Noir

@realBigBrainAI Impressive..

Years of reinforcement learning in one clip: Atlas pulls off a cartwheel-to-backflip combo with perfect balance.

5.5 instant is scary good. You sure this isn't a local LLM? @OpenAI

@github ?? You guys okay

@Nate_Esparza Because testosterone...

I don’t understand why everyone doesn’t drive electric tbh

@Snixtp Didn't get the memo hey? GPU AI is ending.

Another 3090 172GB VRAM✅

Agent architectures haemorrhage their token budget on chain of thought theatre garbage. "I need to think… my plan is… therefore I will…" That's not reasoning. That's annoying narration. Shannon called it noise. Unix calls it a violation of the rule of silence: when a program has nothing surprising to say, it should say nothing, just like humans. McIlroy nailed this in '78. Did we forget? The better execution model: → Agent in a waiting room → Tools bound, task scoped → Atomic execution, zero output tokens → Silent on success → Narration only when reality diverges from the prior. Happy path entropy floor. Exception path signal actually worth transmitting. This is event driven semantics applied to cognition. Predictive coding for agents: encode the surprise, drop the rest. Everything else burns a finite channel. Humans don't narrate every nail. They yell when the wall cracks. Why are we building agents that do worse than carpenters? Silent execution. Exception only reporting. Atomic tool calls. "The waiting room model" @OpenAI

@KettlebellDan Spend it

ok your best tip to lose 10 pounds

@aakashgupta Nice! I'm looking forward to seeing the first AI movie. People are just over paid and need a reality check. AI will do this thankfully.

Rendering this 29-second AI dragon shot in 2019 would have cost over $500,000 at a VFX house. The Chinese model that probably generated it ran the entire shot for $2.90. The Artificial Analysis video arena ranks every public AI video model by ELO score. The top spot belongs to Kling 3.0 from Kuaishou at 1243. Sora 2 sits below it. Veo 3.1 sits below it. Runway Gen-4.5 sits below it. The base model leading the benchmark, and the open-source variants any kid can download tonight, both shipped from companies the export controls were supposed to handicap. Wan 2.2 from Alibaba has been downloaded 5.4 million times from Hugging Face. Open source, runs on a single consumer GPU, ships free. Tencent's Hunyuan Video 1.5 generates clips on an RTX 4090 in 75 seconds at 8.3 billion parameters. Also open source. Also free. Pixomondo built every dragon shot on House of the Dragon. The CGI work on a single dragon scene ran over $1 million. HBO shot the whole show at under $20 million per episode and the dragons were the headline line item. 29 seconds at Kling 3.0 pricing is $2.90. Run the prompt 100 times to nail the take and you're still under $300 for the entire production. Disney spent six years and over $15 billion buying Pixar, Marvel, and Lucasfilm for the pipeline. The pipeline just got commoditized in Hangzhou.

@dangerm00se @aakashgupta Greedy. Like the US has always been.

@aakashgupta do you have a take on why china is currently so open and US so closed?

@luthiraabeykoon @karpathy Called it.. x.com/Entrouter/stat…

We implemented @karpathy 's MicroGPT fully on FPGA fabric. No GPU. No PyTorch. No CPU inference loop. Just a transformer burned into hardware, generating 50,000+ tokens/sec. The model is small, but the idea is not: inference does not have to live only in software 👇

@elonmusk Hitting that blunt again isn't he..

Yes

Mate I don't use codex and can see what they need.. A simple one click package apply, linked with a payment feature. When they max their plan, they click "Plan Multiplier". It bills their plan price again and shows a X2 on the DP or info bar. This continues to stack each multiplier added. Though I'd suggest non persistent stats/features. Just the same initial usage from that plan refreshed.

What are we obviously not getting right with Codex?

Mate I don't use codex and can see what they need.. A simple one click package apply, linked with a payment feature. When they max their plan, they click "Plan Multiplier". It bills their plan price again and shows a X2 on the DP or info bar. This continues to stack each multiplier added. Though I'd suggest non persistent stats/features. Just the same initial usage from that plan refreshed.

@official_taches what do you need for just one plan?

I’ve officially cancelled both Claude Max plans and have 2 x Codex Max plans. Codex - particular GPT5.5 is the best coding model.

@jbhuang0604 Now spend the same again fixing all its mistakes..

Keep getting rate-limited by Claude, so I tried out DeepSeek V4 for the first time. After 10M+ tokens, holy crap the cost is ... 🤯

@dali99288206 @travelingenes Good, I hope not. China and Japan are phenomenal cultures. The west makes me sick with its infestation.

@Entrouter @travelingenes 你不可能毁掉日本文化，因为他们的文化是从中国学的。又超强的民族凝聚力。除非你把他们从地图上抹掉。

i'm 20. i flew to japan alone to make a documentary about a stranger. this is what happened.

@jamesqquick Moat + infra

Dumb question. How do you make money with an open source project…?

@elonmusk Actually insane.. First time I've been speechless in a long time.

Grok Imagine tutorial made with Grok Imagine. These is all AI-generated!

@BrunoLM7 @aileaksofficial @github @GitHubCopilot Not to mention being global rate limited and weekly rate limited.. Junk company, they killed themselves.

@aileaksofficial @github @GitHubCopilot I hope this is not real, 1500 with a 27x, I'm not paying $100 for that.

GitHub is shaking things up! 🚀 A new "Copilot Max" plan is officially in the works as part of a major lineup revamp. While it's mostly a placeholder for now, GitHub is clearly gearing up for a higher-tier AI experience. $99/mo 👀 github.blog/news-insights/…

@GHchangelog Can you show what the Devs are charging you via API so we can see how much profit you earn via each API?

GPT-5.2 and GPT-5.2-Codex will be deprecated across GitHub Copilot experiences on June 1, 2026. • Switch to GPT-5.5 or GPT-5.3-Codex before deprecation. • Copilot Enterprise admins may need to enable new models in settings. github.blog/changelog/2026…