Figma expanded a personal Claude Code setup into PM OS, a shared repository with 31 skills used by more than 60 product managers. U.S. Census Bureau data show AI adoption among large businesses rising from about 30% in February to about 40% in October, but the figures cannot be directly compared with the roughly 12% reported in September 2025 under the old definition. Blacksmith raised $66 million at a $1 billion valuation, and its customer LangChain cut its wait for CI compute from 15–20 minutes to 10–15 seconds. Particle-1 answered 91.4% of 500 memory questions correctly, at a list-price cost per correct answer of roughly 1/23 that of Opus 5.5, but this result comes only from an exploratory test with one attempt per question and a single AI judge.
Figma offers a workflow-level example, while the Census Bureau provides a broader measure of adoption.
Figma growth product manager Ezra Mechaber and his colleagues expanded a personal Claude Code setup into PM OS, a shared repository containing 31 skills for the company's more than 60 product managers. When he joined, almost no one outside engineering used AI to get actual work done; now every product manager at the company uses the repository. He runs one skill for organizing Slack six times a day.
Read original →Data released by the U.S. Census Bureau on October 22, 2026, show that AI adoption among businesses with 250 or more employees rose from about 30% in February to about 40% in October. In November 2025, the survey broadened its definition of AI use from “producing goods or services” to any business function. The old and new measures therefore cannot be directly compared: the roughly 12% adoption rate in September 2025 and the current figure do not represent growth measured on the same basis.
Read original →🔍 Analysis: The Census Bureau's roughly 40% figure measures whether businesses use AI; Figma's 31 skills bring the focus down to specific, reusable tasks. These are not interchangeable measures of deployment. With the survey expanded to cover any business function, adoption rates are even less suitable as a direct measure of changes to production processes. For businesses, the range of shared skills in use and how often they are invoked offer a better test of whether AI has entered daily work than an “adopted” label.
Blacksmith cuts CI wait times, while OpenAI brings grid reliability and consumer electricity costs into its campus plans.
Blacksmith raised a $66 million Series C led by Sequoia at a $1 billion valuation. The company runs GitHub Actions on gaming CPUs to provide faster compute for continuous integration (CI). Its customer LangChain cut its wait for CI compute from 15–20 minutes to 10–15 seconds.
Read original →OpenAI launched the Grid Protectors Program for future compute campuses. The program will work with utilities so that these campuses help maintain grid reliability. Alongside supporting grid operations, protecting consumers from electricity cost increases is also a goal of the program.
Read original →🔍 Analysis: Blacksmith and OpenAI address different constraints on compute delivery: the former focuses on whether tasks can get computing resources promptly, while the latter focuses on coordinating campus electricity use with the public grid. LangChain's CI wait times are not a measure of faster model inference, and the Grid Protectors Program is not evidence of power-supply benefits already delivered. Businesses evaluating compute need to measure task wait times, computation, and power availability separately.
Nebius's customer mix and the Particle-1 test both highlight model-selection questions that rankings of unmodified models do not address.
Nebius CTO Danila Shtan said 95% of customers run modified models rather than unmodified ones. He believes inference for open models remains far from fully optimized. He also noted that prefill-decode disaggregation is still discussed more than it is used in practice.
Read original →Particle-1 answered 91.4% of 500 LongMemEval memory questions correctly, 3.4 percentage points below Opus 5.5; all models used the same retrieval results and prompts. At list prices, its cost per correct answer was $0.0043, roughly 1/23 that of Opus 5.5. This was only an exploratory test with one attempt per question and a single AI judge, not a measure of retrieval itself or agent capabilities.
Read original →🔍 Analysis: Nebius's statement that 95% of its customers use modified models means comparisons of unmodified models cannot directly substitute for those customers' deployment evaluations. The Particle-1 and Opus 5.5 test, meanwhile, held retrieval results and prompts constant to isolate cost differences in a specific question-answering step. For developers, the roughly 1-to-23 cost ratio is a clue for model selection in memory question answering, not a result that can be directly extrapolated into a cost advantage for customized models or complete agent systems.
Jepsen raises questions about correctness, documentation accuracy, and how maintainers demonstrate quality.
In his analysis of ParticleDB 1.1.4, Jepsen's Kyle Kingsbury argues that LLMs are eroding trust in software, documentation, and development organizations. He notes that signals once used to judge system maturity are now easier to fake, potentially making engineers broadly skeptical of new projects. His question for maintainers is: How do they ensure software correctness and documentation accuracy, and demonstrate those qualities to users?
Read original →In about two hours, Karpathy turned historical Hacker News discussions into HackerTube, a gateway to video recommendations.
Karpathy used Claude Opus 5.5 to build HackerTube in about two hours, identifying 6,454 repeatedly recommended videos from 43 million comments spanning two decades of Hacker News. Dissatisfied with YouTube's recommendations, he turned to the community's historical discussions to find videos. Karpathy believes this delivered a noticeably better viewing experience.
Read original →