Google AI Updates: May 24, 2026

1. MIT Technology Review reads Gemini for Science as a strategic pivot away from purpose-built scientific tools

Google. MIT Technology Review’s I/O 2026 wrap-up argues that Gemini for Science marks a deliberate reallocation of DeepMind’s attention away from the AlphaFold-style purpose-built scientific systems toward general-purpose LLM agents that can run hypothesis generation, code search, and literature review end to end. The piece notes that Co-Scientist and AlphaEvolve, which underpin Hypothesis Generation and Computational Discovery in the new suite, are still gated behind an application process but are being talked about by researchers as transformative once available. The framing matters because it positions Gemini for Science less as a product launch and more as DeepMind betting that agentic LLMs, not narrow scientific models, become the default tool for empirical research. Source

2. AI Mode crosses 1 billion monthly users as Search is rebuilt around longer-running agentic tasks

Google. A post-I/O analysis from PPC Land reported that Google AI Mode in Search has crossed 1 billion monthly active users and has been upgraded onto Gemini 3.5 Flash, with the search box itself redesigned to accommodate longer-running agentic queries that span multiple turns and tool calls rather than single-question lookups. The same write-up positions Gemini Spark, the always-on background agent, and the Antigravity SDK as the infrastructure that now powers Google’s own agentic Search features, with the SDK exposing the same production harness externally that Google uses internally. Together, the changes reframe Search from a list-of-blue-links surface into a task surface where AI Mode is expected to delegate, browse, and report back. Source

3. Google open-sources a DESIGN.md spec so models can read design languages directly

Google. Alongside the agentic Search push, Google introduced an open DESIGN.md specification that lets design systems publish a machine-readable description of their components, tokens, and usage rules, intended to be consumed by code-generating models the same way AGENTS.md and SKILL.md are consumed by Gemini’s managed agents. The format extends the pattern Google has been building since AGENTS.md, shifting documentation from human-only prose to artifacts that LLMs can parse without bespoke prompting, and it is positioned as a way to make Antigravity-generated UI conform to a team’s existing design language out of the box. PPC Land frames the file as the first design-systems-grade entry in Google’s growing family of machine-readable markdown specs. Source