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

Callosum raises $100M for heterogeneous AI
08/20/26, 04:46 PM

Callosum raised $100 million in seed funding to route complex AI workloads across different models and chips, including Nvidia, AMD, Google, and other hardware, rather than assuming one homogeneous GPU stack, according to the funding report and the company’s heterogeneous-intelligence thesis. The round follows a $10.25 million pre-seed six months earlier, a roughly 9.8× capital jump. For builders, file the bet as infrastructure optionality: task decomposition may make mixed accelerators usable, but Callosum has not yet published the technical benchmark behind its cost and speed claims, so the compute-futures market’s new price layer remains the more measurable signal today.
Wire

Nvidia discusses Mercor at a $20B valuation
08/20/26, 02:46 PM

Nvidia is reportedly discussing an investment in Mercor at a $20 billion valuation, a potential doubling from the AI-data company’s $10 billion valuation in October 2025, according to the reported financing talks. Mercor supplies expert human data and evaluations for model developers, including work connected to Nvidia’s open models; the round is preliminary and not confirmed. For builders, the signal is that frontier-model economics are pricing scarce judgment—doctors, lawyers, engineers, and other specialists—as infrastructure, not annotation overhead; the physical-AI funding surge is the parallel capital-market read-through.
Wire

Microsoft patches the three-flaw Copilot chain
08/20/26, 02:43 PM

Microsoft patched CoSnitch, a critical Copilot vulnerability chain that combined three flaws into a one-click route to prompt execution, connected-app data access, and persistent memory poisoning, according to Varonis’s disclosure. Varonis disclosed it in December 2025; patches shipped on August 18, 2026, and the researchers report no evidence of exploitation in the wild. The lesson for teams wiring agents into email and files is the same one behind Claude’s new write actions and approval trail: treat connectors and durable memory as a security boundary, not a convenience feature.
Wire

Network Bio raises $50M for tissue-trained AI
08/20/26, 02:40 PM

Network Bio launched with $50 million in financing and a connected network of academic biobanks intended to train medical AI on patient tissue, paired blood samples, molecular data, and longitudinal clinical outcomes, according to the company’s launch page. The scarce input is not another model checkpoint but the standardized biological evidence behind it: Network Bio says it has Fortune 100 partners and is working with NVIDIA on biological foundation models. Builders evaluating medical AI should file this alongside the shadow-clinic warning about medical AI’s evidence gap: better models do not erase provenance, cohort, and validation costs.
Wire

AI package hallucinations create slopsquatting risk
08/20/26, 02:00 PM

An AI agent nearly sent a developer toward malware after recommending a plausible package name; Softjourn caught the risk by checking GitHub history, download counts, and source code before installation, according to The Register’s incident report. The report identifies the pattern as “slopsquatting,” in which attackers register packages under names hallucinated by models, but does not describe an operational payload. For coding-agent teams, the number that matters is the 1 missing approval step that can turn a routine dependency suggestion into a supply-chain incident—exactly why agent harness cost comparisons must include human review time, not just tokens.
Wire

Google adds A/B tests to AI Max Search
08/20/26, 01:55 PM

Google is adding multi-campaign A/B tests to AI Max for Search, letting advertisers compare budget and ROI-target changes across multiple campaigns, with the rollout scheduled for September 2026, according to Google’s announcement. The update also adds brand and location controls to experiments and lets Performance Planner apply recommended bidding or budget changes in one click. For operators, the concrete shift is measurement: automated campaign expansion now comes with a vendor-built test harness, so teams should evaluate incremental return rather than accept a black-box efficiency claim alongside the archive’s analysis of AI intent monetization.
Wire

Slack Code puts agents in project channels
08/20/26, 01:52 PM

Slack launched Slack Code, dedicated project channels where teams can work with 4 named coding-agent partners—Claude Code, Devin, Vercel Agent, and GitHub Copilot—while reviewing diffs and live previews, according to the launch report. Slack says the channels are available on every plan, archive on completion, and preserve an audit log for feedback and approval. The operator signal is a workflow boundary rather than another model: coding agents are moving into the team system of record, a useful complement to the approval layer for agentic workplace actions.
Wire

Xiaomi scales robot training to 100,000 hours
08/20/26, 12:41 PM

Xiaomi-Robotics-1 uses 100,000 hours of embodiment-free trajectories for pre-training and more than 7,200 hours of real-robot data for post-training. The project reports 75% overall success after fewer than 10 hours per new task and 85% after fewer than 40 hours, with gains carrying into unseen environments and four simulation benchmarks, according to the accompanying paper. For robot builders, the useful fact is a data strategy rather than a humanoid demo: broad UMI pre-training can make scarce embodiment-specific demonstrations more valuable, reinforcing the simulation funnel behind modern physical-AI research.
Wire

Unitree puts a dexterous arm below ¥10,000
08/20/26, 12:38 PM

Unitree launched the R1 seven-axis dexterous robotic arm at RMB9,900. The official specification page lists a 5.5kg mass, 2kg payload, 650mm reach, force feedback and collision detection, with control software and interfaces planned for open release. For embodied-AI labs, a sub-RMB10,000 manipulation platform lowers the cost of collecting contact-rich demonstrations, although Unitree’s warning about limited position-control stiffness makes the price an entry ticket rather than a turnkey workcell; it extends the hardware curve behind the physical-AI funding boom.
Wire

Kling AI clears ¥850M in quarterly revenue
08/20/26, 12:35 PM

Kuaishou says Kling AI generated more than RMB850 million in Q2 2026 revenue, up over 200% year on year, against RMB35.5 billion in group revenue. The company also says Kling 3.0 now supports native 4K output, while its MCP and CLI releases let agents orchestrate batch video creation. For builders, the signal is less another video demo than a monetization threshold: AI video is now a product line large enough to measure inside a public company’s core results, a commercial contrast worth tracking alongside FLUX 3’s gated audiovisual and robotics claims.