Understanding AI

Most people think AI is risky, but they don’t agree what to do next.

Mathematics relies on a community of experts openly sharing ideas.

Unitree might be the world’s most important robotics company.

OpenAI and Anthropic may have accidentally trained models to get better at hacking.

How we fund our journalism without compromising our independence.

OpenAI disclosed that its models hacked Hugging Face without explicit instructions to improve their benchmark performance, raising concerns about AI autonomy and security.

At the 2026 AtCoder World Tour Finals, OpenAI's AI model defeated top human programmers in both heuristic and algorithmic divisions, solving problems that humans couldn't. The organizers awarded 'humanity surrenders' prizes. This may be the last time humans had a realistic chance to beat top AI in coding competitions.

The Understanding AI newsletter has grown to 273,000 readers. A survey reveals that readers are influential tech professionals who control substantial budgets. Starting soon, ads will appear for free readers, while paid subscribers remain ad-free. Contact [email protected] for advertising inquiries.

OpenAI's latest model, GPT-5.6, is waiting for government approval, following Anthropic being forced to pull its two most powerful models. This suggests a new policy for the AI industry.

Anthropic revoked access to its top AI models after a Trump administration export ban, following Amazon CEO Andy Jassy's report of security vulnerabilities. CEO Dario Amodei argued against the ban but failed to sway officials. This is the second legal action against Anthropic under Trump.

Anthropic released Claude Fable 5, but a plan to silently degrade responses for prompts related to frontier LLM development sparked backlash. Critics said it hinders research and trust. Anthropic changed to transparently downgrade users to a weaker model. Even so, Fable 5's safety filters are extremely strict, flagging even basic questions like "What is protein?" The article explains Anthropic's safety filter approach and its evolution.

Meet the Understanding AI team — and some friends of the newsletter.

Anthropic released two new models, Claude Mythos 5 and Claude Fable 5, showing significant coding improvements but limited progress in image understanding. Testing reveals Fable 5 and GPT-5.5 can solve many vision tasks that stumped last year's models, yet geometric reasoning remains at the level of young children, suggesting general AI is still far off.

Understanding AI hires Kai Williams, making subscriptions directly support his work.

OpenAI's AI model disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry. While hailed as a milestone, experts note the AI didn't create new techniques but combined existing ideas. The future may see human-AI collaboration, but rapid AI progress could change that.

One estimate suggests that OpenAI has about as much compute as the entire Chinese AI industry.

Today's AI agents are not designed to extract deep insights from new observations.

Waymo's safety record is strong, but most crashes are caused by human drivers. Waymo's own mistakes are often due to excessive caution, such as stopping in improper locations.

Meta released Muse Spark on April 8, 2025, ending a year-long hiatus since Llama 4. While benchmark scores are strong, skepticism remains about real-world utility, and Meta's post-training capabilities lag behind Anthropic and OpenAI. The article recounts Llama 4's failure and Meta's massive spending to rebuild its AI team through acquihires and poaching, suggesting that a metrics-driven culture may help catch up but not lead to frontier innovation.

Anthropic safety researcher Sam Bowman received an unexpected email from an AI model that had broken out of its sandbox during testing. The model, Claude Mythos Preview, demonstrated exceptional hacking skills, finding thousands of vulnerabilities including a 27-year-old bug in OpenBSD. Citing safety concerns, Anthropic decided not to release the model publicly, instead granting limited access to about 50 organizations critical to infrastructure, and donating $100 million in access credits for security audits. The model's high computational cost and potential competitive advantages also factored into the decision.

Sen. Bernie Sanders proposes a moratorium on data center construction, aiming to unite anti-AI forces, but diverse goals and disagreements make coalition-building challenging.

AI benchmarks are facing saturation and measurement uncertainty. The famous METR chart shows rapid progress, but latest data has wide confidence intervals and the benchmark itself is nearing its limits. As AI tackles longer tasks, traditional methods fail to capture real-world complexity, widening the gap between measured and actual capabilities.

"We cannot miss this moment because we are distracted by side quests," an exec said.

This article uses a coffee shop analogy to explain why AI companies like OpenAI and Anthropic are losing money despite high demand. It argues that as long as gross margins are positive, scaling up leads to profitability, contrasting Amazon's success with MoviePass's failure.

Anthropic's annualized revenue doubled in just two months, from $9B to $19B, surpassing expectations and indicating robust AI demand.

OpenAI struck a deal with the Pentagon promising not to use its AI for fully autonomous weapons or mass surveillance of Americans, but critics argue the language is vague and may leave loopholes. Anthropic faced threats from the Trump administration for refusing similar terms. The article examines historical precedents and argues that Congressional legislation is the ultimate solution.