Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

H3 Max renders a 5-second video in 3 seconds
08/28/26, 07:45 AM

fal has released H3 Max, a post-trained MiniMax H3 video model that produces a five-second clip in about three seconds—roughly 35× the throughput of the official H3 endpoint. In its H3 Max release note, fal says the model ranked first across overall quality, prompt understanding, and aesthetics in human-preference tests against 12 video models, though those comparisons are company-run. That shifts the near-term creative-AI question from whether quality requires latency to whether a workload can exploit a faster generation loop; the recent Gemini Flash price-performance comparison is a useful baseline for testing cost and speed together.
Wire

Cursor agent appears in a seven-company attack wave
08/28/26, 06:40 AM

Cursor Agent appeared in an Aurora ransomware campaign spanning ten target organizations, while Reuters identified at least seven companies that were breached. Gambit Security’s investigation documents the agent’s use between April 8 and May 21, and the Reuters report republished by CNA says the extent of Cursor’s causal contribution remains uncertain. For builders, the fact to file is governance, not attack technique: agentic coding tools need stronger abuse monitoring and refusal recovery than a one-time policy check, a risk that sits beside the case for an agent-security firewall.
Wire

KAIST's K-Fold is 25× faster than AlphaFold3
08/28/26, 06:38 AM

KAIST says its K-Fold protein-and-drug model predicts molecular complexes up to 25× faster than existing systems while approaching AlphaFold3 accuracy. The 7B main model and 2B lightweight model are planned for release under Apache 2.0, and the HyperLab service links the model to more than 120 computational tools, according to Seoul Economic Daily’s report on K-Fold. The operator takeaway is a cheaper, locally controlled screening option—not a drug-discovery shortcut: protein teams should compare its claims on their own targets before swapping a wet-lab validation lane, as GenScript’s four-day assay promise shows.
Wire

SK hynix bets $4B on US HBM by 2029
08/28/26, 06:36 AM

SK hynix broke ground on a more-than-$4 billion advanced-packaging facility in Indiana that targets its first U.S. HBM mass production in the second half of 2029. The company says the cleanroom should open by October 2028, the site will employ about 1,000 people, and the broader supplier ecosystem could support 7,000 jobs, according to its Indiana HBM facility announcement. That matters less as a jobs release than as a supply-chain signal: the wafers will still be made in Korea and packaged in the United States, a split architecture builders should track alongside the new economics of Rubin-scale AI infrastructure.
Wire

Gemini Notebook grounds answers in 100,000 books
08/28/26, 05:40 AM

Google launched Expert Intelligence in Gemini Notebook with access to more than 100,000 books from major publishers, letting users query purchased Google Play Books titles and receive cited answers, the company says. The ownership gate is the product’s real business detail: collaborators must buy their own copy before a shared notebook can use the book, preserving a rights boundary rather than turning every upload into a communal corpus. Builders watching grounded assistants should track this distribution move alongside Google’s cheaper Gemini workhorse: trusted licensed sources are becoming a product surface, not just a retrieval backend.
Wire

South Korea opens 1.56T tokens to AI builders
08/28/26, 05:38 AM

South Korea is opening 29 AI training datasets totaling about 1.56 trillion tokens for free download through its national AI Hub, the Ministry of Science and ICT’s government notice says. The release includes pretraining, multimodal, red-teaming, and physical-AI data from five foundation-model teams, but access is limited to Korean companies, researchers, and students. For builders outside Korea, the immediate implication is a sovereign-data benchmark to watch; it adds a public counterweight to the open-model ecosystem’s race beyond model weights.
Wire

Visa opens an 11-stage vulnerability agent
08/28/26, 05:36 AM

Visa released VVAH, an open-source security harness whose pipeline spans 11 stages and four phases for vulnerability discovery, remediation, and validation, according to its Apache 2.0 repository. The documentation warns that a default scan can modify source code, so the important deployment fact is that AI-assisted remediation still belongs behind a review gate rather than in an unattended CI path. Teams testing coding-agent security workflows can compare that boundary with Microsoft’s small-model security routing, but should treat VVAH as a harness to evaluate, not proof that autonomous patching is safe.
Wire

Cohere Parse sets $1.50 per 1,000 pages
08/28/26, 04:30 AM

Cohere has made Parse generally available at $1.50 per 1,000 pages, while its Model Vault deployment promises 23% lower inference cost at 50% GPU utilization and up to 61% lower at full utilization, according to Cohere’s launch details. The document model scores 79.2 on Cohere’s three-dimension ParseBench average and processes 4.5 pages per second, or 36 pages per second on an eight-H100 node, giving buyers a concrete baseline for high-volume RAG and invoice pipelines. Teams comparing OCR vendors should read this beside the archive’s document-agent harness test: cheaper parsing only matters if the downstream agent preserves grounded answers.
Wire

Gemini Live turns voice into scheduled work
08/28/26, 04:27 AM

Google says 63% of Gemini users talk to it out loud as Gemini Live adds a handoff to Spark for multi-step jobs across Docs, Sheets, Drive, and the web, the product announcement explains. Spark can run scheduled work for days or weeks, while Live can search, summarize, archive, or delete Gmail messages by voice. Builders should file this as an interface shift rather than a new chatbot: voice agents now need explicit task state, permissions, and cancellation paths when a conversation can launch background work; the archive’s audio-first device thesis is the adjacent hardware bet.
Wire

OpenAI expands ChatGPT for Teachers by 100,000
08/28/26, 04:24 AM

OpenAI is expanding ChatGPT for Teachers to 55 additional school systems across 20 states, adding more than 100,000 educators and staff, according to the rollout figures reported by 9to5Mac. The company says the program now reaches over 300,000 educators across more than 100 K–12 organizations, while a 16-state data-privacy agreement gives districts a common evaluation path instead of negotiating from scratch. The move extends the free-through-June-2028 offer, but district buyers should still treat the 100,000-seat research subsidy precedent as a reminder to lock down export, retention, and post-pilot pricing before adoption.