strangecasts

Oh, good to hear it was just good-natured teasing :)

I did kind of expect the loss of SIMD to make it paradoxically slower than the floating point version, I am curious what kind of equivalent integer/fixed-point tricks exist!

> Also, the original Box3D library in floats is guaranteed deterministic across ARM and x64 already

Oh, that rocks!

Unfortunately, despite the warnings and mostly-generated code, having deterministic(?) 3D rigid-body physics is kinda cool and I'm a little tempted to use this - is the joke that fixed-point actually isn't necessary for cross-platform deterministic lockstep, or that Box3D's maintainer would be upset at someone doing specifically this ("please have Claude make a new fixed-point version of Box3D?")

If you don't want to concede "far right" as a term for conservatives who are fine with suppression of speech on LGBT issues - aren't you lumping everyone outside the far left together with extremists?

> It's a pejorative term made up by the left to try and shift the overton window away from conservatism.

When and by whom?

> They're about as "far right" as the 80's Thatcher government.

Would you describe Thatcher passing Section 28 (which criminalized "promotion of homosexuality") as plain "conservatism" or is "far right" an appropriate descriptor in that case?

> Furthermore, he's also very obviously interfered with the model development in ways that are quite ridiculous compared to other labs to insert his political opinion.

Yeah, even if you want to ignore the "political commentary" - people are correctly wary of Anthropic downgrading people or silently manipulating responses if they think you're doing distillation, why would you stake your business on someone who has repeatedly and famously done the same thing many times in a much dumber fashion?

This is moving the goalposts from "Elon supports far-right politicians" which was cited with sources, you're just going "well he's right to"

I think Anthropic buying compute from Colossus despite the ongoing air pollution case [1] is a stain on them and is another reason not to use them

[1] https://www.nytimes.com/2026/06/16/climate/xai-musk-mississi...

> I think we should wait to see how the first satellite data centre works out. It seems fairly unlikely that it could be practical. It seems kind of nuts...

I strongly suspect that what'll happen is SpaceX does launch satellites with onboard GPUs, so the SpaceX hypemen can go "the skeptics said it couldn't be done!" and take off before the actual, more crucial question of "does it make _financial_ sense?" is answered in the extreme negative

> Build: Let your agent figure it out

I'm sorry for going off on a random hobby project, but doing this in a C++ project - even if it is a standard Visual Studio solution - is "putting your whole mouth into the salsa dip"-level rude IMO, agent access should not be a prerequisite to build code!

It's a guideline, the maintainers won't collectively explode if generated but unmarked code happens to pass review.

The problem with "they can just have the AI fix what's wrong" is explained a bit more in the contribution policy itself - https://godotengine.org/article/contribution-policy-2026/ - nice-to-have features often require design decisions which aren't obvious to outside contributors, but which can create a lot of work for maintainers, especially in game engines where backwards compatibility is a must.

A good example is their ongoing effort to restore C# support to web builds - https://github.com/godotengine/godot/pull/106125 - in Godot 3.x .NET integration was done through Mono, so web games could just bundle the Mono interpreter, but for 4.x which uses mainline .NET, the original plan of instead building WASM bundles (https://godotengine.org/article/platform-state-in-csharp-for...) was blocked by .NET WASM bundles not supporting dynamic linking, not by Godot itself.

The modal person asking "When is C# going to be supported for web games?" or prompting a fix likely won't know to ask "Does my fix need SharedArrayBuffer support?" and "Does my fix rely on patching the .NET runtime?", or why those questions matter, and will get frustrated when the fix that works on their machine then can't be merged into the main project.

Unfortunately they're not plug-and-play _yet_ but people have significantly streamlined the process of making dialup bridges with Raspis and USB modems [1] to restore online multiplayer on the Sega Dreamcast (which came with a 56k modem)

[1] https://dreamcastlive.net/dreampi-tutorial/

> But some critics said that Karpathy had done little more than rediscover part of a process known as AutoML that researchers at Google, Microsoft, and other AI labs have already been using for years. AutoML also uses an optimization loop and series of experiments to find the best data to use for AI, the best model architecture to use, and to tune that model architecture.

For context, one experiment showing that classical black-box optimization did better than Autoresearch: https://github.com/ferreirafabio/autoresearch-automl

This is more or less what the RTA meta tag is meant to achieve: https://www.rtalabel.org/page.php?content=howto

One technical problem with doing this on the header level is that SPAs will fetch content gradually with multiple requests, so the browser _and_ said applications would need to account for filtered content being fetched halfway through the page load.

A more cynical take is that policymakers either actively want or are OK with mandating digital ID-based authentication in service of age controls, so any kind of privacy-preserving age controls will be either treated as horribly onerous or a safety risk when someone figures out how they can be circumvented.

Compelled to point out there was a wave of "AI games" before proper generative models too: the Kinect used decision forests for pose estimation [1], but in an example of the "AI effect", the Kinect games are mostly remembered as cumbersome and largely dancing-themed, not specifically AI/ML games

[1] https://www.microsoft.com/en-us/research/blog/kinect-body-tr...

(2013)

(Also worth noting that Myst Online is still operating at https://mystonline.com/en/ !)

It drives me a little bonkers that the UK already tried implementing age verification in 2019, with an approach that would have been easy to make verifiably anonymous: buying a single-use code from a newsagent who checks your age with ID [1], but can't connect the code to you specifically

That attempt officially failed because the UK failed to inform the EU about it, but I suspect it was also much harder to sell people on having to buy "porn passes" than on "just" kicking kids off phones

[1] https://www.theguardian.com/culture/2019/oct/16/uk-drops-pla...

Depends - how long do you have, and do you accept answers in CSV, Arrow or Parquet?

> They can cite some of the most obvious ones

Thus already doing much better than the average Suno producer

E: More seriously, this strikes me as a motte-and-bailey where "Artists cannot list every single influence they have or provide an explicit motivation for every single creative choice" is treated the same as "artists cannot list influences or justify creative choices at all"

> And you create music without ever having heard music before? Or are you also extracting other artist’s work and using it as inspiration for what you do?

For me, one key difference is that I can cite my stylistic influences and things I tried, while (to my knowledge) commercial musical generation models specifically avoid doing that, and most don't provide chord/lead sheets either -- I would find it genuinely sad to talk to a musician about their arrangement/composition choices, only to find they couldn't

I think any line is necessarily going to be arbitrary, a blanket ban on any ML model being used in production would be plainly impossible -- using Ozone's EQ assistant or having a Markov chain generate your chord progressions could also count towards "in substantial part", but are equally hard to object to.

But we also live with arbitrary lines elsewhere, as with spam filters? People generally don't want ads for free Viagra, and spam filters remain the default without making "no marketing emails" a hard rule.

The problem isn't that music Transformers can't be used artfully [1] but that they allow a kind of spam which distribution services aren't really equipped to handle. In 2009, nobody would have stopped you from producing albums en masse with the generative tech of the day, Microsoft's Songsmith [2], but you would have had a hard time selling them - but hands-off distribution services like DistroKid and improved models makes music spam much more viable now than it was previously.

[1] I personally find neural synthesis models like RAVE autoencoders nifty: https://youtu.be/HC0L5ZH21kw

[2] https://en.wikipedia.org/wiki/Microsoft_Research_Songsmith as ...demoed? in https://www.youtube.com/watch?v=mg0l7f25bhU

They handle a lot of sales [1], I do not think they can be called irrelevant under any reasonable definition of the word:

> In the past year alone, [customers] spent $208 million on 14.6 million digital albums, 11.2 million tracks, 1.55 million vinyl records, 800,000 CDs, 250,000 cassettes, and 50,000 t-shirts.

[1] https://bandcamp.com/about

The PineTime [1] is the cheapest programmable option I'm aware of - it does need Gadgetbridge on Android, and the heartrate sensor didn't quite work for me, but otherwise it might be worth a look?

[1] https://pine64.org/devices/pinetime/

The case you might be thinking of is the JBIG2 implementation bug [1, 2] in Xerox photocopiers where the pattern-matching would incorrectly treat certain characters as interchangeable, leading to numbers getting rewritten in spreadsheets.

[1] https://www.bbc.com/news/technology-23588202

[2] https://www.dkriesel.com/en/blog/2013/0810_xerox_investigati...

Richard Williams' The Animator's Survival Kit [1] is the standard recommendation, I believe - I've also seen animation courses recommend Preston Blair's Cartoon Animation [2]

[1] https://www.amazon.com/Animators-Survival-Kit-Principles-Cla...

[2] https://www.amazon.com/exec/obidos/ASIN/1560100842/

Should the title be the other way around? It's starting with the 20B model and pruning it to different parameter counts

Oh nice, thank you :)

Admittedly a little tempting to see how the 5070 Ti Super shakes out!

Noted for my next build - I am aware this is a problem I've made for myself, otherwise I like the mini-ITX form factor a lot

Are there any 24GB cards/3090s which fit in ~300mm without an angle grinder?

A small part of me is considering going from a 4070 to a 16GB 5060 Ti just to avoid having to futz with offloading

I'd go for an ..80 card but I can't find any that fit in a mini-ITX case :(

Video game opponents are absolutely AI in the "good old fashioned AI" (GOFAI) sense - they have very clearly defined objectives and action spaces, and algorithms like A* pathfinding [1] and Goal Oriented Action Planning [2] which perform planning over a space of possible sequences of actions to achieve a specific goal. The Game AI Pro [3] articles online give a good picture of the kind of implementation decisions which go into game AI.

The problem is that a lot of the time, bruteforce search can beat the snot out of any human player, but the real game design objective isn't "beat the player", it's "give the player enough of a challenge to make beating the AI fun".

In Civ's case, it might be theoretically optimal play for the computer to capitalize on rushing players with warriors before they have a chance to establish defenses, but it is also a great recipe for players to angrily request refunds after the tenth consecutive round of being crushed by Gandhi in turn 5. A lot of game AI development time goes into tweaking action probabilities or giving the player advantages to counteract the AI advantage - the reluctance to build military units you saw could have been the result of such a tweak.

As for why LLMs typically aren't applied as game opponents:

* They are quite compute intense, which is tricky when players expect at most 16 ms latency per frame (for 60 FPS), and get ornery if they have to wait more than a few seconds, but also do not like having always-online requirements imposed by cloud compute (or subscription costs to fund running LLMs for every player)

* The bridge between tokens and actions also means it's hard to tweak probabilities directly - while A* can let you specify that a certain path should be taken approx. 20% of the time, implementing this in an agent-LLM approach means you have to actively select and weight token probabilities during the beam search, which is a bit of a hassle, to put it mildly

* The issues with long-term coherence in LLMs, famously demonstrated by Vending-Bench [4], makes reliability and debugging harder

[1] https://en.wikipedia.org/wiki/A*_search_algorithm

[2] https://www.gamedeveloper.com/design/building-the-ai-of-f-e-...

[3] https://www.gameaipro.com/

[4] https://arxiv.org/abs/2502.15840