Seedance hypocrisy, DiscoBench findings, and sqlite-utils 4.0

Hollywood secretly uses Seedance while banning it, DiscoBench exposes search agent flaws, and sqlite-utils 4.0rc2 ships mostly written by Claude.

The big picture

A quiet batch overall, but two items punch above their weight: a sharp research finding about why AI agents fail at search, and a Hollywood story that crystallizes the industry’s AI contradiction better than any think-piece could.

Hollywood’s Seedance problem is everyone’s AI problem

ByteDance’s AI video tool Seedance generated a fake Brad Pitt and Tom Cruise clip, prompting the Motion Picture Association to issue its first-ever cease-and-desist against an AI company. The catch: studios are reportedly still using Seedance internally on a quiet “don’t ask, don’t tell” basis, according to Simpsons animation producer Joel Kuwahara. The Decoder

This is the clearest example yet of institutional AI hypocrisy in the wild. Public condemnation, private adoption. If you’re building AI tooling for any creative industry, this dynamic is your actual market reality — not the press releases.

DiscoBench: search agents are failing at the wrong step

A new benchmark called DiscoBench finds that AI search agents don’t really struggle with searching — they struggle with knowing when to ask for clarification instead of just guessing. Models that keep searching on ambiguous queries score only 51.9%, which is actually worse than models that just take a stab at an answer. Strip out the ambiguous queries and accuracy jumps by up to 40 percentage points. Even the top model only hits 43% overall. The Decoder

This is a genuinely useful finding if you’re building agentic search pipelines. The instinct to “just try more searches” is actively harmful. Building in an explicit clarification step before multi-hop retrieval isn’t a UX nicety — it may be the single highest-leverage improvement you can make. The 43% ceiling on the best model also suggests there’s a lot of headroom left, so don’t over-index your architecture on current benchmark scores.

Mistral pitches sovereignty while trailing on performance

Mistral CEO Arthur Mensch made a pointed argument that proprietary AI labs are accumulating customer data and, in some cases, using it to compete directly against those same customers. The framing is sharp and the concern isn’t unreasonable. The Decoder

But the pitch is easier to take seriously if Mistral can close the capability gap with OpenAI and Anthropic, which it currently can’t. EU data sovereignty is a real differentiator for European enterprise customers, less so for anyone else. File this under “valid point, motivated messenger.”

sqlite-utils 4.0rc2 ships, Claude did most of the work

Simon Willison released sqlite-utils 4.0rc2 and wrote up the process — Claude Fable (Anthropic’s newest model family member) did the bulk of the coding for approximately $149.25 in API costs. Simon Willison’s Weblog

The dollar figure is the interesting part. That’s a real-world data point on what a meaningful open-source release costs when you’re leaning heavily on a frontier model. Not free, not prohibitive — somewhere between a dinner out and a software license.

Quick hits

AI private schools like Alpha School are charging up to $75,000/year for two hours of AI tutoring plus workshops, widening the gap between families who can afford personalized AI learning and those stuck with under-resourced public schools still figuring out their device policies. The Decoder