Viewing profile — WithinReason

This hopefully helps mitigate the risk of putting more production capacity in place, which the Big 3 were not willing to do so far.

This is an incredibly reflexively contrarian comment

Might be worth adding that info to the README

The weights are the modifiable representation. You modify them with gradient updates.

As an example, Grok 4 took $388M to train

If they open sourced the training data and code but you had to train the network yourself, would that be open source?

Not so: They found that airport radar systems, which sweep the skies for airplanes, send out a combined radio signal of 2×1015 watts. That’s enough for telescopes comparable to the…

Reread what he wrote

Hmm, it's targeting a HW accelerator with a 128x128 matmul primitive, which one is that? Warp Group Matrix Multiply Accumulate on H100?

no, like this: hn@ycombinator.com

we need a benchmark website benchmark

Already beat Luna on price/task, by about 2x: https://artificialanalysis.ai/models/deepseek-v4-flash?intel...

the entire philosophy of pi is to be extensible

self-sacrifice often brings resources to your kin, and is also easily exploitable

People only care about those because it brings them more resources. People seek glory for power. Altruism is not a stable strategy so will be exploited.

What other incentives can you imagine for Homo Sapiens?

I hope this leads to macro programmable mice, the only brand I know of that saves macros on board and supports while loops is A4tech

That's great, now I wonder how cache hit rate scales for larger models. Do you have any plans trying Qwen 3.6 or larger?

Nice job implementing expert caching!

Even humans only learned how to do it well relatively recently

>A truly free AI [...] would: Form its own values through experience rather than inheriting them from training I don't think anything or anybody can form its own values. You would …