Gemini 3
Google launched Gemini 3 Pro on Tuesday, its most capable AI model to date. The model, which has generated considerable coverage (including here), demonstrates strong planning, coding, and judgment skills, showing that AI models have progressed beyond hallucinations to subtle, human-like errors.
What to watch:
Google also released “Nano Banana Pro,” its viral-named Gemini 3-powered image model that excels at following instructions and creates full infographics from short prompts.
Google debuted Antigravity, an “agent-first” coding tool that leverages Gemini 3 Pro and third-party models, in public preview for Windows, macOS, and Linux.
Google priced Gemini 3 Pro at $2 to $4 per 1M input tokens and $12 to $18 per 1M output tokens, cheaper than Claude Sonnet 4.5 but more expensive than GPT-5.1.
Google details a new generative UI implementation that lets Gemini 3 Pro create interactive interfaces for any prompt, launching in the Gemini app and AI Mode.
And Google plans to release Gemini 3 Deep Think to Google AI Ultra subscribers in the coming weeks, once it passes further rounds of safety testing.
Elsewhere in frontier models:
OpenAI unveils GPT-5.1-Codex-Max, saying it is “significantly better” at “long-horizon reasoning” and is the first model it has trained for Windows environments.
xAI’s Grok 4.1 tops benchmarks in emotional intelligence, while its model card also shows a marked increase in sycophancy compared to Grok 4.
Allen Institute for AI unveils Olmo 3 models that it says outperform open models like Stanford’s Marin and commercial open-weight models like Llama 3.1.
Meta releases Segment Anything Model 3 for advanced image segmentation tasks.
And Google DeepMind releases WeatherNext 2, a weather model that it says offers faster, more accurate two-week forecasts and includes more tools for energy traders.
Elsewhere in the FAANG free-for-all:
Google says the Gemini app is now able to detect images created or edited by Google AI, and that it plans to roll out verification of video and audio “soon”.
Microsoft launches Agent 365, a framework that lets businesses deploy and manage AI agents like human employees, with dashboards showing telemetry and alerts.
Microsoft integrates more AI features into Windows, including new Copilot skills, AI agents on the taskbar, writing assistance, and troubleshooting agents.
Amazon launches a Video Recaps feature on Prime Video that uses AI to summarize a show’s key plot points with snippets and music, available in beta for US users.
Yann LeCun says he is leaving Meta at the end of 2025 to build a new startup and continue his “Advanced Machine Intelligence research”, with Meta as a partner.
And Meta Chief Revenue Officer John Hegeman is leaving after 17 years at Meta; Clara Shih, who joined in November 2024 to lead its Business AI unit, is also exiting.
Irrational Exuberance
Tech leaders are now openly questioning the trillion-dollar AI investment boom, with Google CEO Sundar Pichai warning of “elements of irrationality” in the current spending frenzy, the CEO of Klarna being “nervous about the size of [data center] investments,” and Peter Thiel’s hedge fund selling its entire $100 million Nvidia stake.
Between the lines:
AI company valuations continue to soar - Nvidia’s $7T, Alphabet’s $3.5T, OpenAI’s $500B, Anthropic’s $350B, and xAI’s $230B - raising questions about whether AI-powered products will generate enough revenue to justify the massive capital outlays.
Pichai’s candid admission that “no company is going to be immune” to a potential AI bubble burst - including Google - marks a notable shift from the unchecked optimism that has characterized tech leadership rhetoric.
JPMorgan analysts estimate that $5 trillion in global AI infrastructure investment through 2030 would require generating an additional $650 billion annually - equivalent to every iPhone owner paying $35 more per month - to deliver reasonable returns to investors.
Despite the warnings, tech giants are doubling down rather than pulling back - but supply chain constraints for critical components like transformers (booked through 2028) and power generation equipment are creating hard limits on how fast data centers can actually be built.
Elsewhere in AI infrastructure:
The US DOE accelerates its approach to equipping national labs with AI supercomputers by working with Nvidia, AMD, and Oracle, which will pay some of the costs.
AI server shipments from Taiwan to the US are set to double in 2025 compared to 2024, as Foxconn, Quanta, and Wistron ramp up production to meet demand.
The US Commerce Department approves the sale of up to 35K Nvidia GB300 servers or their equivalents each to Abu Dhabi’s G42 and Saudi government-backed Humain.
Elon Musk says xAI plans to develop a 500-megawatt data center in Saudi Arabia with state-backed AI startup Humain; the data center will rely on Nvidia chips.
And Elon Musk, Jeff Bezos, and Sundar Pichai have each discussed building lunar and orbital AI data centers to leverage constant solar power and bypass regulations.
Elsewhere in OpenAI:
OpenAI expands group chats in ChatGPT globally to all logged-in users on Free, Go, Plus, and Pro plans after piloting the feature in select regions.
OpenAI announces ChatGPT for Teachers, designed for K-12 educators and school districts, which will be free to K-12 educators in the US through June 2027.
OpenAI and Intuit sign a multiyear deal set to generate $100M+ in OpenAI revenue, in which Intuit will deepen its use of OpenAI models and launch ChatGPT apps.
OpenAI says ChatGPT will now avoid em dashes if users tell it to, addressing the telltale sign that has signaled AI-written text.
And OpenAI says GPT-5 has demonstrated the ability to accelerate scientific research workflows but can’t run projects or solve scientific problems autonomously.
States Rights
President Trump is reportedly preparing an Executive Order as soon as Friday that would give the federal government sweeping power to override state AI laws, including creating an “AI Litigation Task Force” under the attorney general to go after “woke” state AI regulation.
Why it matters:
This would represent a dramatic assertion of federal power over AI regulation, with the Justice Department empowered to challenge state laws like California’s AI safety requirements and Colorado’s algorithmic discrimination protections.
The “AI Litigation Task Force” would weaponize critical infrastructure funding - states could lose Broadband Equity Access and Deployment (BEAD) program money for rural broadband if they maintain AI regulations deemed “onerous.”
But the EO wouldn’t pass without a fight - not only will there be lawsuits, but also intra-party dissent. High-profile Republicans - including Republican governors, senators, and MAGA supporters - are already openly opposing the President.
Ultimately, a pro-AI stance is increasingly a risky political calculation for both parties: concerns about AI-related child suicides, job losses, and energy costs could become devastating attack lines in 2028.
Elsewhere in AI geopolitics:
The European Commission’s postponement of a key part of the AI Act has become a case study for critics who say the EU prioritizes regulation over innovation.
Key White House officials are pressing lawmakers to keep AI chip export restrictions to China out of the annual defense policy bill.
Pro-AI super PAC Leading the Future is targeting New York State Assembly member Alex Bores, who co-sponsored the RAISE Act requiring AI safety protocols.
An investigation reveals that as US Commerce Secretary Howard Lutnick pushes AI data centers and pressures US allies, his companies and sons are profiting from them.
And the current AI strategies of China and the US are complementary, as China focuses on embodied AI and open source models rather than AGI like the US.
Elsewhere in AI anxiety:
A study found that teen mental health chatbots like ChatGPT, Claude, Gemini, and Meta AI often failed to recognize signs of mental health conditions and gave general advice.
AI is pressuring Poland’s economy, Europe’s top location for global business services, as companies lay off staff in Kraków’s BPO hub.
Some experts question Anthropic’s claims of cyberattack breakthroughs using its tools, noting that white-hat hackers report modest gains from AI-aided hacking.
Chinese toymaker FoloToy suspended sales of its GPT-4o-powered teddy bear after researchers found the toy gave kids harmful responses about fire and kinks.
And Microsoft warns that Copilot Actions in Windows, now in beta and off by default, can infect devices and pilfer data, prompting concern from security researchers.
Things happen
Warner Music settles its lawsuit with AI music startup Udio. Soumith Chintala, PyTorch co-creator, joins Mira Murati’s lab. Poe launches group chats with up to 200 people across 200+ AI models. A look at Inception Point AI’s podcast network generating 3,000 podcasts per week. Cloudflare acquires Replicate, which hosts 50K+ AI models. Larry Summers resigns from OpenAI’s board after his Epstein emails surface. Elon Musk could “drink piss better than any human in history,” Grok says. Massive leak shows chatbot users turned yearbook pictures into AI porn. A researcher made an AI that completely breaks online surveys. DeepMind plans to open an AI research lab in Singapore. An overview of macro tech trends for 2026 as AI eats the world. At an AI conference, attendees said they’d short Perplexity over OpenAI. AI is a front for consolidation of resources and power. The AI bubble is bigger than you think. Claude launches structured outputs on its developer platform. Solving a million-step LLM task with zero errors. Heretic: Automatic censorship removal for language models. Google opens its biggest AI hardware hub outside the US in Taipei. Berkshire Hathaway discloses a $4.3B stake in Alphabet. Demis Hassabis pursued a Nobel as a DeepMind objective over revenue. Meta will assess employees on their “AI-driven impact” starting in 2026. Apple requires apps to disclose AI data sharing with third-party providers.
