Chris Coates

There is a thin line between insanity and genius. I walk that line every day of my life.

All views are my own.

[Chris Coates @CJCShadowsan]

@owainkenway So then they want to take a lot of generated similar content to keep you looking longer, making them more in ads.

The problem is that it's an amplification loop... If you spend time looking at something that could be bad news? The algorithms don't care. They'll feed you more.

[Chris Coates @CJCShadowsan]

@owainkenway Traditional ad companies get low click through rates.

People can only generate a finite amount of content and then you'd run out of things that keep your interest so you'd stop reading (and stop looking at ads).

The goal is to keep you looking as long as possible so you see ads

[Chris Coates @CJCShadowsan]

@owainkenway Because... If people gave you what you wanted, they'd not make loads of money off it is the short answer.

[Chris Coates @CJCShadowsan]

This is the future of where AI inferencing is headed. AI as a Utility. #AIGrid

[Nscale @nscale]

Your phone crosses a border. It finds a tower, authenticates, and continues. No interruption, no instruction. Enterprise AI can't do the same thing yet 🧵

[Chris Coates @CJCShadowsan]

@owainkenway Don't even bother building your own training infra now. Just get a ton of agents to distill someone else's.

Faster, orders of magnitude cheaper, and just as effective.

🤯

[Chris Coates @CJCShadowsan]

@owainkenway I still remember at CIUK on a panel talking about this very thing, about how we'd got very wasteful with hardware and lazy when it comes to code optimisation.

Bring back the demo scene but this time for accelerators.

[Chris Coates @CJCShadowsan]

@owainkenway The key takeaway here is... No matter what stack you choose? Bad code is bad code.

[Chris Coates @CJCShadowsan]

@dairsie @BBCSport Would agree. They were passive.

But when you win a bunch of games in a row, you start to get complacent and believe your own hype.

Some teams think that just by being there they'll win. They start to forget you've got to earn it.

[Chris Coates @CJCShadowsan]

Can someone explain to @BBCSport that you can't play 3, win 2 and lose 2.

#SixNationsRugby #SixNations2026

[Robert Youssef @rryssf]

🚨 DeepSeek just did something wild.

They built a math model that doesn’t just solve problems, it checks its own proofs, criticizes itself, fixes the logic, and tries again until it can’t find a single flaw.

That final part is the breakthrough a model that can verify its own reasoning before you verify it.

And the results are ridiculous:

• Gold-level performance on IMO 2025
• Gold-level performance on CMO 2024
• 118/120 on Putnam 2024 near-perfect, beating every human score
• Outperforms GPT-5 Thinking and Gemini 2.5 Pro on the hardest categories

What makes DeepSeek Math V2 crazy isn’t accuracy, it’s the architecture behind it.

They didn’t chase bigger models or longer chain-of-thought.

They built an ecosystem:

✓ a dedicated verifier that hunts for logical gaps
✓ a meta-verifier that checks whether the verifier is hallucinating
✓ a proof generator that learns to fear bad reasoning
✓ and a training loop where the model keeps generating harder proofs that force the verifier to evolve

The cycle is brutal:

Generate → Verify → Meta-verify → Fix → Repeat.

The core issue they solved: final-answer accuracy means nothing in theorem proving. You can get the right number with garbage logic. So they trained a verifier to judge the proof itself, not the final answer.

The graph comparing proof accuracy across algebra, geometry, combinatorics, number theory, and inequalities shows DeepSeekMath-V2 beating both GPT-5 Thinking and Gemini 2.5 Pro across the board (this is on page 7).

The wild part sequential self-refinement. With every iteration, the model’s own proof score climbs as it keeps debugging itself without human feedback.

This isn’t “longer chain-of-thought.”

This is “I will keep thinking until I’m sure I’m right.”

A real shift in how we train reasoning models.

DeepSeek basically proved that self-verifiable reasoning is possible in natural language no formal proof assistant, no human traces, just the model learning to distrust its own output.

If you care about AI reasoning, this paper is a turning point.

[Chris Coates @CJCShadowsan]

@COYFCast Not enough go forward in that pack. If we don't pick up some metre makers it'll make no difference whether we have decent backs or not.

[Chris Coates @CJCShadowsan]

I'll be here, hot off the heels of #GTCDC25! See you there 🚀

[Nscale @nscale]

We’re heading to @Supercomputing 2025 next month.

Join us at booth 6351 as we showcase how we deliver high-performance, sustainable GPU infrastructure at scale. ✨

See you in St Louis! 🎉

[Chris Coates @CJCShadowsan]

@CTRLFC For the love of God listen to your fans and stop drinking your own kool-aid.

This is dire.

[Chris Coates @CJCShadowsan]

@CTRLFC Got to be honest... It looks cheap.

Two colours give no detail.

The stripes are the wrong way.

Nobody looked at any other Tigers logo in professional sports, and learnt anything?

Come on now... See some sense.

Even the old logo in two colours looks better than this.

[Chris Coates @CJCShadowsan]

Seriously @EastMidRailway - all I want to be able to do is get some work done whilst on the train. I can get a working internet connection on a trans Atlantic flight, but not on a land train? Come on now... Make it make sense.

[Chris Coates @CJCShadowsan]

@BentleyAudrey

[Chris Coates @CJCShadowsan]

@BentleyAudrey I must have just been asleep or something or attached to back-to-back conference calls all day. What did I miss?...

...Oh. Oh God.

You've got to love YouTube. The closed captions are amazing sometimes. This is The White House's own YouTube channel 😂

[Chris Coates @CJCShadowsan]

@FelixCLC_ This feels more like a horrifying over-egg with voltage curves to bring up the yields that had a knock-on effect of frying chips due to thermals than anything else - The hot-spotting is just the result.

[Chris Coates @CJCShadowsan]

@FelixCLC_ Or you just change the PBO curve, undervolt and you're all good. 9800X3D user here... -30 all core 1.1v VSoC...

[Chris Coates @CJCShadowsan]

@mattdunnell I know it isn't a valid excuse, but...

...There's a lot of blokes putting the cue on the rack because they know the season went months ago.

Blokes clearly being told they're gonna be moved on won't have helped either, nor would sacking mid-season with no replacement.