Daily AI Digest

Meituan uses 64-dimensional LLM vectors to improve long-tail local search

Meituan describes three production phases that compress queries, merchants and products into semantic vectors before injecting similarity into ranking. The first phase used 64-dimensional vectors and raised long-tail NDCG@5 by 2.21 percentage points while reducing bad top results by 2.96 points.

The second phase trained joint query, merchant and product representations with hard negatives drawn from exposed but unclicked results in the same request. Adding triplet loss raised product order AUC by 11.02 points, versus 4.85 points for click AUC. All three phases completed full production rollout.

Verdict: The useful lesson is not merely that LLMs can improve search. Compact vectors, strong representations and business-specific hard negatives can matter more than simply scaling the model.

"Don't paste the AI" tops HN: people ask for your judgment, not a generic answer

dontpastetheai.com argues that a person who can call the same chatbot still asks you because they want your context, taste and judgment. It does not reject AI drafting; it asks users to read, compress and rewrite the result, or quote only the useful part and explain why it matters.

The near-thousand-point debate reframes the issue from whether AI was used to whether the sender filtered the output, accepted responsibility for the judgment and respected the reader's time.

A 125M piano-completion model generates about 108 notes per second on an iPhone 15

Simon Edwardsson trained a 125-million-parameter Transformer to continue MIDI piano performances on-device in real time. Across 14 experiments, the largest gains came from a better MIDI event representation, aggressive data cleaning and DPO post-training rather than a larger model.

The free RollTab app targets iPhone and iPad users with a MIDI keyboard. It shows how task-specific representation and data quality can let a small model deliver genuinely interactive latency.

Huzzah turns prompts into persistent pseudocode files

Huzzah proposes a different coding workflow: prompts become declarative, persistent pseudocode instead of long, imperative, disposable chat messages. Saving the file turns its diff into the prompt and regenerates the affected source code.

The author calls it experimental and notes that cross-file dependencies remain hard to express reliably. The broader idea is to bring intent back into versioned files so it can be compared, reviewed and maintained.

Meta senior AI director Madhu Guru recommends building a failure-mode taxonomy as soon as eval v1 exists. Study 500 to 1,000 production traces and separate wrong retrieval, irrelevant sections, ungrounded answers, failed refusals and missing clarification instead of calling everything a bad answer. Specific names turn failures into targeted tests and an improvement loop.

Vercel CEO Guillermo Rauch argues that AI will push infrastructure toward native optimization. He points to fx, a 6.3 MB Zig-compiled tool that starts in about 10 microseconds, as evidence that agent startup and dependency overhead should be compressed alongside model latency.

Box CEO Aaron Levie says experts still have the advantage in the AI era. Models lower the entry cost, but deciding what an agent should do, when to redirect it, how to verify output and what good looks like remain domain skills; AI may amplify the leverage gap.

Tencent open-sources UI-Mate, a GUI agent that adapts from a demonstrated workflow

Tencent HY Frontier released UI-Mate 9B, 27B and a demonstration-guided 27B checkpoint on Hugging Face. The models read live screenshots and a natural-language task, then emit structured mouse, keyboard, scrolling and waiting actions. The demonstration-guided version lets a user show a workflow once before the agent adapts it.

The official model card reports 77.0 on OSWorld-Verified and 66.2 on WindowsAgentArena for UI-Mate-27B. The Apache-2.0 checkpoints are intended to run with the official prompt, parser and interaction harness, not as ordinary visual-chat models.

Verdict: Demonstrations can reduce prompt-writing overhead, but desktop agents still touch files and application permissions. Test them with isolated accounts, rollback and task-scoped authorization.