Jin Budelmann

@jinthagerman Yeah we’ve been working on improving reliability. It should be feeling better lately, expect it to improve more

@bcherny Classifier is down 90% of the time I try use it. It doesn’t do the thing I want it to do which is allow me to leave and trust safe actions still happen

People often ask what my biggest tip is for getting the most out of Claude Code. These days my #1 tip is: use auto mode Auto mode means no more permission prompts. It is the key building block for multi-clauding: start a session, then while it runs, work on another session in parallel.

@cryptopunk7213 This is a weird take. Why does it matter if Apple uses Claude to build? It has nothing material to do with Apple using Gemini in their products.

this is pretty hilarious latest Apple update included leaked Claude MD files i.e. apple engineers are using claude to create new software at apple apple signed a $1B partnership with google to use gemini for their AI… … but they initially asked anthropic (who refused)

I hate that I feel obligated to do this but this is the most condescending post I’ve read on this platform. I dono how you thought this would land but geez.

@cryptopunk7213 v4 was not trained on Ascend by a long shot. Flash was and only partially. That’s a gross overstatement

china said "we don't fucking need the USA to win the ai race" today. blocking the $2B meta manus deal is half the story: > china released 3 models in the last 7 days that are 95% frontier model capability (opus4.7, gpt 5.4) > deepseek v4 was trained on Huawei chips and will only use chinese chips going forward > china has told all chinese AI labs NOT to use nvidia chips. bytedance has been banned. > china has 3X energy vs. USA to power training of new models. its not even a competition. message is pretty fucking clear: "we have our own sovreign ai stack" either that or china wants to use this as a bargaining chip to convince trump to sell them the best nvidia chips.

@Mikethejohnson @Ric_RTP “its chips were used for part of V4-Flash's training.” Sounds pretty weak

@jinthagerman @Ric_RTP DeepSeek-V4, the Chinese AI model adapted for Huawei chips | Reuters share.google/ENhAxwfaX0OoMf…

China just made Silicon Valley's entire AI industry look like a scam. The US government spent 3 years trying to stop China from building competitive AI. But this backfired HORRIBLY. Here's what happened: Yesterday, a Chinese startup called DeepSeek released a new AI model called V4. It matches the performance of OpenAI and Anthropic's best models. At 1/7th the price. And for the first time ever, it was built on Chinese chips. NOT American ones. That last part is the one that terrifies the west. For context: Since 2022, the US has banned the export of advanced AI chips to China. The entire strategy was built on the assumption that if China can't access Nvidia's best hardware, they can't build frontier AI. But DeepSeek just proved that assumption wrong. Their V4 model was trained and runs on Huawei's Ascend chips. Huawei spent months working directly with DeepSeek to make sure V4 runs across their entire line of AI processors. Jensen Huang even predicted this on a recent podcast: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation." That day was yesterday. And the numbers are crazy: DeepSeek V4 costs $3.48 per million output tokens. OpenAI's latest model GPT-5.5 costs $30. Anthropic's Claude charges $25. Same ballpark performance. 7x cheaper. Uber's CTO just admitted they burned through their ENTIRE 2026 AI budget in 4 months using Anthropic's tools. If Uber had used DeepSeek instead, that same budget would have lasted 7 YEARS. 4 months vs 7 years. Same work getting done. But the pricing isn't even the big thing here. The real story is what DeepSeek did with their technical report: They published the benchmarks where they LOSE. Every AI company cherry-picks the tests where their model wins. DeepSeek ran the full comparison against GPT-5.4 and Google's Gemini, found they trail frontier models by 3 to 6 months, and printed it anyway. They literally don't care because the price gap makes the performance gap irrelevant for 90% of use cases. So the US export controls didn't slow China down. They ACCELERATED China's independence. Because Chinese developers were FORCED to train models with limited resources, they had to figure out how to make AI radically more efficient. That constraint became their competitive advantage. Every generation of DeepSeek has gotten dramatically cheaper to train. V4 continues the trend. Meanwhile US companies are going the OPPOSITE direction: OpenAI's GPT-5.5 Pro costs $180 per million output tokens. That's 51x more expensive than DeepSeek V4 for comparable work. The Commerce Secretary confirmed this week that ZERO Nvidia advanced chip shipments have actually gone through to China despite being approved in January. So China built frontier AI anyway. Without American chips. At a fraction of the cost. And the market response tells you everything: Chinese chipmaker SMIC surged 10%. Huahong Semiconductor jumped 15%. DeepSeek's Chinese AI competitors Zhipu AI and MiniMax dropped 9% because V4 is destroying them too. DeepSeek is making Silicon Valley's pricing model look like a scam. US tech companies spent $650 billion on AI infrastructure this year. DeepSeek just showed the world you can match their output for pennies. The export controls were supposed to be America's ace card. Instead they taught China how to win without American chips, at American prices nobody can compete with. Jensen Huang was right. This is a horrible outcome. But it's the outcome America built for itself.

@realarmaansidhu Where does it say it’s trained on Ascend chips?

DeepSeek just shipped V4. 1.6 trillion parameters. 1 million token context. Free and open source. The kicker: it's the first major model trained almost entirely on Huawei Ascend chips. The moat the US built around AI leadership just evaporated on a single launch. The numbers: V4-Pro runs 1.6T total / 49B active parameters. V4-Flash runs 284B / 13B active. Both ship with 1M context as the default across every API endpoint. The tech report claims open-source SOTA on agentic coding benchmarks. Rich world knowledge trails only Gemini 3.1 Pro. Math, STEM, and coding rival top closed-source labs. Liang Wenfeng's sign-off: "The whole world should use a 1.6T model for free." That's the thesis. Meanwhile OpenAI shipped GPT-5.5 two days ago behind a $200/month paywall, Anthropic gates Claude Opus behind enterprise contracts, and Google's Gemini 3.1 Pro is API-metered. The hardware story is what destroys the narrative. The Biden and Trump export controls on Nvidia H100/B200 chips were explicitly designed to prevent exactly this outcome. Stop advanced chips from reaching China. Slow Chinese AI training by 18-24 months. Buy American labs time to entrench. Two things happened instead. First, Huawei's Ascend 910B and 910C stack matured faster than US intelligence projected. Second, Chinese labs optimized training pipelines aggressively enough that they no longer need H100-tier hardware to reach frontier performance. DeepSeek V4 is proof the wall failed. Kimi K2.6 took Claude 4.7 Pro's crown on SWE-bench last week. Qwen 3.6 27B ties Claude 4.5 Opus from an 18GB laptop. Three Chinese open-source frontier launches in eight days. American labs have gone quiet on X. Chinese labs publicly celebrate each other's releases as a team. The contrast is the whole story. The US bet the AI decade on a chip embargo. The Chinese bet it on shipping open-source on chips they built themselves. One of those bets is already in production. The other is still waiting at customs.

@GeoffreyHuntley @DanielW_Kiwi It me

gm

@Last_Week_in_AI why no turboquant coverage?

On Repo Bench Opus 4.7 scores meaningfully worse than 4.6, which did worse than 4.5 With this score it's no longer in the top 25

Excited for this one meetup.com/prompt-poets-s…

@0xSero No TurboQuant?

Guide to running BIG B0Is on your small hardware. 1. Use REAPs: up to 50% savings 2. Use quantisations: 75% savings - AWQ / GPTQ / W4A16 / FP8 = FAST inference - GGUF / EXL3 = Slow but just works - MLX = Best for apple 3. Use 8bit KV cache: 50-75% savings

@cyprusad @0xSero cerebras.ai/blog/reap

@0xSero What is a REAP and what does observation mean here?

Currently: 1. Framework desktop quantising and benchmarkjng all my qwen3.5-122B reaps on the strix halo 2. 8x 3090s inference for my friend qwen3.5-262b 3. 4x B200s observing GLM-5.1 for reap. Should be done Thursday 4. 8x H100 observing Qwen3.5-397B for pruning done Friday 5. H100 observing Trinity Large Thinking for pruning 135M tokens of observation data, packaged into batches of 16k tokens - cyber security - philosophy - math - reasoning - coding - agentic - terminal - browser - cuda - my own sessions Once done I’m going to start optimising this observation process with autoresearch Cheap AI at home

@traskjd Congrats!

Family Size++ ❤️

My take on Anthropic's new Mythos model: Don't worry about it. It might be amazing, it might be disruptive. But there's nothing you can do about it. Ignore the hype, keep your head down and keep learning, growing and creating. Adopt this: Relentless Optimism.

no no i've definitely been doing a lot of work it's all just too dangerous to release

@sporadica The irony of this post

being a visceral hater is the new meta, apparently an entire genre of twitter personality has just converged on this meme. they’re just doing the meme. and they don’t even realize it lmao

They have a openclaw sized hole in their inference compute budget now, wonder what replaces that...

I say this as a mostly very happy Claude Code user but the last week has been weird and it's been bugging me: - Anthropic are obviously struggling with inference demand and also trading off compute for inference with compute for training - If you need evidence of this, see OpenClaw bans, see pretty consistent model quality concerns from the community, see patchy inference speeds, see off peak 2x limits - There is no world where I believe they are gating Mythos for safety alone. If they could release it publicly, they would - Enterprise is going to pay through the nose for the Mythos inference they have available - They can now delay distillation from 3rd parties indefinitely - And it allows them to build hype and desirability which they can parlay into more compute investment - Don't fall for the hype or the doom, remember they just leaked the source for Claude Code, yet they have the "most sophisticated cyber security model ever" 🤔