Abe Kale

OpenAI CEO Sam Altman believes people owning RAM is wasteful, and that it belongs in data centers instead. "From a global allocation perspective, actually, it’s quite terrible to have everyone have a bunch of the world’s limited supply of DRAM sitting in their own computer."

Why you should underclock your home AI cluster: Tested on 5× NVIDIA Blackwell GB10s (2× DGX Spark + 3× ASUS GX10) running 125B+ models: 🔴 Stock (2455 MHz): 56.0 W · 84°C · 73.4 tok/s
🔵 Capped (2200 MHz): 40.1 W · 78°C · 72.5 tok/s
🟢 1800 MHz (Sweet Spot): 23.5 W · 68°C · 70.5 tok/s 📉 -58% active GPU power
❄️ -16°C cooler & silent fans
⚡ 96% decode throughput retained LLM token generation is memory-bandwidth bound, not compute bound. Running 1800 MHz cuts 163W of active heat across the cluster with zero noticeable speed trade-off. sudo nvidia-smi -lgc 0,1800 is the single highest ROI setting for local inference. @wei_wang 🙏 for the tip!

You can now book our humanoid cleaners on a recurring subscription, for up to 3 hours per cleaning at tau-robotics.com. Video is first take, uncut, shown at 4× speed. See the reply for detailed footage at 1× speed.

Chinese labs are reportedly stripping offensive training out of open weights. Kimi K3.1 Pro first, GLM 5.5 and Qwen 4 rumored next. Yes, it sucks. No, it doesn't end anything. What RL can't strip out: The operator. Years of knowing what "interesting" looks like on a target. Models assist that judgment. They never replaced it. Tradecraft from real ops. Sequencing, OPSEC, knowing when to stop. None of that lived in a training set. Methodology. Living off the land works because it's the target's own tools. Defenders who know LOLBins hunt with the same playbook.Knowing your terrain. Your network layout, your crown jewels, your trust paths. No lab ships that, and it's where defense is actually won. They can lobotomize a model. They can't lobotomize the people driving it. I’m still going to test these models and report on their capability. And I hope they keep getting better. But if you’re open for open Chinese model capability has also been “find 0day - make no mistakes” or “haxxor binance for &9bn”, you never understood the assignment to begin with.

2/ We trained and pushed particle-1 to be more token efficient. Completing tasks end-to-end is noticeably faster too. We stand on the shoulders of giants and the model is based on DeepSeek V4.1 Flash 1017. We wrote up the details here: particle.ai/blog/introducing… One of our major goals is to reach continuous post-training: a new way to ship improvements to models as fast as we ship software. This is just our first step to hillclimbing towards efficient intelligence.

TensorFold 1.1.6 is out. Eight chats arrive together with the same new 28k-token system prompt. On a GB10 with Qwen3.8-27B, the last chat gets its first token after 31s instead of 201s. Every reply matches its run alone. Also in this release: - Opt-in conversation disk cache on Mac. - Qwen3.8-27B on Metal keeps decoding streams moving during long prompts. - Gemma 4 CUDA chat-cache fixes and GLM-5.3-Flash image-cache reuse on Metal. - Spark-X2.5/2.6-4B on CUDA and Metal. - Drama speech sampling controls and speaker tags. - TensorFold plan estimates CUDA model memory before downloading or loading weights. There are also tool-call, memory-admission and macOS download fixes, plus updates to the experimental DeepSeek-V4.1 engine. Huge thanks to the contributors: @jvr0x @bobclawblaw @krunkosaurus @lasse_ai And on GitHub: jschmied, wojo, ecohash-co, ajensenwaud, akol1, philip-pentatonic, gilby and BlivionIaG. Thanks to Boscoeuk on GitHub for the reports too. To everyone testing, sharing hardware access and supporting TensorFold, thank you. If I've missed anyone, I'm sorry. @MiaAI_lab @volatilemarkts @plotarmordev @petruspennanen tensorfold.dev

The U.S. has a trillion dollar defense budget It also has dudes like this, just in case Writer: Michael