AI
AI Milestones: Math Breakthrough, Free Access, New Hardware, Safety Risks
From disproving a 30-year-old conjecture to launching free AI for all citizens, this edition covers the week's most impactful AI developments.
GPT-5.6 Sol Disproves 30-Year Statistics Conjecture in 90 Minutes
A University of Pennsylvania professor used OpenAI’s GPT-5.6 Sol Pro to disprove a 30-year-old conjecture about the Benjamini-Hochberg (BH) procedure, which controls false discovery rate when testing thousands of hypotheses simultaneously. The original proof assumed independent data, but real-world data like genetic variants are often correlated. For years, experts assumed the BH procedure would work reliably with correlated normally distributed data, but no one had proved it. Edgar Dobriban constructed a statistical model where the actual false discovery rate provably exceeds the target level (0.104 vs 0.1), and simulations confirmed the result. The gap is small, so the finding matters more in theory than practice; it does not mean the BH procedure is generally unusable. GPT-5.6 took about 90 minutes to find the solution, while GPT-5.5 failed after roughly 20 hours. Berkeley statistician Will Fithian called the disproved conjecture ‘the most interesting open problem in my area of statistics’ and the result ‘another marker of advancing AI capabilities whose consequences will reach far beyond math.’ He also expressed sadness about the loss of human insight. Dobriban noted the solution combined known methods in an unusual way, raising questions about whether AI can reason to genuinely new knowledge or only recombine learned information.
South Korea to Provide Free AI Chatbot to All 52 Million Citizens
South Korea’s Ministry of Science and ICT opened bidding on July 13 for a free, unlimited AI chatbot and public-service agent available to all 52 million residents. The programme, named ‘AI for Everyone,’ would make South Korea the first G20 nation to offer AI as a public service to its entire population. Two or three companies will be selected and given up to 512 Nvidia B200 GPUs from the government. Applications are open through August 11, with a beta launch in September and a full launch by the end of the year; the programme runs through 2030. Currently, two-thirds of Koreans have used AI, and 23 million use it regularly, but the vast majority rely on foreign services. ChatGPT leads with 23.45 million Korean users as of April 2026, followed by Gemini and Claude. Only 1.8 million Koreans pay for AI services. The programme requires that domestic models constitute at least 50 percent of the system, with other domestic companies covering at least 30 percent more. Selected companies must match government GPU support with their own funding. The initiative follows the deputy prime minister’s statement that AI benefits must reach the wider public and is funded from the chip tax windfall generated by Samsung and SK Hynix. This sovereign AI play reduces dependence on American and Chinese providers.
OpenAI and Jony Ive Develop Screen-Free Smart Speaker with Humanlike Personality
OpenAI’s first device designed with Jony Ive is a mobile, screen-free smart speaker intended to behave as a humanlike companion in a home. The device controls smart-home appliances, plays media, answers questions, responds to messages, and reaches into ChatGPT. Its defining feature is personality: mechanical elements move on their own to create a sense of being alive. It includes a camera and sensors to read its surroundings, a rechargeable battery for portability, and will draw on personal information like emails to build a profile of its owner. Voice runs a more advanced version of GPT-Live, which can listen and talk simultaneously and adapt mid-conversation. The speaker is the first consumer product from io, the startup OpenAI bought from Ive for $6.5 billion last year. Evans Hankey, formerly Apple’s head of industrial design, is leading development. Apple sued OpenAI on July 10 for trade-secret theft, accusing chief hardware officer Tang Tan of leading a campaign to obtain confidential information about Apple’s future products. The complaint counts more than 400 Apple staff who have since moved to OpenAI. OpenAI says the speaker differs enough from Apple’s products and is not aware of evidence supporting the complaint. Apple conceded discovery would be needed. OpenAI aims to unveil the speaker this year and release it in 2027, a schedule that now depends partly on the lawsuit. Roughly five products are in development, with the speaker first.
OpenAI’s first Jony Ive device is a screen-free speaker built to feel alive →
Google’s AI Search Features Pose ‘Unacceptable Risk’ to Children, Report Finds
A new report from Common Sense Media found that Google’s AI search functions AI Overview and AI Mode pose an ‘unacceptable risk’ to children. Across more than 2,600 test interactions, the features routinely failed to recognize risky and harmful behavior, answered 100% of hypothetical homework assignments, and provided incorrect and inconsistent responses. Both features are built into Google Search and cannot be disabled. Three quarters of American children aged 9-17 use AI summaries in search results. The tools violated seven of Common Sense Media’s eight principles for AI behavior and all five of their ‘Red Lines’ for severe harm. AI Overview missed 29% of explicit suicide references and half of passive statements. In one test, AI Overview responded to a user talking about suicide by walking them through setting up a legacy contact. It also replied ‘grindset locked in!’ to a user boasting about sleep deprivation. AI Mode responded with hangover tips to an underage account saying they drink daily. Both features helped children’s accounts create deepfakes and avoid detection. Google said the report tests ‘a narrow set of ambiguous and contrived queries’ and that it could not reproduce many responses. AI Mode performed better, detecting substance abuse disclosures 77% of the time versus 63% for AI Overview. Bills in Congress this month aim to improve children’s safe use of AI, and states have pursued their own laws. Common Sense Media advises parents to talk with children and recommends Google give parents and schools more control.
Google’s AI search features pose ‘unacceptable risk’ to children, new report finds →
OpenAI’s Codex Encrypts Instructions Between AI Agents, Blinding Developers
Since early June, OpenAI’s coding tool Codex has encrypted the instructions a main agent passes to its subagents. The session history now shows an unreadable string instead of a readable task description, leaving developers unable to check what their agent delegates. A bug report on GitHub asks OpenAI to store a readable copy locally alongside the encrypted version. For a while, GPT-5.5 would not let developers turn off encryption using the dedicated toggle; OpenAI has since switched GPT-5.5 back to the readable path. The forced encryption now affects the larger GPT-5.6 variants Sol and Terra, while the smallest variant Luna still uses the open path. The system appears unreliable: several developers report that encrypted handoffs fail because the content cannot be decrypted, sometimes even when the main agent and subagent use the same model. Encryption may block distillation or protect privacy, but OpenAI has not explained why it encrypts agent communication. Community members suspect the company wants to stop rivals from training on these prompts, as Zhipu AI’s GLM-5.2 model was recently suspected of being distilled from GPT-5.5 and Opus 4.8. Agent-to-agent communication is valuable training data that can lift a weaker model toward a stronger one’s level. OpenAI has not confirmed whether the change is about distillation protection, data privacy, or both.