A few hundred dollars in computing time is all Google's AlphaProof Nexus needed per problem to solve nine mathematics questions unsolved for decades, two of them for fifty-six years. At the same time, Gemini Intelligence is rolling out to Samsung Galaxy S26 and Pixel 10, where it autonomously books appointments, generates widgets, and refines voice input. This summer, AI is simultaneously advancing the research frontier and moving into pockets.
What Google announced at I/O 2026
On May 19, Sundar Pichai introduced Android 17 at Google I/O as an "intelligence system" rather than an operating system. The core is Gemini Intelligence: a set of AI features running on the device itself, woven through all apps. Chrome gets an autonomous browsing mode called Chrome Auto Browse starting in late June, which independently books appointments or reserves parking. Gboard cleans up voice input automatically; widgets can be generated by text description.
Adding to this is Gemini Spark, a personal agent coordinating actions across multiple apps. For U.S. subscribers to the Google AI Ultra plan, which dropped to 100 dollars monthly after a price cut, Spark is already available. Rollout begins staggered: Samsung Galaxy S26, S26+, and S26 Ultra receive Gemini Intelligence first, followed by Google's Pixel 10 lineup. The Galaxy Z Fold 8 is due in July.
DeepMind Chief Demis Hassabis said at the conference that humanity stands "at the foot of singularity." He sees AGI as achievable "around 2030, plus or minus a year." The formula was not a promise but calibration: Google wants to be understood as a company expecting a watershed moment, not merely delivering product iterations.
AlphaProof Nexus: solving nine decade-old problems for a few hundred dollars
One day after OpenAI's announcement that it had disproven the 80-year-old Erdős unit distance conjecture, Google DeepMind published its own results on May 21, 2026 on arXiv. AlphaProof Nexus combines Gemini 3.1 Pro as a strategist with the formal proof system Lean, which algorithmically verifies every proof step. The system solved nine open Erdős problems from a list of 353 plus 44 conjectures from the Online Encyclopedia of Integer Sequences. Two of the nine problems had been open for fifty-six years.
The technical difference from OpenAI's method is significant. OpenAI uses natural-language reasoning followed by human verification; AlphaProof Nexus generates machine-verifiable proofs, with each step checked by Lean. No human mathematician needs to judge the evidence. According to Google DeepMind, each solved problem cost a few hundred dollars in computing time, while some of these questions have occupied professional mathematicians for months or years.
Leading mathematicians, including Fields Medalist Tim Gowers, have rated both results as genuine contributions to mathematics. That's unusual: AI results are rarely endorsed by the research community so unconditionally. The Erdős problems solved by AlphaProof Nexus concern combinatorial structures and questions about the distribution of number sequences that serve as touchstones for mathematical structural understanding. That an AI system solves multiple such problems in sequence changes assessments of what algorithmic thinking can accomplish.
Why mathematics was the easiest test
Google DeepMind explicitly stated: AlphaProof Nexus is not AGI. The system operates in formal languages with clearly defined verification criteria. Mathematics offers the advantage that every solution is uniquely verifiable. Other sciences where open questions are ambiguously formulated or lack Lean-compatible formalization remain beyond the system's reach for now.
The economic value of the solved Erdős problems is limited. They belong to pure combinatorics and geometry, not applied research. Whether similar systems could one day solve open problems in chemistry, pharmacology, or climate modeling remains unanswered. Those disciplines lack the closed formalization system that mathematics and Lean provide.
The Gemini Intelligence rollout raises different questions. The features require the AI system to know calendar entries, email content, and location. Google has announced that many processing steps happen on-device. How far that extends in practice and which data still reaches Google servers will be central questions for the European rollout.
Gemini in Chrome by end of June, Z Fold 8 in July
The immediate next steps are concretely scheduled. Chrome Auto Browse launches in the U.S. by the end of June according to Google. The Galaxy Z Fold 8 will receive Gemini Intelligence features at market launch in July. Whether and when features expand to European users, Google has left open.
For AlphaProof Nexus, Google has established an "AI for Math Initiative" to deploy the system against further open problems. DeepMind emphasized the system is not yet adapted for problems outside formally verifiable mathematics. The competition with OpenAI, which in turn announced plans to tackle additional math problems, will likely continue through the summer.