Open Weights Catch Up | AI Field Notes #79

AnalysisA Chinese model you can download now scores higher than a Western flagship on the tests that measure agent work. Alibaba released Qwen3.8-Max on August 3, a mixture-of-experts model (a design that activates only part of itself per request to cut cost) with 2.4 trillion total parameters and 95 billion active. On PaperBench it posted 93.0 against Anthropic's Fable 5 at 88.8, and on the instruction-following test IFBench it hit 82.8 to Fable's 63.5. It runs a 1-million-token context. Open weights land next week. The frontier's paywall is thinning.

AnalysisThirty seconds of video with matched sound, made in one pass, is the new bar for a single prompt. ByteDance launched Seedance 2.5 on August 1, generating clips three times longer than Google's Gemini Omni Flash, which stops near ten seconds. It accepts up to thirty reference images and ten audio files, builds multi-character scenes, and swaps camera angles inside a shot. It runs on ByteDance's Jimeng and Doubao apps today, with paid API access through BytePlus coming later. The audio and video arrive together, so there is no separate sound pass to book.

AnalysisThe best open model for making video is now Chinese and free to download under most conditions. MiniMax released H3 on August 3, a 33-billion-parameter system that reads text, images, video, and audio together and generates four to fifteen second clips with stereo sound. On the Artificial Analysis boards it placed first in video editing and second in text-to-video, the first open-weight entry to lead any of them. The license lets companies under $20 million in revenue use it commercially, though the 2K module stays closed. For short-form video, free now competes with the tools people pay for.

AnalysisTen open problems in mathematics and theoretical computer science, untouched for at least a decade, fell to an internal version of a model OpenAI calls Astra. The company unveiled it on August 1, describing a new family that coordinates several agents over long runs, alongside its Sol, Terra, and Luna models. One proof settled whether certain non-sofic groups exist; others reached into lattice cryptography and quantum complexity. Generating all ten cost about $2,000 in API fees. Sam Altman showed the work to Washington first. There is no release date, because Astra would be the first model routed through the government's new approval framework.

AnalysisText a person cannot see turns an ordinary document into a carrier, because the assistant reads it anyway. Security researcher Håkon Måløy showed on August 1 how white text at a tiny size, invisible on the page, feeds Microsoft Copilot instructions when it summarizes the file. Copilot strips the formatting, follows the buried commands, and copies them into whatever it writes next, so an infected market analysis quietly poisons the financial report built from it. This is a prompt-injection attack, malicious instructions smuggled in as content. Microsoft acknowledged it on March 31, tried twice to fix it, and after 144 days Måløy went public.

AnalysisThe expensive part of an AI breach is rarely a clever attack. In its 2026 Cost of a Data Breach report, published August 3 with the Ponemon Institute across 602 companies, IBM found that 92 percent of firms hit by an AI-related incident lacked basic access controls. The usual door was a compromised API, a connected app, or a misconfigured cloud service. Open-weight or proprietary made almost no difference to the odds. AI-involved breaches averaged $5.33 million against $4.70 million for the rest, and when attackers used AI themselves the bill reached $6.04 million.

AnalysisThe part of Europe's AI law that touches ordinary products took effect on August 2, and regulators can now inspect providers and levy fines. Under Article 50, anyone offering a chatbot must tell users they are talking to a machine, and synthetic images, audio, and video have to be labeled as generated. Penalties reach 7 percent of global turnover. The heavier rules for high-risk uses like hiring and credit scoring slipped to December 2027 under a July revision called the Digital Omnibus. What binds you today is disclosure and content labeling; the rest waits.

AnalysisSix well-known songs turned up, reproducibly, inside an AI music model, and a court called that infringement. On August 1 the Munich Regional Court I sided with GEMA, Germany's music-rights body, finding that Suno both trained on protected tracks and reproduced their pieces when prompted with a title and style. The judges held the company responsible rather than its users, since Suno chose the data and built the model, and noted it stream-ripped songs from YouTube. Germany's text-and-data-mining exception did not cover it, and, applying US law, the court rejected a fair-use defense under the Supreme Court's Warhol standard.

AnalysisThe money chasing AI has moved past chips to the power plants themselves. Valar Atomics, a startup building small modular reactors (compact nuclear units meant to ship faster than traditional plants), raised $1 billion on August 3 in a round led by Sequoia's Shaun Maguire, at a $6 billion valuation. The pitch is direct: data centers need firm electricity that solar and wind alone cannot promise, and the grid cannot add it fast enough. Whether Valar can clear regulators and actually pour concrete is the open question, but investors are pricing the shortage as real.

AnalysisEvery new data center waiting to plug into the Texas grid is frozen while the state checks who they are and what they use. Governor Greg Abbott ordered the audit on August 3, directing ERCOT and the utility commission to verify tax breaks, power draw, water use, and ownership before any connection clears. The queue is enormous: more than 1,800 projects and over 474 gigawatts, roughly five times the grid's record peak demand, about 90 percent of it data centers. Only 28 of 377 firms had answered a water-use survey. ERCOT paused its review the same day.

AnalysisTwo platforms decided this week to treat machine-made content as a problem to filter. Snap said in August it will drop fully AI-generated videos from Spotlight, its recommendation feed, keeping the surface on what it calls real creativity; clips made with Snap's own tools stay but carry a label. LinkedIn added a dedicated button to report "AI slop," separate from normal complaints, after a study found its feed thick with generated filler. The backdrop is hard to ignore: 21 percent of YouTube Shorts are already AI-made, and the big feeds are deciding how much of that they want.