ElFitz

Would also work very well for "open prompts" interpreted and mapped to one of the available options by a LLM. As some sort of constraints.

I am glad to have checked that one out. I don’t know whether to be amazed or horrified.

> Claude Code is the cancer that will kill the patient, Boris is the the Kardashian version of Karpathy, with less business sense.

I am not sure I understand.

That, especially the conclusion, is hilariously bad.

I’ve been experimenting with two things on this:

- multi-model consensus, with multiple cross-review rounds. Obviously, the number of inference tasks explodes with the number of models. Led to some interesting results [^0].

- giving an agent "stray thoughts" produced by the same model, or another, giving the second model a selection of the agent’s context, with different triggers (random, loop detection,…)[^1]. So far has proven very helpful and much cheaper than the first.

[0]: https://github.com/lightless-labs/refinery

[1]: https://github.com/Lightless-Labs/skunkworks/tree/main/flux

> With such eggregious trillions of dollars worth of money (basically the whole economy getting floated by tech), you are bound to see people within this do the grift playbook and talk about themselves and succeed and that has become the playbook.

Reminds me of Pink Floyd’s "Have a Cigar":

> And did we tell you the name of the game, boy?

> We call it Riding the Gravy Train

I’m working on Descartes[^0]. First to help diagnose what’s wrong with a machine.

I’ve started implementing actual background monitoring of the system, and next will be letting an agent build its own layers of tailor-made deterministic rules and statistical models, to "learn" what the system’s normal behaviour is and only "wake up" the agent when something unusual is going on. Either to update its rules and models, or alert the user.

Like the ship’s AI at the beginning of Absolution Gap. Next will be enabling it to serve as the interface for the system. An ops "point of contact" for both the user and their agents for the machine / fleet of machines it’s in charge of.

———

I’m also working on third thoughts[^1], a tool that analyses local agents logs to find patterns and behaviours, identify what works, what doesn’t, how they evolve over time, using deterministic and statistical methodologies and techniques from multiple domains (including, to my surprise, genetics and psychology / sociology), with an agent layer that interprets the results.

I’d like to add a "federative" layer where people can contribute the results, patterns, and findings, without leaking their logs or personal / private data, so that we can all better learn how to identify failure modes, predict them, and see what works and what doesn’t.

———

I’m also having Claude & Codex revive Jasonette[^2], which died off and was turned into some weird paid unrelated thing by those who picked it up. I’d been meaning to but never took the time. But now with agents…

All rebuilt in Swift / SwiftUI on the iOS side, and Kotlin on the Android side. Some features are still missing, but it works quite well! [^3].

———

Oh, and Boucle[^4] is doing its thing on its own. No idea how it got to 100 GitHub star. My "autonomous dog-fooding expensive pet" is apparently more "internet successful" than I’ve ever been.

And quite a few other side projects.

———

[0]: https://github.com/lightless-labs/descartes

[1]: https://github.com/lightless-labs/third-thoughts

[2]: https://github.com/Jasonette/JASONETTE-iOS

[3]: https://github.com/Bande-a-Bonnot/JASONETTE-Reborn

[4]: https://github.com/bande-a-bonnot/boucle-framework

From what I found, current state of the art on modelling "reaction space" with graphs that is to use "hypergraphs" where edges can lead to more than one node[^0].

But I am just someone who got curious; not even an amateur ^^’

[0]: https://pubs.acs.org/doi/10.1021/acs.jcim.5c00265

> Feels complex like solving a Rubik's cube to write down synthesis steps but it is all a sequence of memorized tricks. Do Cannizaro if you want this, Bergmann to do that.

I remember two years ago, when I actually got into using graph data structures, wondering if maybe the "space" of available reactions for any given starter and target molecules could be mapped as a graph, with intermediates as nodes and reactions as weighted directed edges, so synthesis becomes pathfinding through chemical space.

Turns out, it’s a thing! [^0]

Edit: Makes you wonder how much interesting stuff is sitting in plain sight, waiting for someone with the right cross-domain awareness / knowledge / whatever to notice it.

[0]: https://pmc.ncbi.nlm.nih.gov/articles/PMC9574932/

> Ah, I see, sort of like figuring out the boundaries of your knowledge base and seeing if you have missed any connections between concepts?

Yeah. And also resurface relevant old notes, ideas, and web clippings I have forgotten. Two recent example[^0][^1].

> it could be an interesting synthesis/writing exercise to try to connect concepts that are far removed in your own mental model.

As well as different takes on or applications for those concepts, same for sources.

[0]: https://imgur.com/a/jhxUV9i

[1]: https://imgur.com/a/gmaq9Yw

Isn't that what we call public debt?

The graph visualisation itself isn’t much use in my experience.

But being able to tie related notes together, and see at the bottom of one which other notes reference it is interesting.

Even more now that a LLM can take care of the actual tending and pruning.

I think it’s not the first time the US has used that sort of interpretation of the law. There’s this one[^0] but also, I believe, an older case, also involving Microsoft, about data in Ireland. But I can’t find it.

[0]: https://hackernoon.com/the-factors-prompting-a-judge-to-issu...

These kinds of situations are why I gave my AI agents stray thoughts (automated insights / suggestions from a separate llm call with some curated context) that trigger on loop / rabbit hole detection.

Quite a bit of false positives, but it hasn’t had any ill-effect so far. Aside from increased quota usage.

Haha. Yes. Much smaller scale versions of this led me to joke with a coding agent that LLMs tended to converge towards "Large corporation infrastructure best practices" when designing cloud infrastructure, when it was only me working on hobby side-projects with nearly no users and that I wouldn’t be able to put food in my fridge if they kept just spinning up VPCs for no reason.

Which somehow ended up being a very convincing argument for more frugal engineering, leading to a sort of "mind the user’s fridge" policy, "Fridge-Driven Development".

A policy that has been dutifully and scrupulously observed by all agents since, across all projects. Unlike my original clear, comprehensive, infrastructure guidelines.

I’ve been making Codex and Claude get their work reviewed by most recent best performing model of their own family, and each other’s, for months.

On top of that, we have been running multi-model AI reviews on every PR through their respective GitHub integrations (Codex, Gemini, Copilot, Greptile, CodeRabbit).

They never fully overlap, and yet they somehow usually all miss the same things. The most significant improvement came from having agents commit their plan along with their work.

On the upside, it means I get to focus my reviews on different things.

> and I have the feeling that the harness is much more important than the consensus expectation.

Is that really the consensus? There’s been a bit of literature lately on that. Can’t find the one about looking into whether or not the harness had a greater impact than the models (for comparable models), but there’s this one: https://arxiv.org/html/2605.23950

I’ve had a similar experience.

Pointing out past suboptimal / failing behaviours to new opus sessions would almost always actually create a sort of "anchoring bias" that would drive the agents towards exhibiting the failure mode (often while mentioning how it wouldn’t fall for it).

As far as I can recall, Fable has been the first model to discover the documented failure modes, comment on them, and just… keep going, actually avoiding them. Quite a surprise.

I love hard caps, and am tired of cloud services not even offering those. May make sense for large companies. Makes no sense for hobbyists and small companies. But maybe that's the point?

That’s what evals are for.

And there’s no reason evals can’t be done on multi-turn agents in a loop (or not): it’s pretty much what all these benchmarks do.

I remember Windows Phone 8 existed, but that’s pretty much it. And yes, that’s the big question: what’s in it for the app publishers?

:/

A few months back someone reverse-engineered private ANE APIs and shown some significant performance improvements compared to CoreML and Metal, on both inference and training.

- https://maderix.substack.com/p/inside-the-m4-apple-neural-en...

- https://news.ycombinator.com/item?id=47257931

Oh, neat. Totally missed it, thanks!

Yes, but you can have this by providing surfaces as UI components to Siri.

Not sure if they do that (yet), but no reason an app couldn’t expose "Here’s what you can use to present data of shape X", or "here’s a UI for process doing y".

It feels like turning the common approach inside-out. But it works.

Edit: you could even imagine, in that world, apps that only expose surfaces, composable UI libraries, multi-step flows, declaring what they’re for, what kind of inputs they take, and what output they produce. Without ever owning any of the data (eg flights data, hotels inventory, booked trips, financial data, etc) or capabilities (eg book a flight).

I used to wonder what "apps" might become in an "App Intent-first" world.

Bundles that provide data and capabilities to iOS and Siri? And perhaps libraries of UI components to display and interact with said data?

But then, if that works really well, and gets strong adoption, why ever open the app? What’s the point of having navigation flows inside an app? Could one make entire apps solely dedicated to providing a set of data, capabilities, and UI components to the system?

In that world, what drives user retention, for such apps? What even is an app? App engagement disappears as well.

And that’s not even diving into the use-case of Siri, say, planning a trip across five different apps (flights, hotel, restaurants, whatever) using just App Intents. If done well.

In that world, do most apps just become plugins, providers for Siri?

Wait. You worked on Guild Wars, Starcraft, Warcraft, and Diablo?

This place is incredible.

I’m working on Descartes[^0]. First to help diagnose what’s wrong with a machine. Later to help manage and monitor it by letting an agent build layers of tailor-made deterministic rules and statistical models, a bit like the description of the ship’s AI at the beginning of Absolution Gap. And serve as an ops "point of contact" for other agents for the machine / fleet of machines it’s in charge of.

Sometimes I daydream that the end goal demo would be to set one up in a VM and let a sort of ransomware lose in it, and see if the agent can identify what’s going on and react in time.

But for now I’m fighting Apple notarization to enable local notifications on macOS.

[0]: https://github.com/lightless-labs/descartes

Perhaps that’s it. I would tend to agree with his position, I think, but don’t appreciate being preached to. Even less so when I agree with what’s being said.

You can disagree. Sarcastically, or otherwise. But I think you may be reading more into my comment than I put there.

I’m not attacking the piece. I’m not saying it’s right. I’m not saying it’s wrong.

What I’m saying is, the tone made it hard for me to judge the arguments fairly, despite finding some of them convincing. And as much as I dislike it, persuasion does partly depend on how an argument is made.

I find it difficult to separate this piece’s tone from its content. The tone puts me off and makes it hard for me to judge it on its merits, despite some of the arguments seeming sound and well supported.