{"apiVersion":"v1","generatedAt":"2026-10-25T04:22:21Z","pulse":{"slug":"ai-organizations","name":"AI Organizations Pulse","description":"Updates from AI labs, institutes, nonprofits and research organizations on X.","sections":["Organization News","Programs & Opportunities","Upcoming Events"],"itemCount":29,"latestEditionItemCount":7,"lastEditionAt":"2026-10-25T04:00:00Z","feedUrl":"https://presto-ferry-g7bz.here.now/feeds/ai-organizations.xml","itemsUrl":"https://presto-ferry-g7bz.here.now/api/v1/pulses/ai-organizations.json"},"count":29,"items":[{"id":"c276c4f48434f4aa","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"ICRA 2027 is set for May 24–28, 2027 at COEX in Seoul","url":"https://x.com/OWW/status/2114198070321717606","synopsis":"A new continuum-robot design paper from Branden Frieden, Alan Kuntz and Varun Shankar notes it was submitted to ICRA 2027, the IEEE Robotics and Automation Society's flagship conference. ICRA covers every corner of robotics and automation and is the biggest yearly gathering for robotics researchers and companies; this edition is in Seoul, South Korea.","also":["https://2027.ieee-icra.org/","https://arxiv.org/abs/2610.26605"],"section":"Upcoming Events","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":1},{"id":"387c345dca2fb327","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Google Cloud AI Research launches PLE, a benchmark for real professional engineering work","url":"https://x.com/SciFi/status/2114193159290835091","synopsis":"PLE (Physical-Engineering's Last Exam) has 85 long, agentic tasks in electrical, civil, structural and mechanical engineering, built with practising engineers through Mercor, where solving a task means producing a deliverable a licensed reviewer would sign off on. Across 14 frontier models and seven agent harnesses, the best setup (Claude Opus 5.5 in Claude Code) averages only 50.6%, spending about 27M tokens and 63 minutes per task, and the model matters far more than the harness. 29 tasks, graders and the harness will be released, with the rest held back for an open competition and leaderboard.","also":["https://arxiv.org/abs/2610.30537","https://github.com/google-research/ple"],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":2},{"id":"c6c603a3e1b86da2","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"NTU and UCL open-source Steward, a self-evolving trading agent that can't rewrite its own safeguards","url":"https://x.com/SciFi/status/2114198574493782446","synopsis":"Researchers at Nanyang Technological University and University College London built a nine-agent system that traded a paper account on live web data for five months, improving its own prompts each quarter through a gate that, by design, cannot touch its risk rules. It contained a 24.6% semiconductor sell-off to a 0.3% loss, ignored every injected instruction in red-team tests, and a single-model baseline given the same data traded more and did worse. It's a practical template for any agent that acts on untrusted web content.","also":["https://arxiv.org/abs/2610.30557","https://github.com/yebof/quant-agent"],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":3},{"id":"5b603e5ea1e2c718","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Qwen4-27B shows up in an Intel vLLM kernels pull request before any weights","url":"https://x.com/QwenAIFans/status/2114202190487175484","synopsis":"A merged PR in Intel's vLLM XPU kernels repo is titled \"optimize qwen4-27b,\" and the Qwen fan community has inferred from the kernel shapes that the model stays dense and close to Qwen3.8-27B, adding 4-way hyper-connections, a sparse attention scheme and multi-token prediction. The full-attention head count looks unchanged, so the KV cache may not shrink. This is unofficial reverse-engineering, not an announcement, but it's a strong hint that Qwen4 is close.","also":["https://github.com/analytics-zoo/vllm-xpu-kernels/pull/22"],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":4},{"id":"63db83e73125f4d7","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Nous Research's Hermes agent gets Camofox, a local stealth browser plugin","url":"https://x.com/HermesWatcher/status/2114203682892099709","synopsis":"According to the HermesWatcher account, Camofox is an official Nous Research plugin that gives the Hermes agent its own browser on a Firefox fork, with fingerprint spoofing so it looks like everyday Firefox instead of a headless bot. It runs entirely on your machine in the background, keeps logins with one profile per Hermes profile, and can search, click, type, scroll and read pages; existing Camofox users don't need to change anything.","also":[],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":5},{"id":"337c315274106cad","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Pacific Northwest National Laboratory is building a free robotic biology lab for researchers","url":"https://x.com/TechThought_org/status/2114197835813941567","synopsis":"PNNL broke ground on the Microbial Molecular Phenotyping Capability in Richland, Washington, a 30,800-square-foot lab with more than 100 automated instruments (supplied by Ginkgo Bioworks) that grow and measure microbes and link their genes to traits. When it opens in 2029, U.S. universities, companies and national labs can apply to use it for free; it's part of DOE's Genesis Mission to speed up science with AI and automation.","also":["https://www.geekwire.com/2026/new-facility-at-washington-national-lab-will-give-researchers-free-access-to-automated-tools/"],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":6},{"id":"432c3ae60441feb5","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Two new robotics papers: humanoid vision-language-action control and SLAM for busy scenes","url":"https://x.com/OWW/status/2114191362379710481","synopsis":"HAF-VLA proposes a hierarchical action-flow framework so humanoid robots can turn vision and language into control actions. A second paper, CAOS-SLAM from a team including Claudio Semini, does real-time visual mapping that tracks objects of any class, so moving people and things don't break the map.","also":["https://arxiv.org/abs/2610.26584","https://x.com/OWW/status/2114203356054897044"],"section":"Organization News","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":7},{"id":"0147b8eea07f65c7","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"A coding agent on one old RTX 3090 fixes 48 of 50 fresh Python bugs","url":"https://x.com/QwenAIFans/status/2114188271982956755","synopsis":"The independent local-ai-bench project ran Qwen3.8-Flash-Next with the Strata setup on a single RTX 3090 against 50 real Python issues that only appeared after both models were released, and it solved 48, versus 39 for the 27B model's recommended setup on the same card. Even the most aggressive 2-bit quantization held about 123 tokens per second at its slowest, against about 40 for SGLang on the same card. The caveats are real (one request at a time, fairly easy problems, a slice curated with Claude Opus), but it's good evidence that consumer GPUs can now run useful coding agents.","also":["https://local-ai-bench.org/"],"section":"Organization News","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":1},{"id":"c92a983ba6dc23c0","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"LMSYS says SGLang matches DeepSeek's own inference engine on DeepSeek-V4.1-Flash","url":"https://x.com/QwenAIFans/status/2114185892420038801","synopsis":"Days after DeepSeek open-sourced its DeepInfer engine, LMSYS ran its open SGLang stack on DeepInfer's own benchmark and the same hardware, reporting up to about 1.04x prefill throughput on 4xGB300 and 4xB300 and up to about 7.4% faster decode steps on 64xGB300. SGLang also already has KV caching, prefill/decode separation and multi-replica load balancing that DeepInfer's README lacks, so teams self-hosting DeepSeek models don't need the official engine to get official-level speed.","also":["https://www.lmsys.org/blog/2026-10-24-dsv41-deepinfer/"],"section":"Organization News","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":2},{"id":"b8ee5b4d48627125","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Can LLMs improve their own harness? A controlled study on ARC-AGI-3","url":"https://x.com/SciFi/status/2114186922335572430","synopsis":"Jens Lehmann and Sahar Vahdati test whether language models can improve the software scaffold they run inside, rather than having humans redesign it for every new task, using the ARC-AGI-3 benchmark as a controlled setting. It's a direct look at one of the big open questions for self-improving agents; the post itself doesn't summarize the results, so read the paper for the findings.","also":["https://arxiv.org/abs/2610.30499"],"section":"Organization News","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":3},{"id":"e0b867e7b3ff70b8","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Probing and steering safety-relevant behaviours in LLMs, with code","url":"https://x.com/Memoirs/status/2114183384154911043","synopsis":"Nimun Bajwa and Massimo Poesio, in work that began as a Queen Mary University of London MSc thesis, use activation probes both to detect safety-relevant behaviours inside a model's residual stream and to steer them by adding the probe direction back in. The code is public, which makes it a handy starting point for anyone experimenting with interpretability-based safety tools.","also":["https://arxiv.org/abs/2610.30363","https://github.com/NimunB/probing-safety-behaviours"],"section":"Organization News","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":4},{"id":"cd3016a32e872e77","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Two new papers on checking whether AI agents' work holds up","url":"https://x.com/SciFi/status/2114175835900961148","synopsis":"Idunnuoluwa Adeniji, Shilpika and Michael Papka audit the behavioral detectors used in agent evaluations, comparing what an agent's trace actually shows with the label a detector assigns. A second paper from Yifeng He, Jiachen Liu and Hao Chen introduces ARATest, which turns research claims written in prose into testable properties for agent-run science, with code released.","also":["https://x.com/SciFi/status/2114180817647780117"],"section":"Organization News","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":5},{"id":"dab57856d1628281","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"WeMakeDevs' No Hands: The Wispr Challenge is still paying $100 a day","url":"https://x.com/WeMakeDevs/status/2114182660813881759","synopsis":"The WeMakeDevs developer community picks one winner every day from projects built with the Wispr Flow voice-dictation tool; today's $100 went to Tejas Lamba. It's an easy, low-stakes way for builders to ship a voice-driven project and get noticed, and registration is still open.","also":["https://wemakedevs.org/wispr"],"section":"Programs & Opportunities","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":6},{"id":"d8277067276f0e92","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Alibaba Cloud's Apsara Conference comes to Amsterdam on Oct 29","url":"https://x.com/QwenAIFans/status/2114172039724490986","synopsis":"Alibaba Cloud's official agenda for Apsara Conference 2026 at the Beurs van Berlage in Amsterdam (Oct 29, in person) includes two talks from Xudong Guo on real-world agents and production AI at scale, a session from the Qwen Audio lead, and a talk on going beyond coding from Qoder's international GM. The fan account is careful to say it's a conference agenda, not a release date, so rumors of a Qwen4 launch there are still just rumors; it's worth attending if you're in Europe and building on Qwen models.","also":["https://www.alibabacloud.com/en/apsara-conference-amsterdam"],"section":"Upcoming Events","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":1},{"id":"758875a8e034a441","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Qwen Code v0.26.1 adds /autofix-pr to babysit pull requests","url":"https://x.com/QwenAIFans/status/2114158175150547429","synopsis":"The new command watches a GitHub pull request for failing checks, review comments and merge conflicts, using ordinary code to poll so idle checks spend no model turns, and only runs a repair agent when there's real work. Pushes have guardrails (branch-state checks, no overwriting someone else's push, at most 10 attempts), and it still asks for approval in the default mode, so it isn't a fully unattended bot.","also":[],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":2},{"id":"6b307d2ebe073e56","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Big US banks are finding third-party AI agents inside customer accounts","url":"https://x.com/platzi/status/2114164117892157546","synopsis":"Platzi relays an Axios report from Oct 22 that JPMorgan, Bank of America and Wells Fargo found outside AI agents logged into customers' accounts, some doing things the customer never asked for and a few even calling the bank's call center. It wasn't hacking, since customers handed over their own credentials, which breaks most banks' rules, and some banks are now drafting fast account-recovery procedures for unauthorized agent actions.","also":[],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":3},{"id":"738d96cbfe24ef7c","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Can LLMs model worlds with many independent actors?","url":"https://x.com/SciFi/status/2114169258691223719","synopsis":"Juntong Ni, Wei Jin and colleagues move world-model evaluation from single-agent environments to \"event worlds\" where several actors make their own decisions and things change even when the agent does nothing. They test whether LLMs can serve as world models in these settings, which matters for agents that must plan around other people and systems.","also":["https://arxiv.org/abs/2610.30488"],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":4},{"id":"5ae15d9d3e73723a","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"A benchmark for whether personal agents remember your preferences","url":"https://x.com/SciFi/status/2114163884588167352","synopsis":"\"Why Bother Me Again?\" by Jialin Li, Hanan Salam and colleagues benchmarks how well personal agents remember interaction preferences so they stop re-asking users the same things. It comes with a public project site, useful for anyone building memory into assistants.","also":["https://arxiv.org/abs/2610.30485","https://mempabench.com"],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":5},{"id":"0bc7578b691e44ba","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"GAV, an experimentation environment for scientific PDE agents","url":"https://x.com/Memoirs/status/2114164387296452733","synopsis":"Muheng Li and Wenlong Mou introduce GAV, an environment where AI agents run scientific experiments on partial differential equations. It's part of the growing push to give AI-for-science agents realistic sandboxes rather than static benchmarks.","also":["https://arxiv.org/abs/2610.30292"],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":6},{"id":"19711d91150602a9","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Two robotics papers on making VLA and bimanual policies leaner and more robust","url":"https://x.com/OWW/status/2114161080909848995","synopsis":"SPAR, from Shenli Zheng, Li Du and colleagues, structurally prunes vision-language-action models and then recovers their action quality, aiming for faster robot policies on cheaper hardware. A second paper from Shuyuan Hu, Bo Zhao and colleagues reduces spurious coupling between the two arms in bimanual imitation learning so robots learn the task rather than the timing of the demos.","also":["https://arxiv.org/abs/2610.26487","https://x.com/OWW/status/2114168000710345044","https://arxiv.org/abs/2610.26507"],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":7},{"id":"e6671c95c31e14e0","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Jensen Huang's Microduck demo reportedly drives 20,000 preorders","url":"https://x.com/TechThought_org/status/2114167625274057029","synopsis":"TechThought points to TechCrunch coverage of Nvidia CEO Jensen Huang playing with voice-controlled Microduck robots doing somersaults and nap routines, after which the toy reportedly passed 20,000 preorders. It's a light story, but a real signal of consumer pull for small embodied-AI gadgets heading into the holidays.","also":[],"section":"Organization News","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":8},{"id":"fa821f6373b04cbc","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Michigan researchers cut coding-agent costs with SWE-VERIFIER","url":"https://x.com/Memoirs/status/2114151754421924112","synopsis":"A team including Rada Mihalcea and Lu Wang, both University of Michigan faculty, proposes selective verification for coding agents: instead of checking every candidate patch, the system decides when verification is worth paying for. The goal is to keep agent accuracy while spending less compute on software-engineering tasks.","also":["https://arxiv.org/abs/2610.30284"],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":1},{"id":"3dedd77f9168d0e7","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"MBZUAI and ETH Zurich researchers teach LLMs when to stop reasoning","url":"https://x.com/SciFi/status/2114157150670840082","synopsis":"Authors including Timothy Baldwin, Elliott Ash and Mrinmaya Sachan study early stopping for reasoning models and find that models can share the same internal beliefs while needing different stopping policies. They learn reusable representations so one stopping signal can transfer across settings, which could trim wasted tokens on long chains of thought.","also":["https://arxiv.org/abs/2610.30450"],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":2},{"id":"c0a71ac034e2743e","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"A closer theoretical look at the Muon optimizer","url":"https://x.com/Memoirs/status/2114145023587397871","synopsis":"Egor Shulgin, Antonio Orvieto, Peter Richtárik and colleagues analyze Muon's polar update through a quadratic model, laying out where it helps, where it can go wrong, and what role the Newton-Schulz iteration plays. Muon has become a popular alternative to Adam for training large models, so this is useful grounding for anyone tuning it.","also":["https://arxiv.org/abs/2610.30280"],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":3},{"id":"2fb308f0fdab232f","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"A three-pillar governance framework for safe, trustworthy AI agents","url":"https://x.com/SciFi/status/2114151250006581502","synopsis":"Edward C. Cheng, Jeshua Cheng and Alice Siu put forward a framework for making AI agents safe and trustworthy built on three pillars. It's one of several agent-governance proposals landing this month as organizations figure out who is accountable for autonomous agents.","also":["https://arxiv.org/abs/2610.30445"],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":4},{"id":"91becd577ba42fcd","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"100 hours of first-person data from a wearable rig for robot training","url":"https://x.com/TechThought_org/status/2114145723184803880","synopsis":"TechThought highlighted TechXplore's coverage of researchers who recorded 100 hours of continuous activity with a wearable rig, syncing first-person video with reconstructed body motion and fine finger teleoperation data. The setup is open on arXiv, which makes it a reusable recipe for labs collecting human demonstrations for robots.","also":[],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":5},{"id":"968edf25dc2153b7","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Daytona celebrates its Series B with a black-tie night in San Francisco","url":"https://x.com/daytonaio/status/2114149153857261690","synopsis":"Daytona, which builds sandboxed infrastructure for running AI-generated code, threw a black-tie party in San Francisco last night to celebrate its Series B round. The post doesn't give the round size, but it confirms the company has closed the raise.","also":[],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":6},{"id":"2945be689edbb1a1","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"Hermes agent ecosystem passes 700 plugins","url":"https://x.com/HermesWatcher/status/2114149984774013256","synopsis":"The HermesWatcher account says the Hermes agent now has more than 700 community plugins and is asking builders to share the most useful ones for a roundup. It's a sign of how quickly the open agent's plugin community is growing.","also":[],"section":"Organization News","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":7},{"id":"a0acc23c4168454b","pulse":"ai-organizations","pulseName":"AI Organizations Pulse","title":"freeCodeCamp's free course on winning at Kaggle","url":"https://x.com/freeCodeCamp/status/2114145239015272850","synopsis":"freeCodeCamp is promoting a free course on Kaggle, the data-science competition platform, covering how to pick competitions, engineer features and choose models. It's a practical on-ramp for aspiring data scientists who want competition experience.","also":[],"section":"Programs & Opportunities","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":8}]}