AI News - Sunday, 22 November 2026 - Gemini 4 Neon, Genie 4 World Model, Tool Retrieval
In a Nutshell
AI moved from claims to verification this weekend: two groups independently confirmed that Claude helped produce a Lean proof of the Riemann hypothesis, while Google shipped Gemini 4 Neon at a fraction of frontier pricing and opened Genie 4 worlds to subscribers. Markets were less convinced, with US tech in a bear market, Nvidia under subpoena in California, and UN talks on autonomous weapons collapsing.
5-minute AI news update - 22 November 2026
In this AI News
Google releases Gemini 4 Neon, a cheaper frontier model built for agent work
[Models] Neon is priced at 3 dollars per million input tokens and 15 dollars per million output, with a 1 million token context window. Google ranks it third overall on the Artificial Analysis Intelligence Index at its highest thinking level and first on Zapier's AutomationBench. That pricing moves frontier-grade agent work within reach of mid-size institutions, not just hyperscalers. Source: Google DeepMind
Google opens Genie 4 world model with real-time prompt edits and native audio
[Models] AI Ultra subscribers can now rewrite a running world mid-session, changing weather, adding characters or redirecting the action by typing. Genie 4 pairs the visuals with adaptive audio, and a community gallery supports sharing and remixing. Editable generated environments are becoming practical for simulation and training work without a game engine. Source: Google DeepMind
Gemini API adds Tool Retrieval so agents load tool schemas only when needed
[Tools] Developers add a tool_search tool and set defer_loading, so the model still sees every tool name and description but pulls the full schema on demand inside the same request. Google says this cuts input tokens and latency while keeping tool selection accurate as libraries grow. It answers the first scaling problem every agent platform hits. Source: Google
Two groups independently verify that Claude helped prove the Riemann hypothesis
[Research] Levent Alpoge, a mathematician employed by Anthropic, published a Lean formalisation developed over about six weeks, and the Lean Focused Research Organization and Mathlib Initiative confirmed it on Saturday. The verification, not the claim, is the story: machine-checked proof turns model output into an auditable artefact. It also resets expectations for what one researcher with a model can attempt. Source: The Guardian
MIT panel asks what AI mathematics does to the people who practise it
[Research] Mathematicians from MIT, Harvard, Northwestern and NYU described open problems falling in hours to agents deployed in the thousands. The panel split between embracing the tools and defending human understanding as the point of the discipline. Every institution that teaches mathematics now faces the same questions about assessment and authorship. Source: MIT News
Northeastern and Stanford build Shepherd to supervise and roll back rogue agents
[Tools] Shepherd stores every agent action, state and environment as a version-controlled commit, so a meta-agent can intercept an action, revert it or escalate to a human. Outbound actions land in a sandbox before they are real. The paper goes to NeurIPS in December, and it frames agent safety as governance code rather than better prompting. Source: IEEE Spectrum
Text-message agents become a product category as Instinct hits a 10 billion valuation
[Tools] TechCrunch catalogues assistants that live inside SMS or iMessage threads, from Ask Proxy at a dollar a day to work versions at 15 to 23 dollars per employee monthly. Memory and access to calendar, mail and bookings matter more than the underlying model. For institutions the channel removes install friction but adds a data-governance question, since conversations leave the corporate tenant. Source: TechCrunch
Apple is hiring to build an AI agent that operates a Mac like a person
[Tools] A job listing describes an agent that browses, works across apps, runs commands and debugs code, privately. Bloomberg reports the work is early and could ship at the end of next year, while macOS 27.3 adds warnings against granting third-party agents full machine access. Apple is building the capability and the guardrail at the same time. Source: TechCrunch
Amazon nears a 7.5 billion dollar deal for Decart, an Israeli video-model startup
[Funding] The acquisition would give AWS in-house video generation and world-model capability as it competes with Nvidia for compute demand. Decart is three years old, and reports of an Amazon approach first surfaced last month. Cloud providers are now buying model teams outright rather than only renting them capacity. Source: Financial Times
California subpoenas Nvidia over whether restricted chips reached China
[Policy] Attorney General Rob Bonta said the conduct of Nvidia and its former chief executive raises bright red flags, including whether investors were misled, and the probe will examine if confidential information reached China. The subpoena follows Jensen Huang's resignation over multiple affairs, one involving a woman suspected of espionage. Nvidia shares are down about a fifth since the disclosure. Source: The Guardian
US tech stocks end a brutal week as doubts about AI spending bite
[Industry] The Nasdaq fell 5.3 percent on Friday and the S&P 500 3.7 percent, with the S&P and Dow pushed into correction territory. Anthropic closed down 8.5 percent, back near its 83 dollar IPO price after a strong debut. More than a trillion dollars of AI capital spending in a year now has to be defended to public markets every quarter. Source: BBC News
UN talks on a treaty to control autonomous weapons collapse in Geneva
[Geopolitics] A week of negotiation ended near 5am with only a weak consensus mandate after Russian stalling, and the United States declined to commit to a binding protocol. More than 100 states had pushed for new law. Military AI is now advancing without an international instrument, which pushes the question onto national procurement rules and export controls. Source: The Guardian
China moves to standardise how AI tokens are measured, billed and quality-checked
[Policy] China's National Data Administration will set standards on token measurement, billing and quality assessment as Beijing guides the industry. Token accounting is becoming an economic policy instrument, not merely an engineering convention. Vendors selling into Chinese markets will converge on those definitions, and buyers elsewhere will have to map them. Source: Bloomberg
Google tells US schools that blocking AI summaries means losing Search entirely
[Education] Under the policy announced on 3 November, AI tools and Search switch off by default for K-12 accounts in early 2027, and districts that block the AI layer lose Search with it. Fairfax County built a browser plug-in that hides AI answers, and New York, Los Angeles and Chicago are weighing a switch away from Google. Read it as the template for any vendor that bundles AI into a core tool. Source: EdSurge
Outages keep coming as AI-driven bug discovery outpaces patching capacity
[Industry] WIRED reports that this week's failures at Outlook, Apple, several US banks, United Airlines and the NSA had routine causes and no link to a coordinated AI attack. The structural problem is that AI finds vulnerabilities faster than under-resourced teams can validate and ship fixes, so turnaround times defeat patch verification. Resilience planning should assume shorter patch windows, not fewer incidents. Source: WIRED
White House removes national cyber director after a week of tech outages
[Policy] Sean Cairncross left after losing the confidence of the chief of staff, and Secretary of State Marco Rubio named Defense Department chief information officer Kirsten Davies as his replacement. Officials say the outages were not cyberattacks, but Congress demanded a clearer explanation. The vacancy lands as AI-assisted vulnerability discovery starts to shape national patch policy. Source: BBC News
China's biggest tech firms to double AI capex to 139 billion dollars in 2026
[Industry] Rhodium estimates that AI lab revenue covers under 8 percent of that spending, a financing gap that may bind harder than chip export controls. Reporting contrasts Chinese restraint with Silicon Valley raising tens of billions in fresh capital. Compute access is turning into a balance-sheet question on both sides. Source: Techmeme
The world of AI is evolving at full speed.
Become a Fellow at university-365.com
Become Superhuman... In a world of AI... Prompt Smart, Prompt UP!
























Comments