Ilya Chetviorkin

CTO @truvhq (acq by @checkr), PhD in ML&NLP

[Ilya Chetviorkin @chetvil]

Today @TruvHQ is joining @checkr ! I can't think of a better home for Truv and everything we've built over the last 6 years.

Our joint offering will make a complete verification stack more affordable and much more seamless.

6 years in the making for this overnight success!

[Ilya Chetviorkin @chetvil]

@zephyr_z9 seems to be very quick for this big of a jump? Which means getting close to frontier is much easier than people thought?

[Ilya Chetviorkin @chetvil]

@Lon Nice work

[Ilya Chetviorkin @chetvil]

Humans belong to the stars

[Elon Musk @elonmusk]

This is the future we shall bring into being

[Ilya Chetviorkin @chetvil]

@aimalysheva @mostik_ai @generalcatalyst hi!

[Ilya Chetviorkin @chetvil]

Started full DeepSWE run with excellent @MiaAI_lab EXL3 quant of DeepSeek-0731. I think it is the best coding model for 1 Spark.

Currently about 20 tasks in and 53% success rate. And DeepSWE is hard!

Full results updated daily here https://t.co/rHB5foVAyp

[Ilya Chetviorkin @chetvil]

@teortaxesTex Nah, Qwen is pretty bad. Anecdotal but on Spark it can't solve DeepSWE tasks while DeepSeek exl3 can.

[Ilya Chetviorkin @chetvil]

@filicroval What does it matter if qwen3.8 is dumb as a brick?

[Ilya Chetviorkin @chetvil]

@andrewho03 This is what I think as well. You can ride on fully verifiable domains for so long.

[Ilya Chetviorkin @chetvil]

I ran a few experiments to see how capable Deepseek Flash could be on one DGX Spark and compared to a few OpenAI/Anthropic models.

Surprisingly quite good - see full analysis here https://t.co/2eMfNP3I6b

kudos @MiaAI_lab for the ELX3 recipe.

[Ilya Chetviorkin @chetvil]

Will try DeepSeek Flash next...

[Ilya Chetviorkin @chetvil]

Ran Qwen3.8-27B official FP8 on a DGX Spark (SGLang) as local Claude Code, trying to match the card’s DeepSWE 42.2. One meriyah task. Spec decode: DFlash2, then DSpark. The stack worked. The score didn’t reproduce, timed out every time

[Ilya Chetviorkin @chetvil]

Two Claude Code runs, 2.5h timeout each. xhigh+DFlash2: 71k output tokens, started editing token/errors/parser too late. medium+DSpark: 59k tokens, even less of a patch. Both: 0/49 new tests

[Ilya Chetviorkin @chetvil]

@TruvHQ @checkr @kirklokov @ilyaf @kleinerperkins @nyca @AbstractVC Full announcement https://t.co/z0aNNqtniZ

[Ilya Chetviorkin @chetvil]

Today @TruvHQ is joining @checkr ! I can't think of a better home for Truv and everything we've built over the last 6 years.

Our joint offering will make a complete verification stack more affordable and much more seamless.

6 years in the making for this overnight success!

[Ilya Chetviorkin @chetvil]

@TruvHQ @checkr @kirklokov Last, I want to say really big THANK YOU to our amazing team, real SWAT team which can build anything with whatever level of complexity.

And of course @ilyaf and @kleinerperkins for your unwavering support along with other amazing investors @nyca @AbstractVC .

[Ilya Chetviorkin @chetvil]

@kleinerperkins Big news!

[Dwarkesh Patel @dwarkesh_sp]

.@dylan522p lays out how we know the hard upper bound on how much compute can be produced annually by 2030: around 200 GW/year.

That’s a crazy number (there’s about 20 GW of AI deployed in the world right now), but it’s nowhere near enough to satisfy Sam/Elon/Dario/Demis’s ambitions.

Lots of things in the supply chain can be scaled up over 4 years, including things that other people think are bottlenecks, like datacenter power or fab clean room space.

But the thing that’s inflexible over that timeline is the number of EUV tools.

Dylan forecasts that production of ASML’s EUV tools will scale from 60 per year now to about 100 per year by the end of the decade - which means something like 700 total machines running in 2030.

For a fab to make a GW worth of the Rubin chips that NVIDIA is deploying later this year, it needs to make 55,000 3nm wafers, 6,000 5nm wafers, and 170,000 memory wafers.

Each 3nm wafers needs about 20 EUV passes, so about 1.1 million passes per GW. Adding on 5nm and memory, you need two million passes.

Each tool can do 75 passes per hour, so with 90% uptime that’s around 600k passes per year - so a single machine can make less than a third of a GW in a year.

So in 2030, we have 700 total machines, each making 0.3ish GW a year, which means we can produce 200 GW of compute a year.

That’s a lot. But Sam Altman wants a gigawatt a week by the end of the decade. Anthropic and Google will be wanting about the same. And Elon wants to be putting 100 GW in space every year.

Any one of these players could maybe get what they need, but not all of them.

[Ilya Chetviorkin @chetvil]

@MrNQDC it's only 10k per person, this is the catch.