Scene_Cast2

I wanted to switch to DS4.1F from K3 (not even K3.1). Even the 1017 snapshot takes 5x-10x the tokens that K3 uses for the task, and often misses more intricate gotchas that K3 catches.

One issue I've had with swapping fans for Noctuas is the airflow matching. The louder fans aren't louder just because they're lower quality (which they definitely are), but also because (almost always in my experience) they force more air through.

That unit from the article sounds very temperature sensitive. I'd be wary of throwing a precision unit like that out of spec and out of calibration.

FWIW I've noctua-swapped so many things including an oscilloscope.

Really interesting.

On the other hand, this only applies to a limited set of workloads (training is often memory bandwidth bound for example).

I also wonder how this affects power consumption.

I found that surprisingly many k-pop songs do microtonal shenanigans. Nowhere near most, but it's there. Even something as popular as Boombayah (check the 4th note, it's not an on-tune semitone).

There's also more niche stuff like Flying Microtonal Banana.

A quality ECE program will have you working lower level still - all the way from ALUs down to the electrons & holes in semiconductors.

That's what I had lying around when I started the build. PTM is now my go-to recommendation.

This video seems to be about LM on untreated copper (which I thought wasn't really advised).

Now, I personally have had two builds with LM. One with LM as the material for a CPU re-lid, and another for a direct-die application. Neither touched aluminum or untreated copper.

The re-lidded CPU is still going strong, no issues. But with the direct die application, I found the performance to significantly degrade after two years. Apparently there's a pump-out effect.

I ended up switching to PTM. Same temps as LM, but zero degradation so far.

Not all of the builds seem live. Pixel 9 Pro has 24.0, but Pixel 9 has 23.2, for example.

I ended up going with a Synthetic subscription (they host some open-weight models including K3 and GLM 5.3). They're the only subscription that I found that's simultaneously cheaper than token rates, and is extremely transparent about usage and pricing.

The libraries, yes, but the code you write is portable (at least until you get into squeezing the last few percent and switch to Triton / Helion in case of GPU, and even those are decently portable).

There's also Halide, where you write the algo but the framework gets you the scheduling and SIMD.

What about numpy, numba, and torch.compile?

The official page still downloads 1.23.2. I wonder when this will go live, and if the old app will auto-update.

How? I don't see it. (I'm familiar with the ML side, not the combinatorics side.)

How many tokens are you getting, roughly?

I found LLMs circa 4 months ago to be decent at textbook-style electronics, but they choke at anything nontrivial. They approach things like someone that knows the undergrad and grad textbooks, but has absolutely zero experience. No ability to answer questions such as "is this adequate", "do I need a LPF here", "how do I kill resonance without increasing impedance", etc.

I'm guessing that this is due to a lack of RL and data.

Why does it work for Red Bull?

It worsens the supply crunch, no? A unit you use sparingly vs that memory going into a GPU that serves many more people.

If you need that precision, especially for position and not values, why use floats at all?

Explicitly trying to re-identify data that has been de-identified is typically a fireable offence at FAANG.

Accidentally making a machine learning system that happens to (potentially) do it is a different matter.

Oh hey, that's cheaper than Kimi K3! Amusing to see a SOTA OpenAI model be cheaper than a Chinese open weight model.

Fwiw I love K3 and use it as a daily driver. I haven't tried Sol, as I dislike OpenAI.

I made the move some time ago. There are some unpolished parts here and there, but I'm glad I made the move.

Somewhat unrelated criticism of Gmail - during the migration, I found out that there is no way to search for starred archived emails in Gmail. They don't show up in Starred, nor in any search with is:starred. Their star status also doesn't carry over when migrating.

Awesome. I'd love if Flutter could also finally get Impeller on Web (I'm tracking the GitHub issues, seems like there's progress at least).

I personally would like for it to pick up some ideas from Svelte.

In one of their dev surveys, they did float the idea of divesting Flutter.

There's a trade-off between how long you wait and the skill spread of the team. IIRC at the very top, for DOTA2, the pros complained enough for the devs to make a "wait longer" matchmaking mode.

I got a similar looking Stax amp (the SRM series). Also quite satisfying. It has a split volume dial so that you can adjust the volume of each channel separately.

I wonder how much effect the CPU's register remapping has on these results. Also, they ran it on an ancient CPU, but I'm not sure how much that matters.

I love the hackability of Thinkpads circa that generation.

I'm using Kimi K3 + OpenCode. I pay their API pricing, costs about $5 / hour (and chews through ~10 million tokens / hour) during continuous use when I have one or two sessions running and doing their thing.

Can't comment on how it compares to plans (I really don't like the limitations and general shenanigans I see around plans, so I've never tried them).

It is notably slower than Fable / Opus / Gemini, but also vastly cheaper than their API pricing.

Very much agreed. Even more basic than that - memory access patterns are important. The amusing thing is that you end up writing GPU-style code even for CPU. For example - instead of an array of objects, using parquet-style object of arrays is one such trick.

What are some good tools for _authoring_ HDR photography? Lightroom still seems to be mostly SDR. The new Resolve will probably have the tooling, but I haven't tried it.