9 events
The labs reach for voice and the screen, departures continue at OpenAI
On 13 May, OpenAI announced GPT-4o, handling speech, images and text in one model and responding at close to the pace of human conversation. The next day at I/O, Google put AI Overviews into Search proper and showed Project Astra and Veo. On the 20th, Microsoft announced Copilot+ PCs and Recall; Recall, which continuously recorded the screen, drew privacy criticism straight away.
Departures continued at OpenAI: Ilya Sutskever on the 14th, Jan Leike on the 17th. Leike said safety research had been losing out to product work, and the Superalignment team was dissolved. On the 28th, Leike joined Anthropic.
On the 20th, a dispute broke out with Scarlett Johansson over the "Sky" voice. On the 21st, Anthropic published research extracting millions of interpretable features from inside a large model.
This block is written by the editors. It is kept separate from the sourced record below.
OpenAI announced GPT-4o, a new flagship model that handles text, audio, and images in a single model. The "o" stands for omni; the launch demo showed real-time voice conversation responding to speech at near-human conversational speed. Bringing GPT-4-level capability to free ChatGPT users marked a major shift, and a ChatGPT desktop app was announced alongside the model.
At Google I/O, Google announced the broad US rollout of AI Overviews, placing AI-generated summaries at the top of search results. After the previous year's Search Generative Experience experiment, it was the decision to make generative AI a permanent fixture of Google's core business. Also unveiled were Project Astra, a prototype universal agent that understands and responds to live video and audio, and Veo, a model generating 1080p video from text. AI Overviews immediately produced widely shared errors — advising glue to keep cheese on a pizza, among others — and Google had to narrow where it appeared. It exposed both the difficulty of letting an AI answer search queries and the fact that this touches the company's revenue base.
Ilya Sutskever, co-founder and chief scientist, announced on X that he was leaving OpenAI after almost a decade. Sutskever had largely stayed out of public view since playing a central role in the November 2023 ouster of Sam Altman. Jakub Pachocki succeeded him as chief scientist. Sutskever said he would pursue a project that was "personally meaningful" to him.
Jan Leike, who co-led the Superalignment team with Ilya Sutskever, resigned from OpenAI, writing on X that "safety culture and processes have taken a backseat to shiny products." With both leaders gone, the Superalignment team, formed in July 2023, was dissolved and its remaining members absorbed into other research groups. It drew heavy coverage at the time as internal tensions at OpenAI over safety surfacing in public.
At an event on its Redmond campus, Microsoft announced Copilot+ PCs — a new category running AI workloads on the device — along with new Surface Pro and Surface Laptop models. One headline feature was Recall, which continuously saved snapshots of the screen so users could search for things they had seen. The backlash was immediate: Recall captured sensitive material such as passwords and financial information, and security researchers found that early builds stored it in unencrypted plain text, which settled the matter. Microsoft pulled the June rollout to all users back to Windows Insiders only, then layered on mitigations — making it opt-in, requiring Windows Hello authentication, and decrypting snapshots only at the moment of use. General availability came about a year later, in 2025. It was the sharpest concrete statement of the question of how much of a user an on-device AI may watch.
After the GPT-4o demo drew attention to "Sky," a ChatGPT voice said to closely resemble Scarlett Johansson, who voiced an AI in the film "Her," OpenAI paused the voice. Johansson stated she was "shocked" and "angered," saying Sam Altman had approached her in September 2023 to voice the system and that a similar voice was used after she declined; she said she had hired legal counsel. OpenAI countered that Sky was not an imitation of Johansson but the natural voice of a different professional actress.
Anthropic reported extracting millions of human-interpretable features from the middle layer of Claude 3 Sonnet, calling it the first detailed look inside a modern, production-grade large language model. The features ranged from concrete things such as San Francisco and Rosalind Franklin to abstractions such as bugs in code, gender bias in professions, and keeping secrets, with some responding across images and multiple languages. Artificially amplifying a feature changed the model's behavior: turning up the Golden Gate Bridge feature made Claude describe itself as the bridge, and amplifying a scam-email feature overrode its safety training. The work pointed both to a route for monitoring dangerous capabilities and to a means of subverting safeguards. A public demo of the steered model, "Golden Gate Claude," brought interpretability research to a wide audience.
At an online summit co-hosted by the United Kingdom and South Korea, sixteen companies signed the Frontier AI Safety Commitments: to publish safety frameworks describing how they measure the risks of their frontier models, to say at what point severe risks would be deemed intolerable, and to set out what they will do to stay below those thresholds. Signatories included OpenAI, Anthropic, Google, Microsoft, Amazon and IBM alongside companies from Europe, the Middle East and Asia — Cohere, Mistral AI, G42, the Technology Innovation Institute, Naver, Samsung and Zhipu.ai. Governments issued the Seoul Declaration, restating the cooperation begun at Bletchley.
Jan Leike, who had co-led OpenAI's Superalignment team, announced on X that he was joining Anthropic to continue the superalignment mission, saying his new team would work on scalable oversight, weak-to-strong generalization, and automated alignment research. Leike had resigned from OpenAI on May 15 after publicly criticizing its approach to safety, and the Superalignment team was dissolved shortly afterward. At Anthropic he reported to chief science officer Jared Kaplan. The hire drew attention at the time as part of a movement of safety researchers from OpenAI to Anthropic.