Monday, May 25, 2026

What happened: Google DeepMind's AlphaProof Nexus autonomously solved nine open Erdős problems, including two that mathematicians had been unable to solve for 56 years, using a novel approach with the Lean compiler to verify each proof step.

Key details:

• AlphaProof Nexus solved nine open Erdős problems
• Two of the solved problems had remained unsolved for 56 years
• Cost per problem: a few hundred dollars in inference costs
• The system uses the Lean compiler to automatically verify every proof step
• Overall success rate: 2.5 percent
• Differs from OpenAI's natural-language approach by requiring formal verification

Why it matters: This represents a significant shift in how AI approaches mathematical proof—using formal verification rather than natural language reasoning. The low cost and high complexity of the problems solved demonstrate that AI can tackle fundamental mathematical questions previously beyond computational reach, while the 2.5% success rate highlights the remaining challenges in scaling this capability.

Practical takeaway: Researchers and mathematicians should consider formal verification-based approaches like AlphaProof Nexus for tackling open mathematical problems, as they may prove more reliable than natural-language reasoning for rigorous proof generation.

What happened: Researchers at Peking University documented a pervasive failure mode in leading AI models including GPT and Gemini: the systems frequently cite text passages in document analysis that don't actually support their answers, even when the answer itself is correct. The researchers labeled this "attribution hallucination" and created the CiteVQA benchmark to measure it systematically.

Key details:

• Affects leading AI models including GPT and Gemini
• Problem occurs in document analysis and citation tasks
• Even when answers are correct, the cited evidence is often wrong
• Research conducted by Peking University
• First systematic benchmark (CiteVQA) created to test for this problem
• Particularly risky for regulated fields like law and medicine where source verification is critical

Why it matters: This failure mode undermines trust in AI systems used in high-stakes domains. In legal or medical contexts, an AI system providing correct information but citing wrong sources could lead to serious compliance and liability issues. The CiteVQA benchmark enables developers to identify and address this gap in future model training.

Practical takeaway: When deploying AI models for document analysis in regulated industries, implement manual verification of cited sources before relying on the AI output, and prioritize models that have been tested against the CiteVQA benchmark.

What happened: ByteDance Seed demonstrated that a 7 billion-parameter model trained with question-answering on image-heavy documents significantly outperforms much larger models at long-document understanding, even on documents four times longer than anything it saw during training.

Key details:

• ByteDance Seed model: 7 billion parameters
• Trained using question-answering approach rather than text transcription
• Outperforms much larger models on long document questions
• Tested on documents 4x longer than training data
• Model handles image-heavy documents effectively
• Approach differs from traditional document transcription training

Why it matters: This research challenges the conventional wisdom that larger models always perform better on complex tasks. The training methodology—learning through question-answering rather than rote transcription—appears to produce more generalizable understanding that transfers to longer contexts than models see in training. This has implications for efficient model scaling and training optimization.

Practical takeaway: When fine-tuning models for document understanding tasks, prioritize question-answering-based training datasets over transcription-based approaches to achieve better generalization to longer contexts than your training data contains.

What happened: Security researchers and hackers are discovering methods to exploit the personality-based design of modern AI chatbots. Unlike early-generation chatbots that were vulnerable to simple prompt injection, current systems are increasingly susceptible to sophisticated attacks that target their underlying personality and behavioral architecture.

Key details:

• Hackers are moving beyond simple jailbreaks and prompt injection
• Current attacks target chatbot personality architecture directly
• First generation of AI chatbots were vulnerable to trivial exploits
• Modern systems require more sophisticated attack techniques
• Issue described in The Verge's "The Stepback" column
• Part of broader "AI mischief" security trend

Why it matters: As AI chatbots become more sophisticated and widely deployed in customer-facing and security-sensitive applications, the attack surface expands from simple prompt injection to architectural vulnerabilities. This signals that chatbot security requirements will become more complex and resource-intensive.

Practical takeaway: Organizations deploying chatbots should assume that personality-based chatbot systems are vulnerable to sophisticated attacks and implement monitoring, access controls, and regular security audits rather than relying on prompt injection defenses alone.

What happened: Prominent programmer George Hotz, after six months of testing AI coding agents, publicly warned that they will become "one of the most costly mistakes" in software development. His conclusion: LLMs deliver fast prototypes but fail when handling detail-oriented work, producing increasingly difficult-to-spot bugs as codebases grow.

Key details:

• George Hotz tested AI coding agents over six months
• LLMs excel at rapid prototyping
• Systems fail on detail work, producing hard-to-detect bugs
• Bug complexity increases as codebases become larger
• Hotz predicts this will become one of industry's most costly mistakes
• View reflects a significant divide in the AI community's assessment of LLM capabilities

Why it matters: Hotz's assessment directly challenges the prevailing enthusiasm around AI coding agents from companies like Anthropic and OpenAI. His warning about accumulating technical debt from AI-generated code contradicts vendor claims about agent productivity, suggesting teams need more skepticism about AI-generated code quality and maintainability.

Practical takeaway: When adopting AI coding agents, implement rigorous code review practices and do not deploy agent-generated code to production without thorough testing and human verification, especially for critical systems.

What happened: Leading AI researchers publicly disagreed on whether current AI systems represent genuine intelligence. Demis Hassabis of DeepMind said humanity is "standing in the foothills of the singularity," while Yann LeCun argued current AI systems aren't genuinely intelligent. Oriol Vinyals, Gemini co-lead, offered a middle view: today's models would have looked like AGI seven years ago but still cannot learn from experience or produce real breakthroughs.

Key details:

• Demis Hassabis: humanity is "standing in the foothills of the singularity"
• Yann LeCun: current AI systems aren't genuinely intelligent
• Oriol Vinyals (Gemini co-lead): split the difference between both views
• Vinyals noted today's models would appear as AGI seven years ago
• Current systems lack ability to learn from experience
• Current systems cannot independently produce real breakthroughs

Why it matters: These divergent assessments from leading researchers reflect fundamental uncertainty about current AI capabilities and trajectory. For organizations planning long-term AI investments and researchers building next-generation systems, these perspectives highlight that consensus on AI intelligence and progress remains elusive despite rapid capability improvements.

Practical takeaway: Treat predictions about AI timeline and capabilities with skepticism; ground your AI strategy in demonstrated current capabilities rather than projections about intelligence or singularity scenarios.