Topic
Explore the latest artificial intelligence news and research.
GPT-6 Astra reportedly reached Pokémon champion in 18 hours 12 minutes, but stalled in Minecraft after losing supplies. Operator reports and ARC Prize's evaluation show why test conditions, memory handling, and recovery after failure matter.
OpenAI's Codex developer warns against running three or more parallel sub-agents. Meanwhile, Nous Research's large-scale refactor using 1,393 agents produced real results. We examine when the coordination tax grows and how to judge whether parallelization is worth it.
Huawei is dividing its Ascend 960 AI accelerator into separate training and inference chips, accelerating the rollout by up to nine months. Official documents reveal how the new Atlas 960E's proximity-mounted optics aim to cut power use and boost uptime—and where supply and software gaps remain.
A 1941 Enigma message was restored using GPT-6 Astra. We trace how a place name guided the search, how separate records were used to check it, and what the published key does and does not prove.
AMD has notified partners that it will raise prices for AI chips, GPUs, and motherboard chipsets by roughly 10% starting in Q4 2026. The move, driven by rising TSMC manufacturing costs, coincides with Intel's earlier price hike and is set to push up overall BOM costs across the DIY PC market.
Anthropic has completely revamped the Projects feature in Claude Code. By simply providing a goal, multiple threads now work in parallel on independent branches, autonomously handling git merge conflicts and creating pull requests. With shared memory and individual configuration options, the update pushes autonomous development forward.
We checked the Geekbench 7 listings for the M6 and M5 Ultra against the raw data, comparing 20 Apple M-series configurations, eight rival CPUs and eight workload categories in five charts and several tables.
A report claims Apple is considering an enterprise AI server with multiple M8 Ultra chips and NVIDIA's NVLink Fusion. We examine the claim against Apple's current Mac clusters, Private Cloud Compute, and NVIDIA's official materials.
Google has launched early access to Home MCP in the US. Drawing on primary sources, we examine how external AI agents can access home device data and history, and the safety boundaries in place, including a ban on unlocking doors and requirements for additional consent.
Mozilla's partnership with Mistral brings Mistral Small 4 to Smart Window. Here's how the three-model choice, cloud inference, proxy protection, and memory storage actually work.
