# DeepMind’s Math Breakthrough, xAI’s Grok Sprint, and the Compute Squeeze

*By AI High Signal Digest • May 25, 2026*

DeepMind reported a formally verified math breakthrough, xAI shipped Grok 4.3 while preparing a larger V9 model, and multiple signals pointed to compute becoming the core constraint in frontier AI. Also in the brief: long-context training research, new agent tooling, and mixed labor signals as AI adoption broadens.

## Top Stories

*Why it matters: the strongest signals today were verified reasoning gains, faster frontier model iteration, and growing pressure around compute access.*

- **DeepMind reported a formally verified math advance.** AlphaProof Nexus solved 9 open Erdős problems, some unsolved for 56 years, and also proved 44 open OEIS conjectures, resolved a 15-year-old algebraic geometry question, and found a novel optimization parameter [^1]. The system combines LLM reasoning with Lean verification, and one analysis said a simple generate-check loop matched the full system on all nine Erdős successes, underscoring how formal verification can filter hallucinations in hard reasoning tasks [^1].

- **xAI is compressing its model cycle.** Grok 4.3 is now live on the xAI API, with a 1M-token context window, pricing of $1.25/m input and $2.50/m output, and leaderboard claims in tool use, instruction following, and enterprise domains [^2]. Separately, xAI said Grok V9-Medium (1.5T) finished training, with fine-tuning underway, reinforcement learning starting in days, and public release targeted in 2-3 weeks; Elon Musk said it should materially improve harder coding tasks over the current production model [^3].

- **Compute pressure is intensifying.** GPU rental prices are up more than 2x since January 2026 [^4], while one prominent view this week was that critical-path AGI pretraining now effectively requires the compute scale of OpenAI, Google, Meta, or the Anthropic/xAI/Cursor group [^5]. Against that backdrop, Meta cutting 8,000 jobs while spending $100 billion on AI data centers stood out as a stark capital-allocation signal [^6].

## Research & Innovation

*Why it matters: the most useful research updates were about training models more efficiently and measuring their behavior more honestly.*

- **Long-context pretraining still has architectural traps.** An AllenAI/CMU paper found 4k-token pretraining metrics have little correlation with actual long-context performance, and recommended avoiding QK norm, Group Query Attention, and Sliding Window Attention while pretraining on longer sequences [^7][^8]. Paper: [allenai.org/papers/olmpool](https://allenai.org/papers/olmpool) [^9].

- **OPUS moves data selection from static to dynamic.** The ICML Oral paper dynamically selects training data at every pretraining iteration and reported better efficiency and model quality than static selection across language tasks [^10].

- **A large behavior study raised another warning on post-training.** Testing models on data from more than 200,000 participants and nearly 26 million human responses, the authors found post-training made models less human-like; related commentary warned that optimizing narrow objectives can shift behavior in unrelated domains [^11].

## Products & Launches

*Why it matters: launches centered on enterprise deployment, agent infrastructure, and faster local inference.*

- **Cohere open-sourced Command A+.** The 218B/25B-active MoE targets enterprise agentic workflows, adds multimodal reasoning, supports 48 languages, and can run on as little as two H100s or one Blackwell GPU [^12].

- **Cloudflare expanded Think for agent orchestration.** New updates add support for the agentskills.io spec, local/codebase/R2 skill loading, a configurable permission model, and JS/Python/Bash scripts with workspace access; scheduled tasks can run prompts on cron patterns or a DSL [^13][^14].

- **Local inference got faster.** llama.cpp with MTP support pushed Qwen3.6-27B dense generation on an A10G from 25 tok/s to 45 tok/s, a 78% jump that was framed as making local models more viable as daily drivers [^15].

## Industry Moves

*Why it matters: the business story was split between workforce disruption, expanding software demand, and clearer production use cases.*

- **The labor signal remains mixed.** Meta, Cisco, and Intuit were cited cutting 8,000, 4,000, and 3,000 jobs respectively, with over 100,000 tech jobs gone so far in 2026; one analysis argued companies are now more openly shifting spend from headcount to GPU clusters [^6].

- **But AI coding may be expanding software demand rather than shrinking it.** David Sacks said software-engineer postings are rising as GitHub commits grow 14x YoY and AI lowers the cost of writing code, enabling more bespoke software across businesses [^16].

- **AI video crossed another adoption threshold.** Kling is now being used in TV and film production, and *House of David* was described as the first Hollywood production to openly discuss AI video generation at industrial scale; the show reportedly reached 44M+ viewers and hit #1 on Prime Video U.S. [^17].

## Quick Takes

*Why it matters: a few smaller updates sharpened the picture on security, local AI, and semiconductor competition.*

- TrapDoor hit npm, PyPI, and Crates.io with 34 malicious packages and also used poisoned `CLAUDE.md` and `.cursorrules` files to target developers using AI coding tools [^18].
- Gemma 4 has been downloaded more than 120 million times just weeks after release [^19].
- Hugging Face said 300,000 AI builders completed hardware profiles, another data point behind the rise of local AI [^20].
- Huawei claimed a new path to narrow its semiconductor gap with TSMC without cutting-edge equipment [^21].

---

### Sources

[^1]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058673672169107757)
[^2]: [𝕏 post by @xai](https://x.com/xai/status/2051703217697010103)
[^3]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2058787384364265734)
[^4]: [𝕏 post by @AnjneyMidha](https://x.com/AnjneyMidha/status/2058611711867801989)
[^5]: [𝕏 post by @_aidan_clark_](https://x.com/_aidan_clark_/status/2058606927811346435)
[^6]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058585834597818766)
[^7]: [𝕏 post by @gabriberton](https://x.com/gabriberton/status/2058686103721480508)
[^8]: [𝕏 post by @gabriberton](https://x.com/gabriberton/status/2058686107722870950)
[^9]: [𝕏 post by @gabriberton](https://x.com/gabriberton/status/2058686108926710200)
[^10]: [𝕏 post by @jiqizhixin](https://x.com/jiqizhixin/status/2058478937936994611)
[^11]: [𝕏 post by @ValerioCapraro](https://x.com/ValerioCapraro/status/2057787992933114069)
[^12]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2058548830241521686)
[^13]: [𝕏 post by @threepointone](https://x.com/threepointone/status/2058486087094358509)
[^14]: [𝕏 post by @threepointone](https://x.com/threepointone/status/2058639691147104539)
[^15]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2058672394865111544)
[^16]: [𝕏 post by @DavidSacks](https://x.com/DavidSacks/status/2058606722110107970)
[^17]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058490137139413436)
[^18]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058584943052161488)
[^19]: [𝕏 post by @osanseviero](https://x.com/osanseviero/status/2058502294820290848)
[^20]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2058592962708144170)
[^21]: [𝕏 post by @business](https://x.com/business/status/2058724761513918822)