An OpenAI reasoning model has autonomously disproved a famous conjecture in discrete geometry first posed by Paul Erdős in 1946 — the "unit distance problem," which asks how many pairs of points in a flat plane can sit exactly one unit apart. Rather than relying on traditional geometry, the model connected the problem to algebraic number theory and discovered an infinite family of point arrangements that exceed the classical square-grid bound. External mathematicians have verified the proof, marking the first time a prominent open problem central to a mathematical subfield has been solved autonomously by AI.
Google launched Gemini 3.5 Flash at I/O 2026 on May 19, choosing speed over scale — it outperforms Gemini 3.1 Pro on nearly all benchmarks including coding, agentic tasks, and multimodal reasoning, while delivering 4× the output token generation speed of competing frontier models. Available immediately via the Gemini app, AI Search, Google AI Studio, and Android Studio. Gemini 3.5 Pro, currently used internally, is expected to roll out next month.
blog.google techcrunch.comAnthropic signed a massive compute deal with xAI (SpaceX), paying $1.25 billion per month through May 2029 for exclusive access to the Colossus 1 data center in Memphis, Tennessee — over 220,000 NVIDIA H100/H200/GB200 GPUs delivering 300+ megawatts of capacity. The deal totals over $40 billion in potential revenue for xAI. Simultaneously, Anthropic announced an expansion to Colossus 2, and both companies are exploring multi-gigawatt orbital AI compute capacity in space.
Released May 13, Cursor 3.4 brings production-grade orchestration for agent teams. A new "Build in Parallel" mode identifies independent sub-tasks in a plan and runs them simultaneously via async subagents. Teams can now configure fully controlled development environments for their cloud agents, including multi-repo support (multiple repos in a single session), Dockerfile-based environment definitions, and support for build secrets to access private package registries.
cursor.com cursor.com/blogPresident Trump postponed the signing of a long-anticipated AI executive order on May 21 — hours before the scheduled ceremony — saying "I didn't like certain aspects of it." The order would have empowered federal agencies to pre-evaluate AI models for security vulnerabilities including bioweapons and autonomous AI R&D risks. Trump cited concerns about overregulation hampering U.S. competitiveness with China. The White House said revisions are ongoing with no new signing date set.
cnbc.com washingtonpost.comMeta launched Muse Spark, its first large language model developed under Chief AI Officer Alexandr Wang's newly formed Superintelligence Labs. The model delivers competitive performance across multimodal perception, reasoning, health, and agentic tasks. The launch comes as Meta ramps AI capital expenditure to $115–135 billion for 2026 — nearly double last year — to close the gap with OpenAI and Google. Meta and Microsoft collectively cut 20,000 jobs earlier in April, underscoring how AI productivity gains are reshaping workforce decisions at Big Tech.
fortune.comApple is reportedly preparing a major shift for Apple Intelligence: an "Extensions" capability that would allow users to select Google or Anthropic as the AI backend powering Siri and other Intelligence features across iOS 27, iPadOS 27, and macOS 27. Meanwhile, Apple is also approaching early mass-production testing for new AirPods equipped with low-resolution cameras for visual context in Siri interactions — expanding Apple's AI presence beyond the screen.
cnbc.comExecutives at Google, OpenAI, and Anthropic now describe the frontier race as neck-and-neck, with companies making different tradeoffs around cost, speed, and compute. A stark example: running a standard 10-evaluation workload costs $4,811 with Anthropic's Claude vs. $3,357 with OpenAI's ChatGPT. Google's bet with Gemini 3.5 Flash — fastest output, broadest deployment — reflects a strategy to win on price-performance rather than raw capability. The three-way competition is producing rapid iteration cycles measured in weeks, not quarters.
axios.com axios.com/two-hours