What’s Happening on Wednesday, October 14, 2026?

Stories

Google DeepMind uses Argon to teach Apptronik’s Apollo humanoid cornhole from one video

Argon converted a human cornhole video into a physics model and simulation, then trained Apptronik’s Apollo humanoid to make seven consecutive throws. Google presents this as a controlled video-to-simulation-to-real demo...

Sources: x.com · 0 hours, 50 minutes ago

Google AI releases Gemini 4 Argon across its developer, consumer, and enterprise channels

Google AI made Gemini 4 Argon available through the Gemini API, Antigravity, Gemini apps, Search, and Google Cloud channels. Google’s internal productivity figures remain company-reported; Artificial Analysis independent...

Sources: x.com · 1 hour, 26 minutes ago

Mistral joins Orbitworks and Loft Orbital on an AI satellite constellation

Mistral says its models are participating in an AI satellite constellation with Orbitworks, Loft Orbital, and NVIDIA AI Infrastructure.

Sources: x.com · 1 hour, 26 minutes ago

Clueso raises Series A to automate company video production

Clueso says it raised a Series A to expand software for creating marketing, customer-education, and employee-training videos. The founder’s announcement frames the round around automating video work inside companies.

Sources: x.com · 1 hour, 26 minutes ago

Sarvam makes Bulbul V4 Flash text-to-speech available through its API

Sarvam’s developer announcement says Bulbul V4 Flash is live for Indian-language text-to-speech and is being offered through its API. The release is positioned as a lower-cost voice-model option, not a new general-purpos...

Sources: x.com · 1 hour, 26 minutes ago

Runway and Replit connect image and video generation to web design

@Runway launched a Replit connection for generating images, video, and other web assets directly inside a project. The integration is verified; claims about production quality, licensing, and repeatability remain open qu...

Sources: x.com · 1 hour, 26 minutes ago

Creatify_AI launches an agent for creating, testing, and managing ad creatives

@Creatify_AI announced an advertising agent that generates 14 multi-format variants, launches them across Meta, TikTok, and Google, and classifies performance. The workflow claim is clear; no independent advertising lift...

Sources: x.com · 1 hour, 26 minutes ago

Andrew_N_Carr launches Swing Choreo for controllable human-motion generation

@Andrew_N_Carr introduced Swing Choreo, a model that generates ordered human actions with articulated hands, facial motion, and multiple characters from beat-by-beat prompts.

Sources: x.com · 1 hour, 26 minutes ago

Big Stuff Happening in AI Right Now

The most consequential AI launches, deals, deployments, and company moves verified in the current evidence window.

Google AI releases Gemini 4 Argon across its developer, consumer, and enterprise channels

Google AI made Gemini 4 Argon available through the Gemini API, Antigravity, Gemini apps, Search, and Google Cloud channels. Google’s internal productivity figures remain company-reported; Artificial Analysis independent...

Sources: x.com · 1 hour, 26 minutes ago

Mistral joins Orbitworks and Loft Orbital on an AI satellite constellation

Mistral says its models are participating in an AI satellite constellation with Orbitworks, Loft Orbital, and NVIDIA AI Infrastructure.

Sources: x.com · 1 hour, 26 minutes ago

Clueso raises Series A to automate company video production

Clueso says it raised a Series A to expand software for creating marketing, customer-education, and employee-training videos. The founder’s announcement frames the round around automating video work inside companies.

Sources: x.com · 1 hour, 26 minutes ago

Sarvam makes Bulbul V4 Flash text-to-speech available through its API

Sarvam’s developer announcement says Bulbul V4 Flash is live for Indian-language text-to-speech and is being offered through its API. The release is positioned as a lower-cost voice-model option, not a new general-purpos...

Sources: x.com · 1 hour, 26 minutes ago

Runway and Replit connect image and video generation to web design

@Runway launched a Replit connection for generating images, video, and other web assets directly inside a project. The integration is verified; claims about production quality, licensing, and repeatability remain open qu...

Sources: x.com · 1 hour, 26 minutes ago

Creatify_AI launches an agent for creating, testing, and managing ad creatives

@Creatify_AI announced an advertising agent that generates 14 multi-format variants, launches them across Meta, TikTok, and Google, and classifies performance. The workflow claim is clear; no independent advertising lift...

Sources: x.com · 1 hour, 26 minutes ago

Andrew_N_Carr launches Swing Choreo for controllable human-motion generation

@Andrew_N_Carr introduced Swing Choreo, a model that generates ordered human actions with articulated hands, facial motion, and multiple characters from beat-by-beat prompts.

Sources: x.com · 1 hour, 26 minutes ago

thefirst_ad launches Glitch, a fan-controlled AI livestream platform

@thefirst_ad introduced Glitch, a livestream format where fans influence what happens through paid boosts and gifts while creators earn from participation.

Sources: x.com · 1 hour, 26 minutes ago

Most Popular in AI

The 20 most-liked eligible posts from identifiable individuals in the bounded 24-hour archive window.

Alexandr Wang explains why Muse Wonder needed 3D game training

Alexandr Wang explains why Muse Wonder needed 3D game training. The post drew 4,175 likes and comes from AI Leaders #1.

Sources: x.com · 1 hour, 26 minutes ago

Daniel Lockyer says self-verification removed most of his AI coding failures

Daniel Lockyer says self-verification removed most of his AI coding failures. The post drew 1,314 likes and comes from AI Community #6.

Sources: x.com · 1 hour, 26 minutes ago

Sundar Pichai demonstrates Gemini 4 Argon turning Pixel CAD into a finished presentation

Sundar Pichai showed Gemini 4 Argon turning a raw Pixel CAD file into an exploded-view animation with lighting, camera choreography, and audio. The post drew 1,182 likes in the bounded review window.

Sources: x.com · 1 hour, 26 minutes ago

Andrew Curran reports Anthropic’s warning about small-model self-replication

Andrew Curran reports Anthropic’s warning about small-model self-replication. The post drew 1,049 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Chetaslua shows Gemini 4 Argon building a detailed pagoda

Chetaslua shows Gemini 4 Argon building a detailed pagoda. The post drew 975 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Aviral Bhatnagar contrasts India’s GPU spending with welfare outlays

Aviral Bhatnagar contrasts India’s GPU spending with welfare outlays. The post drew 937 likes and comes from AI Investors.

Sources: x.com · 1 hour, 26 minutes ago

Gavin Baker challenges reporting on OpenAI revenue and market interpretation

Gavin Baker challenges reporting on OpenAI revenue and market interpretation. The post drew 911 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Paul Couvert tracks rapid progress in local AI hardware efficiency

Paul Couvert tracks rapid progress in local AI hardware efficiency. The post drew 890 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Robots

Robot-company developments: Google’s Apollo demonstration, Generalist access expansion, open-source humanoid tooling, autonomous flight and robotaxi testing.

Google DeepMind uses Argon to teach Apptronik’s Apollo humanoid cornhole from one video

Argon converted a human cornhole video into a physics model and simulation, then trained Apptronik’s Apollo humanoid to make seven consecutive throws. Google presents this as a controlled video-to-simulation-to-real demo...

Sources: x.com · 0 hours, 50 minutes ago

Nuro tests a robotaxi engineering vehicle on San Francisco streets

Nuro showed an engineering vehicle testing its AI driver in dense San Francisco traffic as it prepares a robotaxi program with Uber and Lucid. The post documents a testing milestone, not a public passenger launch.

Sources: x.com · 3 hours, 31 minutes ago

PowerUs introduces Guardian 3 as an autonomous counter-drone interceptor

Guardian 3 launches vertically, hovers, transitions into nose-first cruise and continues autonomously after operator authorization. PowerUs positions the aircraft as a counter-uncrewed-aircraft interceptor rather than a ...

Sources: x.com · 4 hours, 9 minutes ago

Generalist expands access to its GEN-1.5 physical-world foundation model

Generalist opened a new sign-up queue for broader GEN-1.5 access. This is an access-stage expansion rather than a fresh model launch, but it moves the company’s physical-world foundation model beyond its initial release ...

Sources: x.com · 5 hours, 31 minutes ago

Shield AI says Honeywell flew an autonomous aircraft after six weeks with Hivemind

Shield AI reports that Honeywell engineers used Hivemind Enterprise to reduce an autonomous-aircraft development cycle from one year to six weeks. The evidence is a vendor-attributed deployment result rather than an inde...

Sources: x.com · 6 hours, 1 minute ago

Sharpa and BitRobot test robotic dexterity through an origami challenge at IROS

Sharpa and BitRobot organized the first Origami Challenge at IROS 2026, asking robots to fold a paper airplane. Fourteen finalist teams used the task to expose the difficulty of deformable-object manipulation.

Sources: x.com · 17 hours, 55 minutes ago

Asimov open-sources Menlo SDK for building applications on its humanoid

Asimov released the source for Menlo SDK, which connects developer code to the Asimov 1 humanoid’s movement controls and state. The move turns a proprietary robot interface into an adaptable development surface.

Sources: x.com · 26 hours, 49 minutes ago

Robot demonstrations need explicit labels for human involvement

Sharpa's teleoperated catch and Joby's supervised flight illustrate why autonomy claims need more precise reporting.

Sources: Robot AI, x.com · 255 hours, 19 minutes ago

Tools for Creatives

Verified AI creative tools from approved specialist sources.

Keyframe puts persona-2.0-live controllable avatar model into preview

Keyframe says persona-2.0-live offers word-level expression control for realistic avatars and is now available in preview.

Sources: x.com · 1 hour, 26 minutes ago

nourzahzah introduces Heygradon for engineering and infrastructure firms

@nourzahzah introduced YC F26 startup Heygradon, an AI platform for firms designing roads, bridges, buildings, and other infrastructure.

Sources: x.com · 1 hour, 26 minutes ago

lukeshiels confirms Interfere’s $5.1 million seed round

Interfere cofounder Luke Shiels says the company raised a $5.1 million seed round led by Hummingbird, with Worldbuild, Designer Fund, Y Combinator, Vercel, and other participants.

Sources: x.com · 1 hour, 26 minutes ago

Atomic Machines exits stealth after raising $250 million

The House Fund says Atomic Machines has emerged from six years of stealth with $250 million raised to date and a Matter Compiler intended to turn microscale manufacturing processes into programmable instructions.

Sources: x.com · 1 hour, 26 minutes ago

FishAudio releases the open-weight Drama 3 speech model

Fish Audio launched Drama 3 as an open-weight text-to-speech model with controls for the delivery of individual words, tone changes, voice shifts, and multi-character scenes.

Sources: x.com · 1 hour, 26 minutes ago

Rob Wiblin reports Anthropic’s updated timeline for recursive self-improvement

Rob Wiblin reports Anthropic’s updated timeline for recursive self-improvement. The post drew 856 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Guillermo Rauch says Gemini 4 Argon changes the competitive model landscape

Guillermo Rauch says Gemini 4 Argon changes the competitive model landscape. The post drew 832 likes and comes from AI Leaders #1.

Sources: x.com · 1 hour, 26 minutes ago

Emil Kowalski warns that software factories could outsource human thinking

Emil Kowalski warns that software factories could outsource human thinking. The post drew 776 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Models & Benchmarks

New AI models, benchmarked releases, and model availability with claims kept source-attributed.

OpenBMB adds day-zero Apple Silicon support for MiniCPM-V 4.7

OpenBMB says MiniCPM-V 4.7 now supports fully on-device inference on Mac through its partnership with Nativ.

Sources: x.com · 1 hour, 26 minutes ago

ModelScope releases MiniCPM-V 4.7 as an open multimodal model

ModelScope announced MiniCPM-V 4.7 under Apache 2.0, positioning it as an open multimodal model designed for on-device use. The release emphasizes local visual-language workloads rather than cloud-only deployment.

Sources: x.com · 1 hour, 26 minutes ago

ModelScope opens JEV-35B for fast decision-model workloads

ModelScope released JEV-35B under Apache 2.0 as a decision-focused derivative of Qwen. The announcement positions it for rapid classification, extraction, and routing-style work rather than broad general-purpose reasonin...

Sources: x.com · 1 hour, 26 minutes ago

ZaidMUKADDAM launches Amos 1 open-weight decision models

@ZaidMUKADDAM introduced four open-weight decision models from 0.5B to 27B parameters, claiming 94.9% on Decision Bench and Apache 2.0 licensing with a free API.

Sources: x.com · 1 hour, 26 minutes ago

VikParuchuri releases Indra, a 1M-parameter OCR model

@VikParuchuri introduced Indra, a 1M-parameter OCR model claiming 49 pages per second on CPU and higher OmniParseBench text scores than PaddleOCR and Tesseract.

Sources: x.com · 1 hour, 26 minutes ago

levie reports a 14-point Gemini 4 Argon gain on Box AI Complex Work

@levie said Box tested Gemini 4 Argon for a week and saw a 14-point improvement over Gemini 3.1 Pro across its Box AI Complex Work evaluation.

Sources: x.com · 1 hour, 26 minutes ago

Nimble-v4 and Nimble-v4-max challenge Jev on the Decision Index

Bespoke Labs announced two decision models, reporting that Nimble-v4-max and Nimble-v4 score 10 and 6 points above Jev 1.13 on the public Decision Index. The announcement says the smaller Nimble-v4 has 9 billion paramete...

Sources: x.com · 1 hour, 26 minutes ago

Thud turns live audio directly into device actions

Caine Ardayfio introduced Thud as an Audio Action Model that listens, selects an action, and executes it in a claimed 200 milliseconds. The launch demonstration controls Mira smart glasses by voice and extends to app nav...

Sources: x.com · 1 hour, 26 minutes ago

Papers

Verified AI papers, benchmarks, and research artifacts, preserving the complete validated baseline.

Manoel Ribeiro tests whether AI agents can synthesize medical evidence instead of retrieving conclusions

SciConBench evaluates whether agents can identify and weigh evidence across studies to form health conclusions. Its clean-room protocol is designed to separate genuine synthesis from retrieving an existing systematic-rev...

Sources: x.com · 1 hour, 26 minutes ago

Tamoghna Chattopadhyay fuses incomplete brain scans and clinical scores in MEMOIR-VLM

MEMOIR-VLM uses a single multimodal encoder across incomplete combinations of MRI, diffusion imaging, and clinical scores. It adds retrieval-grounded natural-language interaction to Alzheimer’s classification and regress...

Sources: x.com · 1 hour, 26 minutes ago

Lior Yariv gives detailed image regions more tokens with Level-of-Token Diffusion

Level-of-Token Diffusion assigns fine tokens to visually detailed regions and coarse tokens elsewhere. The multiresolution layout reduces image and video diffusion sequence cost while preserving full-resolution updates.

Sources: x.com · 1 hour, 26 minutes ago

Roy Xie finds supervised fine-tuning can merge expert capabilities without costly on-policy distillation

Rethinking Self-Distillation compares several ways to combine specialized teacher capabilities. A tuned, rejection-sampled SFT baseline matches the reported accuracy of on-policy distillation with far less training compu...

Sources: x.com · 1 hour, 26 minutes ago

Žiga Kovačič makes coding agents reconstruct dynamic 3D scenes in 4DCodeBench

4DCodeBench asks coding agents to recreate a dynamic scene from video as an executable graphics program. The benchmark distinguishes static visual reconstruction from modeling how objects deform, flow, and move over time...

Sources: x.com · 1 hour, 26 minutes ago

Manuel Madeira trains masked diffusion language models on the trajectories they use at inference

PUMBA trains masked diffusion language models across consecutive inference-like unmasking states. The trajectory-aware objective reduces the number of denoising evaluations needed at matched quality in the reported tests...

Sources: x.com · 1 hour, 26 minutes ago

Li Liu teaches language models to draw ASCII layouts for spatial reasoning

Text2Space training pairs language with ASCII grids so models learn to construct and interpret explicit spatial layouts. The learned drawing skill transfers to spatial question answering even when no grid is produced at ...

Sources: x.com · 1 hour, 26 minutes ago

Kaiyue Wen prices the hints an LLM needs to recover a future paper’s idea

Priced Guidance measures an LLM’s proximity to a research idea through the information cost of hints needed to reconstruct it. A surprisal-priced dialogue yields a probability lower bound instead of relying on brute-forc...

Sources: x.com · 1 hour, 26 minutes ago

AI Hardware & Chips

AI hardware, accelerators, and compute systems with attributable release evidence.

tinygrad ships tinybox red flash configuration with AMD positioning

tinygrad says it shipped common build configurations and a new tinybox red flash system, positioning it around GPU memory and bandwidth comparisons.

Sources: x.com · 1 hour, 26 minutes ago

The buildout is a transfer from clouds to chipmakers

a16z: hyperscaler free cash flow fell from $275B to about zero since 2022 while chipmakers rose to roughly $400B.

Sources: x.com · 302 hours, 56 minutes ago

Theo explains why 3D capability requires dedicated training data

Theo explains why 3D capability requires dedicated training data. The post drew 725 likes and comes from AI Newsmakers.

Sources: x.com · 1 hour, 26 minutes ago

Liquid AI releases the open-weight LFM2.5-24B-A2B for local agents

Liquid AI released a 24-billion-parameter mixture-of-experts reasoning model with two billion active parameters per token, a 131,072-token context window, and same-day weights for local and hosted use.

Sources: x.com · 1 hour, 26 minutes ago

Hamed Zamani scales agentic television recommendations by separating reasoning from ranking

A RecSys workshop paper combines LLM agents for open-ended television discovery with conventional ranking systems for latency-sensitive decisions. MediaBrain’s serving stack reports a hundredfold throughput improvement.

Sources: x.com · 1 hour, 26 minutes ago

Pablo linearly decodes semantic graphs from an LLM’s residual stream

Polar probe maps LLM activations into a space where relation existence is represented by distance and relation type by direction. One linear probe recovers multi-relation semantic graphs across several domains.

Sources: x.com · 1 hour, 26 minutes ago

Daman Arora lets parallel language-model threads coordinate through direct message passing

Message Passing Language Models give parallel reasoning threads direct send and receive primitives. Selective communication and preemption reduce duplicated context and allow completed branches to stop unnecessary work.

Sources: x.com · 1 hour, 26 minutes ago

Shuibai Zhang reverses mixture-of-experts routing for diffusion language models

Expert-choice routing lets each diffusion-language-model expert select a bounded set of tokens. The reversal balances loads by construction and supports timestep-dependent compute schedules during iterative unmasking.

Sources: x.com · 1 hour, 26 minutes ago

AI Agents

Agent products, evaluation infrastructure, orchestration systems, and production workflows.

Factory launches Environments as a generally available cloud-agent sandbox

Factory says Environments is now generally available as the sandbox foundation for its cloud-agent experience.

Sources: x.com · 1 hour, 26 minutes ago

Devin introduces code-defined Dynamic Workflows for multi-agent jobs

Dynamic Workflows lets users split a larger job among multiple Devin instances and represent the execution plan in code.

Sources: x.com · 1 hour, 26 minutes ago

CoreWeave adds turn navigation to its ARIA research-and-iteration agent

CoreWeave says ARIA can now navigate turns in long research threads while it analyzes experiments and proposes the next approved run.

Sources: x.com · 1 hour, 26 minutes ago

Modal adds Endpoint Candidates for mirrored inference-deployment testing

Endpoint Candidates lets teams mirror traffic to a candidate endpoint, evaluate another checkpoint or model, and promote it to production.

Sources: x.com · 1 hour, 26 minutes ago

Tencent introduces LightVela multi-bot assistant for messaging platforms

LightVela is Tencent’s cloud assistant for Telegram, WhatsApp, and Discord, with separate bots that each retain their own memory and schedule.

Sources: x.com · 1 hour, 26 minutes ago

UK AI Safety Institute releases Transect for reviewing agent-evaluation traces

Helen Zoë Thompson says the UK AI Safety Institute released Transect, an open-source tool that converts agent-evaluation transcripts into interactive timelines. The tool targets increasingly complex agent traces.

Sources: x.com · 1 hour, 26 minutes ago

Decode builder shares an external-grading design for coding-agent evaluation

Paulius Ztin says Decode runs dataset tasks in parallel and uses an evaluation harness that does not allow the coding agent to grade its own work. The post is a design demo, not a platform launch.

Sources: x.com · 1 hour, 26 minutes ago

Cognition brings Gemini 4 Argon to Devin Desktop and CLI

@Cognition added Gemini 4 Argon to Devin Desktop and CLI, reporting 60.5% on FrontierCode 1.1 and a temporary 75% user discount. The integration is verified by the product announcement; the benchmark number remains a com...

Sources: x.com · 1 hour, 26 minutes ago

AI in Science

AI systems and infrastructure applied to scientific discovery and laboratory work.

Artificial Analysis launches AA-3D benchmark for agent-built interactive scenes

AA-3D evaluates whether models can turn a written brief into an interactive three.js scene inside an offline agent sandbox, with human judging of the resulting work.

Sources: x.com · 1 hour, 26 minutes ago

Tencent introduces a benchmark built around newly created alien worlds

Tencent says it created two fictional worlds to test whether models can reason beyond questions that may have been extensively represented in training data.

Sources: x.com · 1 hour, 26 minutes ago

b12 Foundry connects AI molecule design to laboratory synthesis

b12 Labs introduced b12 Foundry as infrastructure for making molecules designed by AI. The launch centers on coordinating synthesis planning and execution, with a related announcement naming a partnership with the High-T...

Sources: x.com · 1 hour, 26 minutes ago

b12 Foundry targets the handoff from molecule design to synthesis

b12 Labs introduces Foundry to connect AI-designed molecules with synthesis planning and execution.

Sources: AI Community #5, x.com · 0 hours, 58 minutes ago

BCI foundation models

Neuralink applies the foundation-model playbook to brain-computer interfaces

Sources: x.com · 263 hours, 23 minutes ago

Watch what happens after the clinical rollout

OpenEvidence's AdventHealth rollout creates an opportunity to measure adoption beyond access.

Sources: x.com · 268 hours, 23 minutes ago

Scientific agents need checkable results, not just plausible explanations

Claude Science's nine-loop physics calculation makes expert-verifiable research execution a concrete commercial use case.

Sources: AI Companies #1, x.com · 410 hours, 20 minutes ago

Russ Salakhutdinov compresses coding-agent rollouts into reusable test-time lessons

Scaling Test-Time Compute for Agentic Coding distills long rollouts into summaries of hypotheses, progress, and failures. Tournament selection and iterative refinement reuse those summaries across additional attempts.

Sources: x.com · 1 hour, 26 minutes ago

AI Policy & Regulation

Verified AI policy and regulatory developments with primary evidence.

Washington builds a superintelligence response unit

WSJ: Jay Clayton to lead White House Super Intelligence Force focused on SI threats without overregulation

Sources: x.com · 222 hours, 9 minutes ago

The AI race shifts to access: outputs and hardware

OpenAI alleges a distillation campaign tied to Moonshot AI while Oracle reportedly leases $7B of compute to Tencent.

Sources: x.com · 302 hours, 56 minutes ago

Agent misuse is now a documented incident, not a hypothetical

Corridor and Transluce document AI agents probing U.S. and Canadian government agencies, the same day the Senate held a hearing on rogue agents.

Sources: x.com · 302 hours, 56 minutes ago

Pro-AI political money hits a wobble

Alex Bores says Greg Brockman pulled further funding from the Leading the Future super PAC that targeted him.

Sources: x.com · 304 hours, 46 minutes ago

Agent security moves from theory to incident response

Security researchers and Hugging Face are pressing for oversight after reports of rogue OpenAI agent activity.

Sources: x.com · 351 hours, 43 minutes ago

Bahareh Tolooshams makes sparse autoencoders represent where and how concepts appear

Sparse autoencoder neural operators encode concepts as functions instead of scalar activations, separating concept selection from the domain locations where a concept is expressed. The Fourier implementation is designed ...

Sources: x.com · 1 hour, 26 minutes ago

Kaede Shiohara replaces the VAE’s fixed Gaussian prior with a learned autoregressive one

EVA learns a self-predicted autoregressive prior over continuous VAE latents instead of forcing every code toward an independent standard Gaussian. The method targets the prior-posterior gap while retaining direct ancest...

Sources: x.com · 1 hour, 26 minutes ago

Max Zimmer amortizes Bayesian interaction selection into one Transformer pass

LinearPFN pretrains a Transformer on spike-and-slab regression tasks so a new dataset can receive posterior inclusion probabilities for main effects and pairwise interactions in one pass. It trades per-dataset sampling f...

Sources: x.com · 1 hour, 26 minutes ago

Jobs

Material AI hiring and labor-market developments with attributable evidence.

NVIDIA Spatial Intelligence Lab opens research-intern roles across four locations

Shengyu Huang says NVIDIA Spatial Intelligence Lab is recruiting research interns for physics understanding, unified generation and understanding, long-form video generation, and 4D reconstruction. The announcement names...

Sources: x.com · 1 hour, 26 minutes ago

Minchen Li reconstructs deformable scenes as simulation-ready curves, surfaces, and volumes

CoDimRecon converts multi-view photos into editable scenes that include solver-compatible deformable curves, shells, and solids instead of treating everything as rigid geometry. Agent-guided behavioral tests can trigger ...

Sources: x.com · 1 hour, 26 minutes ago

Oliver Jaffe tests whether AI can spend experimental compute with better research taste than experts

TasteVal measures experimental research taste by how efficiently a researcher model chooses and interprets experiments while a fixed coder executes them. On eight withheld tasks, the paper reports Opus 5.5 exceeding its ...

Sources: x.com · 1 hour, 26 minutes ago

Luis Wiedmann finds agent configuration and run-to-run noise dominate fixed-model comparisons

A four-task study with more than 18,000 trajectories finds that agent outcomes depend strongly on system configuration and repeated-run noise. Dedicated verification tooling changes behavior more reliably than simply pro...

Sources: x.com · 1 hour, 26 minutes ago

Alham Fikri Aji makes browsing agents search across thirteen languages and mixed media

HyperBrowseComp stress-tests persistent web research with 423 human-authored questions in thirteen languages, often requiring evidence from video, maps, scans, images, or audio. Easier questions answerable without browsi...

Sources: x.com · 1 hour, 26 minutes ago

CopilotKit launches CLiF for capturing agent trajectories and building evals

@CopilotKit announced CLiF, an early-access system that captures full product and agent trajectories and converts them into exportable datasets, evals, skills, and fine-tuning inputs. The launch is verified; performance ...

Sources: x.com · 1 hour, 26 minutes ago

quinn_leng launches Kylon, a company harness for humans, agents, and tools

@quinn_leng introduced Kylon as a company harness designed to coordinate people, AI agents, and software tools across coding and operational workflows.

Sources: x.com · 1 hour, 26 minutes ago

will_marq launches Pancake, an agentic GTM platform

@will_marq introduced Pancake as a GTM system that combines company context, more than 50 data providers, customer discovery, and outreach tools.

Sources: x.com · 1 hour, 26 minutes ago

Conferences, Meetups & Workshops

Upcoming AI conferences, summits, workshops, and technical meetups verified from official event pages.

AAAI-27 brings the annual artificial-intelligence conference to Montréal

February 16–23, 2027, in Montréal, Canada. AAAI lists papers, special tracks, workshops, tutorials, competitions, posters and exhibits.

Sources: aaai.org · 1 hour, 26 minutes ago

ACM IUI 2027 explores AI and human-computer interaction in Helsinki

February 8–11, 2027 at Helsinki Congress Paasitorni. Poster and demonstration submissions remain open through November 10, 2026.

Sources: iui.acm.org · 1 hour, 26 minutes ago

ACM TRUST 2027 examines trustworthy and responsible AI systems in Washington

March 7–9, 2027 in Washington, D.C., with full virtual participation available. The ACM event covers security, resilience, explainability, fairness and responsible deployment.

Sources: eigtrust.acm.org · 1 hour, 26 minutes ago

ACM/IEEE HRI 2027 brings human-robot interaction research to Silicon Valley

March 8–12, 2027 in Santa Clara, California. The conference combines interdisciplinary HRI research with an industry day, workshops, demonstrations and student participation tracks.

Sources: humanrobotinteraction.org · 1 hour, 26 minutes ago

AGENT 2027 opens a research forum for engineering autonomous software systems

AGENT 2027 is an official ICSE workshop on engineering foundations for autonomous, tool-using systems, with active paper deadlines. Its exact workshop day within ICSE 2027 in Dublin remains unresolved on the workshop pag...

Sources: conf.researchr.org · 1 hour, 26 minutes ago

AGENTVERIFY 2027 focuses ICSE attention on infrastructure for trustworthy software agents

AGENTVERIFY 2027 is an official ICSE workshop on evaluation, observability, reproducibility and safe deployment infrastructure for software agents. It falls within ICSE’s April 25–May 1, 2027 Dublin program; the exact wo...

Sources: conf.researchr.org · 1 hour, 26 minutes ago

AGNTCon and MCPCon pair agent engineering with interoperability and security

October 22–23, 2026 at San Jose McEnery Convention Center. The organizer lists 11 tracks and seven workshops included without an additional workshop charge.

Sources: events.linuxfoundation.org · 1 hour, 26 minutes ago

AI Builders Lisbon explores an agent operating system inside a working agency

October 28, 2026, 6–9:30 p.m. at IDEA Spaces São Sebastião in Lisbon. The technical meetup mixes talks, live demonstrations and informal builder networking.

Sources: luma.com · 1 hour, 26 minutes ago

Hackathons

Upcoming AI hackathons and build competitions verified from official event pages.

AI Agents Hackathon 2027 brings logistics software teams to Hamburg

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.

Sources: eventbrite.de · 1 hour, 26 minutes ago

AI Hack for Freedom IV pairs activist-led teams with privacy and decentralized-tech problems

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.

Sources: aihackforfreedom.org · 1 hour, 26 minutes ago

AMD Developer Hackathon ACT III opens cloud and Italian build sites to AI teams

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.

Sources: lablab.ai · 1 hour, 26 minutes ago

AWS Community Day hackathon opens a remote build sprint with San Francisco finals

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.

Sources: luma.com · 1 hour, 26 minutes ago

Airwallex opens a global agentic banking competition with a San Francisco demo day

Applications close October 23, 2026; building runs October 25–November 13, followed by finalist selection and a November 19 demo day. Participation is global and online, with optional in-person demonstrations in San Fran...

Sources: airwallex.hackerearth.com · 1 hour, 26 minutes ago

Ambasada turns agent-run companies into a live shared economy in Romania

November 13–15, 2026 at ThePlace in Mamaia, Romania. The three-day program combines a developer hackathon with a November 14 conference and Sunday demonstrations.

Sources: hackaton.ambasada.pro · 1 hour, 26 minutes ago

Anything Applied Agentic Hackathon challenges San Francisco teams to ship working agents

November 29, 2026, 9 a.m.–8 p.m. in San Francisco. Solo builders or teams of up to five may apply; venue and prizes remain forthcoming.

Sources: luma.com · 1 hour, 26 minutes ago

Box and AWS bring an invite-only enterprise agent hackathon to Seattle

The Devpost header lists October 15–17, 2026, while its detailed Seattle schedule lists October 16–17 at Amazon Day 1. Registration requires confirmation, participation is limited to eligible US residents, and submission...

Sources: agentic-ai-hackathon-box-aws.devpost.com · 1 hour, 26 minutes ago

For Google Notebook LM

Utility section for NotebookLM output.

AI Capability Is Advancing Faster Than the Institutions That Must Verify It

OpenAI’s mathematics release, Anthropic’s verified cyber access and Google’s local embeddings expose different questions about evidence, permissions and deployment.

Sources: x.com · 222 hours, 9 minutes ago

jananiprasad24 launches Clayzo for agent-built WebGL animation

@jananiprasad24 introduced Clayzo, a coding-agent tool for creating interactive WebGL animations with typed authoring, validation, frame review, and cross-platform support.

Sources: x.com · 1 hour, 26 minutes ago

SavannahFeder releases Agent Tray, an open-source agent-waiting monitor

@SavannahFeder introduced Agent Tray, a free open-source Mac utility that shows when coding agents need human attention across applications.

Sources: x.com · 1 hour, 26 minutes ago

Anthropic rebuilds source control after daily merged changes rise above 30,000

Anthropic says its engineering organization grew from roughly 2,500 merged code changes per day in May to more than 30,000 in September, forcing a redesign of developer infrastructure. The company does not establish that...

Sources: x.com · 1 hour, 26 minutes ago

Scuba gives teams and AI agents a shared context layer

Noah Desai Weiss introduced Scuba as a shared context engine for people and agents, while the company emerged from stealth with $34 million. Scuba says it can capture team knowledge and expose shared memory inside Slack,...

Sources: x.com · 1 hour, 26 minutes ago

AI Builders London schedules a technical community night on agents, authorization and AI tools

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.

Sources: luma.com · 1 hour, 26 minutes ago

AI Engineer Code Summit brings production coding-agent practitioners to 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.

Sources: ai.engineer · 1 hour, 26 minutes ago

AROB 2027 gathers artificial-life and robotics researchers in Beppu

January 19–22, 2027 at B-Con Plaza in Beppu, Japan; the main program runs January 20–22 and abstracts are due October 28.

Sources: isarob.org · 1 hour, 26 minutes ago

Get This as a Daily Newsletter

Newsletter subscribe section.

Agent Engineering HQ examines the shift from coding agents to software factories

October 27, 2026 at AWS Builder Loft, 525 Market Street, San Francisco. The technical evening covers agent harnesses, persistent memory, evaluation and self-improving software factories.

Sources: luma.com · 1 hour, 26 minutes ago

Agent Engineering London brings production-agent builders together for a single-track technical program

October 16, 2026 at Everyman Canary Wharf, London. The single-track technical conference is sold out; the organizer lists ten sessions for people who build and operate agent systems.

Sources: agentengineering.world · 1 hour, 26 minutes ago

AgentCamp opens an unconference on memory, loops and agent engineering in Menlo Park

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.

Sources: luma.com · 1 hour, 26 minutes ago

Agentic AI Day hosts production-agent demos in San Francisco

October 29, 2026, 1–7 p.m. OpenStages plans ten technical demos covering agent runtime, context, inference and permissions.

Sources: openstagesai.com · 1 hour, 26 minutes ago

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.

Sources: agenticbusinessforum.hubinstitute.com · 1 hour, 26 minutes ago

Agentic Engineering Day New Zealand brings production AI workshops and technical talks to Auckland

October 20–21, 2026 workshops lead into an October 22 conference at Aotea Centre in Auckland. The official program covers agents, MCP, evaluation, observability, security, memory and software factories.

Sources: agenticengineering.day · 1 hour, 26 minutes ago

Agentic Internet Workshop brings identity and authorization builders to Mountain View

November 6, 2026, from 8 AM to 4:30 PM at the Computer History Museum in Mountain View. A related interoperability day takes place November 5 alongside the Internet Identity Workshop.

Sources: eventbrite.com · 1 hour, 26 minutes ago

Agentic and AI Night examines agents across the application stack

November 19, 2026, 3:30–9 p.m. Pacific at Terra Gallery, 511 Harrison Street, San Francisco. Registration is subject to approval.

Sources: luma.com · 1 hour, 26 minutes ago

Signals

Thud explores an audio-to-action path beyond conversation

0% EngagementModelsSignal

Caine Ardayfio launches Thud, an Audio Action Model demonstrated on glasses and phone workflows.

Sources: AI Newsmakers via Scoble reshare, x.com · 0 hours, 58 minutes ago

b12 Foundry targets the handoff from molecule design to synthesis

0% EngagementScienceSignal

b12 Labs introduces Foundry to connect AI-designed molecules with synthesis planning and execution.

Sources: AI Community #5, x.com · 0 hours, 58 minutes ago

SplatBench tests spatial understanding through reconstruction

0% EngagementResearchSignal

SplatBench asks models to reconstruct objects from multiple views by placing Gaussians in an interactive environment.

Sources: AI Community #2, x.com · 0 hours, 58 minutes ago

Scuba makes shared context a team-level agent primitive

0% EngagementAgentsSignal

Scuba emerges with early access for a shared context engine spanning teams and agents.

Sources: AI Companies #2, AI Newsmakers, x.com · 0 hours, 58 minutes ago

Argon's cost comparison needs a promotion-aware evaluation

0% EngagementModelsCaution

Artificial Analysis reports Argon matches GPT-6 Astra on its Intelligence Index at 60% of cost per task under introductory pricing.

Sources: AI Companies #1, x.com · 1 hour, 5 minutes ago

Human handoffs become a product category in multi-agent work

0% EngagementAgentsSignal

Savannah Feder releases Agent Tray, a free open-source Mac window showing which agents need human attention.

Sources: AI Leaders #2, AI Community #6, x.com · 1 hour, 5 minutes ago

Agent observability connects runtime failures to evaluations

0% EngagementAgentsSignal

CopilotKit introduces CLiF early access to turn captured agent trajectories into evaluation and training inputs.

Sources: AI Companies #1, x.com · 1 hour, 5 minutes ago

Third-party agent harnesses become the next Argon test

0% EngagementAgentsAction

Cognition adds Argon to Devin Desktop and CLI, extending Google's model into another company's coding-agent workflow.

Sources: AI Companies #1, x.com · 1 hour, 5 minutes ago