Compute & Infra+5Oct 25
Image: airealist.ai
Apple's local model uses about 2GiB in Mac tests
Apple's Core Advanced used about 2GiB of memory, including its safety models, after a long request in Julien Simon's M3 Max Mac tests, versus about 15GiB for a comparable Gemma build. The model has 20 billion parameters (learned values that shape its responses), with 1–4 billion active at a time. It keeps specialized components in flash storage, leaving more memory and graphics-processing capacity for other applications. On 20 math problems, Core Advanced used about one-third the energy of the best publicly downloadable model Simon tested.
Compute & Infra+5Oct 26
Image: datacenterdynamics.com
Greece inaugurates Daedalus supercomputer for science and AI
Daedalus expands Greece's computing capacity for scientific research and AI after EuroHPC and GRNET inaugurated it today. HPE built the system with more than 2,000 NVIDIA accelerators (specialized computing processors). The new 6.3MW data center occupies a former power-station building and uses direct liquid cooling; a 2MW solar installation at the surrounding park helps cover its energy needs. Daedalus sustains 85.69 petaflops and peaks at 115 (a petaflop means one quadrillion calculations per second).
Text+5Oct 19
Image: theverge.com
GPT-6.1 Sol reaches 98% behavior preservation in COBOL tests
GPT-6.1 Sol passed 87% of OpenAI's tests for translating COBOL (an older programming language) into languages including Java, Python and Rust, using a low reasoning-effort setting. A translation counted as successful only when every test of its original behavior passed, OpenAI said on Oct. 19. At the same low setting, GPT-6 Astra scored about 86%, GPT-6 Sol 21% and GPT-5.6 Sol 5%. With higher reasoning effort, GPT-6.1 Sol reached 98%.
Agents+2Oct 23
Image: x.com
Cursor agents gain desktop controls to test their code
Cursor's Computer Use lets coding agents click and type in desktop applications to test features they build. Released on Oct. 23, the feature extends interaction beyond browsers, including testing in the iOS Simulator (a simulated iOS device). Agents can invoke it automatically when needed, or users can request it explicitly.
Agents+2Oct 23
Image: zoom.com
Zoom connects chat agents and meeting notes to app updates
Zoom lets AI agents join Zoom Chat channels, while My Notes can turn meeting action items into updates in tools such as Zendesk. Automated workflows can also trigger follow-up campaigns, Zoom said on Oct. 23. FirstBank deployed Zoom's customer-experience products, including Virtual Agent and AI assistance, across 11 operations teams in 100 days.
Compute & Infra+2Oct 26
Image: blogs.nvidia.com
NVIDIA launches Minerva to investigate AI-cluster slowdowns
NVIDIA launched Minerva today to help operators trace slow AI training jobs to the equipment or software responsible. Minerva maps readings from graphics processors, networks and storage alongside information about how computing jobs are assigned and run. The platform analyzes billions of data points an hour to group related alerts and predict emerging failures, NVIDIA says, and includes an AI assistant for investigations. IREN is using Minerva.
Compute & Infra+2Oct 26
Image: datacenterdynamics.com
O2 reports about 20% better spectrum use in live trial
At O2's London Paddington Giga Site, Virgin Media O2 and Nokia completed a live AI trial that carried more data over the network's existing spectrum and capacity. The companies tested whether predictive AI could automatically adapt to changing customer demand without manual intervention. O2 reported an approximately 20% improvement in spectrum efficiency, meaning the amount of data carried over the same radio spectrum, today.
Science+2Oct 26
Image: The Threshold Report/GPT Image 2.5
Specialist skin model beats general AI in diagnosis test
Dermatologist Magnus Lynch tested 200 biopsy-proven skin lesions, comparing a small specialist model with much larger general-purpose AI models. He used Claude Code for the analysis and first drafts and published the cases, prompts, replies and evaluation code. The specialist Inception v3 model can run entirely in a phone browser without uploading images. It named 125 diagnoses correctly, compared with 84 for GPT-6 Astra and 85 for Claude Fable 5.1.
Video+2Oct 23
Image: The Threshold Report/GPT Image 2.5
Grok releases seven-minute AI-generated Odyssey episode
Grok Imagine published a seven-minute first episode of an Odyssey series made with Wonder Studios and director Dev Walker. Grok had previously promised a full-length film. TechCrunch found some individual shots plausibly realistic while criticizing the dialogue and storytelling.
