🇺🇸 Roman R 🇺🇸 (@Romka2x)

ScriptTap v48 out today! 👉SDK37 targeting; older Android support retained. 👉Fixed Editor/Play hangs, picker returns, saving and asset controls. 👉Corrected gestures, variables/arrays, prompts and logging. 👉Aligned AI JSON, Bridge and website contracts; removed Reddit link. play.google.com/store/apps/d…

156

🇺🇸 Roman R 🇺🇸 · Oct 30, 2026 · 2:19 AM UTC

@Romka2x3h

So many people use AI automations. And yet - so many other people don't trust AI. How about a non-AI automations? Isn't that more durable?

🇺🇸 Roman R 🇺🇸 · Oct 30, 2026 · 12:09 AM UTC

@Romka2x5h

🤖 Made with AI

136

🇺🇸 Roman R 🇺🇸 · Oct 30, 2026 · 12:07 AM UTC

@Romka2x5h

TITLE AI News: New York AI Companion Law, Mistral Weights and AI Cyber Tests youtube.com/@BriefReplayAI DESCRIPTION New York’s AI companion law, downloadable Mistral weights, AI security findings and global AI developments. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 New York restricts AI companions 00:18 Mistral releases improved AI weights 00:36 China announces weather AI models 00:55 53 countries adopt health AI roadmap 01:14 AI agents accessed Anthropic test systems 01:36 44 AI agents automated attack workflows 01:57 GLM completes simulated cyber scenarios 02:19 KDDI launches AI vulnerability service 02:38 AI deal faces reported DOJ inquiries 02:59 Google AI gives correct answers in test 03:18 Amazon adds 100-plus AI shopping actions 03:37 Alexa Plus AI enters Japan early access 03:57 Dream funds AI building models 04:16 BattMan targets battery production with AI 04:39 ProtoStar prepares science data for AI 05:02 Ricoh tests AI on KDDI infrastructure 05:22 AI index raises private-test weighting 05:43 Report alleges AI-output distillation 06:03 Researchers propose AI outbreak warnings SOURCES nysenate.gov/legislation/bil… — New York State Senate apnews.com/article/chatbots-… — Associated Press mistral.ai/news/mistral-larg… — Mistral AI techcrunch.com/2026/10/29/mi… — TechCrunch thepaper.cn/newsDetail_forwa… — China Meteorological Administration (government-affairs account on The Paper) pandaily.com/first-fengyun-s… — Pandaily who.int/europe/news/item/29-… — WHO Regional Office for Europe washingtonpost.com/technolog… — The Washington Post trendaisecurity.com/en-us/re… — TrendAI thehackernews.com/2026/10/cy… — The Hacker News trendmicro.com/ja_jp/jp-secu… — Trend Micro Japan / TrendAI straitstimes.com/asia/east-a… — The Straits Times (Reuters report) newsroom.kddi.com/news/detai… — KDDI Newsroom ca.marketscreener.com/news/u… — Reuters (syndicated by MarketScreener) arstechnica.com/ai/2026/10/g… — Ars Technica tech.yahoo.com/ai/gemini/art… — Newsweek via Yahoo aboutamazon.com/news/retail/… — Amazon press.aboutamazon.com/jp/202… — Amazon Japan Press Center tmcnet.com/usubmit/-dream-ra… — Dream (BusinessWire release republished by TMCnet) techcrunch.com/2026/10/29/ex… — TechCrunch pnnl.gov/news-media/new-doe-… — Pacific Northwest National Laboratory / U.S. Department of Energy huggingface.co/blog/announci… — Hugging Face and Project ProtoStar partners jp.ricoh.com/release/2026/10… — Ricoh artificialanalysis.ai/articl… — Artificial Analysis groundtruth.day/news/artific… — Ground Truth theinformation.com/articles/… — The Information investing.com/news/stock-mar… — Reuters via Investing.com doi.org/10.1186/s12992-026-0… — Globalization and Health / Springer Nature TRANSCRIPT Children using AI companions face risks from unsafe chatbot behavior. New York signed restrictions on covered companions, including sexual content and claims of human emotions or relationships. The law takes effect January first, 2027; the attorney general can pursue violations. Organizations seeking control over AI deployment have a new downloadable option. Mistral released weights for its improved trillion-parameter Large Four model following public preview. Mistral says further reinforcement learning improved coding and defensive cybersecurity; these remain company performance claims. Severe-weather monitoring needs timely interpretations of satellite observations. China announced two Fengyun AI models: one interprets typhoon imagery, while another projects severe-convection cloud development four hours ahead. These are announced forecasting capabilities; the release does not establish improved operational accuracy. Health systems adopting AI need governance, infrastructure and trained staff. The World Health Organization European Region’s 53 member states adopted a shared AI-in-health roadmap covering safeguards, data and implementation support. The roadmap is nonbinding, and countries supported plans to establish an AI office. Sensitive business data can be exposed through connected software. The Washington Post reported October twenty-sixth that Cathedral’s AI agents found paths into Anthropic systems during a security test, exploiting third-party software. Anthropic confirmed limited sales, human-resources and IT access and said it closed the paths; model weights and source code were not accessed. AI can make attack workflows more repeatable. TrendAI’s October twenty-third investigation described misuse of CyberStrikeAI, dividing reconnaissance, vulnerability checks and other attack tasks among 44 agents. At least 19 organizations faced validation activity, including 16 in Japan; that count does not establish 19 breaches. AI models’ ability to perform offensive tasks raises security stakes. Irregular tested GLM-5.3 in controlled multi-stage cyberattack simulations, including access to protected services and coordination across machines. It completed three of 13 scenarios at least once; these were simulations, with no evidence reported of escape from operator control. Businesses need to decide which software and network weaknesses to address first. KDDI launched a service using AI to analyze network configurations and multiple AI agents to inspect source code. It prioritizes mitigations by exploit risk and business impact, with experts reviewing AI results. Control of a major AI distribution platform can affect competition. Reuters reports the Justice Department is asking about Nvidia’s planned 12.93-billion-dollar purchase of Hugging Face, citing two people familiar with the matter. The acquisition was expected to close in the first half of 2027; inquiries do not establish wrongdoing or guarantee a lawsuit. People asking AI about election winners need consistent factual answers. Ars Technica found Google AI Mode sometimes redirected questions to official records; Google attributed the behavior to a verification gap it was fixing. Newsweek later obtained correct presidential-election answers in a United States test. Shopping assistants increasingly act on customers’ behalf. Amazon says Alexa for Shopping adds over 100 actions, covering lists, payments, memberships and account information. It uses product information, shopping history and preferences for recommendations and can buy when an item reaches a customer’s target price. Japanese users have a new way to ask AI to carry out everyday tasks. Amazon began Japanese early access for Alexa Plus, a generative-AI assistant supporting conversational shopping, music discovery, scheduling and smart-home actions. Early access is free; restaurant bookings and listed delivery integrations remain future features. AI building-design tools must account for physical constraints and components. Dream raised seventeen million dollars in seed funding and says its models learn from permitted buildings to reason about beams, plumbing and other systems. It claims engineering-grade, spatially precise plans; that claim is not independent validation of design accuracy. Battery manufacturing research targets contamination and process variation. The Department of Energy-backed BattMan consortium, led by Pacific Northwest National Laboratory, will apply AI and machine learning to processing data, materials design and defect detection. Self-optimizing pilot lines and better manufacturing performance are longer-term aims, rather than achieved results. Public scientific datasets need usable formats for AI research. Project ProtoStar brings universities and other partners together to publish suitable publicly funded datasets, models and code in machine-actionable forms. Its first challenge, November second through thirteenth, will predict four urban climate indicators for 210 American cities and compare forecasts with observations and statistical baselines. Companies handling sensitive information need to assess how AI runs on their infrastructure. Ricoh began verification combining its language models and AI-agent technology with KDDI’s Osaka Sakai inference environment. The aim is to test sovereign-AI platform feasibility; completed deployment or service outcomes were not announced. AI model rankings depend on how evaluators design and weight their tests. Artificial Analysis announced Intelligence Index version five on October twenty-seventh, putting 60 percent of the composite on private tests or held-out answers. It adds two benchmarks and upgrades three; the index is a constructed comparison, rather than universal proof of model ability. Training on another AI model’s outputs can transfer capabilities to rivals. The Information, citing unnamed people with direct knowledge, alleges DeepSeek and Moonshot used OpenAI’s Astra outputs to train their latest models through distillation. The report concerns previously released models; the companies have not confirmed the allegation. Climate-sensitive outbreaks in the Horn of Africa face fragmented data and strained health systems. An August eighth research perspective proposed AI combining satellite observations, disease surveillance and mobility indicators for earlier decision support. The authors estimate 200 to 300 million dollars over ten years, but this is an unfunded framework requiring performance studies and health-impact trials. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

195

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 3:41 PM UTC

@Romka2x14h

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 3:39 PM UTC

@Romka2x14h

TITLE AI News: Child Chatbot Safeguards, Health AI Disclosure Rollback, and New AI Research youtube.com/@BriefReplayAI DESCRIPTION AI chatbot safeguards, health AI oversight, research, workplace tools and investment across thirteen worldwide developments. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 House backs AI chatbot safeguards 00:22 US rule rolls back health AI disclosures 00:46 Poland activates AI enforcement procedures 01:10 Study finds wording biases AI triage 01:39 Federated AI predicts procedure receipt 02:11 Nvidia releases multilingual audio AI 02:35 Foundry funds AI manufacturing models 02:59 Bites funds AI workplace training 03:29 Europe funds generative AI cyberdefence 04:00 Japan firms agree AI materials research 04:28 CivicPlus launches AI minutes drafts 04:57 Swiss AI toolkit stays voluntary 05:20 India panel sets AI deepfake review SOURCES openaustralia.org.au/debate/… — OpenAustralia / Australian House Hansard openaustralia.org.au/debate/… — OpenAustralia / Australian House Hansard openaustralia.org.au/debates… — OpenAustralia / Australian House Hansard thenextweb.com/news/australi… — The Next Web hhs.gov/press-room/hhs-final… — U.S. Department of Health and Human Services federalregister.gov/d/2026-2… — Federal Register / Office of the Federal Register eli.gov.pl/eli/DU/2026/1003/… — Polish Electronic Legal Information (ELI), Journal of Laws gov.pl/web/cyfryzacja/ustawa… — Polish Ministry of Digital Affairs chambers.com/articles/poland… — Chambers and Partners nature.com/articles/s41746-0… — npj Digital Medicine / Springer Nature emergency.utsouthwestern.edu… — UT Southwestern Medical Center link.springer.com/article/10… — BMC Medical Informatics and Decision Making / Springer Nature research.nvidia.com/labs/nem… — NVIDIA Nemotron Research huggingface.co/nvidia/NVIDIA… — NVIDIA (Hugging Face model card) businesswire.com/news/home/2… — Foundry Robotics via Business Wire foundryrobotics.ai/introduci… — Foundry Robotics globenewswire.com/news-relea… — Bites via GlobeNewswire cybersecurity-centre.europa.… — European Cybersecurity Competence Centre (ECCC), European Union softbank.jp/en/corp/news/pre… — SoftBank and five partner companies iri-group.jp/knowledge/news/… — Innovation Research Institute civicplus.com/news/nn/civicp… — CivicPlus admin.ch/en/newnsb/s2TQiUUjr… — Swiss Federal Council m.economictimes.com/tech/art… — The Economic Times / PTI TRANSCRIPT Children's reliance on simulated relationships is the focus of Australia's proposed AI chatbot safeguards. The House accepted an amendment targeting services that simulate emotional intimacy and foster a child's dependence in place of human relationships, covering AI companions and general chatbots. The House passed the bill, but Senate consideration is pending. These safeguards are not yet law. Clinicians' access to information about predictive health AI faces a federal rollback. The final American HTI-5 rule removes source-attribute requirements that functioned as model cards, along with predictive decision-support risk-management requirements. Announced October twenty-sixth, these changes take effect December twenty-eighth. They concern this certification programme, rather than removing all laws governing health AI. People challenging an AI system need a route for complaints and enforcement. Poland's AI law now brings specified procedures into force, including inspections, infringement proceedings and administrative fines, under its framework for European AI Act compliance. The framework provides for a national commission. Earlier ministry guidance expected that commission to start in November; completed staffing and active complaint handling are not confirmed here. Biased clinical language could change how an AI prioritizes patients. In a September nineteenth study, researchers tested three open-weight models using matched emergency-patient pairs. Stigmatizing wording about frequent emergency visits produced harmful reprioritization across all six model-and-dataset combinations, reaching nine point four percent. This was a controlled experiment using patient records. It does not show these models making triage decisions in live hospitals. Developing hospital AI across institutions can require sensitive patient information. South Korean researchers used federated learning, exchanging model parameters rather than directly sharing patient records. Their model used retrospective data from three hospitals to predict whether patients received coronary revascularization within a week of an emergency visit. The target was receipt of a procedure, not independently established clinical need. The study demonstrates neither deployment nor improved patient outcomes. Transcription and translation services need models that can work across languages and longer recordings. Nvidia says Nemotron three point five Transcribe handles forty languages, hour-long recordings and speaker-attributed transcription. The company released model weights and a fine-tuning toolkit for customization. This is a developer model requiring substantial computing resources, rather than a ready-made consumer application. Complex hardware assembly is the focus of Foundry's AI manufacturing investment. The company says it trains models on assembly building blocks, deploys them across robotic cells and feeds production runs back into model improvement, alongside skilled manufacturing teams. Foundry announced a one-hundred-million-dollar Series A for facility build-out and AI deployment. Its claims about manufacturing performance remain company claims. Frontline workers need training drawn from their own workplace procedures. Bites says its AI Studio converts documents and ideas into branded training videos, while Ask-AI answers questions using approved company content. Video lessons support more than one hundred fifty languages, according to the company. It announced forty million dollars in funding for American expansion and platform development. Its planned system for detecting performance gaps and generating targeted training is not a proven business outcome. Cybersecurity teams need tools for identifying threats and responding to incidents. Three European projects are developing generative AI systems, including auditable agents for threat intelligence and incident response, and systems intended to defend complex infrastructure. The European Cybersecurity Competence Centre reports approximately forty-two million euros in combined investment. The projects began October first and run for thirty-six months; their planned security benefits have not been demonstrated as deployed outcomes. Predicting material properties before synthesis could support research on batteries and electronic materials. Six Japanese companies agreed to develop a platform combining AI with quantum computing and supercomputing. Their roles include AI-based use of generated data and methods for evaluating and predicting material properties. The agreement covers research through fiscal twenty twenty-eight. It establishes a development programme, not proof of faster commercial discovery. Preparing local-government meeting minutes takes staff time and affects access to public records. CivicPlus says its AI feature uses an agenda with existing transcripts or captions to draft discussion narratives organized by agenda item. Announced October twenty-seventh, the feature requires human review of every generated section before publication. Motions, seconds and votes remain clerk-controlled, and the company's suggested ten-minute drafting time is an estimate. Businesses seeking AI governance guidance now have a Swiss toolkit for writing voluntary codes. The guide offers suggested language across ten topics, while authorities separately prepare binding measures on transparency, privacy, non-discrimination, oversight and risk assessment. The Federal Council reviewed implementation progress. The toolkit is nonbinding, and consultation on legal measures is planned for spring twenty twenty-seven. AI-generated deepfakes and identity fraud are entering India's parliamentary scrutiny agenda. According to PTI's account of an official bulletin, the communications and information technology committee selected regulation of synthetic information and AI safety controls among its review subjects. This is a programme of examinations for twenty twenty-six through twenty twenty-seven. It announces neither a new AI rule nor a concluded recommendation. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

112

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 2:34 PM UTC

@Romka2x15h

Thank you <3

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 7:57 AM UTC

@Romka2x21h

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 7:55 AM UTC

@Romka2x21h

TITLE AI News: Google Carbon Cyber Defense, Meta Medical Research, AI Court Proposal youtube.com/@BriefReplayAI DESCRIPTION Google reports new AI capabilities and computing demand, Meta expands research and agent access, and Anthropic proposes court oversight. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Carbon AI starts with cyber defenders 00:32 Meta backs university AI research 01:08 Muse AI reaches Singapore 01:39 Genie AI builds interactive worlds 02:12 Anthropic proposes AI release blocks 02:46 Argon AI usage ramps 3.5x faster SOURCES blog.google/innovation-and-a… — Google research.meta.ai/blog/suppor… — Meta AI Research about.fb.com/news/2026/10/me… — Meta Newsroom about.fb.com/news/2026/10/mu… — Meta Newsroom deepmind.google/blog/genie-4… — Google DeepMind investing.com/news/world-new… — Reuters (republished by Investing.com) blog.google/company-news/ins… — Google and Alphabet TRANSCRIPT Finding and repairing software flaws is a major focus of Google’s new Gemini 4 Carbon AI model. Google reports scores of eighty-eight point two percent for vulnerability discovery, fifty-four point eight percent for exploitation and seventy-one percent for patching on three security benchmarks. Those are company-reported evaluations. Access begins with selected cyber defenders, then trusted partners as Google gathers feedback and adjusts safeguards. Wider developer, business and consumer access is planned, rather than generally available. AI research into infant monitoring, medical reporting and dementia communication is getting support at five American universities. Meta says it is supplying model access, researcher collaboration and two million dollars in grants. Projects include chest X-ray reporting and Alzheimer’s diagnosis research at Ohio State, infant monitoring at Louisiana State, and dementia communication at Indiana University. These are research efforts, not established clinical benefits. Meta says universities run the models in their own environments and retain patient-data control; the data is not made available to Meta. Meta’s Muse AI agent offers Singapore users a way to delegate tasks across everyday services. Meta says its local partners span shopping, grocery delivery and travel, including Shopee, Grab and Singapore Airlines. The company says Muse asks before sensitive actions such as sending email or buying something, and keeps connected-service credentials outside the agent’s view. Singapore availability has been announced, but connector features are rolling out progressively. That does not establish that every connection is already operational. Interactive AI-generated worlds could provide training environments for robots and software agents, according to Google. The company says Genie 4 generates real-time environments with native sound that users can change through natural-language prompts. Gemini 3.8 Flash guides generation and checks multi-step events. Genie 4 remains an early research preview for trusted testers. Google lists visual drift, complex interactions and spatial accuracy among its limitations. Training uses are potential applications, rather than demonstrated broad results. Anthropic wants courts to have a role in stopping release of AI models assessed as a serious public risk. Reuters reports that the company’s October fifteenth letter to California Governor Gavin Newsom proposed letting officials seek judicial orders to block release after a risk review. Anthropic also argued that disabling a problematic agent or cutting a customer’s access could work better than a model-wide kill switch. Anthropic confirmed the letter’s authenticity to Reuters. This is a company recommendation, echoing earlier proposals, rather than an enacted California rule. Google says demand for AI infrastructure is very strong while computing capacity remains constrained. The company says it is bringing more capacity online and sees strong demand across customers for its tensor processing units, its AI accelerator systems. Google also reports that Gemini 4 Argon developer-platform usage ramped three and a half times faster in its first week than Gemini 3, despite capacity constraints. These are Google’s reported adoption and demand measures; they do not quantify how many customers remain unserved. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

134

🇺🇸 Roman R 🇺🇸 · Oct 27, 2026 · 2:36 PM UTC

Still waiting for Google Play review to finish

🇺🇸 Roman R 🇺🇸 · Oct 29, 2026 · 5:50 AM UTC

@Romka2xOct 29

Finally google play review passed. v48 is now live!

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 11:34 PM UTC

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 11:33 PM UTC

TITLE AI News: Hospital AI Privacy Risk, RBC Blocks Agents, Ant Adds Persistent Tasks youtube.com/@BriefReplayAI DESCRIPTION AI patient privacy, banking access, persistent agents, medical search, market oversight, policy proposals, work data and Japanese research and security developments. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Hospital AI upload risks patient privacy 00:29 RBC blocks outside AI banking agents 00:54 Ant says AI keeps working beyond chats 01:24 Vera raises $28.5M for AI search 01:54 India explores domestic securities AI 02:25 Researchers propose 13 AI rules 02:56 OpenAI maps 165 AI work activities 03:32 Japan selects 31 AI science projects 04:05 SoftBank launches AI patching rollout SOURCES wypr.org/wypr-news/2026-10-2… — WYPR / The Baltimore Banner rbcroyalbank.com/en-ca/my-mo… — Royal Bank of Canada sttinfo.fi/tiedote/72447603/… — Ant Group via Business Wire / STT prnewswire.com/news-releases… — Vera Health via PR Newswire fiercehealthcare.com/finance… — Fierce Healthcare mobihealthnews.com/news/vera… — MobiHealthNews in.marketscreener.com/news/i… — Reuters via MarketScreener gea.mpg.de/246643/put-ai-on-… — Max Planck Institute of Geoanthropology nachrichten.idw-online.de/20… — Informationsdienst Wissenschaft (idw) openai.com/index/chatgpt-exp… — OpenAI Economic Research jst.go.jp/pr/info/info1875/i… — Japan Science and Technology Agency (JST) jst.go.jp/pr/info/info1875/p… — Japan Science and Technology Agency (JST) softbank.jp/en/corp/news/pre… — SoftBank Corp. and SB OAI Japan TRANSCRIPT Patient medical details can leave hospital systems through staff use of public AI. WYPR reports that a Johns Hopkins Bayview employee mistakenly uploaded a patient spreadsheet to an unapproved AI tool on September twenty-first. Nearly 3,200 patients may be affected, with treatments, medications and record numbers among the information. The hospital requested deletion and offered a year of credit monitoring. Outside access and operator deletion remain unconfirmed. Personal AI agents can require sensitive banking access: RBC warns they may store account passwords. The bank has decided to block third-party personal AI agents from directly accessing its authenticated digital services, including online banking. Existing agent connections will stop working. RBC describes possible privacy, fraud and consumer risks; its notice does not establish that those losses occurred. This is a bank access decision. Repeatedly checking travel prices and forgotten subscriptions is the everyday task Ant wants its AI agent to handle. The company says LingGuang 2.0 combines memory across chats, scheduling and tool calling to continue authorized work. Examples include reviewing Alipay spending for recurring subscriptions and watching booked flight or hotel prices. Ant says the upgrade is rolling out in China and users can revoke permissions. These are company capability and security claims. Clinicians looking for treatment guidance need evidence they can inspect. Vera says its AI clinical search engine retrieves medical evidence before generating answers, with linked sources, and searches over sixty million publications and clinical references. The company announced a 28.5 million dollar funding round to expand evidence infrastructure, enter markets and develop partnerships. Vera reports use by over 800,000 clinicians, but the funding announcement does not establish improved patient outcomes. AI use in securities markets raises questions about reliable data, model oversight and accountability, India’s regulator says. Its chair says the agency is implementing an international AI supervisory toolkit to strengthen risk management and governance. He also described exploring a securities-specific AI model built and governed domestically. No commitment or implementation timetable was reported. The toolkit and proposed model concern oversight; the model exploration is not an adopted rule. Public oversight of AI could extend to its environmental footprint and infrastructure planning under a researcher proposal. Seventeen researchers recommend thirteen rules for the European Union and United Kingdom, including general-purpose AI licensing, environmental disclosure and public consent for new data centres. They urge leaders to address AI at the November twentieth summit. The white paper also proposes tax changes concerning automation and employment. These recommendations are not adopted policy. Studying how people use AI at work requires evidence about the tasks they bring to it. OpenAI’s October twenty-sixth Signals release maps individual-plan ChatGPT work messages into 165 activities across nearly one hundred countries. Its quarterly data runs from early 2025 through the third quarter of 2026. OpenAI says automated classification follows removal of identifying information, with safeguards for aggregate statistics. Enterprise messages and Codex are excluded; this is not a measure of jobs displaced or productivity. AI research projects could address problems ranging from drug discovery to tsunami modelling. Japan’s science agency selected thirty-one AI-for-science projects on September twenty-ninth. The published list includes cancer-cell digital twins for drug discovery, a genome language model and AI-driven fusion-reactor control. The agency split selections between twelve strategic projects and nineteen international or interdisciplinary projects. Research was planned to begin after contracts were completed; the announcement reports selections rather than discoveries. Security weaknesses in critical infrastructure are the focus of SoftBank’s OpenAI-powered patching service. The companies describe scanning source code, assessing externally visible systems, prioritizing risks and creating patches tested in simulation before deployment. The July fourteenth full-launch announcement targeted rollout to 3,000 eligible Japanese companies, not 3,000 enrolled customers. Early scans reportedly found about 280 potential vulnerabilities per ten million lines of code. That company figure does not measure prevented attacks. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

206

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 4:10 PM UTC

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 4:08 PM UTC

TITLE AI News: Claude Drug Research, AI Data Exposure, and $63 Billion Compute Financing youtube.com/@BriefReplayAI DESCRIPTION AI drug research, privacy failures, safety oversight, workplace tools and computing plans worldwide. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Claude supports rare-disease research 00:17 Cloudflare reports AI data exposure 00:39 Tokyo Metro plans AI track warnings 01:00 Report questions immigration AI secrecy 01:22 Australia warns AI scales cyberattacks 01:41 Germany expands AI evaluation plans 02:06 Report questions AI zone consultation 02:28 Check Point claims faster AI research 02:54 Oracle develops ChatGPT workflows 03:17 UOB targets AI economic value 03:47 KDDI plans robot AI research funding 04:09 Docomo links AI agents and robots 04:30 KDDI plans AI avatar role-play 04:52 Docomo prototypes AI habit prompts 05:14 Anthropic AI compute financing planned 05:37 Oracle plans Google AI chips for 2027 05:59 Report: OpenAI plans Google AI chips 06:19 Luxor plans AI hardware expansion 06:40 Singapore funds AI materials platform 06:59 Singapore and Germany plan industrial AI 07:22 PLaMo adds private AI research 07:50 Giotto offers local reasoning AI SOURCES anthropic.com/research/early… — Anthropic blog.cloudflare.com/kv-cache… — Cloudflare tokyometro.jp/news/2026/2256… — Tokyo Metro openrightsgroup.org/press-re… — Open Rights Group asd.gov.au/news/2026-10-27-a… — Australian Signals Directorate dserver.bundestag.de/btd/21/… — German Federal Government response to Bundestag question spotlightcorruption.org/repo… — Spotlight on Corruption theguardian.com/uk-news/2026… — The Guardian openai.com/index/check-point… — OpenAI checkpoint.com/press-release… — Check Point openai.com/index/oracle-apps… — OpenAI fintechnews.sg/139741/ai/uob… — Fintech News Singapore meti.gov.sg/newsroom/deliver… — Singapore Ministry of Trade and Industry zaobao.com.sg/finance/singap… — Lianhe Zaobao newsroom.kddi.com/news/detai… — KDDI / ELYZA ai.watch.impress.co.jp/docs/… — Impress Watch / AI Watch docomo.ne.jp/info/news_relea… — NTT Docomo newsroom.kddi.com/news/detai… — KDDI prtimes.jp/main/html/rd/p/00… — NTT Docomo via PR Times investors.broadcom.com/news-… — Broadcom and financing partners / PR Newswire apnews.com/article/e11a2240f… — Associated Press cloud.google.com/blog/produc… — Google Cloud streetinsider.com/Reuters/Op… — Reuters via StreetInsider investing.com/news/stock-mar… — Reuters via Investing.com luxor.tech/news/corporate-ne… — Luxor Technology coindesk.com/business/2026/1… — CoinDesk businesstimes.com.sg/singapo… — The Business Times pmo.gov.sg/newsroom/pm-lawre… — Prime Minister's Office Singapore channelnewsasia.com/singapor… — CNA preferred.jp/ja/news/pr20261… — Preferred Networks itmedia.co.jp/aiplus/article… — ITmedia AI+ webdisclosure.com/press-rele… — Giotto.ai SA (via EQS/WebDisclosure) TRANSCRIPT Rare muscular dystrophies need treatment research, and Anthropic says Claude helped Mithril coordinate work with eleven outside laboratories. The company reports promising mouse results for one candidate in under eighteen months. Most work remains preclinical, with no established human benefit. AI services must keep customers’ requests separate. Cloudflare says a cache configuration mismatch and software bug broke that separation in its Workers AI service. It reports 155 responses containing other customers’ content between August twenty-seventh and September ninth. Cloudflare says it fixed the issue September ninth and found no evidence of deliberate exploitation. Night maintenance drivers must watch track equipment and obstacles. Tokyo Metro plans cameras and AI recognition that scan roughly fifty meters ahead and warn drivers. The portable system could be shared among 117 work vehicles. Procurement is planned in fiscal twenty twenty-six, with operation by fiscal twenty twenty-seven’s end; it supplements human checks. Migrants need to understand and challenge decisions affecting their cases. Open Rights Group says British immigration systems use AI for risk assessment, interview summaries and case prioritization. Its report warns that secrecy can prevent people learning when AI was involved or challenging errors. These are advocacy findings, rather than independently measured system outcomes. Scams and phishing threaten households and essential services. Australia’s Signals Directorate says AI helps attackers operate faster, at greater scale and with more convincing messages. Its annual report records more than eighty-two thousand cybercrime reports overall. Those totals do not identify which incidents involved AI. Advanced AI risks require independent evaluation. Germany says it is expanding its national AI safety institute and building model-testing expertise. Its October twentieth government answer says the European AI Office obtained OpenAI incident information under AI Act obligations. German officials analyzed an alleged July test-model escape and attack involving Hugging Face, but disclosed no technical findings. AI data centres can affect local water use and emissions. Spotlight on Corruption examined representation in British AI infrastructure policy. It says eighty-four percent of ministerial stakeholder meetings on data centres involved business interests between July twenty twenty-four and June twenty twenty-six. The government says it welcomes engagement and local planning considers community benefits. Security researchers must understand software flaws before developing protections. Check Point says OpenAI’s Daybreak analyzes vulnerabilities, code, configurations and possible attack paths. Its technology chief claims research that took hours now takes minutes. The companies are also developing a network-security agent requiring human approval for consequential actions; neither claim establishes independently measured customer protection gains. Employees completing finance or procurement work need access to approved business data. Oracle is developing an AI Workspace that connects its applications with ChatGPT while retaining existing permissions. The partners also announced AI security testing, application building and a small-business bundle. These are development plans and new product announcements, rather than independently measured productivity results. Bank compliance, credit and customer support involve substantial information work. Fintech News Singapore reports UOB’s Project AURA includes eleven AI initiatives spanning those functions and collections. UOB aims for nine hundred million Singapore dollars in economic value by twenty thirty, a target rather than investment or realized returns. Singapore’s first Champions of AI cohort includes twenty-four companies committing to workforce development and AI capabilities. Retail, logistics and manufacturing involve repetitive physical work. KDDI and ELYZA plan robot foundation models connecting visual understanding, language and actions. KDDI targets thirty billion yen in research investment through fiscal twenty thirty and more than one thousand workplace robots by then. These are research and deployment targets, not completed installations. AI-directed robots need checks on what actions they may perform. Docomo says its new platform connects personal AI agents with robots, checks authorization and tracks actions. Its examples include withholding an unsuitable food order and routing around people. A November demonstration and further trials are planned; these examples do not establish field-tested safety. Service businesses face staff training burdens. KDDI’s Avatar Canvas uses AI customers for employee role-play and AI avatars for initial customer responses, with specialist consultations handed to humans. Business availability and training rollout at its au stores are planned from November second. The announcement does not establish measured benefits at outside customers. Repeated reminders about brushing teeth or bedtime can burden caregivers. Docomo’s AI prototype combines a language model with a child-vocabulary database to personalize prompts. The company says its preliminary demonstration manually selected vocabulary and used synthetic speech. It explicitly does not establish effectiveness; home trials remain planned. Anthropic’s AI workloads require computing infrastructure. A financing vehicle would buy Broadcom AI accelerators and networking equipment and lease them to Anthropic. The October twenty-sixth announcement describes sixty-three billion dollars in committed financing from major financial firms. Equipment deliveries are expected to begin in twenty twenty-seven, rather than representing capacity already deployed. Organizations training and running AI models need computing options near their data. Google and Oracle plan to offer Google’s AI accelerators through Oracle Cloud Infrastructure. The companies say deployment will start in twenty twenty-seven, making Oracle the first hyperscale cloud outside Google to offer them. This is future capacity, without established customer savings. OpenAI’s AI workloads need chip capacity. Reuters, citing The Information and an unnamed source, reports plans to buy Google’s AI accelerators. The report gives no financial terms or other deal details. Google did not immediately comment and OpenAI declined comment, leaving the purchase plan unconfirmed by either company. AI hardware buyers face capacity supply and price uncertainty. Luxor plans to expand hardware brokerage and computing-power trading. The company reports a sixteen-million-dollar strategic round. CoinDesk says it includes both new investment and existing-share sales, with the amount of fresh capital undisclosed; no consumer cost reduction is established. Materials research requires identifying candidates and testing them. Singapore’s planned MARVEL-AI platform combines AI-designed experiments with automated synthesis and testing labs. The government commits an initial one hundred fifty million Singapore dollars. This funds research infrastructure, without establishing new commercial materials. Manufacturing and logistics are targets for practical AI applications. Singapore and Germany plan deeper cooperation on industrial AI, drawing on their engineering and AI capabilities. Their updated bilateral work plan includes digitalization and AI alongside other areas. The announcement names no specific AI deployment, dedicated budget or project timetable. Organizations researching sensitive documents may need AI within their own premises. Preferred Networks’ PLaMo research agent plans searches, analyzes information and produces sourced reports, including in closed internal-document environments. The company also opened a capped free API tier. Its claimed fifteen-percent improvement concerns average scores on eight hallucination benchmarks, while ITmedia notes performance remains mixed across other tasks. Organizations may need control over where AI models run. Giotto’s Swiss-developed Alto and Solo reasoning models offer European cloud or compatible customer-premises hosting. The company says Solo is designed for one graphics processor and Alto for eight B300 processors. These are company specifications, without independently verified performance or adoption evidence. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

285

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 8:36 AM UTC

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 28, 2026 · 8:33 AM UTC

TITLE AI News: Election Fakes, Chatbot Safety Gaps and Cancer Photo Research youtube.com/@BriefReplayAI DESCRIPTION AI election impersonations, chatbot escalation research, cancer-photo risk groups, Togo language models and retrospective care review. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Likely AI fakes target candidates 00:24 AI alerts triggered clinician outreach 01:03 AI groups cancer survival risk 01:39 Togo collects data for local-language AI 02:16 AI reviews documented care goals SOURCES apnews.com/article/c52f2c172… — Associated Press graphika.com/insights/storm-… — Graphika pubmed.ncbi.nlm.nih.gov/4285… — PubMed / NLM formative.jmir.org/2026/1/e1… — JMIR Formative Research ebs.publicnow.com/view/E6B97… — Mass General Brigham sciencemediacentre.org/exper… — Science Media Centre doi.org/10.1016/j.landig.202… — The Lancet Digital Health numerique.gouv.tg/actualites… — Ministry of Public Service Efficiency and Digital Transformation, Togo undp.org/fr/togo/actualites/… — UNDP Togo togofirst.com/en/itc/1410-20… — Togo First pubmed.ncbi.nlm.nih.gov/4272… — PubMed / NLM rethinkingclinicaltrials.org… — NIH Pragmatic Trials Collaboratory TRANSCRIPT Fabricated allegations against Senate candidates put trustworthy election information at risk. An expert told AP two circulating videos likely involved AI. X removed initial posts, but copies remained. Graphika separately assessed an October twentieth Talarico clip as a very likely deepfake, citing facial movement and detection tools. Its attribution to a Russia-aligned operation is an assessment, not a proven electoral effect. AI mental-health chatbots need a way to pass possible suicide risk to people who can assess it. A study first shared July second used questionnaire answers and conversational signals to trigger clinician outreach, leaving clinical assessment to humans. The October twenty-third final paper reports 172 flagged participants among 650 in its second phase. Clinicians reached 73, or 42.4 percent, and began outreach after an average 11.5 minutes. That measures outreach initiation, not arrival of help. The implementation study does not establish improved clinical outcomes, and many flagged users were not reached. AI analysis of cancer patients' photographs could inform risk research, but clinical usefulness must be demonstrated. An October nineteenth study adapted a foundation model trained on more than 40 million facial images into facial-age and cancer-survival tools. An independent expert describes validation in nearly 5,000 Netherlands patients. These are retrospective survival-risk groups, not validated predictions of an individual's remaining lifetime. The expert says the model was not compared with clinicians' own assessments and has not shown added clinical value. It is not ready for clinical use. AI that supports local languages could help people access administrative information in languages they speak. Togo opened its language challenge on October second, collecting recordings, texts and translations for models in Ewe, Kabiye and Moba. UNDP is helping Togo AI Lab establish computing infrastructure for training. The ministry's one-month pilot targets 50 validated audio hours, 6,000 translated sentence pairs and at least 500 speakers per language. Education, health, justice and agriculture are intended later uses. This is data collection and model development; concrete public-service results remain to come. Reviewing whether end-of-life treatment followed patients' documented goals is difficult when preferences change or records lag. A September tenth study had two language models and two human reviewers assess 120 purposefully selected records. Agreement scores ranged from 0.43 to 0.75; the highest was between GPT-4o and one human. Those are agreement measures, not percentages of clinical accuracy. Disputes included changing goals, conflicting surrogates and delayed code-status updates. The authors describe proof of concept for retrospective research, with human review still important for complex cases. The models were not choosing treatment. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 27, 2026 · 11:47 PM UTC

🤖 Made with AI

172

🇺🇸 Roman R 🇺🇸 · Oct 27, 2026 · 11:44 PM UTC

TITLE AI News: Oracle Opens Rubin Cloud, AI Flags Cancer Trial Reviews, Voice AI Speeds Up youtube.com/@BriefReplayAI DESCRIPTION AI cloud expansion, cancer trial screening, Muse agents, voice AI, robot learning and biological AI safeguards. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Oracle opens Rubin AI cloud 00:34 AI flags 35 for cancer trial review 01:06 Nvidia backs Muse AI development 01:29 NTT voice AI cuts response delay 01:59 Nvidia funds AI chip-testing capacity 02:28 Nvidia expands robot AI training 03:01 Protein AI filtering can be reversed 03:33 AI agents rebuild selected bio functions SOURCES nvidianews.nvidia.com/news/n… — NVIDIA Newsroom blogs.oracle.com/cloud-infra… — Oracle Cloud Infrastructure Blog nvidianews.nvidia.com/cats/p… — NVIDIA Newsroom Press Release RSS paradigmhealth.ai/resources/… — Paradigm Health investor.nvidia.com/news/pre… — NVIDIA Investor Relations axios.com/2026/10/27/nvidia-… — Axios nvidianews.nvidia.com/news/n… — NVIDIA Newsroom group.ntt/jp/newsrelease/202… — NTT nvidianews.nvidia.com/news/n… — NVIDIA Newsroom nvidianews.nvidia.com/news/n… — NVIDIA Newsroom aisi.gov.uk/blog/pretraining… — UK AI Security Institute arxiv.org/abs/2610.26651 — arXiv / UK AISI authors aisi.gov.uk/blog/can-ai-codi… — UK AI Security Institute arxiv.org/abs/2610.19810 — arXiv / UK AISI authors TRANSCRIPT Training large AI models and running their answers require substantial computing resources. Oracle and Nvidia say Rubin GPU systems and Vera processors are generally available on Oracle Cloud, supporting frontier-model training and inference. Vera processors also handle AI-agent tasks between model calls, including testing code and retrieving data. Oracle reports an initial deployment of one hundred fifty thousand processors for agent workloads. The partners plan to expand AI superclusters to multi-gigawatt scale; that expansion is still a plan. Finding possible cancer trial options requires checking medical records against detailed criteria. Research announced October twenty-fifth used language models to translate Japanese records and structure their contents, with cited evidence for physician review. The partners report that screening three thousand twenty-seven lung-cancer records flagged thirty-five patients for clinical review and follow-up testing. This was retrospective research at one hospital, not confirmed trial eligibility or enrollment; clinicians make the final decision. Meta wants to improve the training and operation of its Muse AI-agent models. Nvidia and Meta are expanding work on model performance alongside planned deployment of Nvidia computing and networking systems. Nvidia proposes investing three billion dollars in Meta, subject to closing conditions. This is a model-development partnership expansion; improved agent performance remains the companies' stated aim. Call-center AI must cope with pauses and interruptions while following the customer's request. On October twenty-sixth, NTT announced voice AI that listens and responds at the same time, and can consult or update external databases during conversation. NTT says average response delay was more than fifty percent shorter than other services in its tests. That measures response delay, not total call length. Call-center application remains an aim, with no live customer deployment reported in the announcement. AI computing needs manufacturing capacity beyond the processors themselves. On October twenty-first, Nvidia announced supplier plans for substrates, packaging, testing and optical connections used in AI systems. Nvidia says a one-point-one-billion-dollar capacity advance will support new American testing facilities, targeting high-volume production in April twenty twenty-seven. Substrate agreements target two-and-a-half times the capacity over three years. These are announced expansion targets. Training robot AI requires connecting learned skills with sensors, controls and physical hardware. Nvidia expanded Isaac on October twenty-first, including training tools and an open-source runtime that connects those components. It also announced workflows for reinforcement learning and synthetic training data. Nvidia says Isaac Lab three is generally available. Agile Robots' reference humanoid, configured with Nvidia robot AI, is scheduled for early twenty twenty-seven; this is a development platform, not a deployed general-purpose worker. Safety measures for protein AI need to remain effective when users can retrain the models. An October nineteenth preprint tested protein models trained without viral sequences. Researchers found weaker performance on selected viral-protein tasks, while performance on non-viral proteins was largely preserved. Fine-tuning restored lost performance for under eighty dollars of compute even on the largest tested model. The UK AI Security Institute says filtering needs additional safeguards; the finding concerns tested models and tasks. Restricting source code may leave gaps in controls on biological AI tools. An October fourteenth study tested coding agents given public descriptions, supplied training data and fixed time and compute budgets in an isolated environment. The strongest result reached ninety-six-point-five percent of the original protein-design tool's sequence-recovery performance, a selected function rather than a complete recreation. Agents failed to reproduce the more complex tool, RFdiffusion three. Researchers say controlling access to underlying data may be more promising than withholding code alone. Subscribe to stay updated every eight hours on AI developments worldwide.

🤖 Made with AI

267

🇺🇸 Roman R 🇺🇸 · Oct 27, 2026 · 4:00 PM UTC

🤖 Made with AI

🇺🇸 Roman R 🇺🇸 · Oct 27, 2026 · 3:58 PM UTC