Twitxy Saved

Mario Souto

Verified account This account is verified. Learn more

@mariohsouto

CTO and Co-Founder @ Soma Energy
mariohsouto.github.io Joined November 2013
417 Following 314 Followers 385 Posts

Latest posts 5

Pinned post

Mario Souto Verified account @mariohsouto
The Clay Jar and the Data Center: an essay on the orchestration layer between energy and AI compute, and what will determine the $/token economics in the inference era.
mariohsouto.github.io/essays/the-cla…
mariohsouto.github.io The Clay Jar and the Data Center The orchestration layer that co-optimizes energy storage, workload scheduling, and revenue per token will be one of the most strategic layers of the inference era.
2 replies 9 reposts 55 likes 29K views View on X
Mario Souto Verified account @mariohsouto
When power becomes the binding constraint on what an AI lab can produce, the lab starts to look more like a factory.

I think this is one of the most important problems to work on. It’s why we started Soma Energy.

More on this in The Energized Chip:
mariohsouto.github.io/essays/the-ene…
mariohsouto.github.io The Energized Chip AI companies are becoming industrial firms whose expansion depends on usable power and how much useful computation they can produce from each megawatt.
Chamath Palihapitiya Verified account @chamath Nov 1
He who shapes power, data centers, chips and models into an economically feasible endpoint, disproportionately shapes AI/SI right now.

Why?

Power is THE binding constraint in AI/SI. Few really have it.

Data centers don’t work without power. This is not about promises of power tomorrow. Or forecasted power in 2028. Power TODAY.

Chips can’t operate profitably unless installed in a functional data center. If you can install chips, it’s because you’ve solved the power problem.

Models aren’t economically viable unless run on cost effective infrastructure which requires a functional data center with low input costs and reliable management. This can’t happen without power.

AI/SI ROI doesn’t pencil unless models generate profitable outputs which necessarily needs functional data centers as the critical input. Again, power.

Ergo, he who can shape data centers, chips and models together into a tightly coordinated stack is someone who has solved power.

Solving power allows you to disproportionately shape AI/SI right now.
View quoted post on X
Mario Souto Verified account @mariohsouto
The scarce object at the frontier of AI is the energized chip.

AI labs are becoming industrial firms, organized around the power they can secure and the useful computation they can extract from it.

mariohsouto.github.io/essays/the-ene…
mariohsouto.github.io The Energized Chip AI companies are becoming industrial firms whose expansion depends on usable power and how much useful computation they can produce from each megawatt.
5 replies 8 reposts 16 likes 2.3K views View on X
Mario Souto Verified account @mariohsouto
No way, they actually did it! This would settle the optimality of SDP-based approximation guarantees across all finite-domain Max-CSPs, assuming P ≠ NP.

Incredible work. Huge congratulations to @OpenAI on proving the Unique Games Conjecture!

github.com/openai/math/bl…
github.com math/preprints/The-Unique-Games-Theorem-September-23-2026/paper.pdf at main · openai/math Contribute to openai/math development by creating an account on GitHub.
Mario Souto Verified account @mariohsouto Sep 11
I'd love to see OpenAI take a shot at the Unique Games Conjecture next.

If UGC is true, the basic SDP relaxation is the best possible approximation for every finite-domain Max-CSP.

Full disclosure: I spent my PhD on SDP solvers, so I'm biased. Still, this seems like a fascinating target for AI.
View quoted post on X
Mario Souto Verified account @mariohsouto
Reminds me of Licklider’s “formulative thinking” (1960): computers helping us formulate problems, not just solve them.

AI + formal verification could make programming more about refining what we mean, with the machine helping us discover what we left unspecified.
Boris Cherny Verified account @bcherny Sep 22
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.

TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.

I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.

Is formal verification the future of coding (or at least, bug finding)?
View quoted post on X