iwhalen

Also see "Self-Preference Bias in LLM-as-a-Judge": https://arxiv.org/abs/2410.21819

Was going to say the same thing. The announcement blog[1] has the flash benchmarks. AFAICT Artificial Analysis has not added it yet though.

Interested to see if it also beats DeepSeek V4.1 Flash.

[1]: https://mimo.xiaomi.com/mimo-v2-6

If your goal is to get the network as small as possible, you could try using NEAT[1]. A cool example of a small network for a hard task using NEAT is MarI/O[2].

[1]: https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_t...

[2]: https://www.youtube.com/watch?v=qv6UVOQ0F44

I came to mention genetic antennae as well!

Since you beat me to it, I'll add something that relates relates you were saying on "realizing some pretty simple... ideas".

I think a big plus of computer aided design like this is "innovization"[1]. Somewhat awkward term. But, a system like this leading one to deeper understanding of a particular process is the general idea. It's a fun feeling in practice.

[1]: https://dl.acm.org/doi/10.1145/1143997.1144266

A lot of discussion in this thread about traffic stops. James Duane (the law professor from the video in the blog) himself makes a qualifier about traffic stops: https://youtu.be/-FENubmZGj8?si=bL_1vR0xFKQqkmC5&t=505

If you already have a gaming desktop, I can recommend Shield for both. Streaming services work out of the box of course. Then I use Sunshine[1] on my desktop to stream to Moonlight on the Shield. Both have wired ethernet connections. Latency is not noticeable in most cases.

[1]: https://app.lizardbyte.dev/Sunshine/

Implementing Tierra [1] in Zig a learning exercise. Trying to push all allocations to compile time and not use LLMs for coding. Repo: https://github.com/iwhalen/zierra

[1]: https://en.wikipedia.org/wiki/Tierra_(computer_simulation)

uv supports the same for any Python developers out there: https://docs.astral.sh/uv/concepts/resolution/#dependency-co...

To pile on, they're also bad at games that are 2D text based environments.

ARC-AGI-3 shows this: https://arcprize.org/arc-agi/3

I've done some work as well on Rogue (sorry for self-promotion): https://iwhalen.github.io/rogue-bench/

I wonder why they didn't discuss price in the post?

Compare to the GPT-5.5 announcement: https://openai.com/index/introducing-gpt-5-5/

From the full letter[1]:

> Our Free Time is Our Own: Currently, if an employee makes anything creative in their free time, with their own resources, Hasbro may claim ownership. What we do in our free time should not be dictated by the company; neither should what we make in our free time be owned by the company.

How common is this in creative fields?

From my perspective this seems outlandish. Imagine doing FOSS work or a side project on your personal computer and your company tries to claim it. Odd...

[1]: https://unitedwizardsofthecoast.com/letter

What is happening in this issue thread? Why are there 100+ satisfied slop comments?

Makes me wonder if the era of heavily subsidized tokens is ending.

Very cool stuff. Love the focus on CPU-first.

Would also love to see some throughput numbers on basic VM setup.

Edit: there are some latency numbers in the paper https://arxiv.org/pdf/2507.18546