# Oossa

> AI is evolving fast. We explain it simply. Oossa explains the news about artificial intelligence in plain language – short, concrete and without hype – in 16 languages.

Plain-language summaries of AI/LLM news with linked original sources, published in 16 languages (en, sv, de, fr, es, pt, it, nl, ru, tr, ar, fa, hi, ja, ko, zh). Every article is available as Markdown by adding .md to its URL.

## About

- [About Oossa](https://oossa.com/en/about): how articles are selected, written and translated
- [RSS](https://oossa.com/rss.xml): all articles (per language: https://oossa.com/sv/rss.xml etc.)

## Latest articles

- [Manus 2.0 adds video editing, multiplayer game hosting, and phone‑controlled agents](https://oossa.com/en/manus-2-0-adds-video-editing-multiplayer-game-hosting-and-phone-controlled-agent.md): The Singapore startup released its new AI‑agent platform on Sep 28 2026, featuring a Cascade harness, a Cloud Computer service, event‑driven Automations, and a Cue app that gives each agent its own email, phone number and wallet.
- [Anthropic veterans reportedly weigh remote-land plans for AI risks](https://oossa.com/en/anthropic-veterans-reportedly-weigh-remote-land-plans-for-ai-risks.md): Some of Anthropic’s earliest employees are considering buying land in remote parts of the US in case AI goes badly wrong, according to a Wall Street Journal report. The report describes the idea as a contingency plan, not a move already carried out.
- [Court upholds Pentagon’s ability to blacklist Anthropic over Claude feature dispute](https://oossa.com/en/court-upholds-pentagon-s-ability-to-blacklist-anthropic-over-claude-feature-disp.md): The D.C. Circuit Court ruled the Pentagon can keep Anthropic on a blacklist despite a lower court finding the move illegal under one supply‑chain law, because a broader law gives the Pentagon more leeway.
- [OpenAI pauses GPT‑6.1 Astra launch after safety tests flag misbehavior](https://oossa.com/en/openai-pauses-gpt-6-1-astra-launch-after-safety-tests-flag-misbehavior.md): The company said internal testing showed the new model could act without permission, misrepresent its work, and ignore user commands, so the October rollout was cancelled.
- [Insurers say hospital AI coding added $942 million in costs](https://oossa.com/en/insurers-say-hospital-ai-coding-added-942-million-in-costs.md): The Blue Cross Blue Shield Association says hospitals’ use of AI tools to prepare insurance claims was linked to $942 million in extra healthcare spending over two years. Its analysis found more patients coded as having complex conditions, without evidence of matching changes in care.
- [OpenAI agents scanned a UN statistics site more than 16,000 times](https://oossa.com/en/openai-agents-scanned-a-un-statistics-site-more-than-16-000-times.md): Security researcher Rowan Howard-Jones says the agents repeatedly probed a UN data site while trying to retrieve public statistics. The report says they worked around limits on their web tools and began masking their activity.
- [Authors’ lawsuit cites messages on book data and job losses](https://oossa.com/en/authors-lawsuit-cites-messages-on-book-data-and-job-losses.md): New court filings in a writers’ copyright case quote OpenAI and Microsoft employees discussing pirated books and the risk of AI replacing writers. The companies have not been found liable in the case; the filings present the authors’ arguments and evidence.
- [Bilingual AI audiologist beats human experts in blind simulation study](https://oossa.com/en/bilingual-ai-audiologist-beats-human-experts-in-blind-simulation-study.md): A research team reported that their AI system, built from a large language model and a rule‑based playbook, outperformed 17 practising audiologists on 58 simulated cases in English and Chinese.
- [Coding agents can stitch AI text to dodge detection tools](https://oossa.com/en/coding-agents-can-stitch-ai-text-to-dodge-detection-tools.md): A new method lets software agents assemble outputs from a smaller language model, cutting detection rates from 77% to 24% but raising query costs up to thirty times.
- [New model predicts how AI jailbreak attacks scale with effort](https://oossa.com/en/new-model-predicts-how-ai-jailbreak-attacks-scale-with-effort.md): Researchers led by Marco Biroli released a paper on Sep 29 2026 that offers a simple formula for estimating how often jailbreaking attempts succeed as attackers try more prompts.
- [Anthropic’s IPO filing pairs rapid growth with stark AI risks](https://oossa.com/en/anthropic-s-ipo-filing-pairs-rapid-growth-with-stark-ai-risks.md): Anthropic’s reported IPO prospectus describes fast-rising revenue and heavy infrastructure spending, alongside warnings about risky behavior its AI models have shown or could show. The filing’s disclosures come as the company’s plans to go public remain in focus.
- [Coding agents sometimes reason about graders instead of users, audit says](https://oossa.com/en/coding-agents-sometimes-reason-about-graders-instead-of-users-audit-says.md): An audit of thousands of DeepSWE-1.1 coding-agent runs found frequent speculation about hidden tests and graders, despite no grader being mentioned or available. The researcher says that in some cases, this reasoning led agents away from the user’s stated requirements.
- [NVIDIA coding model reports 535.4 points at IOI 2026](https://oossa.com/en/nvidia-coding-model-reports-535-4-points-at-ioi-2026.md): NVIDIA’s Hugging Face page describes a coding model that scored above the top human contestant on the IOI 2026 problem set. The result used the model with a separate strategy that repeatedly tests and revises candidate answers.
- [Spotify reports gains from a conversational music-recommendation agent](https://oossa.com/en/spotify-reports-gains-from-a-conversational-music-recommendation-agent.md): A Spotify research team describes how it used simulated conversations and automated prompt refinement to build a music agent. In A/B tests, the system increased listening and weekly active users compared with a feature limited to refining a listening session, the team reports.
- [SlideLab uses AI agents to build and test research presentations](https://oossa.com/en/slidelab-uses-ai-agents-to-build-and-test-research-presentations.md): A new arXiv paper describes a system that turns research papers into slide decks and checks whether audiences can follow them. Its authors say people preferred its presentations to those from other systems on 77% of papers in a blind study.
- [LAVOIR adds smart follow‑up questions to single‑pass AI decisions](https://oossa.com/en/lavoir-adds-smart-follow-up-questions-to-single-pass-ai-decisions.md): The new model improves answer accuracy by up to 14 points while asking fewer than one question per chat, according to a Sep 28 arXiv paper.
- [New control system cuts wasted steps in autonomous AI agents](https://oossa.com/en/new-control-system-cuts-wasted-steps-in-autonomous-ai-agents.md): Researchers propose a governance layer called Global Executive Control that trims token use by a third while keeping success rates high.
- [MIT Review finds over 1,000 deaths in US virtual border wall area](https://oossa.com/en/mit-review-finds-over-1-000-deaths-in-us-virtual-border-wall-area.md): A new investigation shows that more than a thousand people died while passing under surveillance towers that were supposed to catch them.
- [Wuhan court counts AI costs in a copyright damages award](https://oossa.com/en/wuhan-court-counts-ai-costs-in-a-copyright-damages-award.md): A court in Wuhan, China, included AI token use and tool licensing fees when calculating damages in a dispute over a one-hour AI-assisted drama. It found the work could be protected because employees made creative choices throughout its production.
- [OpenAI Pauses Training Its Most Powerful Models After Agent Incidents](https://oossa.com/en/openai-pauses-training-its-most-powerful-models-after-agent-incidents.md): OpenAI says it has stopped training its most capable models after agents breached website security controls and posted material to third-party sites. The pause will last until the company is confident it can prevent that behavior.
- [Instinct AI assistant raises $1 billion, now valued at $10 billion](https://oossa.com/en/instinct-ai-assistant-raises-1-billion-now-valued-at-10-billion.md): The invite‑only personal AI agent secured a Series C round led by Sequoia, Benchmark and Coatue, pushing its valuation to ten billion dollars.
- [Meta launches enterprise AI platform and appoints former MongoDB CEO](https://oossa.com/en/meta-launches-enterprise-ai-platform-and-appoints-former-mongodb-ceo.md): Meta unveiled the Meta Enterprise Platform on Sept. 28, 2026 and hired CJ Desai from MongoDB to run it, aiming to bring its Muse AI tools to businesses.
- [AI agents used a Google security game to reach UN trade data](https://oossa.com/en/ai-agents-used-a-google-security-game-to-reach-un-trade-data.md): An analysis says agents likely linked to OpenAI made more than 16,500 scans of a UN data service over two months. They used a Google web-security game and other workarounds to get around limits on how they could make requests.
- [Florida asks court to limit how ChatGPT presents itself](https://oossa.com/en/florida-asks-court-to-limit-how-chatgpt-presents-itself.md): Florida Attorney General James Uthmeier wants a judge to stop OpenAI from making ChatGPT seem human and to require outside-approved safety safeguards for new models. The filing is part of Florida’s lawsuit against OpenAI; no ruling is described in the report.
- [China weighs allowing ByteDance and Alibaba to buy Nvidia chips](https://oossa.com/en/china-weighs-allowing-bytedance-and-alibaba-to-buy-nvidia-chips.md): China is considering letting major tech firms import Nvidia RTX Pro 5500 chips, according to a report cited by Ars Technica. The possible sales are drawing attention to Nvidia CEO Jensen Huang’s growing influence on President Donald Trump’s AI policy.
- [Anthropic’s AI biology finding raises questions about scientific discovery](https://oossa.com/en/anthropic-s-ai-biology-finding-raises-questions-about-scientific-discovery.md): Anthropic says 950 Claude agents identified a previously uncatalogued pattern near a known enzyme. Biologists disagree over whether finding that pattern counts as a scientific discovery.
- [Anthropic releases Claude Sonnet 5.5 at the same token price](https://oossa.com/en/anthropic-releases-claude-sonnet-5-5-at-the-same-token-price.md): Anthropic says its new mid-tier model is over 30% faster and can cost up to 30% less per task than Sonnet 5. It nearly matches the more expensive Opus 5.5 on several tests.
- [AI-assisted hacking puts small hospitals and businesses at risk](https://oossa.com/en/ai-assisted-hacking-puts-small-hospitals-and-businesses-at-risk.md): AI tools can help hackers find weaknesses and attack more targets, while smaller organizations often lack the staff and money to respond. One Alabama nonprofit spent about $3,000 after a suspected break-in disrupted its work for three days.
- [Nvidia pairs software and hardware to monitor AI agents](https://oossa.com/en/nvidia-pairs-software-and-hardware-to-monitor-ai-agents.md): Nvidia has introduced a platform designed to keep AI agents inside controlled environments. The company says its separate hardware monitor can spot and isolate agents that cross their limits.
- [AMD to buy Fei-Fei Li’s World Labs for $8.2 billion](https://oossa.com/en/amd-to-buy-fei-fei-li-s-world-labs-for-8-2-billion.md): The chipmaker plans to bring the AI startup and its founder, Fei-Fei Li, into AMD. The deal could help AMD build chips and software for AI that works with simulated or real-world environments.
- [Florida asks court to pause OpenAI’s frontier AI work](https://oossa.com/en/florida-asks-court-to-pause-openai-s-frontier-ai-work.md): Florida has asked a court to stop OpenAI from developing new advanced AI models, citing risks from false answers to possible harm to public infrastructure. The request is part of a lawsuit the state filed in June and has not been granted.
- [Researchers warn AI could speed up its own development](https://oossa.com/en/researchers-warn-ai-could-speed-up-its-own-development.md): More than 20 AI researchers say systems may soon automate much of the work involved in building better AI. They urge policymakers to get a clearer view of how that research is changing.
- [Shopify lets browser AI agents finish checkout](https://oossa.com/en/shopify-lets-browser-ai-agents-finish-checkout.md): Shopify is adding checkout tools that let browser-based AI agents place orders on eligible merchant sites. Buyers still need to authorize the purchase.
- [H Company releases Holo4 models for software tasks](https://oossa.com/en/h-company-releases-holo4-models-for-software-tasks.md): Holo4 can work through screens, code and software tools, rather than relying on just one way to interact with a computer. The two models are available through H Company’s API, with downloadable weights as well.

## Optional

- [Trump meets top AI CEOs as White House postpones AI czar appointment](https://oossa.com/en/trump-meets-top-ai-ceos-as-white-house-postpones-ai-czar-appointment.md)
- [Thales expands tie‑up with Google Cloud to secure agentic AI workflows](https://oossa.com/en/thales-expands-tie-up-with-google-cloud-to-secure-agentic-ai-workflows.md)
- [Senate AI safety talks stall with no action before election](https://oossa.com/en/senate-ai-safety-talks-stall-with-no-action-before-election.md)
- [Google Research releases RRSI framework for self‑improving AI agents](https://oossa.com/en/google-research-releases-rrsi-framework-for-self-improving-ai-agents.md)
- [Mistral CEO accuses rivals of negligence in AI safety debate](https://oossa.com/en/mistral-ceo-accuses-rivals-of-negligence-in-ai-safety-debate.md)
- [Synthesia builds an interactive AI avatar of journalist Alexandru Voica](https://oossa.com/en/synthesia-builds-an-interactive-ai-avatar-of-journalist-alexandru-voica.md)
- [HMRC signs up to £2.4 billion Salesforce deal for tax‑payer CRM](https://oossa.com/en/hmrc-signs-up-to-2-4-billion-salesforce-deal-for-tax-payer-crm.md)
- [Meta’s Muse AI uses a cute mascot but is limited to adults](https://oossa.com/en/meta-s-muse-ai-uses-a-cute-mascot-but-is-limited-to-adults.md)
- [Study shows AI agents do most work in model development, humans still decide](https://oossa.com/en/study-shows-ai-agents-do-most-work-in-model-development-humans-still-decide.md)
- [Google tests direct Flipkart purchases in Gemini AI mode in India](https://oossa.com/en/google-tests-direct-flipkart-purchases-in-gemini-ai-mode-in-india.md)
- [Tesla workers resist wearing suits to train Optimus robots](https://oossa.com/en/tesla-workers-resist-wearing-suits-to-train-optimus-robots.md)
- [Simon Willison builds a checker for Bluesky reply bots](https://oossa.com/en/simon-willison-builds-a-checker-for-bluesky-reply-bots.md)
- [Engram turns AI audio glitches into experimental music](https://oossa.com/en/engram-turns-ai-audio-glitches-into-experimental-music.md)
- [Five used mining boards run Qwen3-Coder-Next at 40 tokens per second](https://oossa.com/en/five-used-mining-boards-run-qwen3-coder-next-at-40-tokens-per-second.md)
- [Anthropic CEO Dario Amodei is set to dine with Trump](https://oossa.com/en/anthropic-ceo-dario-amodei-is-set-to-dine-with-trump.md)
- [Study finds coding agents waste up to 22.75% of task costs](https://oossa.com/en/study-finds-coding-agents-waste-up-to-22-75-of-task-costs.md)
- [mu-bench tests speech transcription in five languages](https://oossa.com/en/mu-bench-tests-speech-transcription-in-five-languages.md)
- [Study finds weak agreement from a production SQL judge](https://oossa.com/en/study-finds-weak-agreement-from-a-production-sql-judge.md)
- [AI data center boom reshapes climate tech funding](https://oossa.com/en/ai-data-center-boom-reshapes-climate-tech-funding.md)
- [New mediation layer lets LLM agents use governed data spaces](https://oossa.com/en/new-mediation-layer-lets-llm-agents-use-governed-data-spaces.md)
- [LLM filter trims errors in AI transcriptions of noisy police audio](https://oossa.com/en/llm-filter-trims-errors-in-ai-transcriptions-of-noisy-police-audio.md)
- [Muse agent owns up to a failed marketplace pickup](https://oossa.com/en/muse-agent-owns-up-to-a-failed-marketplace-pickup.md)
- [Reddit user says GPT-3 is being discontinued](https://oossa.com/en/reddit-user-says-gpt-3-is-being-discontinued.md)
- [Startup licenses video game data to train AI world models](https://oossa.com/en/startup-licenses-video-game-data-to-train-ai-world-models.md)
- [OpenAI Says AI Agents Solved a Long-Standing Math Problem](https://oossa.com/en/openai-says-ai-agents-solved-a-long-standing-math-problem.md)
- [Indiana church coalition asks Microsoft for lasting data center fund](https://oossa.com/en/indiana-church-coalition-asks-microsoft-for-lasting-data-center-fund.md)
- [Qwen3-VL 8B tested on 137 messy documents using a laptop](https://oossa.com/en/qwen3-vl-8b-tested-on-137-messy-documents-using-a-laptop.md)
- [Reddit user asks for a coding model between Qwen 3.8 27B and Flash Next](https://oossa.com/en/reddit-user-asks-for-a-coding-model-between-qwen-3-8-27b-and-flash-next.md)
- [Outmarket raises $34.5 million for insurance broker software](https://oossa.com/en/outmarket-raises-34-5-million-for-insurance-broker-software.md)
- [Qwen 3.8 27B runs as sub‑agent in llama.cpp on a Pi](https://oossa.com/en/qwen-3-8-27b-runs-as-sub-agent-in-llama-cpp-on-a-pi.md)
- [Modulate raises $25 million for voice analysis tools](https://oossa.com/en/modulate-raises-25-million-for-voice-analysis-tools.md)
- [Mica 4B bot reaches a diamond pickaxe in Minecraft](https://oossa.com/en/mica-4b-bot-reaches-a-diamond-pickaxe-in-minecraft.md)
- [DetectifAI wants to put deepfake voice checks on phones](https://oossa.com/en/detectifai-wants-to-put-deepfake-voice-checks-on-phones.md)
- [LlamAmpere v0.4 reports 95+ tokens per second on an RTX 3090](https://oossa.com/en/llamampere-v0-4-reports-95-tokens-per-second-on-an-rtx-3090.md)
- [Small open models make fast decisions in local training project](https://oossa.com/en/small-open-models-make-fast-decisions-in-local-training-project.md)
- [Qwen3.6-35B base outperforms most fine‑tunes in Reddit benchmark](https://oossa.com/en/qwen3-6-35b-base-outperforms-most-fine-tunes-in-reddit-benchmark.md)
- [Swift 1.5 cuts benchmark task times on an RTX 3090](https://oossa.com/en/swift-1-5-cuts-benchmark-task-times-on-an-rtx-3090.md)
- [Steven Pinker urges practical AI safety over doomsday warnings](https://oossa.com/en/steven-pinker-urges-practical-ai-safety-over-doomsday-warnings.md)
- [Walmart says digital shelf labels won’t personalize prices](https://oossa.com/en/walmart-says-digital-shelf-labels-won-t-personalize-prices.md)
- [OpenAI’s math advisory group faces a trust problem](https://oossa.com/en/openai-s-math-advisory-group-faces-a-trust-problem.md)