Lilith.
⌕
Editorial illustration: Microsoft shows where AI cuts work and where humans still matter
Lilith illustration · editorial remix

Microsoft Research has published a five-minute tour of five projects. The results range from mathematics and multilingual text processing to the operation of coding agents, making this a map of current work rather than one sweeping product announcement.

Five projects share a search for bottlenecks

The context-compaction study provides the clearest operating numbers. An analysis of 13.5 million sessions found that fewer than 8% needed compaction, yet they accounted for 44% of total AI processing costs. APEC prepares summaries in the background and, in evaluations, cut costs by up to 25.4% and execution time by 9.6% while improving coding performance.

The roundup also covers the Latent Core Tokenizer tested across 104 languages, a discussion of machine translation and a robotics-focused internship. Its mathematics segment concerns the resolution of the 31-year-old KLS conjecture. Microsoft says the result was independently checked and formally verified with computer assistance.

Savings on long sessions may matter more than another feature

For teams deploying coding agents, the cost distribution is the important part. A small tail of unusually long sessions can consume almost half of the compute bill. Preparing a summary continuously addresses a specific operational spike instead of adding another capability to a demo.

The mathematics result shows a different division of labor. People directed the research process while agents explored candidate approaches in parallel. The value here is cheaper exploration of dead ends, not an autonomous mathematician.

Microsoft's roundup is no substitute for the individual papers

The figures come from a short corporate summary and the authors' evaluations. They do not show how APEC performs across outside agents, models and repositories, or how much overhead it adds to ordinary sessions that never need compaction. Likewise, better tokenization across 104 languages does not establish equal gains for every language and task.

Replication outside Microsoft will settle the practical value

APEC should be tested on identical tasks against standard compaction, measuring total cost, completion time and change quality. Tokenizer results need language-by-language scrutiny. For the mathematics work, the useful signal will be further independent review of the proof and evidence that the research process transfers to other open problems.

Lilith's verdict

These five research displays share one sober scene: the agent searches the corridors, but a person still holds the map and signs off on the result.