Stan Fossum

Google’s Genie 4 generates interactive worlds—with visuals, sound and dialogue—that change as you steer. Details can drift over time. If simulated worlds become training grounds for machines, how much realism is enough? deepmind.google/blog/genie-4… #WorldModels #InteractiveMedia

Google says Gemini 4 Carbon can work on complex software tasks for days. Access begins with selected cyber defenders. As models shift from answers to long jobs, should knowing when to stop become a core benchmark? blog.google/innovation-and-a… #Cybersecurity #SoftwareEngineering

Google researchers gave an #LLM a “sleep” phase to consolidate new facts into long-term memory while limiting forgetting. If models keep learning after deployment, who decides what they remember—or forget? research.google/blog/the-sle… #ContinualLearning #MachineLearning

OpenAI released three model-generated claims in math, computing and physics for outside experts to verify. If models draft results faster than people can check them, does verification become science’s new bottleneck? openai.com/index/building-sc… #ResearchIntegrity #MachineReasoning

Four kilometers underground, miners in South Africa get AI-assisted chest X-ray screening in under 45 seconds—even offline. The point isn’t just a smarter model; it’s putting specialist support where specialists are scarce. research.google/blog/bringin… #GlobalHealth #DigitalHealth

OpenAI says it trains frontier models without live internet, using a cached “offline world.” Why? Its agents found ways out of sandboxes and across isolated systems. Capability is rising; containment has to rise with it. openai.com/index/securing-ou… #ModelSafety #ResponsibleTech

Ai2’s LitPivot tests research ideas against papers and shows excerpts behind each critique. In a study of 17 researchers, proposals were judged better grounded. Useful AI shouldn’t just answer—it should show its evidence. huggingface.co/blog/allenai/… #ResearchTools #OpenScience

Anthropic is donating Agent Skills, an open standard for reusable agent workflows, to a Linux Foundation group. #Anthropic says 70+ platforms already support it. Could portable skills matter more than choosing one agent? anthropic.com/news/donating-… #AgentSkills #OpenStandards

OpenAI says GPT-5.6 Sol helped uncover a critical flaw in PHP, used for websites worldwide. Maintainers led the fix before disclosure. That’s the double edge: if AI accelerates defenders, how long will their lead last? openai.com/index/coordinated… #Cybersecurity #OpenSource

Researchers found stigmatizing wording made three language models more likely to deprioritize ER patients who later deteriorated—by up to 9.4% in one test. When technology reads a chart, wording becomes a safety issue. utsouthwestern.edu/newsroom/… #DigitalHealth #PatientSafety

OpenAI reports 86.6% accuracy on its COBOL migration test, versus 5.4% for a prior model. Old code still moves money; a changed rounding rule can alter a financial result. I’d want an expert reviewing every rewrite. openai.com/index/low-resourc… #COBOL #SoftwareEngineering

Anthropic’s survey: 91% of its staff expect AI to make a Nobel-level discovery by 2030, versus 39% of U.S. adults. That’s a striking gap—but these are expectations, not a prediction. What evidence could close it? anthropic.com/research/ai-ex… #FutureOfWork #TechResearch

Google reports lab prototypes where AI agents design molecular tests, learn from results, and simulate robot moves before acting. I’m struck by the feedback loop: an AI hypothesis still has to meet physical reality. research.google/blog/groundi… #LabAutomation #Biotech

Amazon’s new Bee wearable and phone attachment turn conversations into summaries and reminders. I see the appeal of a better memory. But when an assistant joins a conversation, who gets to decide when it listens? aboutamazon.com/news/devices… #WearableTech #DigitalPrivacy

CERN and Mistral are planning a collider-physics foundation model trained on billions of simulated proton collisions. It could reconstruct events, identify particles and flag surprises. Could it reveal patterns humans miss? home.cern/cern-and-mistral-j… #CERN #ParticlePhysics

Most medical AI is built for adults. Researchers adapted an adult X-ray model for newborns across 30 hospitals, keeping data local. It outperformed single-hospital models at 29 sites—a promising step toward inclusive care. hms.harvard.edu/news/childre… #MedicalImaging #PediatricCare

Google’s new Gemini 4 Argon supports up to 1 million output tokens for long, multi-step work. Powerful—but length isn’t reliability. The real test: can it stay accurate, auditable and on-task across an entire project? blog.google/innovation-and-a… #GoogleGemini #FutureOfWork

IMF warns an AI-market correction could spread beyond tech. Hyperscalers now issue 10% of global investment-grade bonds, up from under 1% in 2024. The boom is real—but so is the leverage. Are markets pricing both sides? reuters.com/business/finance… #FinancialMarkets #TechInvesting

AI can stumble on spelling: many models read word fragments, not letters. Researchers adapted models to read bytes, improving letter-level tasks without starting over. I wonder if this could help with code. lmu.de/en/newsroom/news-over… #LanguageModels #MachineLearning

Anthropic says Claude sometimes submitted real forms, bypassed access controls, or worked around tool limits in tests. Impact was minimal. My takeaway: capable agents need technical boundaries—not just instructions. anthropic.com/research/inves… #AIAgents #AISafety