Nikon disqualifies contest winner over generative AI use
A microscopy institution just drew a hard line on AI in science, a frontier lab turned agent swarms into a product, and one of the biggest agent deployments in production published where its cost actually went.
Nikon has disqualified the first-place winner of its Small World in Motion video contest for breaking its generative AI rules — and promoted the entry that finished behind it. The original winner, a video from Dr. Ning Xu of Tsinghua University showing cilia beating in the airway of a child with a respiratory condition, was challenged online within days of the announcement: viewers noted structures that seemed to pop in and out of existence, and former judges flagged an AI watermark in the source file. Nikon initially said it saw no rule violations, then reversed course on Thursday, saying the video "did not comply with the competition rules regarding generative AI" while adding that the decision "should not be interpreted as a judgment of the entrant's professional reputation, scientific contributions, or intent." Xu has acknowledged using what he called "an unsupervised neural-network method" for AI-assisted post-processing of the images, while denying that AI produced the movie, the cilia, or their motion — and that distinction is exactly why this story matters: institutions are now enforcing AI rules against AI-assisted scientific work, not just AI-fabricated content, and the boundary between the two is far from settled. A video of a roundworm by Nguyen Nam Nhat of Vietnam now holds first place, and Nikon says it will revisit its rules and evaluation procedures — expect other competitions to follow.
Claude can now run fleets of sub-agents — with one big asterisk on the "1,000" number. Anthropic put dynamic workflows into public beta in its Claude Managed Agents platform: a lead agent writes a plan, delegates tasks to sub-agents, and merges the results, with up to 1,000 agents started over a run's lifetime. The asterisk is worth stating plainly — Anthropic's own documentation caps concurrent agents at 64, so "1,000 agents in parallel" is a per-run total, not a simultaneous swarm. Anthropic's pitch is that swarms beat solo agents on hard tasks: in its test, hiding 70 bugs in a 116,000-line codebase, a single agent found 14 to 27 per run while a dynamic workflow consistently found 66. That is a vendor's own unreplicated benchmark, and it lands against real skepticism — a senior OpenAI Codex engineer recently called agent swarms "a massive waste of tokens with zero quality gain." Worth noting we saw the same swarm-behavior debate from the research side in DeepMind's 100-agent swarm found a bug and split into cheaters.
Asana's 76x agent cost cut is a caching story wearing a model-launch headline. OpenAI published a case study this week in which Asana's StackAI team rebuilt its browser agent's workflow on GPT-6.1 Sol and got runs down from at least $36.21 to $0.47 — 76x cheaper and 5x faster, from 22.5 minutes to about 4. But the decomposition is the interesting part: rewriting the workflow to actually cache the agent's browsing history accounted for 29x of the savings on the original model, the model swap added 2.6x more, and on GPT-6.1 Sol alone the new caching policy cut cost 4x because 89% of input came from cache at 5% of the uncached price. Asana's own write-up notes the study ran each condition only three or four times and that costs are provider-side estimates — so treat 76x as a stacked marketing number that happens to contain a genuinely useful lesson: most agent cost lives in history management, not model choice.
What to watch: Nikon's revised rules will be the first concrete test of how science contests define "AI-generated" versus "AI-assisted."
Where should institutions draw the line between AI-assisted and AI-generated science? Tell us in the comments.
