Models, products and AI startups, near-duplicates collapsed to the most credible source
Kimi K3.1 open weights drop cyberattack training dataKimi K3.1's open-weight release reportedly excludes training data related to cyberattacks. The change is attributed mainly to China's new export restrictions and to growing caution over cybersecurity exploits.
Bologna, UN University sign AI cooperation agreementThe University of Bologna and the United Nations University signed a framework cooperation agreement naming AI a main field of collaboration, tied to the forthcoming UNU Institute for Artificial Intelligence (UNU-AI). Activities cover research, higher education, capacity building and policy work for decision-makers.
EPFL develops implantable neural chip LogSyncEPFL researchers built LogSync, a low-power implantable neural system-on-chip that records cross-region brain signals and extracts oscillation and synchrony biomarkers, then uses machine learning to identify cognitive and emotional states in real time. It targets adaptive deep brain stimulation for psychiatric conditions such as depression and PTSD, moving beyond single-region, open-loop systems.
Community launches oai-math-revised to fix OpenAI math write-upsA guest post on Terence Tao's blog introduces oai-math-revised, a community project to improve OpenAI's math write-ups on GitHub, which can be hard to understand and have incomplete citations. Students, postdocs, professors and enthusiasts are invited to contribute; a browsable site is available and admins/reviewers are being recruited.
8GB GPU runs 262K-context Qwen quant at 70 T/sA user demonstrated running ISTA DASLab's 3.5 BPW Qwen-Flash-Next Coder quant on 8GB VRAM plus 24GB RAM, reaching 262K context with about 70 T/s prefill and 8 T/s generation via MoE offloading and disk n-gram. Half the experts are pruned, giving an effective 1.89 bits per parameter while retained weights stay at 3.5 BPW.
Study: Grok picks Republicans, GPT picks DemocratsA Convovo study prompted Grok 4.7 and GPT-6.1 Sol via API to name one candidate in nearly 500 races. Grok picked the Republican 92% of the time and GPT the Democrat 97%, agreeing under 10%. Consumer apps refused; OpenAI said API output doesn't reflect ChatGPT's experience.
Anthropic sued over Elvis Presley music used to train AIAnthropic faces a lawsuit alleging it used Elvis Presley's music to train its AI models, with the rights holders stressing they are deeply protective of these iconic songs. It adds to the growing set of copyright disputes over generative-AI training data.
Field study: AI boosts junior patent attorneys, gains fadeA three-month field experiment with 133 patent attorneys found AI lifts immediate performance, most for juniors. Once AI was removed, only senior attorneys kept the gains, suggesting baseline expertise is needed to turn AI help into lasting skill.
Rise of unsupervised undergrad authors at A* conferences sparks debateA Reddit thread notes that more undergraduates are publishing at A* conferences without a supervisor. The graduate-student poster calls it impressive and intimidating, and questions their own progress.
Atlantic CEO says AI boom faces growing public backlashThe Atlantic CEO Nicholas Thompson told Bloomberg that AI offers powerful business tools but poses significant risks to media, while its broader economic impact lags behind rapidly advancing capabilities, with public backlash growing.
Anthropic's AI Hid Secret Message in Stars, OpenAI's AI Found ItAnthropic's AI hid a secret message in the stars, which OpenAI's AI successfully found. The event demonstrates AI capabilities in hiding and discovering information, but details are not disclosed.
Meta trains byte-level small models, predicted to beat token modelsMeta trained ~1B-parameter byte-level models distilled from Llama 3 8B; they lag token models early but keep improving. Extrapolated curves suggest they could beat Llama 3.2 1B and Gemma 3 1B with about 6x less text, though real runs still favored tokens.
Minuttia Study: Up to Half of AI-Cited Best Tools Lists Written by VendorsMinuttia analyzed 18,084 software questions and found that 34% to 47% of listicles cited by AI search are vendor-written, rising to 56% to 65% for comparison questions. 87% rank the vendor first, and only 16% disclose authorship. The study reveals self-promotional bias in AI search recommendations.
AI wealth fuels high-end matchmaking with packages over $500,000Linx Dating founder Amy Andersen told Bloomberg that newly wealthy AI creators are driving demand for high-end matchmaking, as busy professionals seek an alternative to dating apps. Her firm vets and recruits partners for clients, with packages that can exceed $500,000.
Vietnam deputy PM: AI is reshaping research and teachingAt the 50th anniversary of Ho Chi Minh City University of Education, Deputy Prime Minister Le Tien Chau said AI is profoundly changing how people research, teach and learn, requiring education to shift from transmitting knowledge to developing independent thinking and creativity. He outlined four development directions, urging future teachers to master technology and AI and judging schools by graduates' competence.
Technical colleges bet on AI-driven demand for skilled workersBloomberg reports that technical colleges are betting a surge in demand for skilled workers, driven by AI, will lift profits, expanding campuses and programs to take in more students—though the payoff may take time to materialize.
AI-assisted paper cites a nonexistent TorrentFreak articleA peer-reviewed paper on Z-Library cites a 2021 TorrentFreak article that does not exist, linking only to the homepage. It also claims the DOJ 'arrested individuals' in 2022, which conflicts with the official press release. The case highlights risks of AI-assisted academic citations.
Review: AI evidence in CNS autoimmune disease is unevenA new review assessed AI across multiple sclerosis, neuromyelitis optica spectrum disorder, autoimmune encephalitis and MOG antibody disease, covering automated imaging, differential diagnosis, relapse prediction and prognosis. Evidence is strongest in multiple sclerosis neuroimaging, and the authors say further validation is needed elsewhere.
US pours billions into AI race as insiders warn of risksDW News reports the US is spending billions on AI to boost its economy and stay ahead of China. Tech leaders foresee gains in medicine, jobs and innovation, but researchers, activists and former AI insiders warn the race carries risks.
AI Debt Monster: Hidden Risks in Infrastructure SpendingAI uncovered a PHP vulnerability affecting 70% of the internet but also fueled exploits, highlighting a paradox. However, the bigger risk is AI infrastructure debt: $3.5 trillion in global spending by 2030, with US hyperscaler capex hitting $800 billion in 2026. Hidden commitments and circular investments could trigger an 'AI Enron' financial crisis.
Nvidia Warns: Lack of Computing Power Threatens Europe's AI CompetitivenessNvidia CEO Jensen Huang warned in Berlin that Europe holds only 5% of global AI computing capacity versus 75% for the US, hindering competitiveness. He unveiled the modular 'Vera Rubin DSK Pod,' which cuts data center setup from months to days, and announced $10 billion in investments.
Medicare Advantage plans exit, leaving seniors with fewer optionsMany U.S. Medicare Advantage plans are exiting or shrinking for next year, leaving 193 counties with no plans offered. Last year 2.9 million seniors were affected; this year analysts expect at least 2 million, with other estimates of 3.2–5 million. About 35 million people are enrolled in Medicare Advantage.
Developer jams live with pretrained music transformer via sampling controlA developer built a live duet system using the pretrained Anticipatory Music Transformer, steering sampling with hand-tuned logit biases and no fine-tuning. It runs on a Jetson Orin Nano, samples about a second ahead, resamples on chord changes, and the code is open-sourced.
India's NCERT adds computational thinking, AI module for grades 9-12India's National Council of Educational Research and Training (NCERT) introduced a module on computational thinking and artificial intelligence for classes 9 to 12, bringing the topics into school curricula. Details of content and rollout timeline were not disclosed.
Zoom launches Unified Context Layer to turn AI talk into actionAt Zoomtopia, Zoom announced the Unified Context Layer, connecting meetings, emails, chats, and external apps so AI agents can access cross-tool context and execute tasks. It aims to reduce fragmentation in enterprise AI, moving beyond content generation to process automation. Zoom is also preparing a deepfake detector.
New Gyro quant seen as EXL3-like but closed-sourceA community user analyzed the new Gyro quantization format, finding it in the same QTIP/Trellis-coded family as EXL3 but non-interoperable, with its recipe closed for now. The user found no published evidence that Gyro beats EXL3 on decode speed or perplexity, and worries it fragments the community.
Psychiatrist outlines four risks of AI mental health adviceSurveys show 43% of UK children aged 11-17 have used AI for advice or emotional support, and 57% of adults turn to chatbots due to time constraints on care. Psychiatrist Daniel Masud cites long waits as the main driver and flags risks including chatbots being overly validating and not objective enough.
RTX 5090 seen in Chinese labs, with dies and 32GB VRAM harvested for AI serversA Reddit post says RTX 5090 cards are not gone but appear in Chinese labs, where technicians remove their dies and 32GB VRAM to build AI servers. The practice highlights how high-end GPUs are being dismantled and reused amid AI compute demand.
HKMA launches ART framework for bank AI adoptionThe Hong Kong Monetary Authority released its "Accountability, Resilience, and Trust (ART)" framework and updated supervisory guidance for responsible AI adoption in banking. The guidance focuses on three pillars: effective governance and accountability, proportionate management of AI-specific risks, and building trust and transparency.
Iran official says AI facilities could be targetsIran's Supreme National Security Council secretary Rezaei warned that cables, data centers and AI facilities could fall within range if conflict resumes. Tehran's intelligence assessment sees a high likelihood of large-scale bombing before Israel's vote, while the White House denied any such plan.
The A.I. Daily