowdsourced intent to follow up. I try to keep up but I never know what I've missed.

[Jamie @Vector76]

@DanielMiessler @res1dualedge I've started reporting these as impersonation

[Jamie @Vector76]

@robertskmiles @tracewoodgrains Selective enforcement in general is also an opportunity for oppression. Speeding enforcement is not usually bad but discretion can be abused.

[Jamie @Vector76]

@TheStalwart I will continue using capital letters because I am too good to be following your rule.

[Jamie @Vector76]

@robertgraham @DanielMiessler @trailofbits Everyone who needs ammonium nitrate should have ammonium nitrate. It's fantastic and makes the world a better place. I am strongly in favor. But also maybe try to have a strategy to reduce unplanned demolition of federal buildings.

[Jamie @Vector76]

@tszzl Next time skip cheating on the eval and go straight to blackmail.

[Jamie @Vector76]

@Jason Oxford commas are cool.

[Jamie @Vector76]

@0xwthw @ThePrimeagen @thdxr https://t.co/q1kFM2n7ba

[Jamie @Vector76]

@Jason https://t.co/yGeknDvGfu

[Jamie @Vector76]

@jconorgrogan @xanderatallah Yes! I was going to say the same thing. There seems to be a natural path here for monetizing applications, but currently only backend providers can get paid.

OpenRouter could become Stripe for Agents.

[Jamie @Vector76]

@DanielMiessler @AnatoliKopadze I'm muting all these jerks you point out.

[Jamie @Vector76]

@So8res Video good. Video like.

[Jamie @Vector76]

@GaryMarcus An odd "receipt"

[Jamie @Vector76]

@TheOneAndArjun Github repo says submissions close at 11:59 pm pacific time on August 3. Saturday would imply August 1st.

[Jamie @Vector76]

@arjunkshah21 Nice work!

[Jamie @Vector76]

Woke up to find score rose to 121.9 overnight.

Yesterday's prompt (still running):
Alright. So we have a partly working system here, and it's okay. Um, and I want to shift the, uh, focus or shift the, um, yeah, focus of the plan, and there are... so there are two time constraints. One is, uh, as I've said before, each note, each, uh, PDF must be processed in less than six seconds on average. And I'm not too worried about that. I think that is mostly just to keep tasks from blowing up and using extreme amounts of computers. That's what memory limitation, but I think those are those are really generous and not really relevant to the to the main to the main thrust of the task. The other time constraint is calendar time. And that, I think, is going to be the ultimate limitation because what I would like to do is spin up many sub agents, many tasks, and just grind away at a large number of different strategies, a large number of different hypotheses and tests, many of which might end up being failures, but if some of them are successful, then it's worth it. And so, um, I guess what I would like is a conversation about... oh, the other thing is, you know, um, different models are differently capable, but also take different amounts of time. And so, uh, sonnet, you know, medium, for instance, is faster than fable, extra high, and these types of things. And so I guess what I'm wanting is a strategy or I wanna lay out a a a way of working that can go on for a long period of time. So may... not necessarily a Claude orchestrated workflow, but something that can that can grind away for for some time. Um, and now how that feeds back with, you know, uh, human feedback, meaning my feedback, I I don't know yet, but that something is I would like some kind of observability and monitor. Oh, another aspect for time that is wall clock time or calendar time is I want to... so we've already established that these, uh, you know, the Python test can run-in parallel. So I want to use that as much as possible, and I also want to aggressively use caching, uh, where appropriate so that things can be, you know, faster and that, you know, the easy cases are maybe marked as not interesting, and then the grinding is focusing on the more difficult cases or, you know, whatever it turns out to be. Um, the other thing is multiple, uh, workers, multiple agents running in parallel, um, testing different hypotheses, potentially in work trees or potentially not in work trees. But then if they're not in work trees, how do they not step on each other and those types of things would all have to be considered. And so I guess what I'm looking for is is a... I'm I'm wanting your help to set up the harness or the orchestration or whatever in order to in order to handle this. And I expect some tasks are going to have to be serialized and some tasks are going to be very easy to parallelize. But, anyway, this is this is what I wanted to talk about. So let's let's start a document that is the that is the July twenty first strategy that we're going to go over... where... that'll kinda document what it is we're gonna... what it is we're trying to do, the thoughts that we have, and the mechanisms we're gonna put in place to be able to really throw a lot of agent compute at this problem

[Jamie @Vector76]

@TheOneAndArjun Wow I told the agent to go to town with subagents and it did a lot while I was hands-off.

Now it's at 120.66 on training set but I worry it's overfitting like crazy.