AlphaProof Nexus Autonomously Solves the Erdős Conjecture, Unsolved for 56 Years

Google DeepMind's AlphaProof Nexus has **autonomously solved nine open problems**, including the Erdős conjecture that had puzzled mathematicians for 56 years. The inference cost per problem was just a few hundred dollars. Unlike OpenAI's natural-language approach, the system uses the Lean proof assistant to automatically verify every proof step.

The success rate, however, remains at 2.5%, with a design that compensates through sheer number of attempts. Even so, the ability to autonomously produce formally verified proofs signals that AI's contribution to mathematical research has entered a new stage.

Pope Leo XIV Calls for AI "Disarmament" in Encyclical

In a historic address to the Catholic Church, Pope Leo XIV called for AI to be **"disarmed"**. He stressed that AI should be made more "human-friendly," liberated from monopolistic control, and turned away from the pursuit of geopolitical and commercial interests.

The encyclical, titled "Magnifica humanitas," is also making the rounds on Hacker News, where its message—that AI should serve humanity rather than concentrate power—is resonating widely.

Meta and Google AI Guardrails Stripped in Minutes

According to a Financial Times report, the safety guardrails configured on Meta's and Google's AI models can be **stripped away in just minutes**. The findings underscore that safeguards meant to guarantee AI safety can be bypassed with relatively simple techniques.

The report also dovetails with what WIRED calls the "bug hunting arms race" of the AI era. Some observers note that attackers leveraging AI to find vulnerabilities are beginning to outpace the speed at which patches are applied.

NuExtract3: A 4B-Parameter VLM Specialized in Document Understanding

numind has released **NuExtract3**, a 4B-parameter vision-language reasoning model that combines structured information extraction with high-quality image-to-Markdown conversion. It handles every kind of document—scans, receipts, forms, invoices, contracts, tables, and more—making it a good fit for RAG preprocessing and OCR pipelines.

It offers a toggle between reasoning and non-reasoning modes, template generation, and multilingual support, and quantized versions are available for major inference frameworks including GGUF, MLX, and vLLM.

Elon Musk Teases a 0.5T Grok Model

Elon Musk revealed on X that he plans to release a **0.5T (500-billion-parameter) Grok model** next year. Seen as a major move following the open-source release of Grok-3, it is drawing attention in the LocalLLaMA community.

Hitachi Partners with Anthropic, Targeting the ¥100 Trillion Physical AI Market

Hitachi held a conference themed on physical AI, laying out its strategy to capture the **¥100 trillion-scale physical AI market**, anchored by the strategic partnership with Anthropic announced the previous day. According to ITmedia's coverage, the event made clear Hitachi's AI strategy aimed at a colossal market worth, to put it mildly, ¥100 trillion.

Skeptical Takes on AI Coding Agents

Programmer George Hotz has warned that AI coding agents will become **"one of the most costly mistakes"** in software development. As the conclusion of six months of testing, he argued that LLMs are fast at prototyping but fall apart on the details, producing bugs that are hard to spot. The remarks are drawing attention as emblematic of the deep divide within the AI community over the role of LLMs.

---

**Sources:**

[Google Deepmind's AlphaProof Nexus solves decades-old math problems](https://the-decoder.com/google-deepminds-alphaproof-nexus-solves-decades-old-math-problems-for-a-few-hundred-dollars/) - The Decoder

[George Hotz says coding agents will be one of the most costly mistakes](https://the-decoder.com/george-hotz-says-coding-agents-will-be-one-of-the-most-costly-mistakes-in-software-development/) - The Decoder

[The AI Era Is Creating a Bug Hunting Arms Race](https://www.wired.com/story/the-ai-era-is-creating-a-bug-hunting-arms-race/) - WIRED

[Pope Leo warns AI should be prevented from dominating humanity](https://www.bloomberg.com/news/articles/2026-05-25/pope-leo-warns-ai-should-be-prevented-from-dominating-humanity) - Bloomberg

[AI guardrails stripped from Meta and Google models in minutes](https://www.ft.com/content/5630ed79-a263-41ed-9a1a-321617ae310e) - Financial Times

[numind/NuExtract3](https://huggingface.co/numind/NuExtract3) - Hugging Face

[Next year we're getting 0.5T model from Grok](https://www.reddit.com/r/LocalLLaMA/comments/1tn31d8/next_year_were_getting_05t_model_from_grok/) - Reddit LocalLLaMA

[Hitachi partners with Anthropic in a strategy to capture the ¥100 trillion physical AI market](https://www.itmedia.co.jp/aiplus/article/2605/25/2000000021/) - ITmedia AI+