Mark Kretschmann

By the end of this year, AI coding models will be better than the vast majority of human coders. Mark my words. 🔮

AI safety is a giant grift:

@Damien739254 @GoogleAI No, the problem is very real. Logan has posted about it too.

@mark_k @GoogleAI Google isn't short on TPUs. They train Gemini 3 on TPUv7 and Meta is literally lining up to buy their chips. The limits aren't a hardware problem — it's a choice. Enterprise gets the capacity. Individual users get throttled. Very different things.

Google should build its own chip fab, just like Tesla does with the Terafab. Although @GoogleAI is usually seen as extremely rich, due to some mishaps it is currently short of enough TPUs and GPUs to meet its enormous AI inference demand. Hence the severe usage limits in Gemini.

@JuniperViews @GoogleAI It can be done, see Tesla.

@mark_k @GoogleAI Fab is very very expensive and they don't have the expertise

@DivineRational Nope.

@mark_k Nvidia partnership

Cursor has *one more thing* coming up... 👀 Take a guess in the replies!

@AshdinV @elonmusk @cursor_ai @grok Grok Build is coming

@elonmusk @mark_k @cursor_ai We need native tools to use @grok on. Not third party MCPs. Native support from @grok to action

Tomorrow, @cursor_ai is going to release a super strong coding model. Possibly SOTA. Stay tuned. 🫡

@chikenfacegoat @GoogleAI 😂😂

@mark_k @GoogleAI They should buy NVIDIA.

@ShamsAmin56 @xai Yeah it looks sweet!

@mark_k @xai Cool animation with the mosaic filter 🔥

Here's a nice marketing video made by @xai, explaining Grok 4.20 with its multi-agent feature:

@AFK500595388855 @xai Yeah that sounds like fun, but it's hardly useful

@mark_k @xai lol. that's what I did, but hoped you experimented with them already. I'll give it a try. Let's see. A paranoid one, a scientist, a cynical one and a troll. 🤔

@Neuralithic Google needs to build a fab, like Tesla does.

Google is not going be able to acquire chips fast enough. They’re already behind in capacity unfortunately. TSMC doesn’t have enough capacity to serve them as Nvidia, Anthropic, Amazon, and OpenAI of course all locked in contracts with TSMC earlier than Google. Biggest constraint by 2030 will come back to the lowest level of supply chain - fabs. Max fab output of TSMC by 2030 will be about 90-100 fabs.

@AFK500595388855 @xai No, so far I didn't see the need. In my mind it's probably easy to make it worse instead of better!

@mark_k @xai Have you tried creating your own agents?

@matijagrcic Cursor and Codex are a good combo!

@mark_k why would you be on anything else, Cursor is the best harness even after the December leap, Codex app being the second and catching up.

Are you going to give Cursor another chance, now that Composer 2 is here?

@ItsBadAni @xai Yes, Grok can take over. Just say the word!

@mark_k @xai Are we in good hands? Can I quit my job?

@p4cs_pedro Dream on! 😉

@mark_k $10 composer plan

@faint6005 lol no

@mark_k Opensource

Interview with Jensen Huang on the All-In pod, live at @Nvidia GTC! Check it out here: x.com/theallinpod/st…

@rayanabdulcader @cursor_ai Yeah, pricing is important

@mark_k @cursor_ai Cool to see AI coding tools getting cheaper and smarter at the same time a frontier-level coding model inside the editor that can handle long tasks feels like a big step for everyday dev workflows.

The team at @Cursor_AI has officially launched Composer 2. It introduces a frontier-level coding model directly to the editor, designed to offer an optimal balance of high intelligence and low cost. The new model delivers significant improvements across key benchmarks, including Terminal-Bench 2.0 and SWE-bench Multilingual. By leveraging reinforcement learning on long-horizon tasks, Composer 2 is now capable of autonomously solving complex problems that require hundreds of actions. Pricing is highly competitive, starting at just $0.50 per million input tokens and $2.50 per million output tokens. A faster variant with the exact same intelligence is also available as the default option, providing developers with premium speed without sacrificing quality.

@CalimanuLoredan @leerob 🤐🤐

@mark_k @leerob Did you find out what it is? 🫡

@leerob OMG. Spill it, now! 👀

@tejashaveridev Who's Wilson?

@mark_k hopefully Wilson's long running agents framework