This is the evening edition for October 28, 2026. Here are six key topics from overseas AI news. Today was a day when AI's "scope of play" expanded: a frontier lab stepped into drug discovery, a benchmark index recalibrated the distance between open and closed models, and new agent memory research exposed the field's current limitations.

Anthropic Moves Into Drug Discovery — Acquires a One-Person Bio Company That Used Claude

This is today's top story. On October 27 (local time), US-based Anthropic announced that it is pursuing in-house drug discovery targeting rare pediatric muscular dystrophies and other diseases. The company also revealed that it has acquired a bio company where a single employee used Claude to run experiments with 11 external research institutions, and that it is taking over that research.

The striking part is the scale: one employee. By putting Claude at the center of the research, a tiny team managed to run drug candidate searches across multiple research institutions — and Anthropic has absorbed that track record wholesale through acquisition. As one of the earliest cases of an LLM company making a serious entry into pharmaceuticals, it could redraw the industry map. For a company reportedly preparing for an IPO, it's also a way to demonstrate value creation beyond competing on model performance. The actual drug discovery results are still to come, but as a signal of how far the "AI company workplace" can extend, this move deserves attention.

AI Intelligence Index v5 — The Open-Weight Gap Widens From 3 Points to 12

Artificial Analysis, well regarded for its model evaluations, has shipped a major update bringing its Intelligence Index to v5. What the community is buzzing about is that this revision dramatically widened the gap between open-weight and closed models — from 3 points to 12.

According to the announcement, Kimi K3.1 Pro — whose weights are not public — tops the open category at 46. Among models with published weights, DeepSeek V4.1 Flash and GLM-5.3 are tied at 32, leaving them 26 points behind the leading closed model and rated below even Claude Haiku 5.5 (36). One caveat, though: 60% of the v5 index is based on private test sets and held-out answers, making the results impossible to reproduce externally. Indeed, users going by their own experience are already raising doubts, suggesting that "Fable-5.1 might actually be better than GPT-Terra-6.1." While ranking precision has improved, this is also the flip side of "what can't be measured can't be measured," and reading the index will likely call for continued caution.

The Net Value of Agent Memory — "Memory Tax" and the Harm of Obsolete Entries

Two studies have now landed that shake the foundations of agent implementation. The first is "Memory Tax: Dissecting the Net Value of Agent Memory." In a comprehensive evaluation, many agent memory systems underperformed a simple naive RAG baseline — at higher cost. The upshot: however much you refine your memory mechanism, its net value may be eroding like a "tax."

The second is "When Retrieval Is Not Enough: Obsolete Facts in Agent Memory." By separating retrieval from answering in memory tasks, the study found that even when the correct, up-to-date entry is successfully retrieved, answers fall apart if an obsolete entry gets mixed in alongside it. In the experiments, simply replacing one irrelevant piece of evidence with an outdated conflicting entry dropped Qwen2.5-7B's accuracy from 100% to 44.6% (Qwen3-14B held at 87.8%). Old information proved more harmful than irrelevant information of the same size. The implication: rather than pushing retrieval accuracy ever higher, the real challenge is designing how memory gets updated and discarded.

Qwen3.5-4B-3D-Coder Nears Practical Use — Dataset and FP16 Weights to Be Released

Now for the local LLM crowd. A new version of "Qwen3.5-4B-3D-Coder," a 4B experimental model aimed at generating interactive 3D scenes with Three.js, has been uploaded. According to the developer, one-shot generation results and visual quality have improved significantly, and training is nearly complete.

The recommended build is the Q8 quantization (temperature 0.6, top-k 20, top-p 0.95). A Q4 version is also available, but it hasn't been validated yet. What's notable is that even though this training run did not include a vision encoder, the model is also delivering good results on 3D scene generation from reference images. And the real show starts now: the dataset and full FP16 weights are slated for release. With over 3,500 downloads already on Hugging Face, the project is shaping up to be one answer to small-scale, local coding needs.

Discord Public Server Conversations Become Google-Searchable — On by Default in Japan

Discord has announced that conversations in certain public servers will become indexed by web search. Starting mid-November, AI-generated summaries will be published, and by mid-December the post text itself will become searchable. The aim is to improve discoverability, but concerns about privacy and use for AI training are already being voiced.

What matters for Japanese users is that this is on by default. If you want to be excluded, you have to change the settings yourself. The shift — statements from communities that were once semi-closed becoming reachable via search engines — is significant, and it's probably time to rethink your operations on the assumption that "what you write in a public server can be found by anyone in the world."

Personal Data of 330,000 People Exposed at the Panasonic Group — The Reality That "Prohibition" Can't Prevent

Finally, from Japan. It has come to light that personal data belonging to roughly 330,000 current and former employees of the Panasonic Group and their families had been uploaded to an external generative AI service whose internal use was banned. The group says the data "will not be used for secondary purposes going forward," but where that assurance comes from is exactly what the reporting side has been trying to verify.

Most corporate generative AI usage guidelines center on "prohibition," but this case exposes the gap between the rules and reality. If the reality is that a certain number of people will always paste data into banned services for the sake of convenience, then the response may need to shift its center of gravity away from "enforcing prohibitions" and toward making the approved environment genuinely attractive.

Wrap-up

Today's six stories put both the expanding horizon of AI's ambitions (drug discovery) and the foundations underpinning its reliability (benchmarks, memory, privacy) to the test at once. Anthropic's entry into drug discovery will take time to bear fruit, but it looks like a turning point in how AI companies articulate their vision. The two agent memory studies carry lessons that will pay off directly in tomorrow's implementations. See the reference links for details on each topic.