Skip to content
TodayNov 17, 2026Tue7items
  1. IEEE Spectrum · AI48

    IEEE Spectrum: optical computing startups target AI inference power limits

    IEEE Spectrum reports that optical computing companies now total US $2.22 billion in market value, up from $0.4 billion in 2022, as AI inference strains data-center power. UK startup Optalysys embeds passive optical elements in the networking path to compute during data transit. Others, including Lumai, Neurophos, Q.ANT, Lightelligence and LightSolver, pursue free-space or on-chip photonics, targeting 10- to 80-fold energy gains over GPUs.

  2. MarkTechPost58

    Sarvam launches Token Factory, hosting 8 models on GPUs in India

    Sarvam opened Sarvam Token Factory, an inference platform serving its own and open-weight models from NVIDIA B200 clusters in India through one OpenAI-compatible API. The launch lists eight models, including Sarvam 105B, GLM 5.3, DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro and Gemma 4 31B. Sarvam 105B is priced at ₹10 input and ₹25 output per 1M tokens, introductory through 2026. All benchmarks are self-reported.

  3. Pandaily40

    METASTONE launches multimodal inference service, reports 2.48x video speedup in self-tests

    Chinese computing infrastructure provider METASTONE has launched a multimodal model service for production workflows, combining video-generation acceleration, cache and memory management, token compression and a Context Engine that turns reference material into structured generation instructions. The company reports an average end-to-end speedup of 2.48 times over a dense-computation baseline with quality maintained, but these are vendor self-tests, not independently verified.

  4. SiliconANGLE · AI78

    Alibaba launches Qwen4-Max with training and attention optimizations

    Alibaba released Qwen4-Max, its newest flagship LLM, built on a Qwen4 architecture that made training more efficient. Its backbone has 2.4 trillion parameters but activates only 71 billion at a time, supported by 328 billion N-gram parameters acting as a cache. Alibaba says it used up to 35% fewer tokens than Qwen3.8-Max on an internal benchmark and 49% fewer on Terminal-Bench 4.0.

    Why it matters: Token savings of this size on coding and knowledge-work tasks bear directly on inference economics, and Alibaba's plan to open-source smaller Qwen4 models could widen adoption beyond its cloud.

  5. IEEE Spectrum · AI78

    TypeSafe AI's Jev sparks wave of decision models from Cloudflare, OpenAI and others

    TypeSafe AI released Jev, a decision model it calls a "System One model," on 15 September, claiming 70–500 millisecond latency, $0.042 per million input tokens with free output, and no hallucination. It raised $870 million at a $7.5 billion valuation. Cloudflare, Amazon, OpenAI and Microsoft have since released similar models. Researchers say decision models cut errors but still trail LLMs in medicine, finance and law.

    Why it matters: Cheap, low-latency decision models could absorb much routine LLM work, but domain-accuracy gaps and hybrid designs mean they complement rather than replace frontier models.

  6. NVIDIA Technical Blog42

    NVIDIA says Vera Rubin NVL72 leads MLPerf Endpoints v1.0 agentic inference

    NVIDIA's technical blog states that its Vera Rubin NVL72 system leads agentic inference performance on MLPerf Endpoints v1.0. The benchmark uses coding and enterprise workflows through an API to measure the balance between serving capacity and responsiveness, reflecting agent workloads that carry growing conversation histories and mix reasoning, tool calls and responses.

Nov 16, 2026Mon
  1. MarkTechPost85

    Alibaba's Qwen team releases Qwen4-Max, a 2.4T-parameter MoE model with 1M context

    Alibaba's Qwen team released Qwen4-Max, the first Qwen4-generation model: a proprietary sparse MoE with a 2.4T-parameter backbone, 71B active per token and a 1M-token context window, accepting text, images and video. It is hosted only via QwenCloud, with no open weights. Alibaba cut API prices to $1.50/$4.50 per million input/output tokens and reports a 62% drop in average cost per task.

    Selected coverage of this event: 《阿里发布 Qwen4-Max 基座大模型,智能体与编程能力进阶》

    Why it matters: Alibaba is competing on agentic work and price rather than coding or vision, where it trails Claude Opus 5.5 and Kimi K3.1; the closed, hosted-only release keeps its frontier model under Alibaba's control.

  2. 量子位 QbitAI42

    Beijing AGI institute spin-off Tongxin raises hundreds of millions for AGI inference chip

    Tongxin Intelligent, founded in April 2026 and incubated by the Beijing Institute for General Artificial Intelligence, has raised several hundred million yuan across seed and angel rounds in half a year. The company is designing an AGI inference chip with Vector, Tensor and AGI Core units, compatible with mainstream models while adding dedicated acceleration for logic, causal and value reasoning. It completed architecture definition in three months and plans FPGA prototyping in early December.

Nov 15, 2026Sun
  1. MarkTechPost62

    Cmpnd releases dspy-wsp, a DSPy optimizer that edits model weights directly

    Cmpnd has released dspy-wsp, an MIT-licensed prototype optimizer that plugs weight-space programming into DSPy: a frontier LLM reads a student model's weights in a sandbox and patches them, with no gradients or training data. In tests on 0.6B–9B decision models, 25 writer passes cost $5.42–$7.63 and took 78–167 minutes; d1-omni-600M rose from 50.9% to 69.1% with WeightSpace plus ReAnchor.

Nov 14, 2026Sat
  1. NVIDIA Technical Blog40

    NVIDIA shows serving GLM-5.3 across two DGX Stations with vLLM

    NVIDIA's technical blog describes serving a frontier-scale model across two DGX Stations linked by ConnectX-8 RDMA, exposing one endpoint while tensor parallelism splits weights and computation across both GPUs. The model used is NVIDIA GLM-5.3 NVFP4, a quantized version of Z.ai's mixture-of-experts model, which no longer fits on a single machine.

Nov 12, 2026Thu
  1. Google AI (The Keyword)82

    Google launches Gemini Decisions API powered by Gemini Decision 1

    Google introduced the Gemini Decisions API, built on its new Gemini Decision 1 model and available through the Gemini API in Google AI Studio. It scores up to 1,000 yes/no, multiple-choice and rubric questions in a single pass over text, images, audio, video and PDFs, returning probabilities at 150-250 ms end-to-end latency and $0.10 per 1M input tokens, with no output tokens billed.

    Why it matters: A cheap, low-latency scoring endpoint for routing, policy checks and grading could replace custom classifiers and cut inference costs for teams building agent and multimodal pipelines.

Nov 11, 2026Wed
  1. Meta Engineering · ML Applications58

    Meta details Your algorithm, letting users steer Facebook Feed and Reels with natural language

    Meta described Your algorithm, a system that lets Facebook users state in their own words what they want in Feed and Reels. A page shows an LLM-written Lately Summary and prefilled Interested In and Not Interested In lists; edits are validated by a smaller model, mapped to clean topic names and saved to a shared store both surfaces read. Changes apply through retrieval, reranking and on-device cache invalidation.

Nov 10, 2026Tue
  1. One Useful Thing (Ethan Mollick)78

    Ethan Mollick calls Claude Fable 5.2 'almost AGI' after hands-on tests

    Wharton professor Ethan Mollick, given pre-launch access to Anthropic's Claude Fable 5.2, tested it on filmmaking, hardware hacking, launching a business, recreating an Auden literature course, original research and building 450-plus 3D worlds. He judged it close to his limited AGI definition — as good as most human experts at most computer-based intellectual tasks — but still 'jagged', because it rarely notices when its own work is dull, bloated or over-regulated.

    Why it matters: A prominent practitioner's field test suggests frontier models can now do expert-level digital work across domains, but the remaining gap is self-critique — a signal for how enterprises should deploy agents with human oversight.

Oct 24, 2026Sat
  1. Ahead of AI (Sebastian Raschka)58

    Sebastian Raschka releases 8-part video course on building reasoning models from scratch

    Sebastian Raschka published eight lecture videos, about 11 hours total, accompanying his book Build a Reasoning Model (From Scratch). Starting from a pre-trained Qwen3 0.6B model, the course implements evaluation, inference-time scaling, reinforcement learning with verifiable rewards via a simplified GRPO, and distillation in Python and PyTorch, evaluating checkpoints on MATH-500.

Oct 21, 2026Wed
  1. Interconnects (Nathan Lambert)62

    DeepSeek releases V4.1 Flash and its inference stack as US and China weigh open-model controls

    DeepSeek released DeepSeek-V4.1-Flash-1017, a 552B-A16B model with image inputs under the MIT license, and published its internal inference stack DeepInfer. China's MOFCOM proposed pre-approval for sharing frontier AI models internationally, with automatic approval for open-source technologies, while the US DHS barred DeepSeek as a vendor in federal civilian procurement.

Oct 19, 2026Mon
Oct 16, 2026Fri
  1. Ollama Blog80

    Reflection's open-weight Beam model launches on Ollama with limited-time free access

    Reflection released Beam, its first open-weight model, on October 16, 2026, and Ollama is offering it free for a limited time as a day-zero partner, with zero data retention on its cloud endpoint. Beam is a mixture-of-experts model with about 501 billion total parameters and 23 billion active per token, a 1 million token context window, and an Apache 2.0 license, aimed at coding, reasoning and long-running agentic tasks.

    Why it matters: A 501B-parameter Apache 2.0 model with a 1M-token context, distributed free through Ollama, lowers the barrier for teams building coding and long-horizon agent workflows without local inference infrastructure.

Oct 13, 2026Tue
  1. Amazon Science75

    AWS priming cuts hybrid model pretraining cost by over 99%

    Amazon Web Services researchers introduced priming, a method that reuses pretrained Transformer weights to build hybrid state space models, cutting pretraining cost by more than 99% and using under 0.5% of the source model's token budget. A 32-billion-parameter Gated KalmaNet hybrid matched its source Transformer on reasoning while decoding up to twice as fast at 128,000-token context. Code and models were released under Apache 2.0.

    Why it matters: Cheap, controlled architecture experiments could accelerate the shift away from pure Transformers for long-context agents, where KV-cache memory costs dominate serving economics.

Sep 29, 2026Tue
  1. Ahead of AI (Sebastian Raschka)73

    Raschka traces text classification history and explains Jev's appeal

    Sebastian Raschka reviews text classification from bag-of-words with naive Bayes and logistic regression through RNNs, CNNs and BERT- and GPT-style transformers, then places TypeSafe AI's Jev in that lineage. In his IMDb test, Jev's Choice API reached 96.47% accuracy on 25,000 reviews in about 22 minutes for $0.65, versus roughly 95% for a fine-tuned ModernBERT.

  2. Ollama Blog65

    Ollama 0.35 adds Jev-style decision models via /v1/systemone endpoint

    Ollama released version 0.35 with support for decision models built on TypeSafe's Jev API. A new /v1/systemone endpoint sends text as state plus named questions, and a locally running model answers all in one request. Three models launch: Bespoke Labs' open-source 9B nimble, plus Together AI's experimental 4B tev1 and 0.8B tev1:0.8b. Ollama says nimble averaged 91ms per decision locally on an M5 Max.

Sep 19, 2026Sat
  1. One Useful Thing (Ethan Mollick)42

    Ethan Mollick says GPT-6 Astra and Fable 5.1 can do weeks of human work

    Ethan Mollick demonstrated GPT-6 Astra turning the 1977 text adventure Zork into a playable 3D action game, and Fable 5.1 reconstructing Umberto Eco's roughly 5,000-book library in 3D from videos and catalogues. He argues a large "capability overhang" separates what current models can do from how most people use them, and urges pairing AI with deep knowledge, wide knowledge, taste and agency.

Sep 9, 2026Wed
  1. Ahead of AI (Sebastian Raschka)72

    Raschka analyses GPT-6 Astra, looped transformers and hidden reasoning

    Sebastian Raschka says OpenAI's GPT-6 Astra is the best model he has used, especially at 3D rendering and computer use, and reports 99.9% on ARC-AGI-3 versus 7.8% for GPT-5.6 Sol. He explains looped transformers, which reuse blocks to add depth without new weights, and argues they are unlikely to be the main cause of shorter or hidden reasoning traces.

Jul 18, 2026Sat
  1. Ahead of AI (Sebastian Raschka)58

    Raschka explains how LLMs control reasoning effort settings

    Sebastian Raschka's newsletter article explains how reasoning-effort settings in models like OpenAI's GPT-5.6 are trained and implemented. It describes effort-conditioned supervised fine-tuning and reinforcement learning with token-length penalties, and compares disclosed recipes from six open-weight models including DeepSeek V4, Kimi K2.5, GLM-5, Qwen3 and Nemotron 3 Ultra. Higher effort generally means longer responses and higher cost.

Jun 10, 2026Wed
  1. One Useful Thing (Ethan Mollick)78

    Mollick tests Claude 5 Fable, first public Mythos-class model

    Ethan Mollick says he had early access to Claude 5 Fable, the first Mythos-class AI model released to the public, and found it outperformed every other public model he has used by a considerable margin. It worked up to a dozen hours on multi-page specifications, building an isochrone map and research software, though he says the process became a black box.

    Why it matters: Mollick's hands-on account suggests agentic models are shifting users from steering work to commissioning outcomes, with token cost and opacity as the trade-offs.