AI This Week
Caitlin Kalinowski resigned from her role leading OpenAI's robotics team in response to the company's controversial Pentagon agreement. The departure rocks Silicon Valley. She warned that "surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got". Kalinowski stepped down on "principle" after OpenAI revealed plans to make its AI systems available inside secure Defense Department computing systems. She clarified that the announcement was rushed without guardrails defined, calling it "a governance concern first and foremost". OpenAI defended the deal, stating it "creates a workable path for responsible national security uses of AI while making clear our red lines". The controversy exposes growing tensions between AI innovation and military oversight.
OpenAI has released GPT-5.4, its first general-purpose model with native computer-use capabilities that enable agents to operate computers and carry out multi-application workflows. The model scored 87.3% on internal spreadsheet modeling benchmarks for investment banking tasks, up dramatically from GPT-5.2's 68.4%. OpenAI paired the launch with ChatGPT for Excel, which can build, update and analyze spreadsheet models directly inside workbooks. The spreadsheet product is rolling out in beta to ChatGPT Business, Enterprise, Edu, Teachers, Pro and Plus users in the U.S., Canada and Australia. This positions OpenAI more directly against Anthropic and Google in the enterprise market, where Anthropic introduced Claude for Financial Services in July 2025.
OpenAI is developing a new code-hosting platform to rival Microsoft's GitHub after engineers encountered a rise in service disruptions that rendered GitHub unavailable in recent months. The project is in its early stages and likely will not be completed for months. Employees working on it have considered making the code repository available for purchase to OpenAI's customer base. The move could pit OpenAI directly against Microsoft, which owns GitHub and holds a major stake in the AI company. GitHub reported a 58% year-over-year increase in incidents during the first half of 2025, rising from 69 cases to 109 with 17 classified as "major," totaling over 100 hours of disruption. Microsoft currently holds roughly 27% of OpenAI and acquired GitHub in 2018 for $7.5 billion.
Chinese AI company DeepSeek is preparing to unveil its mostadvanced artificial intelligence model yet, trained on Nvidia's cutting-edgechips. The launch could happen as soon as March 2026. Timing is everything. Therelease is strategically scheduled before China's "Two Sessions"parliamentary meetings, which kick off March 4. This positioning could cementDeepSeek's status as a national AI champion. While the company has been rollingout incremental updates, competitors like Alibaba have been aggressive. They'recapturing significant market share in the lower-cost, open-source AI space.DeepSeek's new model represents a major step forward, leveraging the mostpowerful hardware available. The move signals China's continued push todominate the global AI landscape. All eyes will be on this March debut.
The AI boom is reshaping global infrastructure at unprecedented scale. Meta announced a staggering $60 billion commitment to U.S. data centers and computing infrastructure through 2028. The social media giant is building two massive new data centers and has locked in a $10 billion deal with Google Cloud. But Meta's spending spree pales beside Oracle's coup. The database company secured a jaw-dropping $300 billion, five-year compute contract with OpenAI beginning in 2027. The deal briefly catapulted Oracle founder Larry Ellison to the top of the world's wealth rankings. These massive infrastructure investments underscore the computing power required for advanced AI development. They signal that tech companies believe the AI revolution demands unprecedented physical and financial resources to maintain competitive advantage.
Block said it will lay off more than 4,000 employees, shrinking its staff from over 10,000 to just under 6,000, according to a letter from co-founder and CEO Jack Dorsey. Shares surged as much as 24% in extended trading and were up about 18% premarket Friday. CFO Amrita Ahuja said the cuts position Block for its next phase of long-term growth, as the company shifts to smaller teams and uses AI to automate more work. Dorsey said many companies may make similar structural changes as “intelligence tools” drive efficiency gains. Block announced the move alongside fourth-quarter results: adjusted EPS of 65 cents on $6.25 billion in revenue, with gross profit up 24% to $2.87 billion.
Google has released Nano Banana 2, an updated version of its fast-spreading AI image generator. The original Nano Banana launched in August and went viral, followed by Nano Banana Pro in November, built on Gemini 3 Pro. Google says Nano Banana 2 brings stronger world knowledge by pulling real-time information from Gemini to produce more accurate images. The update also speeds up generation, tightens instruction following, and improves text rendering for uses like marketing mockups and greeting cards. Nano Banana Pro will stay available for high-fidelity work that needs maximum factual accuracy, while Nano Banana 2 targets rapid output plus image-search grounding. The new model replaces the prior version across Gemini’s Fast, Thinking, and Pro options.
A new Harvard Business School working paper suggests artificial intelligence can replicate many mutual fund managers’ trading decisions, often without placing a single trade. The research, titled “Mimicking Finance,” reports that 71% of managers’ trade directions can be predicted, based on analysis of data from 1990 to 2023 that factored in fund size, economic indicators, and investor flows. For some managers, predictions rose to nearly all trades in a quarter, especially for veterans and those in less competitive categories. Yet the study found important splits: managers with larger ownership stakes acted less predictably, and less predictable managers outperformed peers while highly predictable managers lagged. Even within portfolios, harder-to-predict positions beat easier ones.
Anthropic is scaling back parts of its flagship AI safety policy as competition among top labs accelerates. The company said it will no longer pause work on a model simply because it could qualify as dangerous if a competitor releases a comparable or stronger system. The shift marks a sharp change from guardrails published about 2 1/2 years ago that helped define Anthropic as one of the sector’s most safety-focused players. Rivals including OpenAI, xAI, and Google continue to ship new tools at a rapid pace. Anthropic also faces a separate fight with the Defense Department over how Claude can be used, with officials pressing the company to relax usage limits by Friday. Anthropic says the safety update reflects fast AI progress and limited federal regulation.
OpenAI appears to be preparing a new subscription tier called ChatGPT Pro Lite priced at $100 per month, based on a checkout-page discovery shared by engineer Tibor Blaho, who has previously spotted ChatGPT features ahead of release. The plan would sit between ChatGPT Plus ($20) and ChatGPT Pro ($200), adding a middle step in OpenAI’s consumer lineup. The move could target users who hit Plus rate limits but do not need Pro’s top-end access, including freelancers, researchers, and developers. The timing aligns with OpenAI’s hiring of Peter Steinberger, creator of the agent framework OpenClaw, as CEO Sam Altman points to “extremely multi-agent” systems. No launch date or final feature list has been confirmed.