There is a large graveyard of JITs and JIT-adjacent projects for Python.

By now it should be clear to anybody working on Python JIT that the probability of failure is 90%.

The future is probably rewriting performance critical Python code in Rust instead of trying to fix Python.

Or maybe a future LLM could add a JIT to Python in an effort-run.

Whenever I needed a color for something digital (website, ...) I would use the Pantone color picker in Photoshop. It had multiple lists of colors (some more vivid, some muted, some thematic - only reds) and I would browse the color I wanted to pick a suitable shade.

I didn't need the Pantone aspect specifically (real world printing), these were strictly digital uses, but I found browsing shade lists much better than trying to use a regular analog color picker (RGB, HSV, ...). Maybe because you see a large color swatch, maybe because seeing 10 different shades at once is and choosing is faster then randomly moving the mouse through the analog picker.

Screenshot: https://www.pinterest.com/pin/how-to-find-and-add-pantone-co...

This is why I like the web colours list - there's usually something close enough to what I want that helps avoid the combination cognitive trap of a colour picker and choice paralysis.

You also have the sonar/radar thing on the camera which can measure distance. Can that be hooked into focus so it's automatic?

If you’re going to go through all that trouble you may as well just use a lens or cam body with a quality, built in AF. Way cheaper and the tech has matured greatly. Plus even with lower to midgrade cameras it’s easy to find a capable system with a solid codec you can color match enough with the rest of the movie. Even my GH6 I travel with can shoot 5.7k pro res

I was talking about movie-grade setups, I was wondering why they need the sonar thing if they focus manually anyway. Maybe it's useful for post-production, as a depth field or something.

Oh sorry I misunderstood your point. That is so the AC can get an accurate measurement from the sensor/gate to the focus point, which they then mark on their follow focus. Often it’s a few marks because a shot will have different items of interest or the camera is moving. When you’re shooting on film this is particularly important, but a lot of people also use it for digital sensors so you’re not eyeballing it or depending on focus peaking (though plenty of quality focus pullers can absolutely eyeball it most of the time). Boils down to the shoot you’re on and the process they want, or the importance of the shot, such as something with huge explosions and a flipping car where are you want to reduce any possible margin for error.

How does evolution learn the form-fitness relationship?

It's the same thing here, you randomly try various token-relationship values and the ones which are slightly better will be favoured.

That's easy to say AFTER you know the architecture.

Einstein special relativity is taught these days in high-schools. Doesn't mean it wasn't the very hard part at some point in time.

As they say, shoulders of giants.

Isn't that over-simplifying it a bit too much?

You can go another step - a FFN can be simulated on a Turing machine, thus it just exemplifies the incredible semantical power of the Turing machine model of computation. (in fact you don't even need a Turing machine, since there is no looping in one forward pass).

In theory you can run a huge FFN on the tiniest Turing machine, in practice it's much better to run a Transformer on the latest NVIDIA hardware. Or as they say "quantity (performance) has a quality all its own"

I was about to post your last point / quote. Going multigpu is relatively not so though but once you go multi-node you have distributed storage/io/compute system which is highly non trivial. Add that the long training times now you have robustness/fault-tolerantness concerns with hardware failures and restarts. Today’s training systems are engineering marvels.

Good point!

There is also the case for Markov chains being theoretically able to do these if tuned well. Or even SAT problem.

"LLM is just fancy autocomplete"

LLM is an Oracle

Google already released specialized drafters for Gemma 4.

The E2B ones? Or what do you mean by specialized drafters?

They have -assistant in the name, so e.g.: https://huggingface.co/google/gemma-4-31B-it-assistant

Thanks

The “-assistant” models released by Google are specialised tiny MTP draft models :)

31b-it-assistant is what enables MTP

Google already did

https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-un...

This is safetensors. Is there any way to run these on a Mac paired with the MLX QAT?

(Pardon my ignorance; this stuff moves so fast)

Did you see this?

https://point.free/blog/gemma-4-on-a-2016-xeon/

Xeon, but could be useful for MTP on Mac.

I hadn't seen this, thanks.

I do have the Qwen 3.6 (35B) MTP implementation running (in LM Studio; it doesn't need a separate drafter), along with non-MTP Gemma 4 26B, and I can see that Unsloth Studio can run the new QAT, but I can't see how you can run the assistant/drafter. Yet.

It's just a constantly changing landscape. Don't get me wrong, it's fascinating and for various reasons I am pleased I can keep up even slightly, but eeeehhh :-)

Yeah — that is the base QAT model, and there are safetensors weights for the QAT version of the MTP drafter, but there are no MLX/GGUF versions. I think the answer is a combination of:

1) Gemma 4 MTP is too fresh for off-the-shelf software to use anyway

2) "you can convert them yourself" which is fine, obvs

Agent can get tricked into using a malicious library in your project, commit and push that, which you then run outside the VM.

So if you ever run the repo code outside the VM and don't review everything committed, you are still at danger.

It doesn't have any credentials inside the VM though, not even for git, so it could commit but not push. And I manually review/commit/push outside of the VM since I don't want to just dump stuff without reading it first.

But good call-out if someone uses a different workflow.

The un-quantized MoE outperforms it.

But between same (V)RAM requirement 4 bit 26B-A3B and 8 bit 12B it's unclear which one will win, especially given one is MoE and the other dense.

All the launch benchmarks are at 16 bit.