📊 Today's Overview
The LLM price war went loud: OpenAI is reportedly offering up to 90% discounts to poach Anthropic's biggest customers, with Rippling's coding share flipping from 16% to 52%. Meanwhile, a closer cost breakdown shows cache reads are 96-99% of agent token spend — the line everyone forgets when reading a pricing table. On the hardware side, AWS CEO Jassy says compute is sold out through 2026 and mostly through 2027. And in robotics, Figure 03 sorted 12,732 packages in 10 hours — just 192 short of its human rival.
🔥 Trend Insights
- LLM price war escalates: OpenAI's 90% discounts and DeepSeek-style cuts are shifting competition from raw capability to cost per token — but cache-read pricing is the real lever.
- Compute stays sold out: AWS says 2026 capacity is gone and 2027 nearly gone, while SemiAnalysis projects 50GW of new build in 2027 alone.
- Agents that learn from failure: Meta's Self-Direction distills each mistake into a tiny LoRA adapter, turning one-off retries into a persistent lesson library.
🐦 X/Twitter Highlights
📈 热点与趋势
- Figure 03 vs. humans in a 10-hour package sorting race — just 192 items short - Humans sorted 12,924 items; Figure 03 sorted 12,732. The robot never took a human-style break. The gap is long-horizon stability, not peak speed @ctorobotics (subject: Figure AI)
- Eka Robotics claims 1,500 grasps per hour using force sensing - Humans manage about 600, with over 95% success. a16z led a $150M round. The pitch: let robots sense force, not just rely on vision @LondonRealTV (subject: Eka Robotics)
- Interconnects podcast unpacks Kimi K3.1 Pro's "high scores, mixed reception" - Topics: strong benchmarks but muddled response, whether open-weight cybersecurity risks are underestimated, network vs. AI experts disagreeing on the same events, and Anthropic no longer hiring junior staff @natolambert (Nathan Lambert, Interconnects author / Ai2 researcher)
🔧 工具与产品
- Nanbeige4.3-33B launches, targeting the 30B-class open-source agent ceiling - The team says it hits open-source SOTA at its parameter count on agentic coding and collaboration benchmarks; plugging into mainstream agent harnesses lets it run real tasks cheaply. It will also serve as a teacher model to distill the upcoming Nanbeige4.3-3B @nanbeige (Nanbeige, open-source LLM team)
- Together AI adds MiniMax M3.1 Flash post-training - Users can run SFT or RL with their own data, reward functions, and environments, keeping the training loop in their own Python code — tuned for agent workloads @togethercompute (Together AI, open-source model infrastructure company)
- Ethan Mollick says Opus 5.5 basically rebuilt his book website on its own - The previous version, built with Opus 4.8, needed heavy human effort to get a distinctive static site. This time the model did most of the work alone @emollick (Ethan Mollick, Wharton professor)
⚙️ 技术实践
- Resin splits the LLM inference engine into a "shared core + model forward" - Author Eric Zhang's idea: reuse all the generic parts, and write only the forward pass and kernels per model/hardware. On DeepSeek-V4.1-Flash it lands within 5-10% of DeepInfer's speed, and it's self-contained with metrics, an HTTP server, and prefix caching. Codex generated a GLM-5.3-Flash engine (including MTP) overnight, hitting 60K input tok/s and 16K output tok/s on B200 EP4 @ekzhang1 (Eric Zhang, Resin author / independent developer). vLLM core author Woosuk Kwon responded: vLLM already isolates model-specific code and is considering moving more of the stack to Rust @woosuk_k (Woosuk Kwon, vLLM core author)
- OCRia-1 drops a 37-page technical report; its 4B model takes multiple firsts in document parsing - ParseBench 87.4, olmOCR-bench 83.5 (tables 92.9), PulseBench-Tab 84.1, fr-bench-pdf2md 72.1, benchmarked against Infinity-Parser, LightOnOCR 3, Chandra 2, and dots.mocr. Decoding is 2-4x faster than the latest open-source models, plus a 24-page appendix detailing data and training pipeline experiments @jerryjliu0 (Jerry Liu, LlamaIndex CEO)
- MiniMax H3 ecosystem adds 11 community contributions - FastH3 implements 4-step distillation and runs on DGX Spark and Apple Silicon; NVIDIA SANA team's Sol-H3 generates 15-second 768p video+audio in 6.6 seconds on 8×B300; VDN reworks attention; Alibaba PAI turns PDD distillation into 8-step Acc-LoRA and integrates with ComfyUI; LightX2V offers 4/8-step Turbo LoRA @MiniMax_AI (MiniMax, video generation model company)
- Tsinghua, AgiBot, and Infinite Light Year train with 96 robots sharing experience - The Collective Experience Learning (CEL) framework lets robots check each other's successful trajectories to update policies. Average success on five real tasks rose from 81% to 97.3%; phone packaging went from 79.2% to 98.3%. Scaling the fleet from 16 to 96 robots pushed success from 86.7% to 98.3%. AgiBot shipped over 8,600 humanoid robots in H1 2026 @CyberRobooo (CyberRobo, robotics media account)
- A pure-code prompt gets Claude Opus to replicate the Kurzgesagt animation style - Output is a single HTML file where each frame is a pure function of time, rendered headless with Playwright and voiced with ElevenLabs. It includes shot choreography for a sun collapse, an Earth sunlight countdown, and an astronaut mascot, plus sound-effect alignment rules. The original Kurzgesagt video took about 1,200 hours to make @finnardley (Finn, AI content creator)
⭐ Featured Content
OpenAI and Anthropic spark an LLM price war: up to 90% discounts to poach big customers, and cache-read cuts are the real cost story | Business-side pricing intel + cost structure breakdown
Two complementary reports paint the full picture. The business side, via The Information: OpenAI is offering up to 90% discounts to poach Anthropic's big customers, and it's working — OpenAI's share of Rippling's coding workload rose from 16% to 52%, while Anthropic's fell from 57% to 18%. Replit switched after getting 30%+ off on a contract worth over $100M annually. DataCamp says quotes cut costs by 80-90%. One US bank doubled ChatGPT seats to 500 for free, at under $7 per user. Max Woolf's technical breakdown delivers the more important insight: in agent scenarios, cache reads account for 96-99% of total token spend, with 90-100% hit rates. So a small cut in cache unit price equals a 70%+ drop in overall cost — the column most easily missed when reading a pricing table. The piece closes with Vercel gateway data showing unit token cost has fallen to 43.3% of April 2026 levels, yet total spend rose, because models get better faster than they get cheaper. Teams doing API budgeting, multi-vendor strategy, or agent cost modeling should read both.
Sources: the-decoder.com | minimaxir.com
Scientific American reconstructs the inside story of OpenAI's million-dollar math proof: the model took a path completely different from what its own mathematicians expected, and the community is still reverse-engineering it | A first-hand narrative of "opaque models doing science"
NYU's Tristan Buckmaster gave his first public talk after publicly accusing OpenAI of plagiarizing his work. The venue was packed like a rock concert, with audience laptops running Claude Code and Codex to write proofs. The article confirms OpenAI's internal model did find the correct answer, but the proof path was completely different from what the company's own mathematicians initially assumed, and the community is still reverse-engineering its opaque logic. The core insight: opaque models are making scientific history wrapped in trade secrets, and humans (along with priority disputes) can barely keep up. This is first-hand material for understanding the new normal of "AI doing science" and the priority controversy.
Sources: scientificamerican.com
Mundo: from a few room photos to an explorable, editable, simulatable world — using explicit 3D proxies + a simulator to avoid "looks right but physics doesn't hold" | A clear sample of layered world-model architecture
The pipeline has four stages: first, a 3D foundation model recovers camera pose and geometry; a 3D detector gives initial bounding boxes; then a VLM agent iteratively refines the proxy world in Blender through an analysis-by-synthesis loop of "edit scene program — render — compare with original image." At runtime, a code world maintains explicit 3D state, a simulator computes object motion and interaction, and finally a video model renders the result using the input image as visual context. The core idea: hand "physical feasibility" to explicit proxies and the simulator, and "visual realism" to the video renderer. Teams working on world-model physical grounding and robot simulation can borrow this layered approach directly, and it supports simulating robot tasks inside.
Sources: huggingface.co
llm-d 1.0 released: the first stable version under CNCF governance, making PD disaggregation truly operable and adding agent-aware serving | A platform capability checklist you can benchmark your self-hosted inference stack against
Co-founded in 2025 by Red Hat, IBM, Google Cloud, CoreWeave, and NVIDIA; now 25+ organizations and 500+ contributors. The 1.0 additions focus on Day-2 production operations: multi-tenant priority and fairness flow control (latency-sensitive requests and bursty agent loads coexist without starving each other), router hot-standby high availability with zero-downtime rolling upgrades, coordinating prefill/decode replicas as a single logical unit for upgrades, a v1 config API, and agent-aware serving — session-aware routing pins multi-turn conversations near cached context, paired with turn-priority scheduling and native Responses API state support to reduce recomputation. Under the hood it still reuses prefix-cache/load-aware routing, KV cache offload, and wide expert parallelism, supporting both vLLM and SGLang. A companion quick-reference guide covers deployment dependencies (K8s + Helm + Gateway API Inference Extension), hardware examples, and a decision framework for "when you need cluster-level routing vs. single-machine vLLM is enough."
Sources: thefastmode.com | aicybr.com
NVIDIA open-sources the full Run:ai GPU orchestration platform (Apache-2.0), with four independently adoptable modules | A selection index for building shared GPU pools
Open-sourced on 2026-10-13, including the control plane, cluster services, installer, CLI, and MCP server, with a free community edition on NGC. Four separable modules: KAI Scheduler (gang scheduling, fair-share, topology awareness, preemption), GPU Fractioning Operator (CUDA-native memory partitioning with per-pod metrics), KAI Resource Management (project/department/node-pool quota hierarchy), and Workload Map (a unified contract for heterogeneous workloads). Suited to infra teams building shared GPU pools, who can choose by cluster scale and operational burden — but the secondhand writeup is limited in depth, so read the official announcement and repo directly.
Sources: aicybr.com
Google open-sources TorchTPU: PyTorch tensors, eager execution, and torch.compile run directly on Cloud TPU | The first checklist for evaluating a TPU migration
Open-sourced on 2026-10-20 at PyTorch Conference (google-pytorch/torch_tpu, Apache-2.0, 0.2.0), covering TPU v5e/v5p/v6e/v7 Ironwood, with integration paths for HuggingFace Transformers/Diffusers, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. The writeup turns versions, Python 3.11+, `torch.device("tpu")`, and compilation/distributed caveats into a quick-reference table. It lacks measured performance and operator coverage data — usable as a starting point for migration evaluation, but depth requires the official docs.
Sources: aicybr.com
Open-source project GPU Unlock boosts RTX 30-series AI performance by up to 60%: a kernel module rewrites the speed-select register, bypassing Ampere consumer-card instruction dispatch limits | A hardware bonus for local inference/image-gen users
A Linux kernel module rewrites the hardware speed-select register before driver load and performs an FLR, with no driver changes or VBIOS flashing. Measured: RTX 3080 Ti BF16 from 71.6 to 114 TFLOPS, TF32 from 35.7 to 65.3; in ComfyUI, MOSS-Image/Qwen-Image 2.1/Z-Image Turbo throughput rose 56-58.6%, but INT8 saw almost no gain; vLLM prompt processing +33% while decode stayed flat (limited by memory bandwidth). Limits are clear: no added bandwidth, no unlocked CUDA cores, results vary by card, and the Windows UEFI version is still experimental. Narrow audience (Ampere consumer cards, local deployment only), and the source is a secondhand crypto-news relay — verify against the original GitHub project.
Sources: kucoin.com
AWS CEO Jassy: compute is sold out for 2026, basically sold out for 2027, and a good chunk of 2028 too; SemiAnalysis projects 50GW of new capacity in 2027 alone | A supply-side signal of compute scarcity
John Rath's Digital Infrastructure weekly roundup: beyond Jassy's sold-out remarks, there's Armada's containerized "superintelligence factory" based on NVIDIA DSX POD, Crusoe Spark Liquid modular data centers (3.5MW per rack, 5x density, closed-loop liquid cooling), Crusoe's $12B partnership with Paraguay's Itaipú hydropower, and TikTok's US joint venture seeking several GW of compute. It's a weekly link aggregation, one or two sentences each; the value is the "2027 compute basically sold out" scarcity signal, with details requiring a trip to original sources.
Sources: digitalinfrastructure.substack.com
🎙️ Podcast Picks
AI 时代，PhD 路上的 10 个关键问题｜对谈李磊：港大 PhD、小米 MiMo 研究员
📍 Source: 十字路口Crossing | ⭐ 3/5 | 🏷️ LLM, Research, Interview | ⏱️ 00:51:50
Li Lei, a PhD at HKU and post-training researcher on Xiaomi MiMo, walks through 10 key questions about doing a PhD in the AI era: fast vs. slow research, choosing schools and advisors, cold-starting internship hunts, and using open source and social media to spread research influence (the "big fan theory"). The second half covers Xiaomi MiMo's live post-training, data company valuations, RSI self-training, and Video-MME benchmark design.
💡 Why Listen: If you're weighing a PhD or already in one, the career advice is practical and grounded. The MiMo post-training and benchmark design bits are the technical payoff — just don't expect deep dives.
How I'm Actually Using AI Right Now (Including Claude Duet)
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Agent, Product, LLM | ⏱️ 00:32:28
Tied to the Claude Duet launch, the host maps out his own AI stack across strategy, research, coding, content creation, and data analysis, comparing how different models perform on each task. He then contrasts ChatGPT Dots with Claude Duet and asks why proactive personal agents are finally fitting into real workflows.
💡 Why Listen: A quick, honest look at model selection in real use — useful if you're deciding which tool to reach for. Light on depth, but the Duet vs. Dots comparison is timely.
📄 Paper Highlights
Training an MoE LLM in MXFP4: How We Closed the Gap to BF16
AMD ｜ 🏷️ Quantization, Training, MoE
An open recipe that trains a 20B MoE in FP4 and cuts the convergence penalty vs. BF16 from +20% to +5% — plus honest negative results, making it a practical reference for low-precision training.
Kolibri: A Sovereign European Model on the Pareto Frontier
Aleph Alpha ｜ 🏷️ MoE, Training, Fine-tuning
A 78B-parameter English-German MoE with 3.46B active params, Apache 2.0 weights, trained on 24T tokens with over 20% German data — a full-stack sovereign model play worth studying for data pipeline design.
Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons
Meta, Princeton ｜ 🏷️ Agent Memory, LoRA, Agent Framework
Turns each agent failure into a tiny LoRA adapter (~0.02% of params) stored in a persistent lesson library, boosting success by 8.9-17 points — a neat answer to agents forgetting their mistakes across sessions.