Pick a story. Find a new idea.

CAR T-cell therapy uses modified immune cells to attack cancer. It has helped patients with blood cell cancers. Generate aims to reengineer part of the treatment so that it can also work against solid tumor cancers, which present different challenges.

The team focused on the genetic instructions carried by these cells. Shortening one section could leave space for extra instructions, giving researchers more room to adapt the treatment. Six scientists used Claude Code to explore alternative designs for that section.

Anthropic reports that one candidate was around 70% shorter than the conventional design. Laboratory experiments found that cells using it produced roughly ten times as many receptors, attacked tumor cells more effectively, and stayed active longer.

These findings concern an early design tested in a laboratory. They do not establish how a finished treatment would perform in patients. The example shows how AI can contribute a testable proposal while scientists build it, measure its behavior, and decide what to investigate next.

When mathematical waves concentrate, their most extreme behavior may last only briefly. The local smoothing conjecture describes how averaging over time limits that concentration. OpenAI reports that its models have produced a proof covering every dimension, with consequences for other longstanding questions.

The announcement also describes results in computer science and physics. Across these fields, a new challenge follows the initial output: helping other researchers understand the argument, check its reasoning, and work out what it changes.

For the mathematical result, OpenAI released both a written argument and a version formalized in Lean, a system for checking mathematical proofs. For the other results, outside researchers are organizing collaborations, with different arrangements for publishing the underlying materials.

The proposed process emphasizes independent review. Participants can discuss their work and publish their own conclusions. A checked proof and a useful explanation serve different purposes: one verifies formal steps, while the other helps a community see the ideas, assumptions, and implications behind them.

A language model handling a long document must carry information forward as it generates an answer. A one-million-token context makes that challenge much larger. The DeepSeek-V4 report presents two models designed to make this scale of input more practical.

The central change is a hybrid attention system. One component compresses stored information and selects relevant entries; another uses stronger compression while retaining dense attention over the compressed entries. Together, they reduce the resources required to handle long sequences.

The report compares DeepSeek-V4-Pro with DeepSeek-V3.2 at a million tokens. For generating one token, it reports inference FLOPs of 27% and a KV cache of 10% relative to the earlier model. These figures describe computation and cached attention data under the stated comparison.

That distinction matters when reading efficiency claims. Lower computation and memory requirements can make extended tasks more feasible, but they do not by themselves specify response speed, price, or answer quality. The paper connects these savings to future work on longer reasoning and tasks spanning many documents.