# Google/DeepMind’s Agentic Push, the Compute Scramble, and Wider AI Governance *By AI News Digest • May 27, 2026* Google/DeepMind led the day with a broad agent push, a 2030 ±1 AGI estimate, and more concrete drug-discovery progress. Elsewhere, scarce compute reshaped industry strategy, provenance tooling moved into consumer products, and AI governance broadened from geopolitics to the Vatican. ## What stood out today The clearest pattern today was AI moving from broad capability talk into operating questions: who controls the agents, who gets the compute, and how trust gets built into the stack. ### Google/DeepMind makes the agentic era its main product story Google used I/O to spotlight Omni, Spark agents, and Flash 3.5 inside Antigravity 2.0, with Demis Hassabis saying consumer agents are already showing productivity gains [^1][^2]. In a new interview, Hassabis said his AGI estimate is now “2030 plus or minus a year,” described the moment as the “foothills of the singularity,” and said DeepMind is still on track with its original 20-year mission [^2][^3][^2]. > “These days I’m thinking it’s 2030 plus or minus a year.” [^2] He paired that optimism with a call for security, robustness, standards, and international cooperation as the agentic era expands [^2]. He also said Isomorphic Labs has raised new funding, now has pre-clinical test compounds starting in oncology and immunology, and aims to compress parts of drug discovery from roughly ten years to months or weeks [^2]. *Why it matters:* Google is tying product launches, frontier-model direction, safety, and long-horizon science into one story rather than presenting them as separate bets. ### Compute shortages are redrawing industry strategy Ben Thompson said Anthropic’s xAI compute deal makes sense because Anthropic has the highest willingness to pay in a constrained market and was always likely to find new capacity as demand outstripped supply [^4]. He argued xAI increasingly looks like both a model company and an infrastructure company, and that the infrastructure side may be more valuable if it sells compute broadly instead of limiting it to Grok [^4]. On the hardware side, NVIDIA said its Vera CPU is built for agentic AI workloads, pairing 88 custom Olympus Armv9.2 cores with 1.2 TB/s of memory bandwidth; in Phoronix testing it showed a 1.6x geometric-mean gain over Grace and a 1.5x advantage over a latest-generation 128-core x86 processor, with first customer deliveries already underway [^5]. *Why it matters:* The day’s compute news was less about one deal or one chip than about a broader shift: scarce infrastructure is shaping partnerships, product availability, and where value sits in the AI stack. ### Provenance and containment are moving closer to default infrastructure Google DeepMind said SynthID has watermarked more than 100 billion pieces of content and that Gemini users have run AI-origin verification more than 50 million times [^6][^7]. The company is extending content authentication into Search and Chrome, adding edit trails for Pixel video, and partnering with OpenAI, ElevenLabs, and Kakao to add SynthID watermarking to their models [^7][^8][^6]. Anthropic, meanwhile, said agent permissions should evolve with capability growth and described sandboxing in its own products as a way to limit potentially destructive actions [^9]. *Why it matters:* Leading labs are trying to make authenticity checks and operational containment part of everyday tooling, not just after-the-fact policy. ### AI governance is showing up in both religion and geopolitics Big Technology reported that Pope Leo XIV issued a 55-page encyclical on AI warning about war, disinformation, surveillance, algorithmic addiction, and democratic degradation, while arguing that AI cannot offer genuine human connection and that pauses can be compatible with progress [^10]. Separately, reporting highlighted by Nathan Lambert said China is expanding travel restrictions on top AI talent at key private firms beyond earlier rumors focused on DeepSeek [^11][^12]. *Why it matters:* Different institutions are arriving at the same conclusion from very different angles: AI is now important enough to trigger moral arguments and tighter controls over people and systems. ### New research is a reminder that AI helps science most when humans can still verify the work Sakana AI and collaborators proposed CUSP, a benchmark covering 4,760 scientific events to test whether models can forecast scientific progress [^13]. Their result: frontier models can identify promising directions, but they still struggle to predict whether or when breakthroughs will happen, and the gap is not explained by training-data volume alone [^13]. The authors conclude that science remains open-ended and that AI works best as a collaborative explorer rather than a predictor [^13]. That fits with Terence Tao’s warning that AI could create a “traffic jam” in math by generating more proofs than humans can verify, increasing the need for better scientific infrastructure [^14]. *Why it matters:* Better models may widen exploration faster than existing validation systems can keep up. ## Also notable - **Gemma 4 looks like a meaningful open-model shift.** swyx called it dramatically better than Gemma 3, Interconnects said equivalent-size Gemma 4 models tie or outperform Qwen 3.5/3.6 under an Apache 2.0 license, and Nathan Lambert said adoption is already outpacing Qwen at similar sizes [^15][^16][^17]. - **Agents are being pitched as operators, not just assistants.** Perplexity Computer was demonstrated managing Shopify workflows such as market research, product-image generation, and theme design in parallel, matching Sarah Guo’s argument that customers want agents that resolve issues by taking actions inside real systems [^18][^19][^20]. --- ### Sources [^1]: [𝕏 post by @rowancheung](https://x.com/rowancheung/status/2057491344697012384) [^2]: [Demis Hassabis on AGI by 2030, Curing Every Disease, Life After AGI, and More](https://www.youtube.com/watch?v=4tVCHeAv0D4) [^3]: [𝕏 post by @rowancheung](https://x.com/rowancheung/status/2059307613485940950) [^4]: [SpaceX's Anthropic Deal and the Endgame for xAI | Sharp Tech with Ben Thompson](https://www.youtube.com/watch?v=tZFevpOrB0c) [^5]: [NVIDIA Vera CPU Is ‘Packing a Heavy-Hitting Punch’ Against Competition](https://blogs.nvidia.com/blog/vera-cpu-phoronix) [^6]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2059235181274202500) [^7]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2059235184130535436) [^8]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2059235187003642154) [^9]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2059351260243919269) [^10]: [The Pope Takes On AI](https://www.bigtechnology.com/p/the-pope-takes-on-ai) [^11]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2059274019959128287) [^12]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2059274064297075009) [^13]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2059166749761872342) [^14]: [𝕏 post by @SAIRfoundation](https://x.com/SAIRfoundation/status/2059384274164621648) [^15]: [𝕏 post by @latentspacepod](https://x.com/latentspacepod/status/2039962128208322769) [^16]: [Some ideas for what comes next, May 2026](https://www.interconnects.ai/p/some-ideas-for-what-comes-next-may) [^17]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2059299103964811512) [^18]: [𝕏 post by @Shopify](https://x.com/Shopify/status/2059289769490678072) [^19]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2059327697298559415) [^20]: [𝕏 post by @saranormous](https://x.com/saranormous/status/2059301017251180656)