Edited from five posts published in June 2025, covering May 26–June 29.
June was the month AI’s effect on work moved from forecast to policy inside companies. Anthropic’s Dario Amodei warned that AI could eliminate half of entry-level white-collar jobs within five years, and three weeks later Amazon’s Andy Jassy told staff to expect a smaller corporate workforce. Meta, which had lost most of Llama’s original authors, invested in Scale AI and brought its founder into its AI effort; Google named a chief AI architect.
The month’s verdicts came from courts and markets. Two federal judges ruled on whether training generative models on books is fair use, both largely in the developers’ favor, with differences over market harm. AI dominated closing arguments over remedies in the government’s search case against Google. The New York Times licensed its content to Amazon, and Nvidia regained the title of the world’s most valuable company.
The tools themselves were changing shape. Apple opened its on-device model to developers, Meta released the V-JEPA 2 world model, Midjourney released its first video model and WhatsApp began showing ads. Agents moved into business software at Datadog and Salesforce, and an Anthropic experiment that let Claude run a small office shop showed how a capable model could still lose money over weeks of work. Public institutions adopted the technology as well: the FDA put an assistant into its review work, while Yoshua Bengio launched LawZero to pursue a safer design.
The argument about jobs ran on through the year: Satya Nadella explained more than 15,000 cuts at Microsoft in July, OpenAI announced a jobs platform and certifications in September, and Sam Altman put AI’s share of the economy’s tasks at 30 to 40 percent. The fair-use rulings left Anthropic’s pirated library for trial, and in September the company agreed to settle those claims for at least $1.5 billion. Licensing grew alongside litigation: Cloudflare’s pay-per-crawl test in July, Meta’s license of Midjourney’s technology in August and Universal Music’s collaboration with Stability AI in November. Meta’s Scale investment led directly to Superintelligence Labs at the end of the month, and Nvidia’s market value kept climbing until it passed $5 trillion in October.
Business Insider reported that 11 of the 14 researchers credited on the 2023 paper introducing Meta’s Llama models had left the company, several for rival laboratories; one became a founding member of Mistral.1
The departures were an early sign of the competition for researchers that Meta answered with large hiring offers later in the summer.
Salesforce announced an agreement to acquire Informatica for approximately $8 billion in equity value, net of its existing investment. The proposed combination would bring Informatica’s data integration, cataloging and governance services into Salesforce’s enterprise platform.2
The deal tied AI deployment to the quality and organization of business data; it still needed regulatory approval to close.
Anthropic chief executive Dario Amodei told Axios that AI could eliminate half of entry-level white-collar jobs within one to five years.3,4
The forecast, from the head of a leading developer, became a reference point in the year’s debate over AI and work.
The New York Times Company and Amazon announced a multiyear licensing agreement covering editorial material from the Times, NYT Cooking and The Athletic. The permitted uses included summaries and short excerpts in Amazon services, links to Times products, and training Amazon’s foundation models.5
The agreement set out a commercial route for using journalism in AI products while the paper’s lawsuit against OpenAI and Microsoft continued.
In closing arguments on remedies in the government’s search monopoly case against Google, Judge Amit Mehta spent much of an eight-hour session on AI’s implications for search. The Justice Department sought a sale of Chrome, an end to paid default placements, and sharing of search data that AI rivals such as OpenAI and Perplexity could use. Mehta said he hoped to rule by August.6,7
Generative AI had become central to deciding how much of Google’s search business the court should change.
Mary Meeker and colleagues at BOND published Trends – Artificial Intelligence, a chart-led survey of adoption, computing costs, investment and competition. It brought company disclosures and other datasets together to examine the pace and economics of AI development.8
The report offered a broad frame for reading individual announcements, assembled from many sources and measurement periods.
The US Food and Drug Administration launched Elsa, an internal generative AI tool for reading, writing and summarizing. The agency described uses including adverse-event summaries, label comparisons and assistance with scientific review, within a government cloud environment.9
The agency said the tool arrived ahead of schedule; it put an AI assistant inside the work of a regulator.
Yoshua Bengio launched LawZero, a nonprofit devoted to AI safety research. Its proposed Scientist AI approach aimed to build systems that explain and predict rather than pursue goals through action. Possible applications included scientific discovery and oversight of AI agents.10
The initiative set out a research program for separating understanding from agency.
Mistral introduced Mistral Code in private beta for JetBrains environments and Visual Studio Code. Built on the Continue project, the offering combined coding models, an editor assistant, deployment options and administrative controls for enterprise teams.11
The product addressed the conditions for using AI with proprietary code: where it runs, what it can access and how an organization oversees it.
US Senator Edward Markey held a roundtable opposing a proposed 10-year moratorium on state AI regulation in the House-passed reconciliation bill. Participating advocates argued that it would weaken civil-rights and consumer protections without an adequate federal replacement.12
The debate concerned which level of government could respond to AI-related harm; the Senate struck the moratorium on July 1.13
At WWDC, Apple introduced the Foundation Models framework, giving developers access to the on-device language model behind Apple Intelligence. It also previewed features including Live Translation and expanded visual intelligence. Developer testing began that day, with wider user availability planned for the fall.14
The framework brought generative functions into apps that could operate offline, on supported devices.
OpenAI’s annualized revenue run rate reached about $10 billion, nearly double the $5.5 billion run rate at the end of 2024. The figure covered ChatGPT subscriptions and API sales and excluded licensing and large one-off deals with Microsoft.15
The growth showed how quickly paid use of chatbots and developer access had expanded.
At DASH, Datadog introduced Bits AI agents for incident response, development and security. Bits AI SRE entered limited availability; the Dev Agent and Security Analyst were in preview. Their intended work included investigating alerts, proposing code fixes and assessing security signals.16
The tools connected models to operational data and established engineering workflows, starting with limited releases.
Sundar Pichai named Koray Kavukcuoglu, chief technology officer of Google DeepMind, to the new post of chief AI architect, reporting directly to him. Pichai wrote that Kavukcuoglu would “accelerate how we bring our world-leading models into our products.”17
The appointment put a senior researcher in charge of turning Google’s models into products faster.
Meta released V-JEPA 2, a world model trained primarily through self-supervised learning from video. Its encoder and predictor worked with abstract representations of scenes. Meta also demonstrated an action-conditioned version for robot planning and released benchmarks for physical reasoning.18
The work connected visual prediction with planning, and supplied models and tests for further research.
Scale AI announced a Meta investment valuing the company at more than $29 billion. Founder Alexandr Wang was joining Meta’s AI effort while remaining on Scale’s board. Scale appointed Jason Droege interim chief executive and said Meta would hold a minority equity stake.19
The arrangement combined an investment, a commercial relationship and a leadership move; Wang went on to lead Meta Superintelligence Labs.
Meta announced that WhatsApp would show ads in its Updates tab, alongside Status and Channels, and add paid channel subscriptions and promoted channels. Personal chats would remain ad-free and end-to-end encrypted; ad targeting would use limited information such as location, language and channels followed.20
The change ended WhatsApp’s long-standing ad-free model and brought Meta’s advertising business into its most private service.
Amazon chief executive Andy Jassy told employees that generative AI and agents would change staffing needs. He expected fewer people in some roles and more in others, with a net reduction in the corporate workforce over the next few years.21,22
The message made workforce change an explicit management expectation at one of the largest US employers.
Midjourney released V1 as an image-to-video model. Users could animate a generated or uploaded image, supply a motion prompt and extend a resulting clip. The release included automatic and manual animation controls, with options for higher or lower motion.23
The workflow extended still-image creation into moving scenes; Midjourney described real-time worlds and 3D interaction as its longer-term direction.
Salesforce announced Agentforce 3, including plans for a Command Center to monitor agent activity and support for Model Context Protocol connections. The release combined available features with previews and capabilities scheduled for later delivery.24
Monitoring and integration became product requirements alongside the models’ ability to perform individual tasks.
In Bartz v. Anthropic, Judge William Alsup held that training Claude on books was fair use, calling it “exceedingly transformative,” but that keeping a library of pirated copies was not, leaving that claim for trial. Two days later, in Kadrey v. Meta, Judge Vince Chhabria granted Meta summary judgment on Llama’s training, calling his ruling limited to the plaintiffs’ arguments and record, and pointing to market dilution as a potentially stronger claim.25
Both rulings favored the developers on training, while differing on how much harm to authors’ markets should count.
Senators Rick Scott and Gary Peters introduced the No Adversarial AI Act. The bill proposed a federal list of AI produced or developed by foreign adversaries, agency review for exclusion or removal, and a process for exceptions.26
The proposal made the origin of AI systems a question of federal procurement and security.
Nvidia’s shares rose more than 4 percent to a record $154.10, lifting its market value to about $3.76 trillion and past Microsoft’s.27
Investors priced continued demand for AI data-center chips into the company’s value.
The Qwen team introduced VLo, a preview combining multimodal understanding with image generation and editing. Through Qwen Chat, users could request an image or upload one and describe changes in natural language.28
The preview connected image interpretation with revision in a conversational workflow.
The Information reported that mass production of Microsoft’s next in-house AI chip, code-named Braga, had slipped from 2025 into 2026, after design changes and staff turnover.29
The delay left Microsoft relying on Nvidia’s hardware for longer.
Anthropic and Andon Labs reported an experiment in which Claude Sonnet 3.7 managed a small office shop. The agent selected stock, set prices and interacted with customers, while people handled physical tasks. It made persistent pricing and discount errors, and the business failed to make money.30
The experiment tested sustained commercial judgment in a small, supported setting, where individually useful actions coexisted with poor overall management.
Sources checked October 26, 2026. Dates distinguish announcements, reports and later updates.