Future Technology[AI Weekly] AI cracked a 27-year-old maths problem
OpenAI just got an unreleased model to solve ten maths problems that had beaten professional mathematicians for decades, and a US court just ruled that when an AI agent shops on Amazon for you, you're the one doing the shopping, not the software. Two very different stories about the same question: what happens when we hand AI the wheel.
The Big 3
OpenAI's next model solved ten maths problems nobody had cracked in decades
OpenAI let an internal, unreleased model called Astra loose on a stack of genuinely unsolved problems in group theory, quantum complexity and combinatorics. It came back with ten new results, including the first explicit construction of a non sofic group (open since 1999), a disproof of Connes's rigidity conjecture, and solutions to three problems from Paul Erdos's famous catalogue. Every proof ships as a machine checked Lean 4 certificate, published openly on GitHub. The whole run reportedly cost about 2,000 dollars in compute.
Why it matters: this isn't a model acing a maths exam, it's a model doing original research that stumped experts for a quarter of a century, and that's a genuinely new category of thing for AI to be useful for.
A US court ruled your AI shopping agent isn't "hacking" Amazon
The Ninth Circuit vacated Amazon's injunction against Perplexity's Comet browser, which uses an AI agent to buy things on Amazon for its users. Amazon argued this violated the Computer Fraud and Abuse Act, the same law used in hacking prosecutions. The court disagreed, ruling that it's the user, not Perplexity, who "accesses" Amazon's servers when the agent completes a purchase on their behalf.
Why it matters: there's almost no case law on who's responsible when an AI agent acts for you, and this ruling just became the first real precedent, one that favours letting agentic AI keep operating rather than boxing it in.
Alibaba's Qwen3.8-Max: a 2.4 trillion parameter model, and it's open weight
Alibaba released Qwen3.8-Max, a mixture of experts model with 2.4 trillion total parameters that activates only 95 billion per query to keep costs down. It leads on several vision and document benchmarks, trails the strongest closed models on coding, and lands close behind Claude Opus 5 in frontend code arena rankings. Open weights, frontier scale, released as a genuine competitor rather than a demo.
Why it matters: the gap between "open" and "frontier" keeps shrinking, and that's mostly good news if you'd rather not depend on one lab's API.
Quick Hits
The EU AI Act has teeth now. From 2 August, chatbots and AI systems operating in the EU must tell users they're talking to AI, and deepfakes must be labelled. Fines run up to 15 million euros or 3 percent of global turnover.
Enterprise AI is in production, mostly ungoverned. New industry survey data puts 72 percent of firms with AI agents in live production, but 60 percent say they still have no formal governance for them.
The open-weight race keeps accelerating. GLM-5.2 and DeepSeek V4 both pushed further into long-horizon coding and agentic benchmarks this week, adding to a genuinely crowded field of serious open alternatives to the big commercial labs.
AI funding hasn't slowed for summer. 83 new funding rounds landed this week, up 9 percent on the week before, with money still concentrating in AI infrastructure and physical AI over consumer apps.
Tool of the Week
n8n is a self-hostable workflow automation platform that now wires AI agents directly into your automations, connect it to your email, calendar, CRM or a local model and have it act on triggers without you touching a spreadsheet. Free and open source to self-host, with a paid cloud tier if you'd rather not run your own server.
Who it's for: anyone tired of manually stitching apps together, from solo founders to ops teams.
One More Thing
The Perplexity ruling quietly answers a question nobody had a good answer to: when an AI agent does something on your behalf, who did it, legally speaking. The court said it's you. That's convenient for AI companies today, but it also means every mess an agent makes on your account is yours to clean up too. Worth remembering before you let one loose with your card details.
Forwarded this? Get your own AI briefing at futuretechnologyhq.com/newsletter
Until next Wednesday. Nath, Future Technology
Some links in this newsletter may be affiliate links. We only recommend products we genuinely think are worth your time.