Daily AI Digest

OpenAI says its next Astra model solved ten long-open mathematics problems

OpenAI released ten results spanning high-dimensional sphere packing, group theory, quantum complexity, lattice cryptography and combinatorics. The company says an internal version of its next major model, Astra, generated the arguments, with solution-search tokens costing roughly $2,000 at Sol API rates. Humans then helped prepare manuscripts, and the model produced a Lean certificate for each argument. The papers, certificates and reasoning records are public, but results this consequential still require line-by-line review by the mathematics community.

Verdict: The meaningful advance is not just generating answers but pairing manuscripts with machine-checkable certificates. Independent validation matters more now than the headline count of problems.

Aaron Levie argues that the harness around a model, including work decomposition, tool use and model routing, is becoming the next most important variable after model capability. As tasks grow from millions to tens or hundreds of millions of tokens, decomposition and routing directly affect accuracy and cost.

Swyx still actively uses /loop and /goal. He finds them useful when a user wants both steerability and autonomy, or when the desired end state is an open-ended loop that creates new loops. Early instability, he argues, was not a reason to abandon the pattern permanently.

Peter Steinberger observes that model 5.5 can accept new instructions while it is working without becoming confused, making a strict queue less necessary. That behavior moves long-running agents closer to continuous collaboration instead of one-shot job submission.

Cursor confirms self-serve plans no longer show per-request dollar costs

Cursor staff confirmed on its official forum that individual and self-serve Teams plans now report usage in tokens. The Usage page, CSV exports and the endpoint that reads historical events no longer return per-request dollar costs. Aggregate on-demand charges remain visible. Teams administrators can see each member's on-demand total, but not the old per-model dollar breakdown. Enterprise plans and supported admin interfaces retain some cost fields.

Verdict: Teams that used request-level costs to choose models lose a useful feedback loop. Before changing plans or tools, verify which billing fields are actually available and record tokens, model, task outcome and final charge together.

US lawmakers investigate DoorDash's use of China's Kimi model

ITHome, citing CNBC, reports that the chairs of the US House Homeland Security Committee and China Select Committee asked DoorDash to explain its use of Moonshot AI's Kimi. DoorDash co-founder Andy Fang previously said the company gives the hardest work to a frontier model and delegates lower-level tasks to Kimi K2.6 for lower cost. DoorDash says it will discuss how it safely uses both US frontier and open-weight models. The inquiry requests information and is not a finding of wrongdoing.

AI makes prototypes faster without shortening the road to a reliable product

An engineering essay newly popular on Hacker News argues that AI can produce a convincing app in minutes without supplying error handling, access control, load tolerance, observability or durable data design. Syntax is becoming easier, but understanding how systems fail and which choices create long-term costs still requires judgment. AI amplifies experienced engineers; treating it as a substitute for understanding shifts the risk into production.