Edition: October 1, 2026

Researchers from Carnegie Mellon, MIT, NYU and Stanford have created an AI named Ataraxos that defeated top Stratego champion Pim Niemeijer 15‑1 with four draws, solving a long‑standing challenge in imperfect‑information games. The system was trained on modest hardware—just 16 GPUs and a few thousand dollars—demonstrating that even games with massive hidden information can be mastered without massive resources.

AllenAI unveiled Olmo-core 3, an open‑source training stack that dramatically improves the scalability of mixture‑of‑experts models, enabling trillion‑parameter MoEs with minimal throughput loss. By switching to distributed data parallelism and combining expert, pipeline, and optimizer parallelism, the system boosts token‑processing speed up to 2.7× over previous versions, lowering compute costs for large‑scale AI research.

OpenAI has dismissed three safety researchers after an internal probe found they mishandled confidential information, allegedly sharing it with an external AI safety group. The firings, reported by the Wall Street Journal, come amid broader concerns about OpenAI's security practices and recent incidents involving its AI systems.

OpenAI unveiled Dots, a new GPT‑6‑powered AI agent at its DevDay, positioning it as a direct competitor to Meta’s Muse platform. The colorful, customizable assistant promises advanced capabilities like building virtual worlds, but faces the challenge of matching Muse’s free, widely adopted service.

Researchers have unveiled an AI system that can reconstruct visual images from brain scans and predict brain activity from images, a step toward reading thoughts and potentially aiding locked‑in patients, though it raises privacy concerns. Meanwhile, startups are deploying small, distributed battery units in everyday settings to sidestep community resistance to large battery farms, offering a scalable way to bolster grid resilience.

Researchers at the Weizmann Institute have created an AI system that can reconstruct visual images from high‑resolution fMRI scans, accurately reproducing what a person was looking at and even predicting brain activity from images. While hailed as a breakthrough for neuroscience and potential communication aids for locked‑in patients, experts warn the technology could also enable non‑consensual mind‑reading, raising ethical concerns.

The article critiques the recent AI safety accord signed under President Trump, arguing that self‑regulation by AI firms is insufficient and likening it to a false analogy with seat‑belt legislation. It warns that without enforceable rules, risks—from accidental misinformation to broader existential threats—remain poorly managed, especially as global competition continues.

OpenAI has launched two new ChatGPT shopping tools: a virtual try‑on feature that lets users upload a selfie to see clothing and accessories on themselves, and a Favorites function to save products for later. Powered by the new ChatGPT Images 2.5 model, the updates aim to make the AI a more practical shopping assistant, competing with existing visual search services.

Google launched its first orbital compute satellite, equipped with a Tensor Processing Unit, to test AI workloads in space and assess the feasibility of future orbital data centers. The company’s white paper estimates that SpaceX’s Starship would need about 1,800 launches over the next decade to drive launch costs down enough for a network of 81 satellite data centers, highlighting the long‑term ambition and scale of the project.

Shopify launched Canvas, an AI‑driven site builder that lets merchants create and edit their online stores by chatting with the Sidekick assistant, which updates the actual store code in real time. The tool aims to simplify web design for non‑technical users, though its initial release lacks third‑party theme support and is desktop‑only.

AI startup Photon staged a tongue‑in‑cheek funeral for mobile apps, arguing agents will replace them, and has raised $4.5 million in seed funding to build a managed platform for developers to create messaging‑based agents across iMessage, WhatsApp and more. The company reports over 40,000 developer sign‑ups, 10× revenue growth in four months and a suite of tools aimed at enterprises in fintech, health and consumer AI.

The Verge’s Josh Dzieza details how Kevin O’Leary’s ambitious 40,000‑acre AI data‑center project in Utah, dubbed Stratos or Wonder Valley, collapsed after fierce local opposition, political backlash, and a defamation lawsuit over accusations of Chinese espionage. The saga highlights the growing tension between AI infrastructure expansion and community consent.

Satlyt, founded by ex‑Google and SpaceX product manager Rama Afullo, has secured an $8 million seed round to develop software that runs AI workloads across multiple satellites, effectively creating a cloud‑like platform in orbit. The company’s technology, already demonstrated on two missions and set to launch on a SpaceX rocket alongside Google’s Project Suncatcher, aims to cut costly downlink needs and enable shared computing services for satellite operators.

Barclays is deepening its partnership with Anthropic to roll out the Claude AI system across its global operations, aiming for half its developers to use Claude Code by year‑end and to modernise legacy platforms, boost software quality and speed up customer service. The bank cites early successes such as a knowledge‑assistant used by 16,000 staff and AI‑driven email triage handling 120,000 messages daily, all under strict governance and security controls.

The Uncanny Valley podcast examines the controversial “morally binding” AI accord signed after a meeting between Donald Trump and top tech CEOs, which pledges self‑regulation for AI labs, and discusses OpenAI’s launch of always‑on AI agents aimed at everyday users and workplaces. The episode also highlights the surge of extremist political candidates ahead of the midterms, including a Pentagon influencer with extremist views.

A Graphite study identified dozens of distinctive phrases and constructions that give away AI‑generated prose, revealing that models like Anthropic's Opus 5.5 overuse terms such as “dependable” and the phrase “this matters,” while OpenAI’s Astra favors “another dimension” and corrective framing. Although labs have trimmed well‑known tells like em‑dashes, new quirks persist, showing the difficulty of fully humanizing large language models.

Google has introduced Guided Vision, a new feature in Gemini Live for Android that provides real‑time audio descriptions of anything captured by the phone’s camera, helping users read fine print, identify objects and navigate surroundings. Aimed at people with visual impairments, it mirrors similar accessibility tools from Apple and expands AI‑driven assistance on mobile devices.

Amazon Web Services has open‑sourced Strands Decider 2B, a decision‑making model that picks among predefined options and provides confidence scores, built on a Qwen3.5‑2B backbone and designed for low‑latency, cost‑effective workflows. The release follows a surge of similar decision models like TypeSafe’s Jev, highlighting growing demand for specialized, smaller AI tools beyond full‑scale LLMs.

Airbnb unveiled a new AI‑powered search and CEO Brian Chesky explained why chatbots aren’t sufficient for travel planning, advocating for hybrid and multiplayer AI interfaces that let groups collaborate. He also warned that future consumer AI agents will need a robust SDK and dedicated operating system to integrate smoothly with Airbnb’s app, blending deterministic and generative UI elements.

Legato, a hearing‑tech startup, has released its AI‑powered Legato Frames – glasses that embed hearing‑assistance technology into the arms, offering open‑ear sound amplification with AI‑driven voice isolation and near‑zero sound leakage. Priced from $999, the lightweight frames provide up to 12 hours of battery life, prescription lenses, and Bluetooth streaming, and are now purchasable in the U.S. with vision‑insurance coverage.

Startups are sidestepping New York’s strict regulations on large utility‑scale batteries by deploying small, plug‑in batteries in devices like e‑bike swapping stations, food‑cart power packs, and battery‑integrated induction stoves. These distributed solutions require minimal permitting, cut emissions and costs, and could collectively add gigawatts of capacity to the grid. Industry leaders claim this shift could be the most significant change in power‑grid infrastructure in two decades.

According to a Time magazine report, President Trump allegedly consulted Elon Musk’s AI chatbot Grok before a secret meeting in December 2025, seeking advice on how Venezuelans would react to a planned capture of President Nicolás Maduro. The article claims Grok encouraged the operation, which was followed by a U.S. invasion, and notes the Pentagon’s growing use of AI models like Grok in military actions.

A federal judge dismissed antitrust lawsuits brought by Chegg and Penske Media, which claimed Google’s AI-powered search overviews were siphoning web traffic and forcing publishers to provide content for free. Judge Amit Mehta ruled the claims did not meet antitrust standards, clearing Google of the allegations.

SpaceXAI's AI-driven encyclopedia Grokipedia has rolled out a v0.3 update, adding a new logo and redesigning its homepage with featured and most‑read article lists and a live edits tracker. The refresh aims to make the platform more engaging as it resumes user edits.