GitHub - antithesishq/fluke: Tracing-based Nix evaluation cacher Tracing-based Nix evaluation cacher. Contribute to antithesishq/fluke development by creating an account on GitHub.

here's a lil tool i wrote at work that may be of interest to some of you: github.com/antithesishq...

Hypothesis, Antithesis, synthesis Introducing Hegel, our new family of property-based testing libraries.

So I joined Antithesis a little while back. When I did, I pitched them on a crazy idea. Antithesis... Hypothesis... Hegel!

And here we are. Hegel is our attempt to do Hypothesis-grade PBT everywhere. Today: Rust. Tomorrow: The world!

(Well. In a week or two: Go)

antithesis.com/blog/2026/he...

I've largely got out of the habit of posting on social media to make number go up, but it feels somewhere between appropriate and darkly ironic to do that for my review of @add-hawk.bsky.social's new book, "The Score".

drmaciver.substack.com/p/how-to-be-...

I'm definitely going to be shoving copies into several people's hands.

Equally awkwardly I've logged onto Bluesky for the first time in months to tell you that the book is great.

It's like... 80% familiar ground for me, as I've read a lot of the source material (both yours and others), but even the 80% is clarifying a lot of details.

Yeah! I think I pass the tests with toilets (at least UK ones. I know that US ones do different things with water that I'm not totally clear on the details of, and there's a whole separate category of mains pressure toilets that work differently again) but lots of things I don't.

Many LM applications may be formulated as text generation conditional on some (Boolean) constraint.

Generate a…
- Python program that passes a test suite.
- PDDL plan that satisfies a goal.
- CoT trajectory that yields a positive reward.
The list goes on…

How can we efficiently satisfy these? 🧵👇

I do not feel more zen now Dave.

Post a photo you took with no context to bring some zen to the timeline

(I don't think godlike intelligence is literally impossible, but I do think my probability that we ever get there is significantly under 50%)

But if "superintelligence" means something closer to godlike intelligence, I might be at never. I guess my working definition is "clearly superhuman on some axis, or peak human on multiple axes you'd not normally find together, and not clearly subhuman in easy to notice ways"

I'm honestly not sure how to answer this, and I think a lot of that is weighted on my uncertainty on what "unambiguous superintelligence" means. Erring on the side of stricter criteria I think 2050ish is probably the closest I'd confidently put at >= 50% probability.

Shrinkray has very sensible rules that apply well to languages it has no specific knowledge of -- we make a *lot* of use of this and it's brilliant! -- and it scales near-linearly to the number of cores. If you need a reducer, I highly recommend it! Thanks @drmaciver.bsky.social!

I wrote some notes on how LLMs work, aimed at non-programmers, that you might find interesting. notebook.drmaciver.com/posts/2025-0...

Feedback, corrections, and follow-on questions very welcome.

Started reading this yesterday. It's very good so far, thanks!

Reminded how my first ever conference talk was a lightning talk about how microservices were bullshit. IIRC you were involved in pushing me into doing it, but I may be conflating events.

Mostly that bluesky is in the odd position of being both decentralised identity and also having a canonical identity that most people will use. ciphergoth.bsky.social is much more plausibly legitimate than ciphergoth.someother.domain

That's good, although it sortof feels like this is a nontrivial threat model for scams even without renaming. e.g. this wouldn't have been much better if @ciphergoth.org had always been at the .org rather than previously having the .bsky.social address.

Looks like the account has been suspended at least, but yikes.

Possibly part of the difference is that my balance just isn't very good, so I end up relying a lot more on visual cues to keep me upright, and if I were better at balance I wouldn't need that compensatory strategy.

I genuinely liked the righteous mind and still find bits of it useful. I also straightforwardly disbelieve most of its empirical claims.

My mental categorisation of Haidt is "often interesting, rarely correct"

Every time you wonder how something in Hypothesis works, you can usually reverse engineer it from the following design logic:

1. What's the obvious, natural, way to do it.
2. What are the problems with that?
3. Can those be fixed? Not without doing something ridiculous and unreasonable.
4. Do that.

Right, it's not that replies are excluded, it's that whether or not it's a reply is irrelevant, it's only explicit mentions that count.

Yeah, it's that.

@defenderofbasic.bsky.social I think that (possibly in a few minutes) this message is going to show up in your search.

Actually I have a theory. It's a slightly ridiculous one but I think it's possibly right.

I think replies don't count. I think only explicit mentions do. I'm going to test this shortly one second.

Yeah on closer inspection, you're right, sorry. It's missing a lot of posts.