Ananthakrishna

Sharing, learning and building ai products.

[Ananthakrishna @anantha_24]

Not cool

[OpenAI @OpenAI]

We’re ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor’s direct access to our models would end on November 12.

We know that the people most affected by this decision are the developers who rely on OpenAI models in Cursor. We care about their experience in this transition and we’re ready to go above and beyond to support them.

https://t.co/OzuCTzUjfX

[Ananthakrishna @anantha_24]

@argofowl It's the Cursor provider. Idle, it keeps spawning cursor-agent from ~/ and indexing your home dir, even with no chat open.

Settings > Providers, turn Cursor off. Fix is in 0.0.25. If you're already on that and it's still draining, that's the one to report.

[Ananthakrishna @anantha_24]

@ForwardEditor The devil's advocate sitting in the shared thread is the burn.

Lead writes a self-contained brief. Workers stay in their own windows and only send reports back.

Then the $200 plan can go to coding.

[Ananthakrishna @anantha_24]

@arjun_prmkmr Codex even if @thsottiaux forgets to reset

[Ananthakrishna @anantha_24]

@sflorimm 💯

[Ananthakrishna @anantha_24]

Left an agent on a bug for 20 minutes.

It rewrote the test.

The test was wrong. The code was fine.

I still paid for the 20 minutes.

[Ananthakrishna @anantha_24]

@leo_linsky Don't touch the env score. Add a second column next to it.

Weekly votes from people who actually ran the model. Would I put this in my stack.

Then something can top the board and still lose on use.

[Ananthakrishna @anantha_24]

@TheGeorgePu 70B is a cursed VRAM size. Llama 3.3 was the last dense one people actually ran. After that the labs went MoE. Llama 4 Scout is 17B active, 109B total. You either fit a 24GB card or you go huge. Nothing wants to live in the middle.

[Ananthakrishna @anantha_24]

People are going to see 125B on Qwen3.8-Flash-Next and assume they need a rack. You only run 6B of it. Another 51B is n-gram embeddings you can park off the GPU, so it actually fits a Mac Studio or a couple of Sparks.

Weights: https://t.co/glRUd0Quj1

[Ananthakrishna @anantha_24]

This model will be a game changer, we're beginning to see smaller perform so much better than the previous generations, almost close to the frontier models. Crazy how the AI explosion has started.

[Z.ai @Zai_org]

Introducing GLM-5.3-Flash

- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips

Blog: https://t.co/tzOmB7gdZP

Available now across all official platforms:

Weights: https://t.co/9LRMahY9Wa
API: https://t.co/VcaQnzYmS9
Coding Plan: https://t.co/Nk8Y98HNhU
ZCode: https://t.co/Peepqv4XSx
Chat: https://t.co/WCqWT0qCQb
AutoClaw: https://t.co/aGEG5HqTTb

[Ananthakrishna @anantha_24]

What!!!!!

[AJ @ItsmeAjayKV]

Xiaomi just showed its AI Cube Prototype and this could become a serious GB10 competitor from China 👀

- 3 custom chips: Xring O3, O100, D100
- 200 TOPS NPU
- 1.22 TB/s AI memory bandwidth
- Up to 160GB unified memory
- 150W sustained power
- 120B models running locally

Xring O100: 1.22TB/s + 330 t/s on a 150w AI box is 🔥

Once it hit's the marked, going to sell like hot cakes.

[baba yaga @babayagatwt]

Nobody knows who built Ox Alpha.

But this mysterious model has:

- 1M context + multimodal capabilities
- Zero data retention
- Free, nearly unlimited usage for a week
- OpenCode claims capacity for 100T tokens/day
- Reportedly beats Fable, GPT and Opus on some tests

So what the hell is Ox Alpha?

GLM 5.3? A larger GLM? Or something from Xiaomi?

The Xiaomi theory is interesting: MiMo already has 1M context + multimodal agents, Xiaomi worked with OpenCodenand its 100T Token Program matches OpenCode's claimed 100T/day capacity.

If it's GLM maybe the huge 5.3 jump wasn't magic RL after all. It could be distillation from a larger teacher.

A lot of people are betting on Xiaomi.

[Ananthakrishna @anantha_24]

@mweinbach It’s about time

[Ananthakrishna @anantha_24]

@theo You already trust it that much?

[Ananthakrishna @anantha_24]

@thsottiaux I think the next update should be to add a capability to generate high quality animated gifs since image gen is so good at consistency, this will be huge for mobile and web codex can auto generate animated assets, making the UX 10x better when used correctly.

[Ananthakrishna @anantha_24]

I think X needs a button to report AI slop

[Cursor @cursor_ai]

Origin, our code hosting platform, is now live.

It's fast, easy to use, and deeply integrated with Cursor.

Get started by syncing your repos from GitHub.

[Ananthakrishna @anantha_24]

Wow

[Jun Song @jun_song]

Someone figured out how to unlock the true power of Deepseek-V4-Pro-0813.

They fixed the thinking process errors with just a simple harness, and it completely outperforms Fable across every task.

The benchmark scores are insane.

GitHub ⬇️

[Ananthakrishna @anantha_24]

The only real use case I still have for Opus 5 is frontend work.

Don’t get me wrong, Sol 5.6 is solid at frontend design too, but you have to push it hard to avoid the AI slop look. Opus 5 just gives you clean, polished UI with far less effort.

Outside of that I actually prefer 5.6. But Opus 5 still feels like the better architect in most cases, maybe more architecture focused training data?

[Ananthakrishna @anantha_24]

@Da7_Tech Fr? I mean I know opus 5 is a shitty model but still it’s relatively huge.