The third week of May 2025 brought a research milestone from Google DeepMind, a model upgrade inside ChatGPT, new audio tools and the first concrete signs of generative AI reshaping television advertising. Here are the developments that mattered.
DeepMind’s AlphaEvolve improves on a 56-year-old algorithm
Google DeepMind on May 14 unveiled AlphaEvolve, a coding agent that uses Gemini Flash and Gemini Pro to propose candidate algorithms, test them automatically and evolve the best ones over many rounds. Its most striking result, according to DeepMind, is a method for multiplying two 4×4 complex-valued matrices using 48 scalar multiplications, improving on the 49 required by Volker Strassen’s 1969 algorithm, a bound that had stood for 56 years.
DeepMind said AlphaEvolve had also been applied inside Google, including a scheduling improvement that recovered about 0.7 percent of the company’s global computing resources and optimizations to the kernels used to train Gemini itself. The system is notable less for any single result than for the approach: a language model generating code that is checked by an automated evaluator rather than by people, which lets it explore far more candidates than a human team could.
In a related research thread, a paper from a Tsinghua University-led group earlier in May described the Absolute Zero Reasoner, a model trained to generate its own coding and math problems, solve them and verify the answers without external training data. The authors reported results on coding and math benchmarks competitive with models trained on curated human examples.
OpenAI puts GPT-4.1 in ChatGPT; Windsurf ships its own models
OpenAI on May 14 made GPT-4.1, previously available only through its API, accessible inside ChatGPT for paying users, with GPT-4.1 mini replacing GPT-4o mini for free users. The company positions GPT-4.1 as stronger at coding and instruction following, with faster responses and a longer context window than GPT-4o.
Windsurf, the AI coding editor, announced on May 15 a family of models called SWE-1 trained for software engineering tasks rather than general chat. The company said the models were designed to handle the full flow of engineering work, from reading a codebase to editing and reviewing, and that the largest version performed close to frontier models on its internal evaluations. The announcement landed amid reports that OpenAI was in talks to acquire the company.
Netflix and YouTube bring AI to advertising
At its upfront presentation for advertisers on May 14, Netflix said it would use generative AI to create ad formats that blend into the shows they interrupt, for example by placing a brand’s product against a backdrop drawn from a series such as Stranger Things. Netflix said the formats would roll out in 2026, according to Ad Age and other outlets covering the event.
The same day, at its Brandcast event, YouTube introduced Peak Points, which uses Gemini to identify the moments in a video when viewers are most engaged and places ads immediately after them. Both announcements drew mixed reactions: advertisers welcomed the targeting, while some viewers and critics questioned whether ads timed to emotional peaks would feel intrusive.
New audio tools
ElevenLabs released SB1, which it calls an infinite soundboard: a tool that generates any sound effect from a text description and is aimed at streamers, game developers and podcasters. Stability AI released Stable Audio Open Small, a compact open model that can generate short audio clips and is small enough to run on a smartphone, developed with chip designer Arm.
Tesla’s Optimus dances
Tesla published new videos of its Optimus humanoid robot performing dance routines, which the company said were run on the physical robot in real time rather than in simulation. The clips were intended to show improvements in balance and whole-body control. As with earlier Optimus demonstrations, the footage was produced by Tesla and has not been independently verified, and dancing is a long way from the factory and household work Tesla has promised for the robot.
Taken together, the week showed AI moving in two directions at once: deeper into research problems that had resisted human effort for decades, and further into the commercial plumbing of everyday media, where its effects on viewers will be felt long before anyone notices a faster matrix multiplication.
