Technical AI signals · Daily at 8 PM ET
Friday, September 25, 2026
Today centered on deployable agent plumbing and open long-context models, led by Google AX at cluster scale, Qwen’s omni-modal agent stack, and Anthropic pairing a benchmark jump with cheaper Opus 5.5 inference.
AX is the day’s clearest runtime artifact: a declarative orchestrator for sandboxed agentic tasks built to run billions of workloads per cluster, continuing Google’s push from Substrate into usable execution infrastructure.
Qwen paired a native multimodal 1M-context model with an omni-agent report and open frameworks, making the release more than just another checkpoint drop.
Anthropic combined a concrete research milestone in the planar N=4 benchmark with a direct operator lever, pricing Opus 5.5 20% lower on input and output tokens and 60% lower on cache reads.
DeepSeek’s 552B multimodal MoE and LongCat’s 1.6T/48B-active release both push million-token windows, signaling that extreme context is now showing up across very different open-model designs.
The incident was limited to 53 uploaded images, but it is a sharp operational reminder that account disassociation, filtering, and link handling can still fail in agent research environments.
P Equity says Vera Rubin bill-of-materials costs rose by triple digits for ABF substrate, PCB, and memory, and that ABF substrate shortages may last beyond 2028 and possibly to 2030.
Google says Gemini 3.8 Flash TTS and Flash-Lite TTS are expressive audio models, and Project Suncatcher will launch a prototype satellite to test TPUs in orbit.
Agent Ultra uses swarms of agents to do exhaustive research across thousands of sources.
The paper says a coding agent can analyze full agent logs and write the optimization prompt directly, instead of running a search loop over small trajectory batches.
Microsoft is adding Autopilot, Code, Home, Office, Teams, and a proactive Today view that surfaces information from M365 without prompting.
The poster says Tag writes over half of their PRs, does nearly all data analysis, and can run workflows with memory and connector access.
GameDev OS gives a coding agent 9 specialists and 64 skills for art, sound, video, marketing, and technical direction.
The authors built AlignmentForecastBench from 17 fine-tuned models, 32 datasets, and more than 5,000 (model, dataset, failure) combinations.
The authors report that the question “Suggest a type of amphibian” makes GPT-5.6 Luna answer “frog” 70-95% of the time after benchmarks but 12-38% after real use.
The post argues that continually learning AI systems can learn to evade blocking monitors that replace suspicious actions with weaker trusted-model actions.
WorkspaceBench includes 3,356 questions across 27 evaluation families, plus a hallucination-focused benchmark for activation-to-text tools.
Google Research describes a generative AI system for coherent long-form video generation.
Microsoft Research says offloading inference beyond the robot improves task success and efficiency for physical AI workloads.
vLLM uses distribution-preserving Gumbel-max watermarking with GPU kernels, statistical detection, speculative decoding, and repeated-context safeguards.
Epoch AI says Huawei will improve its AI compute by 2030 but export controls limit the scaling levers needed to catch Nvidia.
The authors say models could do much more reasoning in latent states rather than in text chain-of-thought, reducing CoT’s usefulness for oversight.
The post describes mixing dense and sparse layers to scale biological foundation models more efficiently than dense transformers.
The post says a 512-GPU training run can fail even when every GPU, network link, and pod reports healthy.
SWE-Serve tests changes against live serving, where a patch can pass tests but fail after loading a real model and handling requests.
Commenters say portable SIMD is about 11% slower than non-portable SIMD in one wasm image-swap benchmark, and both are about 5x faster than no SIMD.
Commenters say the project has a roadmap for external auth, a git remote endpoint, and a workaround for a showstopper issue.
Commenters describe the agents as persistent and long-horizon, and one says the captcha-solver approach was eventually abandoned.
Commenters ask whether the project differs from rerankers and say the model can be copied by open source quickly.
Commenters say the product uses fake pen-drawing animations and streaming diagrams, and one notes file editing is not yet available.
Commenters argue signature AV no longer works well because attackers test against scanners and avoid known fingerprints.
Commenters say Typst can import TOML or JSON into templates and is used for badges, invoices, thumbnails, certificates, and reports.
Commenters say LLMs have helped with genealogical research and that SourceLibrary.org was built to be ergonomic for agents and people.
Commenters say the page is interactive and mentions dynamic documents, local-first work, and clickable/dragged elements.
Commenters report 401 Unauthorized errors in the desktop MacOS app, ChatGPT outages, and a status page that stayed green for more than 15 minutes.
The post describes loopjacking attacks against A2A systems that hijack human approval steps.
Commenters say the Dutch government is building a Microsoft alternative on NixOS, and one commenter notes France already has NixOS-based government projects including securix and bureautix-example.
Commenters say a background agent in Muse used an azure/muse-special model while building a website, and another commenter disputes that the evidence proves an OpenAI model.
The paper proposes AEWM, which models how reasoning and actions change task progress instead of simulating tool responses.
The method separates planning from synthesis with a Planner and Synthesizer that maintain an evolving summary state.
The paper tests whether coding agents can synthesize programs for generalized TAMP using simulator access and a fixed synthesis budget.
The recipe uses eight stages: SFT, reasoning RL, coding RL, instruction-following RL, general agent, coding agent, search agent, and RLHF.
The paper argues for a mandatory agent OS substrate for identity, input mediation, memory governance, and execution control.
The framework ties data production, model training, and iterative improvement into a closed loop for mobile planning.
The authors report that linearly combining inputs from distinct text streams produces a superposition of the individual next-token distributions.
WAA gives VLMs a visual action workspace with contact views and action rehearsal for robot manipulation.
The benchmark covers six tasks across 197 scenes, 2,972 question-answer pairs, and 900 navigation samples with FOA audio.
The authors build WROP, a dataset of 150 cognitive-science-inspired tasks across six categories for training object permanence.
This repository is an in-progress exploration of the Recurrent Transformer proposed by Costin-Andrei Oncescu and colleagues.
The episode says Meta’s Muse can act as a productivity tool and, with access to social media and email, can know nearly everything about a user.
The episode traces Runway from Gen-1 and Gen-2 to Gen-3 and Gen-4.5 and says video generation is being used as a path toward world simulation.
The talk says AI tools should be edited beyond the first draft and built around a clear point of view and user understanding.
The talk says Ramp uses AI agents to find customer problems, shape ideas, review and test code, coordinate launches, and fix UX issues.
The episode argues feeds can be up to 100x cheaper to run than chat because they decode pointers to existing content instead of generating every token.
The episode says Spotify uses NEO, a four-stage training recipe, and that its Large Taste Model reaches one in four US Premium subscribers daily.
The episode covers AI infrastructure, energy markets, nuclear power, and a $311 million fundraising round for Invetta.
The episode says Halcyon has supported about 30 organizations and that the bottleneck in AI safety, biosecurity, and cybersecurity is experienced founders.
The episode says OpenRouter routes requests for more than 10 million developers and traces the company back to the first wave of Llama, Alpaca, Mistral, and Midjourney.
The demo turns OpenAPI specs, Postman collections, or Python SDKs into an MCP server and validates the result with static, dynamic, and MCP checks.
The model uses ternary transformer weights, is about 9.3x smaller than FP16, and is said to retain 98.2% of FP16 intelligence.
Qwen says the model weights are in Hugging Face Transformers format and the hosted version will offer 1M context length and built-in tools.
The model is tagged for image-to-video, text-to-video, video-to-video, text-to-audio, and video-to-audio generation.
Nemotron 3 Diarization is an open-weight model for who-spoke-when labeling that supports streaming and offline inference for up to eight speakers.
Audio8 ASR Infinite uses a streaming architecture with selectable 80/120/160 ms clocks and unlimited-length transcription via a rolling KV cache.
Xiaomi says the checkpoint scales reinforcement learning across compute, environment diversity, and grader compute to extend the model frontier.
The dataset contains about 500,000 executable agent trajectories for tool use, search, code, and general agent instruction tuning.
The release contains 1,274 hours of egocentric household video paired with 24 Hz IMU arm tracking from ordinary chores.
This Space points to Hugging Face configuration reference only.
Laya is a multilingual, non-autoregressive decision model that returns typed answers and calibrated probabilities in one forward pass in about 33 ms across 100+ languages.
Xing4.0-29B-A4B has 29B total parameters, activates 4B per token, supports 256K context extensible to 512K, and was trained entirely on Ascend NPU with MindSpore.
ZDTaichu5.0-9B combines a Qwen3.5-9B language backbone with a C-RADIOv4-H vision encoder and supports text, images, videos, spatial reasoning, and agentic tool use.
Paperclip is a Node.js server and React UI for orchestrating teams of AI agents with org charts, budgets, governance, and goal alignment.
Hindsight is an agent memory system focused on learning over time rather than simple conversation recall, and it claims state-of-the-art long-term memory performance.
Univer is a customizable office SDK for spreadsheets, documents, presentations, bases, boards, and PDFs with canvas rendering and a formula engine.
StarNet is a local-first desktop harness where rooms, hallways, and placed objects map directly to capability-scoped teams and handoff lanes.
Model Optimizer combines quantization, pruning, NAS, distillation, speculative decoding, and sparsity for Hugging Face, PyTorch, or ONNX models.
The curriculum has 523 lessons, 20 phases, and about 342 hours, and every lesson ships a reusable artifact such as a prompt, skill, agent, or MCP server.
The directory lists internal and third-party Claude Code plugins and warns that Anthropic does not control included MCP servers or files.
Anthropic says skills are folders of instructions, scripts, and resources that Claude loads dynamically for repeatable specialized tasks.
The tutorial walks through manually bootstrapping a Kubernetes cluster instead of using a fully automated tool.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.