The day's tech, sifted: Jul 04, 2026

What matters today: Midjourney is pushing to widen discovery in its copyright fight with Disney, Universal, and Warner Bros., asking a judge to overturn a ruling that limited the studios to disclosing only consumer-facing AI use, so it can argue that unlicensed AI training is common industry practice. Alongside it, four separate AI-adjacent funding rounds (chips, robotics, quant trading) and a fresh AMD inference benchmark point to the same story from the hardware and capital side: money keeps piling into AI infrastructure and the funds that trade on it.

AI / LLMs

• GLM 5.2 running on AMD's MI355X, benchmarked at 2,626 tokens/sec per node, over 2x cheaper than on Blackwell: a concrete data point (per the submitter's own benchmark) in the continued push to loosen Nvidia's grip on inference pricing.
• Simon Willison links to an "Open Source AI Gap Map": a new attempt to track how far open models trail closed ones, worth a look for anyone weighing open versus proprietary weights.
• Mistral released Leanstral 1.5: the latest version in its Lean theorem proving model line (the announcement page did not load, so specifics beyond the title are unconfirmed).
• Zvi Mowshowitz reports a restricted Anthropic model line, nicknamed Fable/Mythos, had access restored after negotiation with the government: per his account Anthropic did not concede on its stated principles to get there, though the underlying restriction and negotiation are only described secondhand in this roundup.

Devtools & Infra

• SearXNG: a free, AGPL licensed metasearch engine (33k+ GitHub stars) that aggregates results from Google, Bing, DuckDuckGo, Brave and others into one interface without tracking or profiling users, a self-hostable alternative for anyone souring on mainstream search.
• FreeBSD ate my RAM: a developer's writeup diagnosing unexpectedly high memory usage on FreeBSD (the page did not load for a fuller read, so this is title only).
• Dan Luu posts field notes on agentic coding workflows, "from Galapagos Island" (title only, page did not load): another entry in the growing pile of developer accounts of working with AI coding agents.
• US residents angry that datacenters were "shoved down our throats" are recalling local officials over them (title only, page did not load): a sign the AI buildout's local politics are starting to bite back.
• Apple's Doug Brooks explains why Mac minis have become the preferred machine for running AI agents inside frontier labs: a Q&A with Apple's silicon product manager on the on-device AI reasoning behind that choice.

Security & Privacy

• Epoch AI data shows a spike in serious CVEs coinciding with the release of Claude Mythos Preview (title only, page did not load): worth tracking whether this holds up as correlation or turns into a real finding about model releases and vulnerability disclosure timing.
• "Odin, Wikipedia and engagement farming" (title only, page did not load): flagged for its angle on AI-generated content interacting with Wikipedia's engagement incentives.
• A GitHub issue reports possible session or cache leakage in Claude Code, with an Enterprise ZDR workspace session unexpectedly surfacing unrelated Minecraft content: a single report so far with no maintainer response, but one worth watching given the workspace isolation stakes for Enterprise customers.

Startups & Industry

• Midjourney wants Disney, Universal, and Warner Bros. to disclose their own internal AI use: the studios sued Midjourney in 2025 for copyright infringement, and Midjourney is now asking the court to overturn a ruling that limited the studios to disclosing only consumer-facing AI, arguing broader disclosure would show unlicensed AI training is standard industry practice.
• Chinese quant funds have more than doubled assets under management to about $384B in under a year: Bloomberg reports the surge is being credited to AI-driven trading models beating human stock pickers, pulling capital away from discretionary managers.
• Montreal's Stathera raises a $55M Series B led by Maverick Silicon, taking total funding to $75M, for MEMS-based silicon timing chips it says are needed to keep pace with AI compute demand.
• Prague's EquiLibre raises a Series A at a $500M valuation: founded by three ex-Google DeepMind researchers who built a poker-beating AI, now applied to trading stocks.
• Singapore's dConstruct Robotics raises a $125M Series A for spatial tech letting autonomous robots operate in GPS-denied environments.
• Micron breaks ground on a ~$9.3B factory expansion in Hiroshima, targeting HBM shipments from summer 2028 as part of its global ramp-up to meet AI memory demand.
• Hong Kong now accounts for over 50% of China's $239B in chip imports through the first five months of 2026, a record share (up from about a third a decade ago) that Bloomberg says reflects its role as a conduit in the wider chip supply chain.
• ByteDance's Seedance video generator is making inroads with Hollywood, the Los Angeles Times reports, on the strength of low pricing, realism and timeline-based prompting, despite the backlash a viral fake Brad Pitt and Tom Cruise video sparked in the industry.

Research

• A profile of Google DeepMind philosopher Iason Gabriel: his work since 2017 has tracked, and in several cases predicted, the ethical challenges that came with LLMs' success.

Elsewhere

• The Life and Times of Maxis, Part 1: SimEverything: the first installment of a history series on the studio behind SimCity and SimEarth.
• Africans Are Turning to Starlink (paywalled): The Economist on growing satellite internet adoption across the continent.

Hacker News

Security cluster: MSI Center, How to gain SYSTEM privileges in seconds (103 points, 39 comments) walks through a local privilege escalation in MSI's bundled Windows software. Soatok's Informal Guide to Threat Models (102, 22) offers a practitioner's framework for figuring out what you're actually defending against, and why. The CVE severity spike tied to Claude Mythos Preview's release, and a separate report of session or cache leakage in Claude Code, both get full treatment elsewhere in this digest: see New serious vulnerabilities spiked around release of Claude Mythos Preview and Potential session/cache leakage between workspace instances or consumer accounts (154, 56).

Devtools and learning cluster: David Beazley, Programming Courses (110, 30) rounds up his hands-on courses on Python internals and language implementation. Applied Category Theory Course (2018) resurfaces John Baez's UCR materials on category theory for working scientists. Synthesis is harder than analysis argues building coherent designs is intrinsically harder than pulling them apart, and Maybe you should learn something pushes back against overoptimized, credential-driven skill acquisition. Software, from First Principles (112, 17) lays out a set of mental models for reasoning about software design, and Explanation of everything you can see in htop/top on Linux (187, 27) walks through every field the two monitors report. Dan Luu's Agentic coding notes from Galapagos Island is covered in full above.

Lighter fare: FIDE's ethics commission handed down sanctions in International chess federation sanctions Kramnik. Exeter researchers explain why Giant trees have no trouble pumping water to top branches, and The firefighting system of the Van der Heyden brothers in 17th century Amsterdam traces the invention of the modern fire hose and department. Astrophysicists Puzzle over Webb's New Universe (129, 74) covers findings from the telescope that keep confounding cosmologists. On the hardware side, Steam Controller Auto-Charge, pilot to magnetic charging puck using CV uses computer vision to dock a controller for charging, and The bottleneck might be the air in the room makes the case that CO2 buildup in meeting rooms degrades decision-making before anyone notices.

Also on the front page, already covered in full above: The Life and Times of Maxis, Part 1: SimEverything, the datacenter recall elections story, FreeBSD ate my RAM, SearXNG, Africans Are Turning to Starlink, GLM5.2 on AMD MI355X, Leanstral 1.5, and Odin, Wikipedia and engagement farming.

Threads

• The AI hardware buildout shows up from several angles: Micron's $9.3B Hiroshima HBM expansion, Stathera's timing-chip raise, GLM 5.2's AMD MI355X benchmark, and Hong Kong's record share of China's chip imports all point at the same scramble to feed and diversify AI compute supply, while Apple's Mac minis becoming the preferred machine for running AI agents shows the same pressure reaching the client side.
• AI is displacing human judgment in finance on two fronts at once: China's quant funds doubling AUM and EquiLibre's $500M-valuation raise both credit AI trading models with beating human stock pickers.
• An "alternatives to the default stack" thread runs through the day: GLM 5.2 on AMD's MI355X (an alternative to Nvidia for inference), SearXNG (an alternative to mainstream search engines), and the Open Source AI Gap Map (tracking how open models compare to closed ones).
• Scrutiny of how AI actually gets built keeps surfacing: Midjourney's discovery fight over training data, Iason Gabriel's ethics work at DeepMind, and the CVE spike tied to Claude Mythos Preview are three different angles on the same underlying question of accountability.
• Anthropic's Claude models keep drawing security scrutiny: the CVE severity spike coinciding with Claude Mythos Preview's release and the separate, unconfirmed report of session or cache leakage in Claude Code both raise open questions about how these systems are deployed and isolated.