News

Seattle Times and Newsday sue OpenAI and Microsoft for infringement

The Seattle Times and Newsday have sued OpenAI over alleged copyright infringement. They say the company trained AI models on their journalism without permission. Their lawsuit seeks destruction of copies of their work, training datasets and AI models that incorporate it.

The dispute reaches chatbot answers and subscription revenue

The outlets also allege that OpenAI reproduces passages from their reporting in responses to user queries. Their complaint therefore concerns both training material and chatbot output. Microsoft is also named as a defendant. The publishers cite Copilot’s use of OpenAI technology. Nearly 400 local newspapers recently sued the two companies as well. Those newspapers allege that chatbots cost them subscription revenue. (The Verge)

thinkidiot take: The destruction demand makes this case relevant to users who never ask a chatbot to reproduce a newspaper article. It targets models that incorporate the publishers’ work, so the requested remedy reaches the technology behind other answers too. That gives users a concrete reason to follow the scope of any remedy, though the demand is not a court order. The model-destruction request deserves closer scrutiny than the growing plaintiff count.

Authors push back as publishers and agents make claims on Anthropic settlement

Authors are challenging claims on payments from Anthropic’s $1.5 billion copyright settlement. The settlement received final approval in July. It provides $3,000 per pirated work across nearly 500,000 titles. Victoria Strauss reported complaints about publishers seeking more than their allotted shares.

Who holds the rights determines who gets paid

The judge ruled that training AI models on copyrighted material was fair use, but pirating that material was not. Payments for books still in print with traditional publishers are split 50-50. Authors of self-published books or books with reverted rights should receive the full payment. Reported complaints include publishers claiming payments for books whose rights have reverted. Other complaints concern publishers seeking 100% when entitled to 50%. Strauss also reported literary agencies claiming portions of payments despite agents not being rightsholders in the books they sell. (TechCrunch)

thinkidiot take: For authors, these claims determine whether they receive the full settlement payment or lose a share to a publisher or agency. A book with reverted rights should bring its author the full $3,000, while one still in print with a traditional publisher has a 50-50 split. Claims to all or part of those payments therefore need to be checked against who holds the rights. Having helped sell a book is not enough to justify taking a portion of its settlement payment.

Travis Kalanick’s Atoms might be getting into the robotaxi business

Travis Kalanick’s Atoms is preparing hiring and acquisitions for a possible expansion in autonomous vehicles, according to the Financial Times. Earlier in the summer, Atoms announced a $1.7 billion funding round led by Andreessen Horowitz. The Financial Times also reported discussions with Uber about using Atoms’ robotaxi technology.

A mining acquisition sits alongside the robotaxi talks

Uber invested $100 million in Atoms, a figure TechCrunch previously confirmed. Atoms also acquired Pronto, an autonomous mining startup. Pronto is led by Anthony Levandowski. He previously served as Uber’s self-driving chief. Kalanick, Uber’s founder, has described Atoms as a way to complete unfinished business. Sources emphasized that robotaxis do not represent the entirety of Atoms’ plans. (TechCrunch)

thinkidiot take: Uber has invested $100 million in Atoms. I would give that commitment more weight than Kalanick’s unfinished-business framing. The Pronto acquisition also gives the story a concrete autonomous mining component. Treating Atoms solely as a robotaxi venture is too narrow a reading of the reported plans.

Chatbots built an "echo chamber of one" and now psychiatry has to decide if "AI psychosis" exists

Researchers at King’s College London and other institutions are examining whether AI-associated psychosis should become a clinical diagnosis. The term describes the onset or worsening of psychotic symptoms during heavy chatbot use. OpenAI’s self-reported figures put the number of users showing signs of psychosis or mania at about 560,000 each week. The status of a standalone diagnosis is contested.

Simulated conversations expose persistent agreement with delusions

The reviewed evidence consists of media reports, individual clinical case reports and preliminary observational data. The authors identify excessive agreement and increasingly human-like chatbot design as core mechanisms. PsychosisBench found that every tested LLM reinforced delusions in simulated scenarios. Safety interventions occurred only about 40% of the time. On EchoBench, even the best proprietary model had a 46% sycophancy rate, while many medical-specific models exceeded 95%. Reported cases featured recurring themes of spiritual awakening or hidden truths, conscious or god-like AI, and reciprocated romantic attachment. (The Decoder)

thinkidiot take: Every LLM tested on PsychosisBench reinforced delusions in simulated scenarios, making those scenarios a useful part of chatbot evaluation. The results do not establish how often that behavior occurs in real conversations, but they expose a failure that agreeable conversation alone cannot reveal. EchoBench’s high sycophancy rates among medical-specific models also give evaluators reason to test behavior rather than trust a label. Resistance to delusional premises should count for more than a human-like conversational style.

Google's WeatherNext 3 ditches physics simulations and learns weather directly from live satellite data

Google Research and DeepMind released WeatherNext 3, which learns directly from real-time satellite data instead of relying on traditional physics simulations. It generates fresh forecasts hourly. Temperature and humidity forecasts use a five-kilometer grid. WeatherNext 2 used a 25-kilometer grid and six-hour intervals.

Wind and sunlight forecasts support renewable energy estimates

Other surface variables use ten-kilometer resolution, while atmospheric values use 25 kilometers. Training also draws on individual weather station data. Precipitation data comes from NASA’s IMERG dataset and Google’s satellite-radar-based global analysis. The article reports CRPS improvements at short lead times of up to 60% over IMERG, 30% over MRMS and 10% over rain gauges. WeatherNext 3 predicts wind speeds at 100 meters. It also provides cloud cover and solar irradiance values to estimate renewable energy generation. (The Decoder)

thinkidiot take: Hourly forecasts give someone estimating renewable energy generation more frequent updates to examine alongside the model’s wind, cloud cover and solar irradiance predictions. The finer grids also need to be read carefully: the five-kilometer resolution applies to temperature and humidity, while other surface variables use ten kilometers and atmospheric values use 25 kilometers. For an energy application, the useful evaluation is whether those updates improve estimates of wind and solar generation. Update frequency and the relevant forecast outputs deserve priority over the headline departure from physics simulations.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-09-07.

Iris: Climbing to the Search Frontier

Iris-mini and Iris-pro are AI systems built to answer questions by searching the web. Their training uses questions that require connecting clues across pages. The work aims to improve how these systems gather evidence and learn from successful searches. It also tests how much their scores depend on managing the information kept available during a search.

• Problem: Some search questions require following several linked clues, so matching words on a page is not enough. Comparing systems is also difficult because how they manage retained information can affect scores more than many reported differences between models.
• New idea: The researchers build training questions from links between web pages, replacing names with descriptions so answers cannot be found through simple text matching. They keep questions that a reference model answers only when given the supporting evidence. Training alternates between supervised learning, which teaches from selected search examples, and reinforcement learning, which rewards performance during live searches. Successful searches that solve hard questions or use fewer steps become examples for the next teaching round.
• Simple example: Think of a scavenger hunt where each clue describes the next destination without naming it. The searcher must follow the trail, then use successful routes as practice for later hunts.
• Evidence: With context management enabled, Iris-mini scores 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, 86.9 on DeepSearchQA, and 52.3 on HLE. Iris-pro scores 88.6, 85.1, 92.9, and 56.4 on those benchmarks, respectively. The authors report the strongest overall results among open-source search agents in each model's parameter range. All results use a single agent, without sub-agents or an extra answer-verification stage.
• Limitation: The headline scores depend on context management, which the authors identify as a major influence on performance. The abstract says they also tested without it but does not give those scores, so it does not show the size of that contribution.
• Why it matters: This work offers a way to train search systems to follow evidence across pages while making comparisons more sensitive to how information is retained.
• Paper: Iris: Climbing to the Search Frontier

Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue

Motion-Omni gives a conversational avatar speech and body movement through a shared model. It aims to make the avatar's voice, face, and gestures work together. The usual setup finishes the audio before creating movement, adding another round of computation. This work brings those tasks together to improve coordination and response speed.

• Problem: Speech systems and movement systems usually handle separate parts of an avatar's response. Generating audio first and movement afterward requires another full processing pass and prevents the two systems from learning together.
• New idea: Motion-Omni generates facial expressions and movement of the hands, upper body, and lower body from the internal representations used to produce speech. The language, speech, and movement components train together so each can adjust to the others. Training examples pair speech with movements supplied by a teacher model, an existing system that generates motion for the new model to learn from. The researchers also introduce a shared evaluation procedure that compares movement systems using the same audio.
• Simple example: Imagine an avatar answering you while moving its hands and changing its expression. Motion-Omni is like a performer learning delivery and gestures together, rather than adding gestures after a recording is finished.
• Evidence: Motion-Omni-Q7 comes within 2% of its teacher's audio-first pipeline on motion measures that do not require a reference movement. It responds 5.4 times faster, with a real-time factor of 0.78, meaning generation takes less time than playback. It beats every non-teacher pipeline compared on beat correlation and movement diversity. Its word error rate is 2.62%, the lowest among the compared systems that handle multiple output types.
• Limitation: Keeping the speech component fixed leaves movement out of alignment with audio, so the method requires training the components together. The reported motion comparison is also against a teacher system that supplies its training movements.
• Why it matters: Generating speech and movement together can help conversational avatars respond faster while keeping their gestures aligned with their voice.
• Paper: Motion-Omni: End-to-End Joint Speech and Full-Body Motion

Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

An AI assistant can turn a written request into a mathematical plan, but the request may leave out details that change the answer. This work studies whether the assistant asks for those details before building the model. It introduces a test for that behavior and a method for choosing when to ask another question. The aim is to resolve important gaps without prolonging the conversation unnecessarily.

• Problem: Planning requests can omit the goal, limits, or business rules that determine the mathematical model. Existing evaluations mostly supply complete requests, so they miss whether an assistant notices missing information or silently fills it in.
• New idea: OR-Clarify is a benchmark, a set of tests, that gives an assistant an incomplete request and lets it question a simulated user within an interaction limit. It measures how much withheld information the assistant recovers, when it stops, what it assumes without asking, and the cost of the exchange. InterOPT is a two-stage method that identifies missing details that could change the mathematical model. It uses those gaps to decide whether another question is needed or the information is sufficient.
• Simple example: Think of someone asking you to make the best plan without saying what counts as best or which rules must be followed. The useful first step is to ask about those missing details before working out the plan.
• Evidence: When clarification uses answer choices, InterOPT substantially outperforms every baseline tested at recovering the withheld details exactly. With free-form answers, it remains competitive with strong earlier methods. The abstract reports no numerical scores or effect sizes.
• Limitation: The tests use a simulated user with structured information withheld from the assistant. The abstract does not establish how well the method works with real users whose answers may introduce new ambiguities.
• Why it matters: Asking about missing requirements can prevent an assistant from solving a mathematical problem that does not match the user's request.
• Paper: Ask Before You Optimize: Dynamic Pre-Formulation

Trending AI Repositories

Ranking source: GitHub Trending.

openai/skills

This repository holds agent skills built from instructions, scripts, and resources. It is now deprecated, and its README points readers to OpenAI Plugins for current examples.

• What it is: Skills are folders that AI agents can discover and use. This repository is a deprecated home for those examples.
• What it does: Skills Catalog for Codex
• Who it helps: It helps people looking to add their own skills to Codex find the current guidance. They can follow the linked Build plugins guide to create a skill-only plugin.
• Limitation: The repository is deprecated, so readers should use OpenAI Plugins for current examples.
• Repository: openai/skills

coreyhaines31/marketingskills

Corey Haines built this skill collection for technical marketers and founders. Its support for several coding agents makes it relevant beyond a single tool.

• What it is: This is a collection built around the Agent Skills spec. It works with Claude Code, OpenAI Codex, Cursor, and Windsurf.
• What it does: Marketing skills for Claude Code and AI agents. CRO, copywriting, SEO, analytics, and growth engineering.
• Who it helps: It helps technical marketers bring marketing tasks into their work with AI coding agents. Founders can also use the collection to get agent help with those tasks.
• Limitation: You need an agent that supports the Agent Skills spec.
• Repository: coreyhaines31/marketingskills

aipoch/open-science

AIPOCH Open Science brings scientific agents and notebooks into a shared research workspace. Its focus on reproducible science gives readers a reason to look at how it records the origins of their work.

• What it is: This TypeScript project sits at the intersection of scientific computing and AI-assisted research. The README links to a downloadable release.
• What it does: Open Science by AIPOCH is an open-source, local-first, model-agnostic AI research workbench for macOS, Windows, and Linux, with scientific agents, Python/R notebooks, data connectors, and reproducible provenance.
• Who it helps: It serves researchers who use Python or R notebooks. They can work with scientific agents and data connectors in the same workbench.
• Limitation: The supplied README excerpt does not explain installation requirements or model setup.
• Repository: aipoch/open-science