Category
Anthropic CEO Dario Amodei has proposed slowing AI development and embedding external evaluators inside the company. We compare the plan with METR's post-incident investigation, looking at access to training, publication rights, the reality of self-improvement, and the difficulty of international coordination.
Goldman Sachs has rewritten its 2035 humanoid robot market forecast twice, from $6 billion to $38 billion to $138 billion. Meanwhile, the Bank of America Institute estimates per-unit parts costs at $35,000. Dividing that by 30 actuators reveals what's really setting the pace of price declines.
Google's TimesFM-3 is a time-series forecasting model that handles sales, related series and planned promotions without extra training. We look at how it differs from the XReg-based TimesFM-2.5, how it forecasts, what its benchmarks compare, and its non-commercial weight license.
Antares says it has been selected for a $161 million Strategic Breakthrough award to demonstrate a space reactor. This explainer separates Mark-0 criticality, the 2027 Mark-1 power test and R1-S flight qualification, and covers heat rejection and government targets.
Roblox plans standalone game apps and Chrome support by the end of 2026. The shared platform stays, but how safety settings carry over is unclear. We look at Australia's review and ongoing age estimation to see what protections will be needed.
OpenAI reportedly asked members of the US Congress about the legality of jointly slowing AI development with competitors. This piece examines the company's public stance alongside US bill S.5105, breaking down the requirements for advance notice, the burden of proof placed on companies, and the challenges of verifying an international pause.
Toshiba's U.S. subsidiary and DIMAAG are teaming up on SCiB-based storage for AI data centers. We look at the cooling behind continuous 10C charging and discharging, how it fits with 800V power delivery, and the conditions behind the sub-1% claim.
OpenAI has released the Agents API, which exposes the execution infrastructure behind Codex. This article explains how it differs from the Agents SDK, the lifespan of sessions and working files, and the data retention conditions that apply even when running in your own environment.
Graphics card shipments rose in Q2, according to JPR data and TrendForce's memory supply analysis. We look at the rush-buying hypothesis, discrepancies in the published figures, and what the card-to-CPU shipment ratio can and cannot show.
ClickFix tricks users into running code via fake verification and repair steps. Primary research on compromised-site delivery, Windows and Mac intrusion paths, and in-browser crypto theft shows where defenses fall short.
