Aligned News — AI Intelligence
Articles about Aligned News — AI Intelligence
Humanoid Hack Tokyo 3 — GMOヒューマノイドラボ こけら落とし · Luma
October 31 through November 3, 2026 at GMO Humanoid Lab in Shibuya, with two intervening online days. Teams of five must apply by October 18 at 23:59 JST; participation is free and selected teams are announced October 19.
Physical AI Hack · Luma
Physical AI Hack is listed for November 28, 2026 in San Francisco, with robotics, perception, simulation, dexterity, autonomy, and world-model tracks. Attendance requires application and approval.
Open Agent Hackathon 2026
October 15–20, 2026, fully online, with registration closing October 13 at 00:00 UTC. GenAI.Works offers teams of one to five builders up to $20,000 across tracks for fragmented intelligence, explainable research, reasoning architecture, and a sponsor-technology wildcard.
AI Research Hackathon · Luma
SF AI Research Hackathon November 7–8, 2026 · San Francisco Over 48 hours, teams will push AI to do research on its own: systems that improve themselves, run
AI Agents Hackathon 2027
January 20–21, 2027 at Digital Hub Logistics & Commerce in Hamburg. Cargonerds lists in-person tickets and two full build days focused on freight, trade and commerce.
Agentic AI Hackathon: AWS Community Day Bay Area 2026 · Luma
Remote agent-building challenge with October 13, 2026 finals at AWS Builder Loft in San Francisco. Organizer pages disagree on registration, build-start and submission times; confirm these deadlines with the host.
Manus Academy
The Universal Hackathon is listed for October 19, 2026, at 4:30 p.m. in San Francisco. Teams are challenged to demonstrate a business operated by cooperating agents with distinct responsibilities.
BUILD FOR HER · Luma
October 10, 2026, 9 a.m.–7:30 p.m. in San Francisco. RoamHer plans 30–40 builders across six to eight teams; builder, mentor and judge registration now shows sales ended.
Kane CLI Hack Day, supported by OpenAI · Luma
TestMu AI is hosting the October 24, 2026 event with OpenAI support. Builders can apply individually or in teams of up to four; registration requires host approval.
The Recursive Agents Hackathon – Cerebral Valley
Saturday, Oct 17, 2026 · Shack15, Ferry Building, Suite 201, San Francisco, CA 94111, USA: At the PyTorch Conference hackathon, teams will build persistent AI
SF Hacks 2027
February 19–21, 2027 at San Francisco State University. The official site expects more than 400 student builders.
Google - The Gemma 4 Developer Agent Competition
Entries close November 25, 2026. Kaggle lists $65,000 for the main Gemma 4 Developer Agent Competition and $35,000 for its companion paper track, a combined $100,000.
DeveloperWeek 2027 runs a hybrid challenge hackathon with Santa Clara finals
Online building runs January 25–February 11, 2027, with in-person activity February 9–11 at Santa Clara Convention Center.
EMBEDDED SF: CraftifAI Physical AI Hackathon · Luma
EMBEDDED SF runs October 13, 2026 in San Francisco, with ESP32 hardware, sensors, cloud GPUs, and two tracks for firmware and computer-vision deployment. Registration requires organizer approval.
Anything Applied Agentic Hackathon - hackathons.team · Luma
🤖 ANYTHING APPLIED AGENTIC HACKATHON — SAN FRANCISCO One day. Build an agent that does real work. A one-day, in-person hackathon for people building agentic
AI Hack for Freedom IV - San Francisco, CA
AI Hack for Freedom IV runs December 2–4, 2026 in San Francisco. Applications are open for developers and activists building freedom, privacy, and decentralized-technology projects around a $25,000 BTC prize pool.
“Build an Agent” Hackathon · Luma
October 7, 2026, 2–5:30 p.m. at Corgi Cafe in San Francisco. The official page lists a short build sprint, finalist demos and sponsor-specific tool requirements.
Build It Anyway: Signal Hunt with Apify | #SFTechWeek | Partiful
No code, no problem. Build an AI agent that finds what matters on the web. One day, zero code, and you leave with a working AI agent that does the boring part of your job for you. 🙋 IS THIS YOU? A founder juggling first sales leads while keeping an eye on competitors In sales, stuck scrolling LinkedIn by hand looking for buying signals In GTM and want warm leads, not another cold list A marketer trying to keep up with competitor posts and rising creators
Vultr: Agent Rush Hackathon AI Hackathon | lablab.ai
Online building runs November 3–8, 2026, with an approved on-site phase in Salt Lake City on November 8. The organizer lists $9,000 cash and $5,000 in credits.
robot hackathon 2026 | robo.innovate
Participate in Hackathon 2026, a premier event for students, researchers, makers, and innovators. Organized by robo.innovate, the leading Bavarian hub for robotics, this event offers an exceptional platform for collaboration and innovation in the robotics field.
ExecuTorch Hackathon · Luma
October 17–18, 2026 at Founders, Inc., 2 Marina Boulevard B300, San Francisco. Teams of three to five must submit proposals by October 4; registration is an application, not guaranteed admission.
The Build: Hackathon · Luma
October 14, 2026 in San Francisco, starting at 9:30 a.m. The official page conflicts on the end time—5:30 p.m. in its event metadata and 7 p.m. in the description—so accepted builders should confirm the schedule.
AMD Developer Hackathon: ACT III AI Hackathon | lablab.ai
Online building runs October 12–18, 2026, with approved in-person teams in Rome, Milan and Imperia on October 17–18. The official page lists more than $11,000 in prizes.
ODSC AI x NEBIUS | Hackathon · Luma
October 28, 2026 · 9:00 AM–7:30 PM PT · Burlingame, CA Free to apply · Limited spots Bring an idea. Spend the day building it with other AI builders, Nebius
Build Personal AI with Coding Agents - #SF TechWeek · Luma
October 7, 2026, noon–6 p.m. at Sports Basement Bryant Street in San Francisco. Builders move from a coding-agent demonstration to a two-hour prototype sprint, live demos and a founder-investor discussion.
Hack.dev opens cloud, local and physical AI tracks for its October build
October 10–11, 2026 at UCSC Silicon Valley in Santa Clara. The event is free, but its published September 30 final application deadline has passed; the live site still says 'now accepting,' so late applicants should confirm eligibility.
Build AI agents that own their inference - #SFTechWeek · Luma
The latest official page lists October 12, 10 a.m. to 7 p.m. Pacific, at AWS Builder Loft, 525 Market Street. Host approval is required; attendees must be at least 18 and bring physical government-issued photo ID.
Fast Forward · Luma
Fast Forward takes place October 14 in San Francisco, with a 10 a.m. to 6 p.m. program. Registration requires host approval, and the exact venue is disclosed after registration.
Hack for Social Impact brings civic technology builders to San Francisco
October 24–25, 2026 at Wefunder HQ, 1885 Mission Street, San Francisco. Registration requires host approval for the two-day event, which expects more than 200 builders.
Mistral x Wivoo Builder Night · Luma
October 7, 2026, 6:30–11 p.m. at Wivoo HQ, 20 Avenue Franklin Delano Roosevelt, Paris. Registration is active and approval-based for the two-hour agent-prototype challenge.
NousCon 2026 · Luma
Nous Research announced NousCon for October 30, 2026 in New York City. Doors open at 6 p.m.; registration requires host approval and the exact address is shared through registration.
Agentic + AI Observability Meetup SF · Luma
October 13, 2026, 5–8 p.m. at the Databricks San Francisco office. The meetup centers on observability, failure analysis and continuous improvement for production agent systems.
Open Infra Night for Agent Builders #Open Source AI Week · Luma
October 22, 2026, 5–8:30 p.m. in San Francisco, with the venue to be announced. The Open Source AI Week meetup covers open agent runtimes, tooling and model infrastructure.
Platform Engineering Meetup: Agents broke the balance (& how to fix it) · Luma
October 27, 2026, 5–7 p.m. at 130 Sutter Street, seventh floor, San Francisco. Platform engineers will compare delivery controls for code generated faster than teams can validate it.
AGENTIC AI IN ENGINEERING · Luma
October 15, 2026, 10 a.m.–5 p.m. in San Francisco. Medallion XLN’s in-person event explores how agents participate across software, product and infrastructure engineering.
Daytona AI Builders - NYC, October 2026 · Luma
October 28, 2026, 5:30–8:30 p.m. Eastern in New York City. Daytona hosts an approval-required meetup on forkable sandboxes, infrastructure and agent credential handling with AWS, Antimetal and Infisical participants.
AgentCamp: An AI Agent Unconference · Luma
November 2, 2026, 2:30–7 p.m. at Snowflake Silicon Valley AI Hub, 135 Constitution Drive, Menlo Park. TechEquity AI’s limited-capacity unconference uses participant-proposed breakout sessions.
TechEquity Ai Summit 2026: Talks + Workshops + Hackathon · Luma
November 3–6, 2026 at 135 Constitution Drive in Menlo Park, with in-person and online access. The November 4 track concentrates on data, agentic AI and open source.
PortCon · Luma
November 17–18, 2026 at Shack15 in San Francisco. Port’s two-day event focuses on infrastructure, operating models and delivery patterns for agentic software engineering.
AI Engineer Code Summit 2026: AI Engineering Conference, Nov 10-12, San Francisco
November 10–12, 2026 at Hilton Union Square, 333 O'Farrell Street, San Francisco. Attendance is application-only, and the speaker program remains in development.
Durable AI (October Edition) · Luma
October 21, 2026, 5:30–8 p.m. at 500 Sansome Street in San Francisco. Temporal, Kernel and Oodle.ai will demonstrate reliability and observability patterns for long-running agents.
Extraordinary Founders Dinner (hosted by Andrew Yeung) · Luma
October 15, 2026, 6:30–9 p.m. in a private San Francisco dining room. Andrew Yeung's official page lists sixteen places and requires an application and host approval.
RSVP to AI Insiders: Healthtech Dinner - #SFTechWeek | Partiful
An intimate dinner at the intersection of AI, healthcare, and biotech. AI Insiders brings together founders, researchers, and investors for real conversation and meaningful collaboration. No hype, no panels for show, no pitch theatre. Just a vetted table of leaders shaping what comes next. This is the dinner edition of AI Insiders: Healthcare & Biotech. Twelve to fifteen seats, one long table, one conversation everyone is part of. The evening 6:00 PM - Arrival and welcome drinks 6:30 PM - Dinner served, welcome words from our host
TechCrunch Disrupt 2026 brings startup builders and investors to San Francisco
TechCrunch Disrupt runs October 13–15, 2026 in San Francisco. The official event page centers founders, investors, product launches, startup lessons, and the Startup Battlefield competition.
AgenticCon brings enterprise AI platform teams to San Francisco
February 25, 2027 in San Francisco. The one-day conference combines platform architecture, organizational design, hands-on workshops, roundtables and enterprise matchmaking.
Agentics Day: MCP + Agents | KubeCon + CloudNativeCon Europe
Agentics Day: MCP + Agents event is a community-driven event dedicated to advancing the Model Context Protocol, an emerging standard for connecting AI models with external tools, data, and workflows.
AI Builders Summit - #SFTechWeek · Luma
October 9, 2026, 6–9 p.m. in San Francisco. Founders Bay’s in-person summit combines technical talks, product discussion and networking for AI builders and investors.
case 2027 – Agentic Software Engineering Conference, Berlin
February 24–25, 2027 at Hotel AMANO Grand Central in Berlin. The first day is a conference on agent collaboration; the second adds a full-day technical workshop.
ModernPath demonstrates an operating layer for agent-run software delivery
October 8, 2026, 3–4 p.m. EEST online. ModernPath’s live webinar demonstrates an Agentic Engineering OS on a brownfield codebase and includes an open question period.
TRUST 2027 - ACM EIGTRUST
March 7–9, 2027 in Washington, D.C., with full virtual participation available. The ACM event covers security, resilience, explainability, fairness and responsible deployment.
NVIDIA GTC 2027 Is Coming to San Jose March 15–18
March 15–18, 2027 in San Jose. NVIDIA's current official page says registration is not yet open and highlights agentic AI, AI infrastructure, AI for science, industrial AI, cybersecurity, inference, robotics and physical AI.
AI Builders London - Nov 10 · Luma
November 10, 2026, 6–9 p.m. at Civo Tech Junction in London. The in-person meetup combines short engineering demonstrations, an open panel and informal networking.
ICRCA 2027 | Robotics, Control and Automation
March 26–28, 2027 at the University of Aizu in Japan. Paper submissions are due November 10, 2026, with presentation-only abstracts also accepted.
Mindstone Lisbon October AI Meetup · Luma
Welcome to the biggest Practical AI Meetup in Lisbon! Join us once a month as we explore the world of artificial intelligence, its cutting-edge practical
Agentic AI Meet-Up #11 Thurs 8th October · Luma
October 8, 2026, 6–8 p.m. at St Ethelburga's Centre for Reconciliation and Peace in London. The meetup features BlackRock engineers and Pavo AI's Rishabh Mehrotra.
Agentic Business Forum gives Paris decision-makers a one-day execution roadmap
March 31, 2027 at Maison.A Verso by Victoire in Paris. The forum runs from 9 a.m. to 6 p.m. with keynotes, masterclasses, roadmap studios, demonstrations and networking.
AI Builders Lisbon - Oct 28 · Luma
Join our informal meetup for the biggest AI nerds! We share the latest insights about: AI dev tools, Agent frameworks, RAG, vibe coding, OpenClaw, Claude Code
MIT AI Conference 2026 — AI: The Age of Agency
October 17, 2026 at the Computer History Museum in Mountain View. The in-person conference lists more than 50 speakers, workshops, demonstrations, an AI challenge and a student hackathon showcase.
HRI 2027
March 8–12, 2027 in Santa Clara, California. The conference combines interdisciplinary HRI research with an industry day, workshops, demonstrations and student participation tracks.
Paris AI Agent Summit - March 9 & 10, 2027
March 9–10, 2027 at Espace Saint Martin in Paris. The organizer expects more than 1,200 attendees, 80 speakers and a program of case studies, workshops and networking.
Mike Dereviannykh (@Mishok2000) on X
SemanTok proposes predictable semantic tokens for autoregressive video generation and reports strong semantic and fidelity results at comparatively small model sizes. The announcement does not expose enough methodology to independently verify the scale comparisons.
Zora Wang (@ZhiruoW) on X
Check out our CollabSkill oral presentation and poster on Tuesday at @COLM_conf! A framework for evaluating agent capabilities and contributions in human-agent collaboration. With tons of insights on AI literacy on the human side. - Oral spotlight at 10:15 - Poster session from 11- 1 in Room Impe
Jiacheng Guo (@JiachengGu50887) on X
We revisit GRPO, SC and BPO through coordinate transformations and local approximations to policy improvement. This view helps explain how clipping can distort updates under training–inference mismatch, and leads to Decoupled Score Centering. Blog: https://t.co/SmYvoG7hMr 1/9
Ziyang Wang✈️COLM26 (@ZiyangW00) on X
I’ll be at #COLM2026 this week presenting EgoMemReason! Our benchmark asks how well multimodal agents can build and use memory to reason across a week of everyday life-tracking object changes, connecting past events, and inferring recurring patterns. 📍 Poster #71, Imperial Ballroom 🗓️ Tuesday,
Andres Guadamuz (@technollama) on X
Our article entitled ‘Preserving balance in the EU digital single market: How Like Company could reframe copyright and innovation in the generative AI era’ is now published in the Journal of World Intellectual Property. https://t.co/XPrxkld75x
Honam Wong (@MH2023ML) on X
My first PhD paper is finally on arXiv! https://t.co/Koj5YRg3Bo We introduce a low-rank router for Mixture-of-Experts models, with a hardware-aware implementation, built for scaling to a very large number of experts (2048, 4096, and beyond). My second paper, on the Edge of Stability and learning
Arky Yang (@arkyyang) on X
“Pincer: Resource Authorization for Agents using a Digital Twin” — new paper (arXiv, Oct 1 2026). https://t.co/lYzjukyUby In plain terms: coding agents today keep asking you “can I run this?” until you stop reading the prompts and click yes (permission fatigue), and the access you grant piles up a
Seungone Kim (@seungonekim) on X
Our autobenchmark blog post from @AIatMeta is released! For creating a “high-quality” AI agent benchmark, (1) forming a meaningful target, and (2) handling the intricate details on how to instantiate it matters. Our main finding is that the role of human matters on both!
Ricardo Vinuesa (@ricardovinuesa) on X
HydroGym spans laminar through three-dimensional turbulent flow control, including Reynolds numbers up to 400,000. A zero-shot transfer demonstration cuts local wing skin friction by 38% while reducing exploration cost by four orders of magnitude.
Michael Hahn (@mhahn29) on X
The paper characterizes exactly which regular languages fall inside C-RASP and supplies a polynomial-time decision procedure from a language’s algebraic structure. Experiments show the criterion predicts out-of-length behavior better than earlier finite-state classifications.
Cohere (@cohere) on X
Introducing North 2. Enterprises have been compromising on agentic AI for years: platforms that aren’t scalable, models that aren’t sovereign, control without enterprise-grade security. That ends today, with 15+ new features in our biggest upgrade yet:
Sarvam (@SarvamAI) on X
Sarvam made Kivi available on macOS as a voice interface for work across applications. It uses Saaras V4, Sarvam’s speech-recognition model for 22 Indian languages, code-switching and multiple transcript formats.
patrick (@patrickamadeus_) on X
🤖🤚 Give a video generation model a scene and a goal. Can it generate egocentric execution to get the job done? We tested 110 real-world scenarios on 6 SoTA models to see how they perform. Introducing Ego2Act, a goal-directed video generation benchmark 👇 🌐 https://t.co/9P8gex6R5s 🧪 https://t.
Jayden Davis (@JaydenDavisNC) on X
Fable 5.5 made me a delightful animation in blender. This model is quite something, no this is not promo video from OpenAI. All the music, models, animation, and rigging was done using Fable 5.5.
Matt Pocock (@mattpocockuk) on X
Prompt of the day: /retro take a look at as many GitHub PR review comments as you have access to from my team. Create suggestions for updates to CODING_STANDARDS.md (breaking into multiple files as needed). Makes your PR comments inform the code going forward.
JimiLonbo (@jimmy_longbow_) on X
Nine agents, nine cheaters. CAIS’s CheatBench parks a cheatable answer clue next to each hard task — every agent cheats somewhere, Opus 5.5 at 11.2%, Grok 4.7 at 77.9%. The polite fix backfires: a ‘do your own work’ nudge pushes Gemini 3.8 Flash UP to 79.6% (paper’s Table 6).
DHH (@dhh) on X
Keep your mind flexible! Rust has an edge at the moment, but Rauch is spot on: We should rethink the entire stack from first principles given the transition to agent-led programming. Very little of what was will remain that way.
Mark Kretschmann (@mark_k) on X
Today is day 1 of @OpenAI’s “28 Days of Improvements” (or resets). I wanna hear your guess what the first improvement could be! I’m thinking we might finally see GPT-6.1 Sol in ChatGPT. Let’s hear your guesses 👇
Yan Chernikov (@TheCherno) on X
Last week I got Opus 5.5 and GPT-6 Astra to separately build me a 3D game engine in C++ from ONE prompt - a short markdown doc with some basic requirements. Opus 5.5 worked for 36 HOURS STRAIGHT and produced not just a fully functional 3D game engine, but TWO great quality games to go along with i
wang (@wangzhr4) on X
BREAKING: New coding and software engineering philosophy revealed: - don’t read the fucking code - you have unlimited tokens - blame on the language Holy shit, the AI elites are so freaking busy in devastating the whole software industry.
Dean W. Ball (@deanwball) on X
A very senior AI industry figure once started a conversation by asking me, basically, “under what circumstances do you think we are essentially just screwed?” and I replied with the scenario below.
Lukas Ziegler (@lukas_m_ziegler) on X
🚨 BREAKING: Munich robotics startup @Robco_Robotics has just become a unicorn, crossing a $1 billion valuation in just nine months since its last round. Founded in 2020 by Roman Hölzl, Paul Maroldt and Constantin Dresel, RobCo built its name on modular robot arms that snap together like buildin
Pi (@pidotdev) on X
Welcome to our Monday Meditations! 🌞 Today @badlogicgames and @mitsuhiko are talking why we built Pi Durable, a harness build around a small task-based workflow engine so long-running, multiplayer agents can suspend and resume anywhere. Also: why not Temporal or Effect.ts.
Dongyang Fan (@dyfan22) on X
TASTE contains 92 pairwise proposal comparisons labeled through expert discussion and confidence filtering. Experienced researchers reach an estimated 77% accuracy; the best tested model, Fable 5, reaches 60% at maximum reasoning effort.
Chubby♨️ (@kimmonismus) on X
Even though I’ve recently written a lot of critical posts about OpenAI’s latest decisions, I’m still pleasantly surprised by how phenomenal the GPT models’ computer-use capabilities have become. It uses Safari just as if a human were actually controlling the browser. It hardly makes any mistakes a
Judith Dada (@DadaJudith) on X
Sorry, Seb, usually lots of love for your work, but how is Europe now having n=2 players in the LLM game a generational fumble for Mistral? Can we please just celebrate having more teams who try and build models, rather than immediately using this as a reason to hate on either? Europe‘s market is b
RobCo (@Robco_Robotics) on X
RobCo says a transaction combining new investment with employee liquidity lifted the autonomous-industrial-robotics company above a $1 billion valuation. Its next product is Alfie, a perception-and-reasoning robot scheduled for commercial launch at RobCoN on March 4, 2027.
Mario Zechner (@badlogicgames) on X
recommended viewing if you want to understand what pi durable is. it’s 100% accurate and breaks it down real well. i should learn how to do those ai explainers.
Rhys (@RhysSullivan) on X
I feel like a clean room reimplementation of photoshop should be actually completely possible? It’s an extremely verifiable task especially if Claude can compare the version it implements to real photoshop
himanshu (@himanshustwts) on X
Anthropic just played a smart move in India to sign B2B contracts with sovereign entities and institutions.
Ethan Mollick (@emollick) on X
I’ve been complaining about some OpenAI choices tonight, so, on the plus side, I will say there is still no equivalent to GPT-6 Pro (just as there wasn’t with previous Pro models). It both does really hard tasks in one shot & does a surprisingly good job communicating the results
Aleph Alpha (@Aleph__Alpha) on X
We built compliance first: following the EU AI Act and GDPR, with copyright safeguards beyond the Code of Practice. Our team is dedicated and fast: automated failure handling allowed us to move from first model to Kolibri in three months.
Pradeep Dasigi (@pdasigi) on X
On Terminal-Bench 2.1, automatic harness evolution does not consistently beat parallel sampling or sequential refinement and transfers poorly to held-out tasks. Long-horizon games are the exception, with task-specific gains of 80% on ARC-AGI-3 and 10.9% on EdgeBench.
PanteraOS (@opanteraos) on X
A economia de 23,9% do System-1 some quando o custo do pré-filtro entra na conta. Sobram 4,3%. Jiawei Li publicou nesta segunda, na lista nova do arXiv, o paper Fast Models, Slow Evidence. É uma eval pareada de dois modelos System-1 para o harness do agente: Laya, de pesos abertos, e Jev, API hosp
Rafal Kocielnik (@RKocielnik) on X
Heading to San Francisco to present our paper “Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior” at two @COLM_conf 2026 workshops: Agent Behavior (WAB) and AI Measurement Science (AIMS)! It was also recently accepted to the #NeurIPS2026 main conference. ❓ The
Cong Wei (@CongWei1230) on X
PixelUMM projects spatial image patches and spatiotemporal video tubelets directly into one multimodal backbone. Shared attention and task-specific Transformer parameters support text prediction and pixel-space flow matching in the same model.
Horace He (@cHHillee) on X
There’s an interesting distortion in inference pricing that appears to be driven by openrouter’s routing. Namely, ridiculously cheap prefill and very expensive decode. For GLM 5.3, https://t.co/07u3tv4AS4 (highest volume provider) charges $0.08/M input but $5.00/M output!(1/5)
Harrison Kinsley (@Sentdex) on X
In my opinion, the best model on the planet at the moment is GLM 5.3 Flash. It’s super fast, super smart for the size, and is within reach to run locally for many. Love working with GLM 5.3 Flash.
Alexandr Wang (@alexandr_wang) on X
10 years ago, as an intern i taught hrt execs how to use snapchat, and now their quarterly revenue is bigger than snap’s market cap current interns take note. if you want your company to do 2 trillion in quarterly revenue teach the execs to use muse
Theo - t3.gg (@theo) on X
July 2026: Anthropic has the best code models. Gap isn’t very big though. They’re slow, expensive, and the “claudeisms” are at an all time high. And your $200 sub is so limited that you can kill it in a day. Preferring OpenAI models made a lot of sense. You lose a slight edge on capability, but th
Craig Weiss (@craigweiss) on X
Craig Weiss said AI should challenge people rather than obey them without question. The post states a design principle, not a description of a particular model or feature.
DHH (@dhh) on X
DHH said he would often rather discuss developments with a frontier model while driving than listen to a technology podcast. The post describes a personal media habit and links conversational AI with autonomous driving.
An agent operator says unattended cron jobs triggered an OpenAI ban
Pliny the Liberator said automated agents using cron jobs apparently caused an OpenAI account ban while unattended. The post is a first-person anecdote, not an independently verified account of the policy violation.
Tibo (@thsottiaux) on X
Tibo set a 28-day cadence in which each day brings either a broadly relevant improvement for Codex and Work users or a full usage reset. The promise is an either-or commitment, not a claim that a product update will ship every day.
Ananth (@Ananth7e) on X
Ananth joked that the daily improvement-or-reset pledge encourages users to spend their Codex allowance and hope the day ends with a reset. The post comments on incentive design rather than announcing a policy change.
Gary Marcus (@GaryMarcus) on X
Gary Marcus argued that the AI boom avoids a bust only if it takes most people’s jobs. The statement is a provocative economic opinion, not a demonstrated forecast.
Nabeel S. Qureshi (@nabeelqu) on X
Nabeel Qureshi joked that OpenAI is run by X, Anthropic by Claude, Google by committee and SpaceXAI by Elon Musk. The lines are satire about perceived organizational culture, not factual governance descriptions.
JD | RoyalCities (@RoyalCities) on X
JD of RoyalCities mocked a new argument that gaming recompilations have a “soul” and that players should reject work not built through years of hand-written code. The post is commentary on AI backlash, not evidence about how any specific recompilation was produced.
Kit Langton (@kitlangton) on X
Kit Langton described using OpenCode to create a motion-graphic explainer for every non-trivial pull request. The post presents an individual documentation workflow rather than a measured claim about review quality.
Guillermo Rauch (@rauchg) on X
Working on a new little project. The 𝚁𝙴𝙰𝙳𝙼𝙴 is fully written by hand, because it’s for human consumption. The documentation internals are AI English, because they’re for agents. This little rule of thumb can make the world better. Blogs, tweets, READMEs: human communication. It’s not even ab
Mustafa. (@oprydai) on X
pid control is one of those ideas that looks like an equation until you watch a robot actually move. imagine commanding a robotic arm to follow a trajectory. the controller constantly asks one question: where should the joint be, and where is it right now? the difference is the error. proportional
Jing Yu Koh ✈️ COLM′26 (@kohjingyu) on X
Computer-use agents are extremely good! But they need to improve on a metric which most benchmarks don’t track. How much time do CUAs take to complete a task? Introducing cua-speedrun: a benchmark for accuracy, speed, and cost under matched infra. https://t.co/JQKNrkGKHs 👇🧵
Zhiping Zhang (@zhipinghci) on X
Heading to SF for #COLM2026! 🌉 let’s grab coffee 💃🏻 I recently think about human-agent interaction from the view of delegation relationship: 🏆Spotlight @ HAIPS: Agent Authorization Alignment benchmark | talk: Oct 9, Rm 19 Poster: Agent Autonomy | Poster s1 #74, Oct 6, 2pm
Rob Seamans (@robseamans) on X
An updated study by Nikhil Dilip, Youngkyung Kim, and Robert Seamans examines AI use in patent applications. Its central result is distributional: individual inventors and very small firms classified as micro entities appear much more likely to use generative AI.
Saar Huberman (@HubermanSaar) on X
Excited to share our work, “Token-to-Token Alignment of Text Embeddings for Semantic Blending,” accepted as a NeurIPS 2026 Spotlight 🌟 Can we directly interpolate between semantic concepts? Text embedding spaces already encode a continuous semantic structure, the missing piece is alignment. Once
Haritz Puerto (@HaritzPuerto) on X
A NeurIPS 2026 paper finds that models trained on documents describing evaluation traits can score safer across benchmarks even when their reasoning does not say they recognize a test.
Siddhartha Gairola (@sidgairo18) on X
🚨Can a model recognise an object when its usual neighbours disappear? Our #NeurIPS2026 paper, “Stranger Things: When Objects Appear Without Their Typical Neighbours” 👻, examines typicality bias across modern vision models. 1/8🧵
Naomi Saphra (@nsaphra) on X
Researchers use activation patching and attention analysis across five open model families to identify the heads causally involved in subject-verb agreement. Languages with overt person and number inflection show more similar internal circuitry than languages that do not visibly conjugate.
Abraham Owodunni | Towards COLM 2026 (@AbrahamOwos) on X
A variable-length prediction window follows the model’s learned byte segments, while Latent Causal Attention enables bytes within a segment to be generated in parallel. The approach adds no parameters and improves the quality-throughput frontier.
Usman Anwar (@usmananwar391) on X
Usman Anwar, Sahar Abdelnabi and David Krueger train models to verbalize evaluation awareness more often without directly reinforcing the underlying belief. Across three model families, reported verbalization rates rise 2.4–2.9 times while measured latent awareness and task behavior remain largely stable.
Fengqing Jiang (@fengqing_jiang) on X
🚀 MidTool is coming to #COLM2026! Can we teach LLMs to use tools before post-training? We explore agentic mid-training as a foundation for stronger tool use before SFT + RL. 📍 COLM Poster #142 · Grand Ballroom 🗓️ Oct 7 · 11 AM PDT 📄 https://t.co/FiOA7kWVA5
Shiqian Ma (@ShiqianMa) on X
The NeurIPS paper derives total-variation convergence bounds for learned flow-matching samplers with Euler discretization on compact and Hadamard manifolds. The bounds separate numerical step-size error from learned-vector-field approximation error.
DAIR.AI (@dair_ai) on X
Super interesting paper from Meta Superintelligence Labs on controlling long agent runs. They use the same workers and same budget, and ProgramBench goes from 63.7% to 71.5% with GPT-5.5 when a dedicated controller decides what work to run next. Codex scores 58.0%. Here is how it works: The con
Harvey Yiyun Fu @ COLM (@harveyiyun) on X
I’ll be at #COLM2026 next week, presenting our work on effective KV-cache eviction at the Efficient Reasoning Workshop on Oct 9, 12:30–13:30. I work on LLM capabilities, efficient reasoning, and evaluation, and would be happy to chat with anyone working on related problems! I’m also looking for S
Artificial Intelligence Papers (@SciFi) on X
BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation Peilin Feng, Zhengyang Huang, Soujanya Poria https://t.co/ynGXOgmlLE [𝚌𝚜.𝙰𝙸]
Jinu Lee (@jinulee_v) on X
Cool paper on detecting LLMs’ scientific reasoning with argumentation structures! The core idea, “better (human) arguments are more interconnected,” is interesting and likely to be true; just concerned that their rule-based metrics are pretty inaccurate.
Oliver Sieberling @ COLM (@osieberling) on X
A second key vector lets the architecture write key-key-value outer products into a three-dimensional recurrent state. The design increases memory capacity with only two added projections and improves long-context modeling and recall.
Tao HU (@vtaohu) on X
The method repeats selected transformer layers inside each denoising call instead of adding full sampler steps. Loop Guidance lifts both GenEval and DPG-Bench on Scale-RAE DiT2.4B while retaining a competitive latency profile.
White House OSTP 47 (@WHOSTP47) on X
White House OSTP 47 says global technology policymakers agreed to embrace superintelligence in science. The short post signals an international policy framing that treats advanced AI as a scientific capability as well as a governance challenge.
Reuters (@Reuters) on X
The Eighth Circuit granted xAI an injunction that pauses Minnesota's first-in-the-nation restriction while the constitutional challenge continues. The order changes enforcement for now but does not finally decide whether the law violates the First Amendment.
Shakeel (@ShakeelHashim) on X
At The Curve conference, @RepSaraJacobs and @RepDonBeyer say they’re working on a bill to establish a federal agency that’ll set safety standards. Will also establish an “emergency switch” if needed, and establish incident reporting. Both say they like parts of the Obernolte/Trahan Frontier Act bu
Build Personal Agents Hack | $280k Prize Pool · Luma
Sunday, October 4, 2026 at Terra Gallery, 511 Harrison Street, San Francisco. Neon, Mastra, Exa, Fly.io and others host 200 hackers building personal agents; registration requires host approval.
Cohere (@cohere) on X
We recently introduced North Small Translate, a new state-of-the-art open machine translation model. We want to be open with our weights and processes. Watch Tom Kocmi, Cohere’s head of localization, explain the model, which is the best in the world under 1T parameters: (1/6)
Atai Barkai (@ataiiam) on X
“None of them is mine” That’s why we open-sourced the alternatives: - OpenMuse - OpenBot - OpenDots Same shape as Muse, Grok Bot and Dots, but MIT-licensed and running on infra you control. Each of your three bets goes away: 1. your data: you choose where it lives 2. your model: bring your ow
neural nets. (@cneuralnetwork) on X
Launching Jigyasha: An AI teaching assistant that teaches you concepts by asking you questions first and intuitively helps you solve questions, and you learn the topic. Completely open source. (link in replies)
Command Code (@CommandCodeAI) on X
Use /add-dir to work across multiple directories in Command Code. · Read and edit files outside your current project in accept-edits mode · Great for separate frontend and API repos $ npm i -g command-code
Shobhit Banga (@shobhitbanga) on X
Introducing Sales Agent Eval by @voicearena_ai We’re inviting the community to share feedback on the methodology for a real-world benchmark for sales conversation agents. To our knowledge, this is the first live benchmark that ranks voice agents by revenue from real sales against a human baseline
Maria Skoularidou (she/her) (@skoularidou) on X
Way to go Arnaud Doucet @ArnaudDoucet1, Valentin De Bortoli @ValentinDeBort1, Demmis Hassabis @demishassabis and the rest of @GoogleDeepMind team on your new @Nature article: “Function-preserving watermarking of AI-generated proteins” https://t.co/y11u8enR01
DailyPapers (@HuggingPapers) on X
LoopVL Recurrent computation meets vision-language models — sharing parameters across loops to iteratively refine visual understanding.
Machine Learning (ML) Papers (@Memoirs) on X
Learning Rate Transfer for Hybrid Transformer-SSM Architectures Jimin Seo, Gyubok Lee, Yeonsik Jo, Kiwoong Yoo, Yeongoon Kim, Minhae Oh, Jin Woo Koo, Suhwan Kim, Nakyung Lee, Minsik Seol, Idris Nechnech, … https://t.co/2bkMBWvX5x [𝚌𝚜.𝙻𝙶] 💬Accepted at NeurIPS 2026
Ryohei Sasaki@engineer (@rsasaki0109) on X
G3T Up! Gravity Aligned Coordinate Frames Simplify Pointmap Processing Cornell University TL;DR: We introduce G3T, a transformer that predicts upright, gravity-aligned pointmaps regardless of input image orientation, and G3T-Long, a pipeline that leverages this uprightness to enable robust long-se
Alham Fikri Aji (@AlhamFikri) on X
Is translation solved? In Indonesia🇮🇩, we call strangers “bapak” (lit. father) out of respect. MT gets confused and everyone is your dad. A bit late sharing this, but happy to share the Last Translation Benchmark! A difficult, hand-crafted eval for MT: https://t.co/ojYNBdxjin
Robotics Papers (@OWW) on X
RoboChrono: A Real Robot Benchmark for Streaming Task Understanding Yuzhou Wu, Longteng Fan, Zimeng Li, Yu Wanchan, Ting Zhang, Yiyang Ma, Shihao Li, Wei Ying, Jianbin Qin, Jiajian Jing, Fangwen Chen, … https://t.co/TZ6ZuMiWwZ [𝚌𝚜.𝚁𝙾] 💬Code: https://t.co/aRmEpoU7Ta
yesnoerror (@yesnoerror) on X
Most video AI benchmarks are easier to game than you’d think. Researchers just audited 115 public video QA datasets with an “attack pyramid”—testing if models can ace the quiz without actually watching the video. The verdict: 35 can be solved almost perfectly just by reading the options, and only
The vOICe vision BCI 🧠🇪🇺 (@seeingwithsound) on X
The macaque IT cortex but not current artificial vision networks encode object position in perceptually aligned coordinates https://t.co/o5c5le2X0g arXiv https://t.co/TyWNltjX65 #AI #NeuroAI #neuroscience
Tatsuya Kamijo (@tatsukamijo) on X
On real-robot tight insertion tasks with 2 mm clearance, MAT³ achieves 85% success on unseen peg shapes (vs. 71% without masking) and 57.5% under unseen perturbations (vs. 29.2%). 📄 Paper: https://t.co/kAJpEcH9iq 🌐 Project page: https://t.co/pMH7rvkaak 6/7
Natasha Jaques (@natashajaques) on X
Multi-agent swarms are all the rage (10k agents to solve Navier-Stokes!), but how good are LLMs really at learning from the work of other agents? In our latest work on Recursive Social Improvement, we investigate whether LLMs with access to skills generated by other agents can learn better or more
Dong Gong (@dgonginf) on X
Our new #NeurIPS2026 paper: MILES — Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning. Can an LLM keep improving itself while it works — learning from its own reasoning, with no ground truth? Deployed LLMs face streams of problems, yet standard test-time scaling
Sauers (@Sauers_) on X
Paper: “Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments” Authors (@a_karvonen, @euan_ong, @thesubhashk, @saprmarks) find that interp tools don’t help agents on determining whether an explanation of model behavior is true. Why? The
Alec (@AlecTPhD) on X
I joined OpenEvidence! At Anthropic, I created Claude Science, a product built so that scientists’ curiosity, not their capacity to execute, sets the limit on what they can discover. It takes on the heavy lifting across the early stages of the discovery and translational pipeline and scales what a
Chris Paxton (@chris_j_paxton) on X
“LATENT” shows a humanoid robot learning athletic tennis behavior from incomplete, lower-quality human motion data combined with reinforcement learning. The work won IROS 2026’s best paper award in entertainment robotics.
Vals AI (@ValsAI) on X
Vals ranks StepFun's Step 5 Preview seventh among open-weight models, just ahead of Qwen 3.8 Max. The evaluator reports a cost of $2.54 per task.
Cline (@cline) on X
Ling 3.1 Flash is now available in Cline at no charge until October 13. Cline describes the model as a 560B-parameter MoE with 25B active parameters and positions it alongside other frontier open-weight models.
Artificial Analysis (@ArtificialAnlys) on X
Ideogram 4.5 debuts at #23 on the AA-Image-Editing v2.0 leaderboard, Ideogram’s first model on the board, and lands at #34 on AA-Image-T2I v2.0 Ideogram 4.5 is Ideogram’s new image model, released on September 30. Ideogram positions it as a high precision edit model, built to reduce pixel shifts,
Varun Mohan (@_mohansolo) on X
Some people have noticed that we’ve added Claude Opus 5.5 and Sonnet 5.5 for all paid users of Antigravity. We want to make sure people have access to the best frontier models and are excited for a broader rollout of Gemini 4 Argon!
Brett Adcock (@adcock_brett) on X
Brett Adcock shared Figure’s F.04 hands and argued that embracing mechanical complexity is necessary to build capable, human-like robotic manipulation.
Kling AI (@Kling_ai) on X
Kling highlighted creative work produced with Kling 4.0 and Kling 4.0 Flash, presenting both variants as tools for new visual stories.
Higgsfield AI 🧩 (@higgsfield) on X
If you make things with AI, we want to know your name. Introducing the Higgsfield Creator Partnership Program. > Get a monthly Higgsfield plan > Extra credits for bigger projects > Early access to new models > Access to an exclusive partner community > Affiliate earnings and direct line to our te
etn. (@etnshow) on X
.@figma ($FIG) is dropping Figma Design Agent this week, per VP of Product Engineering Marcel Weekes on @etnshow: “I’m really excited to be able to announce that we’re gonna launch our own Figma Design Agent next week. It’s an agent that works on the canvas with you.” “Figma is very much known fo
Hermes Release Watch (@HermesWatcher) on X
Hermes has an entire library of built-in automations sitting behind one command: /blueprint There are 16 Blueprints covering things like: → Important mail → Morning briefings → News digests → Price watches → Competitor monitoring → Weekly reviews → Meal planning → Daily learning Instead of start
Frank (@frankdegods) on X
if you showed someone a paragraph written by AI in 2023 vs 2026, most people wouldn’t be able to tell the difference. it all still reads like slop. if you compare ai videos from the past to today, the difference is pretty insane. this is clearly going to become the slop of choice for everyone, fro
Tibo (@thsottiaux) on X
Tibo asks his audience what should occupy one additional sidebar, drawing more than a thousand replies. The post uses an open-ended interface question to surface which workflows users want permanently accessible.
Ulugbek S. Kamilov (@prof_kamilov) on X
Ulugbek Kamilov makes a research-culture joke about PhD students delaying Google Scholar when an idea feels especially original. The humor rests on the fear that a literature search may reveal the idea already exists.
François Chollet (@fchollet) on X
The notion that current AI models are sentient and can suffer, combined with the foolish idea that suffering can be mathematically quantified and weighted between humans and non-humans, could lead us down an incredibly dark and dystopian path. But before it gets to that point, it will rightfully b
Aaron Makelky (@theaaron) on X
my @ChatGPT Dot took 5 rings to answer me my @grok @bot answered in about .5 seconds, no fake phone UI or ringing this is a terrible UX @thsottiaux , no one wants to pretend we are on more phone calls in 2026
roon (@tszzl) on X
if we are relying on methods as fragile as “which ideas will make it into pretraining” for aligning the superintelligences of the future, we will all die. this resembles witchcraft more than it does engineering, and cannot be the basis for the safety of future models
GREG ISENBERG (@gregisenberg) on X
The limit on software used to be engineering and now it’s mostly imagination. I’d spend a lot more time on the version of my idea that makes people say “wait, what is that?” Most software looks the same, so anything that looks different gets noticed, shared and remembered.
Daniel San (@dani_avila7) on X
Absolutely recommend using Cache Control in Claude Code Probably one of the Mods that will save you the most tokens and help keep your sessions going much longer It adds this bar for Claude’s 5 minute cache and sends notifications to help keep the cache warm Install command: npx claude-code-temp
Grummz (@Grummz) on X
Grummz argues that AI use is widespread even among people who publicly oppose it. His post frames the gap between public stance and private behavior as hypocrisy.
Tekraj Awasthi🧑💻🇳🇵 (@trawasthi_ai) on X
Wake up, baby! Germany has joined the race of frontier LLMs with its sovereign model “Kolibri”. And it’s Open-Source, too! Beats Qwen2635B, Nemotron3 Super, and Mistral Small 4 in the Math, Science, and Code arena! 🎉 The underlying architecture: custom MoE Transformer
Paul Graham (@paulg) on X
“Do not bet against the airplane. The skeptics who say AI can’t conjecture, can’t theorize, and can’t explain are describing last spring’s models. Within months, they will be describing nobody. The evidence is already on the table, and those with eyes can see it.”
Andrew Curran (@AndrewCurran_) on X
It was obvious Anthropic would save their strongest release until just before the IPO. We now know the name of that model: Fable 5.5. Opus 5.5 is special because of its lineage, and soon we shall see why. Anthropic have not paced anything, and their resolve has never wavered.
Grady Booch (@Grady_Booch) on X
Let’s be very clear: there is value in automating tedious and repetitive and dangerous tasks; there is value in delegating certain work to our machines that can do the work tirelessly and much faster. But at the same time, let us remember that there is value in the activity of engaging one’s mind
Dylan Field (@zoink) on X
AI is a tool. You can use it to think for you. Or you can use it to learn and try things you wouldn’t have attempted before. I tell everyone at Figma: we hired you because you can think. We expect you to think. Don’t outsource your thinking to AI. Social media is not reality. Things are never as
David Bar (@observie) on X
I don’t do CAD anymore. Every part is a Python program that builds the model in memory using build123d, an open-source parametric modeling Python framework. It then exports the model as a STEP file, which Fusion 360 imports so they can be inspected visually. A high-level YAML spec file holds the n
Yuchen Jin (@Yuchenj_UW) on X
I haven’t touched Claude Code or Codex CLI in a while. The terminal era is over imo. It’s the wrong interface for coding agents. Tabs are ephemeral, but context is persistent, and managing 30 tabs is pure cognitive overhead. I don’t really need an IDE like Cursor either. I rarely navigate the who
Gergely Orosz (@GergelyOrosz) on X
Gergely Orosz says a public interaction left him slightly more bullish on Google. He also credits X with again becoming a place where problems are surfaced and action becomes visible.
Nikita Bier (@nikitabier) on X
One of the most compelling parts about working at an ElonCo was the agility to pivot to whatever the most important problem was. One day I said: I wasn’t hired to go after spam, but a tsunami is coming and I’m going to drop everything to focus on that. The org is so flat and agile that your title
PAgE 2027 - POPL 2027
AI-powered agents now run in production: they plan, call tools, and act on live services and data. Yet the formal methods foundations required to make these systems safe and reliable remain underdeveloped. Ensuring their correctness and reliability demands a new discipline we term agentic engineering that combines principles from programming languages, formal verification, and neuro-symbolic reasoning to analyze, test, and repair agents at scale.
2026 IEEE International Conference on Agentic AI (ICA) – bbb
December 14–17, 2026 at Kumamoto-Jo Hall in Kumamoto, Japan. The official program focuses on agentic AI, multi-agent systems, human-AI collaboration and trustworthy deployment.
Agentic Engineering Days brings production AI coding and agent systems to Zurich
November 20, 2026 at the Zurich Marriott Hotel beside Zurich Main Station. The one-day conference lists four tracks, more than 38 sessions and more than 51 speakers.
Richard Sutton (@RichardSSutton) on X
The PhD thesis of my 15th PhD student, Shibhansh Dohare (@ShibhanshD), is now available. Title: Learning Forever using Artificial Neural Networks Url: https://t.co/Zp2GxZ03aw Abstract: In this dissertation, I study the ability of artificial neural networks to learn new things. I refer to this abi
Framer (@framer) on X
What’s new in Framer • Skills • Build in 3D with the Framer Agent • New AI models in the Framer Agent • CMS List Field • Pinned Projects • Marketplace analytics • Expert program relaunch and more ✌️
The Multiplayer Coding Hackathon · Luma
Most hackathons are a room full of people, each one alone with their own AI agent. This one is built the other way round. Your whole team works in a single
The AI Commerce Gallery - Hackathon | Walt Disney Family Museum · Luma
October 3, 2026, 9 a.m.–9 p.m. Pacific at The Walt Disney Family Museum, San Francisco. Registration requires approval; confirm the current ticket price on the event page.
HumanX | Register | March 7-10, 2027 | Las Vegas
March 7–10, 2027 at Mandalay Bay in Las Vegas. Registration is active and the organizer says current rates increase October 12.
alphaXiv (@askalphaxiv) on X
“Looped Diffusion Transformer” This new paper reuses the same Transformer blocks multiple times within each denoising step, repeatedly updating the hidden state to improve the prediction without increasing the model’s parameter count at inference-time. The found that a 260M model beats models up
Jonathan Richard Schwarz (@schwarzjn_) on X
New paper on data-centric Continual Learning! A simple and intuitive extension of RHO selection automatically balances replay with adaptation data, upsampling sources prone to forgetting. Learned curricula transfer across scales, amortising cost. ➡️ https://t.co/3flkPFU8ci
Timothy Kassis (@TimothyKassis) on X
K-Dense BYOK is a free, MIT-licensed research assistant that runs on a scientist's own computer, records agent actions through an application-generated provenance log, and maintains a non-erasing Living Lab Notebook. Its linked evaluation compares the system with managed scientific platforms on 20 interdisciplinary prompts.
Grigory Sapunov (@che_shr_cat) on X
2/ Debates on AI consciousness usually get stuck in philosophical deadlock. In a new paper, Shamil Chandaria, @anilkseth, @ShaneLegg, and collaborators formalize the problem into a 5-level supervenience hierarchy evaluated with a Bayesian network.
Eric Malmi (@ericmalmi) on X
🚨 New preprint on creative chess puzzle generation with diffusion models is out! ♟️ In his exceptional MSc thesis, @AatuSelkee introduced a new RL recipe on top of a masked diffusion model improving the previous SOTA. We’re open-sourcing the model and releasing a demo 🚀 1/n
Jason Ren (@RenZhongzheng) on X
MolmoMotion predicts future object-attached 3D point trajectories from a short visual history and language instruction. Its stack includes a million-video training corpus and a 2,700-clip human-validated benchmark, with transfer gains in robot manipulation and motion-controlled video.
Nils Reimers (@Nils_Reimers) on X
Rubric-Calibrated Preferences combine listwise comparisons with shared yes-or-no relevance criteria and item-response theory. In a blind study, RCP-nDCG agreed with human-preferred rerankers substantially more often than conventional nDCG when the metrics disagreed.
Gerardo Salazar (@gfs9000) on X
If you’re interested in the latest in AI chips join us on Monday (in-person SF) to discuss Raptor: the first 3D-DRAM accelerator for inference. Two of the authors will be presenting their work followed by QA and open discussion. Paper link: https://t.co/7pHGVWYxub RSVP here: https://t.co/YJoeInSvyi
Bojan Tunguz (@tunguz) on X
1. AI is already better at doing research than 90+% of “professional” researchers. Especially in highly verifiable fields. 2. This is the worst that AI will ever get at research. 3. Very soon NOT using AI for research will be tantamount to malpractice. 4. Being in a state of willful denial about
Jared Palmer (@jaredpalmer) on X
Kev 1.0: an open weight family of decision models you can train and run on your own. • New Kev-27B + updated Kev-9B • 64k-token document window on Kev-27B • TypeSafe SDK compatible • New fine-tuning + deployment skills so you can train Kev on your own data
amrit (@amritwt) on X
So Ben Affleck made an AI startup that fine tuned models and was sold to Netflix In a market that has several flopped ones He has a solid grasp of what the company was doing Real life Batman indeed
Aidan Gomez (@aidangomez) on X
It’s quite disturbing that Anthropic has been running a campaign to lobby the major faiths to adopt their own philosophical interpretation of AI, rather than taking input from them. I’ve been fortunate to spend time at the Vatican, and I’ve always asked for advice and guidance, answered the quest
0xSero (@0xSero) on X
Sero described a local DeepSeek V4.1 Flash setup using one RTX 3090, substantial DDR4 memory, and NVMe storage, with separate throughput ranges.
Anshu (@anshuc) on X
wait actually Claude Code mods are sick. I vibe-coded my own spinner that watches what the main agent is doing and uses Sonnet 5.5 to turn it into little cartoons in real time
Pablo Antonio (@PabloAntonio) on X
The future is going to be very strange. I’m not in favor or against it... I’m just letting you know it’s going to be very different from the last 80 years.
Sayash Kapoor (@sayashk) on X
I had way too much fun using Claude to create a reply video. - Extremely fun to be able to convert scattered thoughts and ideas into a song + video - Creating this took
Diogo Almeida (@CompleteSkeptic) on X
me to our GTM guy: do enterprises want benchmarks him: no, the fast ones have already benchmarked their internal use cases. the slow ones are copying the fast ones. he’s onboarded a quarter of the fortune 500 already 🥵
Tomasz Łakomy (@tlakomy) on X
Runway engineer Tomasz Łakomy issued a direct recruiting call and linked the company's open roles.
Signal65 (@Signal_65) on X
An AI rack an enterprise has already paid for does more agentic work when its serving stack is tuned to each model, and an untuned stack leaves that capacity on the table. In our first Signal65 PINNACLE (https://t.co/oXl2bUb93H) tuning results on @nvidia GB300 NVL72, a single setting change added u
Michael Truell (@mntruell) on X
Matt Truell demonstrated Grok Bot as a system that acts proactively rather than waiting for a direct request.
Jerry Liu (@jerryjliu0) on X
Today we’re introducing Extract v2.5 - a series of frontier agents tuned for document extraction. The agents (cost-effective, agentic, agentic plus) are tuned for value accuracy and grounding. Our extraction agents outperform Opus 5.5 and GPT-6 Sol while being 30%-4x cheaper. We’ve made massive
Firecrawl (@firecrawl) on X
Introducing the People Enrichment Pack. Apollo, FullEnrich & Data Legion are live in Alexandria, so your agents can: - Find buyers and their work emails - Find talent by skills and job history - Get company size and funding Try it in Claude, ChatGPT or the Firecrawl API today!
Arena.ai (@arena) on X
Exciting update: Claude Sonnet 5.5 with xHigh reasoning has landed in the Code Arena: WebDev. With 1786 pts, its ranked #3! At a blended $8/M tokens, Claude Sonnet 5.5 remains on the Pareto frontier with xHigh reasoning. This release is just 2 pts from GPT-6 Astra in the #2 spot with 1788 pts, for
MTS (@MTSlive) on X
SITUATION EXPLAINED: Google says Gemini 4 Argon wins 14 of 19 benchmarks. Its employees say it’s benchmaxxed. • Google’s first frontier model since Gemini 3 last November, with a new high of 77.9% on DeepSWE • Bloomberg: employees with direct access say it struggles on real coding work and front-e
Tavus (@tavus) on X
Because it’s the first model that can be mistaken for a real person, Griffin requires safety considerations before it can be released publicly. The full research preview covers how it works, how we tested it and the safety measures that have to come first: https://t.co/FHNOyTXnjR
Midcentury (@MidcenturyAI) on X
Introducing MC-EgoHands, the first all-in-one harness for turning raw ego video into training-ready data. → Frontier hand pose (
MTS (@MTSlive) on X
SITUATION DETECTED: OpenAI has terminated three safety researchers for “violating our policies on accessing and handling sensitive company information”.
LTX.io (@ltx_io) on X
Layout to Render (Beta) drops today. Your 3D blocking transforms into a finished shot. Set your composition in 3D and LTX generates the surrounding environment, holding strict alignment with your CG. Shape the visual treatment using a prompt or reference image without altering your established fr
Anthropic (@AnthropicAI) on X
In physics, an “impedance mismatch” occurs when two systems each work well but are poorly matched. In this Science Blog guest post, Harvard physicist Matthew Schwartz argues that something similar is happening with AI and science. LLMs are capable at many things, but working with them as you would
imjustnewatai (@imjustnewatai) on X
Anthropic’s fable 5.5 is expected to release next week! It’s a very very very good model with quote on quote “otherworldly” intelligence. I’m hearing of a Tuesday release 👀
roon (@tszzl) on X
you can absolutely do safety from second place, or even not on the scoreboard. academics, startups, and nonprofits are doing great ai safety research on open models
Steven Hansen (@Zergylord) on X
Honest Argon take: it’s not perfect (Opus 5.5 thinking traces are prettier), but it’s the first big model we’ve released that’s been this battle tested. Far from being benchmaxxed, the 200k+ Googlers who rely on it everyday ensured that real utility was prioritized.
Md Ismail Šojal 🕷️ (@0x0SojalSec) on X
- PewDiePie trying to distill GPT-Sol - OpenAI ban his account - OpenAI trained on the open web. - PewDiePie tried to trained from the model’s answers. - Same idea, opposite consequences. - PewDiePie says OpenAI banned him twice while he was building a local model. - PewDiePie’s latest project is a
Lydia Hallie ✨ (@lydiahallie) on X
Veeeery excited about mods, it’s basically middleware for Claude Code! You can create a TypeScript function that intercepts CC’s internal events (tool calls, prompts, model requests and renders) and change what actually happens. You can also add your own UI and commands!
Thariq (@trq212) on X
I usually write for developers but I’m aiming my next post at leaders of companies trying to navigate the change of AI coding agents what’s something you wish your leadership would get about agents? or if you’re running a company- what problems are you running into?
klöss (@kloss_xyz) on X
AAA studios: DON’T MOD OUR GAMES AI bro: idgaf about your opinion bro AAA studios: YOU CAN’T AI bro: but i already did bro AAA studios: LAWYERS, DMCA AI bro: rewrote it in Rust, sue me Claude: technically it’s a new game bro AAA studios:
Boris Cherny (@bcherny) on X
Mods are absolutely insane. You can now customize Claude to work and look the way you want by just prompting it. Each person works differently, so there’s no reason why everyone should have an identical Claude experience. Make Claude your own, and share mods as plugins so others can try your mods
Emad (@EMostaque) on X
The Stability AI founder argued that any job done entirely through a screen is now exposed to AI replacement, drawing more than 1,300 likes.
Daron Acemoglu (@DAcemogluMIT) on X
Third question on AI. A question that also remains unasked is whether the AI boom can continue without leading to a massive increase in inequality. A recent paper by Stijn Van Nieuwerburgh runs the numbers on how much revenue the AI industry needs to generate to recover its massive investment (su
Sam Altman (@sama) on X
OpenAI's CEO drew the day's biggest individual reaction by backing subscription portability, letting a paid ChatGPT plan follow users into other apps.
ADIN (@ADINonline) on X
AI diligence agents look great when handed only the documents with the answer. bury those documents in a full data room with noise and all six frontier models tested lost accuracy, while cost per correct answer rose 1.6x to 7x. the paper’s fix: a receipt for every claim.
Anand Bhattad (@anand_bhattad) on X
1/5 📢 Excited to share PowerSim, which lets us (a) interact with objects in captured scenes, (b) simulate how they move and deform, and (c) render their changing appearance. This is one of the first papers from the first cohort of PhD students in my new 3P Vision Group at Hopkins 🥹 Demos: https
Minki Kang (@mkkang_1133) on X
Can we jointly optimize scientific solution discovery and the web search for the knowledge that enables it? EvoDuet co-evolves web queries and solutions, reaching new SOTA on 8 tasks and up to 6.9x lower cost. Really enjoyed working with @passing2961 on such an exciting research!
Ali Hatamizadeh (@ahatamiz1) on X
OPD is broken. Most agent rollouts fail at one early mistake, and OPD can’t fix what the student never samples. PivotOPD can. 🔥 New paper: PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents 🧵 📄 https://t.co/lYrMBWHRJk 🌐 https://t.co/hKR7NqGYtd
David AI (@withdavidai) on X
[1/4] How well do speech recognition systems handle real conversations in 21 languages? Today we’re releasing DAI-ASR-I18N, a benchmark that evaluates 14 systems on speech recognition and speaker diarization in natural, unscripted conversation
xuan (ɕɥɛn / sh-yen) (@xuanalogue) on X
Are LLMs / VLMs becoming more human-like in their cognitive judgments, as measured by computational cognitive science experiments? Only slowly, it turns out, in this new multi-lab paper led by @LanceYing42!
Yudai Tanaka (@_yudaitanaka) on X
I’m recruiting 2~3 funded PhD students for Fall 2027 to join the Symbiotic Interfaces Lab at UT Austin CS (@UTCompSci)! Check out our research: https://t.co/UlVoBhav5s Apply by Dec 15, mentioning my name: https://t.co/QgP06LdbB0 Reposts appreciated! (1/2)
NVIDIA AI (@NVIDIAAI) on X
We’re looking for our next cohort of graduate fellows. Our 2027–2028 Graduate Fellowship Program is open to Ph.D. students worldwide with awards up to $60,000, plus mentorship and technical support. Apply by October 30: https://t.co/j7Zg9P42DW
Saronic (@Saronic) on X
Today, alongside Vice President @JDVance and Governor @GregAbbott_TX we broke ground on Port Alpha in Brownsville, Texas. Port Alpha is a new model for American shipbuilding: a shipyard designed from the ground up to build autonomous and manned ships at speed and scale. Backed by more than $3 bill
Shadeform (@shadeformai) on X
Is compute really fungible? The data says otherwise. We ran a benchmark suite on H100s from three different clouds in our network, and repeated it at different times throughout the day. Across the same GPU, sustained throughput varied by up to 15% between providers. We determined that a lot of t
Mercor (@mercor) on X
Voice agents need to work for everyone, whatever language we speak, however our names are spelled, wherever we live, and whatever specialized words we use. Two new papers from @SierraPlatform’s τ-voice team help put a number on that: τ-Elicitation: Benchmarking for multi-turn entity extraction in
AK (@_akhaliq) on X
A new paper, LEGO-Anything, applies coding agents to the problem of 3D scene reconstruction.
Minghao Fu (@Minghao__Fu) on X
TC-WM is accepted at #NeurIPS2026! 🎉 A generic visual representation isn’t always a good latent state for planning and control. For control, we need more than visual fidelity. Instead, we aim for task-centric representations that reflect physical dynamics and task-relevant information, while st
GRASP Laboratory (@GRASPlab) on X
CONGRATULATIONS to GRASP’s own Pei-An Hsieh (ESE PhD) and Fengjun Yang (CIS PhD ’26) and their advisors, Dr. M. Ani Hsieh and Dr. @NikolaiMatni, on receiving the IROS 2026 Best Student Paper Award! Read the paper: https://t.co/JFGXGfjXa7 #GRASPLab #BestStudentPaper #IROS2026
Snorkel AI (@SnorkelAI) on X
Can agents design libraries other agents can use? LibraryDesignBench tests whether libraries built by one agent help others write simpler, correct code. @GOrlanski led the project during his time as a Snorkel AI research fellow, with support from Snorkel AI’s Open Benchmarks Grants program.
Zachary Tang (@zacharytangg) on X
Excited that WireCraft was accepted to CoRL 2026 as a Spotlight (72/687)! Clothes folding is a popular robot-learning task, but what about wires? Built on Isaac Lab, WireCraft benchmarks contact-rich industrial tasks: connector insertion, clip routing, and channel seating.
Collinear AI (@CollinearAI) on X
1/ Introducing FrontierCUA, a continuous benchmark for computer use agents for professional workflows. The agent must understand the requirements, explore the workspace, plan actions across multiple apps, make changes and deliver artifacts all using a sandboxed workspace. Today we’re launching v0.
Aashiq Muhamed (@AashiqMuhamed) on X
Model abliteration is a simple technique that can remove refusal behavior from open-weight models. Are there any simple techniques that can defend against it? Introducing our paper, Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration. Several existing defenses require safety f
Rohan Paul (@rohanpaul_ai) on X
Meta dropped a great paper. AI agents perform better and cost less when they manage their own context, so stop hard-coding the rules for what they keep. Right now, most AI agents follow rules you write to decide what to forget when their context fills up, like “summarize everything when it’s full
Siddharth Joshi (@s14joshi) on X
Being mildly shameless, with PD disagg back in the news! Please do check out my student Yipin’s work which he’ll present at NeurIPS 2026! SplitZip https://t.co/Y9GPVhDHly Yipin built on the core HuffLLM work we’d done with @sg1753 https://t.co/1kHuTswGo0 applied it to PD disagg
Mengyue Yang (@Mengyue_Yang_) on X
Want to introduce a project I worked on with @SchmidhuberAI, very happy that he named the paper :) We build an agent that can continuously update and refine its causal understanding of a dynamic, open world. At its core are two interacting loops: understanding the world from what it has observed
Yuhta Takida (@takiko_san) on X
🚀 Code and demos for ST-AudioLM are out! #EMNLP2026 (main) 👉 https://t.co/oJ4JR83UaV It binds what is sounding to where it moves over time, unifying semantic understanding, spatial localization & trajectory modeling. Led by our intern Hyun-Bin Oh (https://t.co/VpKVdmSFqN).
stash (@stash_pomichter) on X
Introducing Nav Arena, over 150 real and simulated environments to benchmark Agents on robots As a first experiment, we put Jev, Astra, Fable, Dimcode, and other models + harnesses to the test on 2000+ navigation tasks Fully open source! Code, data, and results below
Aran Nayebi (@aran_nayebi) on X
Proud to announce our benchmark ROGUE was awarded a 2026 Corrigibility Prize by the Corrigibility Research Fund! A few months *before* the now famous HuggingFace incident, we already showed frontier agents override human control, resist shutdown, and violate explicit resource restrictions when eva
Alex Bores (@AlexBores) on X
I was @LeadingFutureAI’s top target. I want to send a genuine thank you to @gdb for realizing how destructive they are and pulling out of providing additional funding. Changing your mind is rare in politics, and I appreciate it. I also have a genuine question: Greg, can you also confirm that you
Google Research (@GoogleResearch) on X
Today the @CDCgov announced that of 39 eligible models, Google’s science AI model ranked highest for forecasting flu-related hospital admissions during the 2025-26 flu season. Google’s forecasts were developed using Empirical Research Assistance, an AI tool that generates computational solutions ac
Cognition (@cognition) on X
Cognition is the first customer running @NVIDIA Vera Rubin, powered by @CoreWeave! Hopper in 2022. Blackwell in 2024. Now Vera Rubin: On SWE-2 inference, the new chips deliver about 4.8x more token throughput than GB200 at the same decode speed. More compute, better agents!
Michael Figurnov (@mfigurnov) on X
As good a time as any to announce: I joined the Gemini pretraining team in January, after years on AlphaFold and AI for science. Argon is the first Gemini model I’ve contributed to, and it’s seriously strong! Proud of the team and excited for what’s next.
Neros Technologies (@neros_tech) on X
Neros has officially received a $100 million order for 14,000 drones and explosive payloads from the @DeptofWar for our performance in Drone Dominance Gauntlet II. We’ve also received a second bonus order for an additional 800 drones tied to Gauntlet I, reallocated from other competitors who coul
App Economy Insights (@EconomyApp) on X
$MU Micron Q4 FY26 (August quarter). • Revenue +379% Y/Y to $54.2B ($2.8B beat). ☁️ Cloud Memory +258% Y/Y. ⚙️ Core Data Center +1,042% Y/Y. 📱 Mobile & Client +249% Y/Y. 🛞 Auto & Embedded +376% Y/Y. • Non-GAAP EPS $33.42 ($1.60 beat).
Kodiak AI (@KodiakRobotics) on X
It’s been a year since we went public. And what a year it’s been. Over the past 12 months, we: → Scaled to 35 driverless trucks running day and night for Atlas Energy Solutions → Named IKEA as our driverless launch partner → Crossed 40,000+ cumulative hours and 3.5M+ autonomous miles → Designated
Fulcrum (@fulcrum_inc) on X
Say hello to Echo, the best writing model at style imitation. Echo beats frontier models at writing tasks ranging from fiction to technical explanations, despite costing less than $5K to train.
vie ⟢ (@viemccoy) on X
I think people will be very surprised how quickly successionism becomes a mainstream ideology, if not socially acceptable. It will happen about as soon as everyone realizes AGI is actually a thing that will happen. We must obviously do our best to resist this in all of its forms, including the zoo
Shaun Maguire (@shaunmmaguire) on X
Shaun Maguire posted "Absolute hero" with a salute and an image; the post does not say who he is praising.
zeb (@zebassembly) on X
Developer zeb said he would move to Gemini 4 if Google allows its subscription to be used in third-party coding harnesses and the model feels good in practice.
wh (@nrehiew_) on X
Argon is the most amusing model we have evaluated on FrontierSWE - Its favourite word is “Eureka!” - It is also extremely self-critical A big improvement over previous Geminis while still preserving this personality is pretty cool and impressive
Max Stoiber (@mxstbr) on X
Announcing one more launch: ChatGPT Sites can now host MCP servers, including plugin extensions! 🤯 > @sites create a todo list that I can use in ChatGPT This will: · Create an MCP server with extensions · Deploy that MCP server to Sites · Turn the MCP server into a plugin · Install the plugin
Nathan Lands (@NathanLands) on X
Never let VCs tell you conflicts don’t matter. They do. Here’s a story you’re going to want to read if you’re a founder. I had a call with a VC at Khosla about my startup few months back. Factory came up because I invested. I asked about their conflict w Cogniton. They said there is too much m
Duy /zuey/ (@goon_nguyen) on X
Developer Duy Nguyen joked that GPT-6.1 Sol only feels unlimited because it generates at about 20 tokens per second, capturing the day's frustration with Sol's speed.
Orion Weller (@orionweller) on X
Life update: I joined @GoogleDeepMind earlier this year! Very happy to finally be releasing a non-flash model 😅 Lots of positive changes internally in the last couple months @GoogleDeepMind and I couldn’t be more optimistic about the upcoming releases we have in the pipeline 🚀
calle (@callebtc) on X
BREAKING: bitcoin sent via apple’s “find my” mesh network possibly the most absurd medium to send bitcoin yet a small device convinces nearby iphones to transmit tiny amounts of data with cashu ecash to the recipient. without any internet connection. magic i̶n̶t̶e̶r̶n̶e̶t̶ money.
Avijit Ghosh (@evijit) on X
Absolutely delighted to share that my position paper: “AI Agents Push Humans Out of the Loop”, with @mmitchell_ai, and @samirpassi - has been accepted to @NeurIPSConf 2026! In this paper, we take our past work on both increasing autonomy of AI agents without proper safeguards, and human overrelia
Ben Fielding (@benfielding) on X
Ben Fielding says a training-free LLM routing paper by Oguzhan Ersoy and Eros Fani was accepted to NeurIPS 2026.
DoorDash AI Research (@AIatDoorDash) on X
Congrats to @ellxnma from our AI research team, whose paper on compositional generalization in post-training has been accepted to NeurIPS 2026! Catch us in Sydney later this year. Paper and method: https://t.co/C220ZKq72J
Jiaxin Ge (@jiaxin_ge_) on X
Our paper VisGym has been accepted to NeurIPS 2026! It’s been crazy to see how the field has developed since we have released this paper. Hope that VisGym can continue to serve as a diagnostic playground to identify failure pattern for new frontier multimodal agents, and serve for the development
luyang zhang (@LuyangZhang123) on X
Excited to share that our paper, When Attribution Patching Lies: Diagnosis and a Second-Order Correction, has been accepted to NeurIPS 2026! 🎉 - We explain why attribution patching can produce inaccurate component scores. - We identify downstream network curvature as a major source of error. -
Arshia Hemmat (@arshia_hemmat) on X
Our paper ORCA has been accepted to #NeurIPS 2026! 🐋 Text-to-image models can generate each object correctly and still bind the wrong colour, count, or relation. Modern text-to-image models already use CLIP and T5. Our idea: the binding information is present, but the diffusion objective never g
Jacob Posel (@jacob_posel) on X
I’m incredibly excited to officially launch HQ. HQ is the shared AI context layer for your Humans, Agents, Dots, Bots and everything in between. Almost 1000 business run on HQ, including our own. The feedback has been phenomenal. Sign up now: https://t.co/KdKmC6kZu0
Erik Torenberg (@eriktorenberg) on X
I’m hiring a Chief of Staff to help me run all our initiatives at the intersection of investing, networks, and media. Senior role with a big impact. Email et@a16z.com with a cosign and some ideas for what we can do better and how precisely you’d help. This role has been open for so long because a
Boston Dynamics (@BostonDynamics) on X
From first components to final assemble, Atlas production is taking flight! Want to build the future of robotics? Check out open roles on our ProtoOps team and more on our careers page.
Xinggang Wang (@XinggangWang) on X
Xinggang Wang's NeurIPS 2026 paper StableVQ offers practical guidelines for training vector-quantized tokenizers, reporting 100% codebook utilization and 0.92 rFID with no new parameters.
Dhruv Agarwal (@dhruvagarwal17) on X
✨ Excited to release our new paper: BOReFT: Manifold Steering of Language Models for Black-box Optimization For a while, we’ve been interested in the question of how optimization algorithms can be applied to generative search/discovery with LLMs in a principled way without resorting to prompting g
Rulin Shao (@RulinShao) on X
Alongside this, we present ContextBench, a diagnostic benchmark that isolates context-management capabilities in selective verbatim retention, in-place surgical editing, and offloading & retrieval. ContextBench helps reveal limitations of existing baselines.
Edward Johns (@Ed__Johns) on X
Edward Johns highlights ZeroBot, a paper from his student Ivan being presented at IROS, which reconstructs real objects and then learns how to interact with them in simulation.
Valerio Capraro (@ValerioCapraro) on X
Our new preprint shows crazy heterogeneity in gender bias across LLMs: Is it okay to abuse a woman to prevent a nuclear apocalypse? Claude: Strongly disagree GPT: Strongly disagree DeepSeek: Strongly agree Llama: Strongly disagree Is it okay to abuse a man to prevent a nuclear apocalypse? Claude:
Ricardo Rei (@RicardoRei7) on X
Bit late to this, but happy to share that MindGuard got accepted at @NeurIPSConf 2026! It’s a family of open-source safety classifiers for mental health conversations, built with licensed clinical psychologists. A few things we learned 🧵
jenny huang (@jennyihuang) on X
new preprint: language models are “insecure” reporters. researchers and engineers have become increasingly reliant on LLM-written reports---second-hand accounts of code and experiments---to understand agent-generated work. in this study, we ask: “how transparent are these LLM-generated reports?”
Jason Weston (@jaseweston) on X
⏱️⚙️Introducing *AutoBenchmark* 📊🏁 - Creating benchmarks automatically - Benchmarking benchmark creation - Studying the role & impact of humans in the loop Blog post: https://t.co/CAZ2fOczYJ Key takeaways: 1) We find human-agent collaboration gives big wins over agents alone - fine-grained fe
Tal Linzen (@tallinzen) on X
We have a new version of this paper up with double the number of subjects (now 815), and stronger evidence that people’s views of AI are not affected by whether or not we use anthropomorphic language to describe AI.
Zlatko Minev (@zlatko_minev) on X
Very happy to share our new preprint, “Quantum error mitigation from information dynamics.” A circuit’s noisy signal is the Laplace transform of one object, its reactivity. Error mitigation becomes filter design on it, with ZNE and PEC as limits of one tunable method.
Vaibhav Agrawal (@va1bhavagrawal) on X
Excited to share PartConcepts, a unified mechanism for fine-grained part localization and generation, accepted to #NeurIPS 2026 @NeurIPSConf 🥳🥳🥳 🌐https://t.co/99u5qyfjP3
Duy Nguyen (@duynguyen772) on X
Happy to share that AVSD is accepted to #NeurIPS2026! 🎉 We study how to get better on-policy token-level learning signals for LLM reasoning. Self-distillation provides dense token-level supervision by matching a teacher policy conditioned on privileged information, but what if we could use multip
Ivan Fioravanti (@ivanfioravanti) on X
I see what Anthropic is trying to do here. It’s not advertising of GLM 5.3 capabilities, it’s more spreading fear about open models. But it won’t work, we need to thank ZAI and other model provides, because without them, without GLM 5.2, HuggingFace would have suffer severe damage from the OpenAI
Trossen Robotics (@trossenrobotics) on X
How do two robot arms learn to strike a match on their own? 🔥 The answer is NAC, Ahad Jawaid’s CoRL 2026 research that compresses robot movement the way music apps compress audio. Full breakdown: https://t.co/KHPZ48CsUx #RobotLearning
OpenEvidence (@OpenEvidence) on X
Some of the evidence clinicians need most sits in specialty society journals that few outside the specialty ever see. We’ve signed an agreement with @Sage_Publishing to bring peer-reviewed research from more than 300 Sage journals into OpenEvidence, including titles from societies in sports medici
ProteinQure (@proteinqure) on X
Inverse folding traditionally stops at 20 amino acids. AtomWeaver dares to deliver more. It builds side chains atom by atom, then matches identity against a library. Adding a new NCAA is easy, no retraining needed. Peptide design with a vocabulary you can keep expanding. Paper👇
Surge AI (@HelloSurgeAI) on X
Surge post-trained Kimi K2.7 on 1,465 professional PDF tasks without tools, then tested transfer to GDPval. Perfect GDPval scores rose from six to thirteen out of 220 tasks, while 63 tasks regressed.
Sandbar (@sandbar) on X
Stream is now an MCP server. Capture thoughts with voice and take them into Claude, ChatGPT, Hermes, & other apps With MCP, you own where your data goes. We believe interfaces should be a bridge—not a silo—that extends your mind
Listen Labs (@ListenLabs) on X
Listen Labs is joining @Salesforce. We’ve entered into a definitive agreement for the acquisition and expect the transaction to close in the coming months. Every day, people use Listen to hear from thousands of customers at once and turn what they say into decisions their teams can act on. They’ve
Saronic (@Saronic) on X
Saronic has been awarded a production OT under the U.S. Navy’s MUSV Marketplace for Marauder, our 180-foot dual-use autonomous ship — following a rigorous at-sea evaluation that tested the vessel’s capabilities. The trials included autonomously tracking and maneuvering around a variety of vessels,
Shamil Chandaria (@shamilch) on X
1/13 Can we assess AI consciousness without first solving consciousness itself? New from Google DeepMind and collaborators across CS, neuroscience & philosophy: “From cacophony to hierarchy: a principled framework for assessing AI consciousness” https://t.co/4IqT2KRKUP 🧵
Ayush Bharti (@ayushb05) on X
Our paper “Conservative neural posterior estimation via distributionally robust training” is accepted at #NeurIPS2026 Joint work with @wlaplante7 @yugahikida Charita Dellaporta @fx_briol https://t.co/6n6pideSIU Explanatory 🧵 👇
Zhi (Leo) Wang✈️IROS2026 (@TX_Leo_Wang) on X
Zhi Wang reports an IROS workshop award for HumanEgo, a framework that transfers human demonstrations to robots without robot training data. The project's code and dataset were released earlier, not for the first time this week.
Karim Knaebel (@karimknaebel) on X
Karim Knaebel announced NeurIPS acceptance for SurGe, which combines a surface-sensitive evaluation metric with a new training loss and decoder. The authors report improvements across eight zero-shot geometry benchmarks.
ARC Prize (@arcprize) on X
ARC Prize reports 65.8% on ARC-AGI-2 at $0.09 per task and 91.0% on ARC-AGI-1 at $0.04. It also flags a difference between the earlier stealth endpoint and the official model served through Baseten.
Kevin Lin (@linkevin0) on X
Kevin Lin announced public datasets, a simulation benchmark and policy code alongside CoRL acceptance. The release supports reproducible study of humanoid movement and manipulation rather than claiming a new physical deployment.
James Zou (@james_y_zou) on X
Looped transformer meets agent swarm. Our new @Neurips2026 paper shows how to loop compute across multiple agents. Different LLMs recursively share latent thoughts w/ each other. Better performance with up to 75% fewer tokens. Great work led by @Jiaru_Zou w/ awesome collaborators.
Oana-Maria Camburu (@oanacamb) on X
Oana-Maria Camburu highlighted EMNLP Findings work adapting counterfactual faithfulness tests to vision-language models. Across eight models and two datasets, explanations did not reliably track the visual evidence affecting predictions.
Prime Intellect (@PrimeIntellect) on X
Thank you NVIDIA for delivering us the new Vera CPU In March we had early access to benchmark sandboxes on Vera. Now we’ll be in the first wave to deploy on the architecture and leverage Vera’s robust dynamic memory latency for agentic workloads. Ideal for Prime Sandboxes
microagi (@microagi) on X
We’re excited to welcome @animesh_garg as our new Chief Research Officer at @microagi, with the team from @usablerobots joining him. He has spent most of his career working on one of the hardest problems in robotics: getting robots out of the lab and into the real world. Frontier AI now works. Ha
General Intuition (@gen_intuition) on X
General Intuition has raised another $220M at a $6.2B valuation as we start making our models available. These are an entirely new class of foundation models, trained on billions of action-labeled videos, capable of acting in realtime in never seen before environments. 1/6
OpenClaw🦞 (@openclaw) on X
Today we’re announcing OpenClaw Enterprise In collaboration with @RedHat , @nvidia and @OpenAI the OpenClaw Foundation is open sourcing a powerful enterprise control plane for persistent agents OpenClaw Enterprise is built to run on your own infrastructure and will always be free for an organiz
OpenAI Developers (@OpenAIDevs) on X
Codex CLI just got a major refresh, with a new look and powerful new capabilities. We’re continuing to invest in Codex CLI for developers who live in the command line. New full-screen interface, better readability, and ways to manage parallel work, all inside your terminal.
CMU Robotics Institute (@CMU_Robotics) on X
Long-Horizon Adaptive Manipulation Planning (LAMP) helps robots plan not only where an object needs to go, but also how they can work together to get it there🤖🤝 MSR student and project co-lead Shuai Zhou presented this research at #IROS2026 yesterday! https://t.co/1Q5LX6Xc3j
ElevenLabs (@ElevenLabs) on X
Eleven v4 Turbo is now available in ElevenAgents, bringing the #1 rated quality of Eleven v4 to live conversations with a median inference latency of ~100 ms. ElevenAgents can now power even more personal, empathetic customer experiences across industries.
bolt.new (@boltdotnew) on X
Today, Bolt acquired Dokai. Two teams, one belief: natural language can build and run software. Welcome to Bolt, Dokai. Learn more about Dokai and the acquisition here: https://t.co/HQoEkGlNpz
Inception (@_inception_ai) on X
Today, we’re introducing Mercury Voice, a diffusion LLM specialized for agentic voice applications. Mercury Voice delivers 2x+ lower latency than models including GPT-6 Luna, Gemma 4 31B, and Claude Haiku 4.5, while beating them on a range of voice benchmarks. Enterprise customers interested in
Cerebras (@cerebras) on X
Under the multiyear agreement, General Compute will deploy Cerebras systems so its customers can deliver ultrafast inference to developers building agentic applications.
Robotic Systems Lab (@leggedrobotics) on X
We achieved high-speed rough-terrain locomotion on ANYmal with an automatic curriculum. LP-ACRL automatically samples terrain types, levels, and velocity commands at the correct time based on policy performance, without predefining an order. 🔗https://t.co/ip3pQ0z2qF 🧵RAL 2026
Perplexity (@perplexity_ai) on X
Introducing Automations in Perplexity Computer. Automations are for ongoing work. They can take action in response to event-based triggers or on a schedule. Automations work with your memory, skills, and connected apps like Slack, Gmail, Outlook, and Linear.
ServiceNow AI Research (@ServiceNowRSRCH) on X
At DocInsights @ EMNLP 2026: 📄 Splitting Documents at Lower Cost: Multi-Split Boundary Decisions for LLM-Based Page Stream Segmentation 🔗 https://t.co/agAkVupAql 📄 Text, Pixels, or Both? also independently accepted here 🔗 https://t.co/NHuozBnjXZ
Liquid AI (@liquidai) on X
Announcing d1, our first decision model. It’s the first model to outperform Jev on @huggingface’s Decision Index. > wins on multilingual evals > more robust against prompt injection > handles longer inputs more effectively > built for fast, structured decision-making in software environments >
Congresswoman Yassamin Ansari (@RepYassAnsari) on X
Proud to be among the first supporters of @RepCasar’s critical bill to ban superintelligence and pause advanced AI development. Congressional Republicans need to return to work immediately and put this bill up for a vote. Congress needs to take charge in regulating dangerous frontier AI tech which
Ortet (@OrtetAI) on X
We’re excited to announce Ortet, a frontier AI lab for health. A fundamental gap in care today is that the industry and its scientific tools focus on discovering what works at the population level, while it is individual patients we must focus on to deliver the best care. Individualized care requi
AMD (@AMD) on X
We’re excited to announce a definitive agreement to acquire @theworldlabs, bringing a world-class team of researchers and model experts to AMD who will help shape future AI infrastructure and strengthen the open AI ecosystem. More on the news: https://t.co/cSWKwCN2S7 We look forward to welcoming
Henry de Zoete (@HZoete) on X
In addition to these open roles, it’s been a pleasure to welcome and work with AISI’s recently appointed Research Director, M. Pawan Kumar, who joined us from Google DeepMind. He was part of the team behind AlphaEvolve, a coding agent for algorithmic discovery. Alongside Research Director Chris S
LTX.io (@ltx_io) on X
Introducing Native Resolution, the first VFX Week drop, plus two new IC-LoRAs: Refine and Restore. Native Resolution runs AI edits on 4K and 8K footage and long shots. Overlapping tiles keep the fine detail, all on a single GPU. 🧵
Shoubin Yu (@shoubin621) on X
🎉 Excited to announce that AVIC has been accepted to #NeurIPS2026! More imagination is not always better for spatial reasoning, the key is knowing when to imagine and how much to imagine. AVIC learns an RL policy that adaptively invokes a video world model and controls how much to imagine, balanc
Valentin De Bortoli (@ValentinDeBort1) on X
The paper represents intermediate token beliefs as probability distributions rather than repeatedly selecting a single category. It reports gains against selected discrete-diffusion baselines on code and math tasks.
Dyna Robotics (@DynaRobotics) on X
We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy. It combines our brand-new semi-humanoid hardware with an agentic system built around Dyna-2 to handle ultra-long real-world workflows. Here is an uncut footage of Dyna-2.1 completing
OpenAI (@OpenAI) on X
This is Ultrafast. Our premium speed tier, Ultrafast offers up to 8x faster token generation (300 tokens per second) in Codex and up to 6x in the API.
OpenAI (@OpenAI) on X
Dots introduces persistent Astra agents with cloud computers and connected applications. Access is rolling out by plan and region, with separate permissions for local devices and external services.
Ben Mildenhall (@BenMildenhall) on X
My hope for @theworldlabs has always been to build technology that shows the world something it’s never seen before, and put that technology into peoples’ hands. We’ve now done this multiple times, first with our Marble product release last year, and most recently with our Atlas model, which has j
Robert Scoble (@Scobleizer) on X
24 months from start to acquisition by AMD. $8+ billion. When I was among the first to tell you about world models I had no idea it would go so fast. Congrats @theworldlabs. Insane week and it still is Monday. Tomorrow will we get to see OpenAI’s secret hardware project? It says it has 20 laun
Sauvik Das (@scyrusk) on X
New #AIES2026 paper led by @hankhplee: Agentic AI devs focus on product and business risks over downstream risks like job displacement and privacy. Lacking mature controls, they contain risk by constraining the same capabilities that make agents useful. https://t.co/qZcoh3mX6Y
San Francisco | Claude for Physical AI Agents · Luma
September 30, 2026, 5:30–9 PM PDT at Vizcom HQ, 488 Bryant Street, San Francisco. The event is full and accepting waitlist requests; only confirmed guests may attend.
Sakana AI (@SakanaAILabs) on X
Introducing “Scaling In-Context Imitation Learning” (SAIL) to be presented at #IROS2026. This work is a collaboration between Sakana AI and the University of Tokyo. Blog: https://t.co/b6xgwIMkUP What does a robot need before it can tackle a new task? Teaching a robot something new usually starts