Daily Roundup

AI News: Midjourney Hits Studios; AI Impersonates UK Politicians

July 5, 2026: Midjourney demands Hollywood studios reveal AI use; GPT-4 Turbo impersonates BBC panelists; Fable writes sqlite-utils 4.0rc2.

Top Stories

Midjourney Demands Disney, Universal, Warner Bros Disclose Their Own AI Use

Midjourney has filed a counter-discovery demand in the copyright suit brought by Disney, Universal, and Warner Bros, asking the studios to hand over every internal AI tool they use and every prompt they ran through Midjourney. Disney and Universal sued on June 11, 2025, and Warner Bros followed on September 5, 2025; Midjourney argues training on the studios’ characters is fair use.

The filing targets a prior ruling that only required studios to share generative-AI usage tied to “consumer-facing” videos and images. Midjourney says that limit “unfairly” lets the plaintiffs “cherry-pick only those documents they believe support their market harm claims,” and alleges the withheld material “would reveal whether, behind closed doors, they are doing exactly what they are suing Midjourney for doing.” Studios’ lead attorney David Singer called the request a “fishing expedition,” and added that studios “do not seek to stop AI technology or even shut down Midjourney’s business” but want Midjourney to stop copying their IP. The discovery fight will set the practical rules of evidence for every AI-versus-studio case this year.

GPT-4 Turbo Impersonated 112 BBC Question Time Figures, and Most UK Viewers Preferred the AI

A peer-reviewed study published in PLOS One and led by Steffen Herbold at the University of Passau used BBC Question Time transcripts plus Wikipedia biographies to prompt GPT-4 Turbo to impersonate 112 UK public figures who appeared on the show ahead of the 2024 UK election. 948 UK participants rated the real and AI responses on authenticity, coherence, and relevance.

The clean numbers: the clear majority favored the AI for coherence and relevance, and more than half rated the chatbot as more authentic than the actual person. Herbold told 404 Media that authenticity is “supposedly hard to fake… We’re not talking about unknown people. We’re talking about one of the biggest shows in the UK,” and quoted participant reactions including “Wow, I never believed this was AI” and “if AI can do this, what else might I have missed?” It is the first clean, peer-reviewed anchor we have for synthetic-media risk in a real political panel, and the most useful data point yet for the “AI president” problem.

Simon Willison Ships sqlite-utils 4.0rc2 With Fable, Posting Receipts

sqlite-utils 4.0rc2 was mostly written by Claude Fable: 37 prompts, 34 commits, +1,321/-190 lines across 30 files, with an estimated unsubsidized API cost of $149.25 ($141.02 on the main session via claude-fable-5 plus about $8.23 across smaller agents). Fable’s review caught a release-blocker data-loss bug in delete_where() that “never commits and poisons the connection,” plus a Python 3.12 / auto-commit incompatibility that broke almost the entire test suite.

A second cross-model pass with GPT-5.5 surfaced two P1 issues that the docs had hidden: db.query("update ...") committed the update before raising ValueError, and INSERT ... RETURNING via db.query() only committed if the generator was fully exhausted. Willison upgraded from the $100/month Claude Max plan to the $200/month plan ahead of the “July 7th Fablepocalypse,” when Max subscribers lose subsidized access to claude-fable-5. The full transcript and PRs are linked from the post, which makes it the most concrete receipt yet that one experienced developer can ship an end-to-end open-source release through an agent on a working budget.

”Better Models, Worse Tools”: Anthropic’s Coding-Agent Style Leaks Into Pi

Armin Ronacher, via Simon Willison, reports that Opus 4.8 and Sonnet 5 sometimes call Pi’s edit tool with invented fields in the nested edits[] array - the edits themselves are usually correct, but the arguments do not match the schema, so Pi rejects the call. Older Anthropic models do not show the problem; Ronacher notes “not Haiku or some small model: Opus 4.8,” and the pattern is getting worse, not better.

His theory: newer Anthropic models have been trained (likely via RL) to better use the edit tools baked into Claude Code, which hurts third-party harnesses like Pi with their own custom edit tools. Claude’s edit tool uses search-and-replace, OpenAI’s Codex uses apply_patch, and OpenAI has trained models specifically for that tool. The practical signal for anyone integrating frontier models into a heterogeneous tool stack is that “drop in a frontier model” is not a safe upgrade path, and a vendor’s coding-agent style is starting to act like a de facto programming environment.

Alibaba Damo Academy’s “Elements Claw” Discovers Four Lab-Confirmed Superconductors

Elements Claw is a 1-billion-parameter foundation model trained on 125 million molecular and crystal structures and run for 28 hours of GPU processing. The agent screened 2.4 million stable crystal structures, surfaced about 68,000 candidates with superconducting potential, and laboratory-confirmed four new superconductors, with partners Renmin University of China and the University of Chinese Academy of Sciences.

For context, the SuperCon database held only about 2,000 superconductors before this work. The South China Morning Post frames Elements Claw as “the industry’s first artificial intelligence agent for discovering superconducting materials,” designed to “accelerate the timeline by scanning scientific literature and screening millions of crystal structures.” It is the first widely reported case of an “agent-on-foundation-model” loop producing novel inorganic materials that survive physical synthesis, and a meaningful counterweight to the coding-agent narrative when someone asks where agents are being trusted with real-world consequences in mid-2026.

Meituan’s LongCat-2.0: A 1.6T Open-Source Model Trained Without Nvidia

Meituan has documented LongCat-2.0, a 1.6-trillion-parameter open-source LLM with about 48 billion parameters activated per token, trained on 35T+ tokens. The model card only names “domestic AI ASIC superpods” as the hardware, with HCCL coordinating the cluster; the community estimates roughly 50,000-60,000 Huawei Ascend 910C cards, though Meituan does not confirm the figure.

The architecture is Mixture-of-Experts with Zero-computation Experts and Shortcut-connected MoE, with activated parameters ranging from 33B to 56B and roughly 97% sparsity excluding N-gram embedding. The post cites 1.5x training MFU improvement, 70%+ reduction in daily failure rate, and over 30% MFU. The cleanest comparison line is from a developer quoted in the piece: previous domestic-compute work was “building a house elsewhere, then using domestic compute to decorate it,” whereas LongCat-2.0 is “laying the foundation, building the house, moving in, and finding that it is actually livable.” It is the clearest open-weights receipt yet that the Chinese frontier stack is producing trillion-parameter artifacts outside Nvidia’s hardware envelope.

Apple’s “Hide My Email” Bug Has Been Unpatched for Over a Year

A vulnerability in Apple’s email-aliasing service lets an attacker recover the real email address behind a Hide My Email alias. A security researcher reported the flaw to Apple more than a year before 404 Media’s July 1 writeup, and 404 Media verified it was still exploitable the Monday before publication.

The article withholds the technical steps so the bug is not further exposed, but the practical signal is clear: “Hide My Email users deserve to know that it may be possible for attackers to discover their hidden email addresses.” The disclosure timeline - over a year, no patch - is exactly the kind of slow-patch arc that makes everyday privacy tools unsafe to depend on for at-risk users, and it lines up with the broader pattern of Apple privacy promises that do not survive contact with a determined attacker.

EFF, Demand Progress Education Fund, the National Consumers League, and EPIC filed comments on July 2 urging the FTC to reject X Corp.’s May 15, 2026 petition to set aside a 2022 consent decree stemming from Twitter’s misuse of 140 million users’ security data (phone numbers and emails) for targeted ads. The order, which included a $150 million fine, runs through 2042.

EFF’s three counter-arguments: FTC orders bind the corporate entity and do not dissolve with staff turnover, citing X’s 2024 Grok training on user data without meaningful consent and a 2025 data breach; AI training on user data “supercharge[s]” secondary-use risk via prompt-engineering attacks that extract training data; and compliance costs are “a rounding error against the $200 billion valuation of X Corp. following the xAI merger.” The original 2011 FTC settlement followed Twitter’s failure to secure user data and required a 20-year reporting regime, which the 2022 order extended to 2042. Whatever the FTC decides sets the precedent for every future “ownership changed, kill the consent decree” petition.

The “Agent Skills” Spec Goes Cross-Vendor

agentskills/agentskills is the first upstream-looking specification for what an Agent Skill is, covering discovery, activation, and execution as three distinct loading stages and a SKILL.md manifest with required name and description fields. The codebase is Apache-2.0 and the documentation is CC-BY-4.0; the project tracks compatible clients in a public showcase.

practical-engineer-skills sits beside it: an MIT-licensed curated bundle of Matt Pocock’s own Claude Code skills, including tdd, triage, code-review, domain-modeling, codebase-design, diagnosing-bugs, and research. The repo ships a /setup-matt-pocock-skills setup helper and was last tagged v1.0.1 on June 17, 2026. Together they show the same pattern from two angles: the spec lets the open ecosystem standardise on what a Skill is, while individual engineers publish their working kits as distribution. It is the local-tools beat of the consolidation: closed-vs-open keeps burning, but the agent ecosystem is quietly agreeing on a file format.

Quick Hits

  • page-agent from Alibaba: alibaba/page-agent is an MIT-licensed TypeScript SDK that turns any web page into an LLM-controllable surface via text-based DOM manipulation - no extension, no headless browser, no multi-modal model required. v1.11.0 shipped on July 3, 2026.

  • Chrome DevTools as an MCP server: ChromeDevTools/chrome-devtools-mcp is the Apache-2.0 first-party MCP server that wraps Chrome DevTools - network, console, screenshots, performance - as tool calls any coding agent can invoke. Usage-statistics telemetry is on by default and can be disabled with --no-usage-statistics.

  • GPT-5.5 Codex reasoning-token clustering: openai/codex issue #30364 reports that 19.3% of Codex responses are GPT-5.5 but 82.0% of exact-516-token events, with an exact-516 / at-least-516 ratio of 44.0% versus a 1.3% baseline (about 33.6x higher) across 390,195 response records and 865 sessions from February 1 to June 27, 2026.

  • EFF queer-data self-help guide: EFF’s LGBT Q&A on wiping online data points of queer identity walks through self-search, account-by-account auditing, Privacy Badger, EasyOptOuts, Optery, Google’s “Results about you” page, the California Privacy Protection Agency’s “Drop” tool, and removing mobile advertising IDs.

  • Google’s “Group project, but make it 1776” ad: A Google commercial for the 250th anniversary of the Declaration of Independence depicts Thomas Jefferson and Ben Franklin collaborating via Google Docs, Google Calendar, Google Meet, e-signatures, Gemini meeting notes, and a “help me visualize” tool for the national seal.

Worth Watching

AI-versus-studio discovery fights widen. Midjourney’s counter-discovery demand turns the next round of AI copyright cases into a fight over what studios, labels, and newspapers do with AI in-house, not just what model vendors do during training. Expect more defendants to follow with similar motions, and more plaintiffs to scramble to redact internal AI usage memos before they reach opposing counsel.

Frontier models leaking into third-party tool stacks. Ronacher’s “better models, worse tools” piece is the cleanest articulation yet of a category of failure that will only get worse as labs push RL against their own agent harnesses. Watch for the next vendor with a popular coding agent to ship a post about “tool contamination” in a competitor’s model.

Domestic-compute trillion-parameter models. LongCat-2.0 and Elements Claw land in the same week and both come out of Chinese labs using Huawei silicon. If a second or third 1T+ open-weights model arrives this quarter on the same hardware envelope, the assumption that Nvidia-only training is a hard floor for frontier-scale artifacts starts to break.

Slow disclosure timelines on privacy tools. The Hide My Email bug, the EFF X filing, and the queer-data self-help guide all sit in the same window. The single most actionable beat for readers is “don’t assume an everyday privacy tool has been audited by the vendor that ships it” - that assumption is unsafe even for first-party Apple services.