X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault Sottiaux expects 2026 to be the year enterprises truly prioritize model efficiency and reliability. The reason is that models are moving from assistive tools to critical enterprise infrastructure. Once embedded in critical business processes, their ability to work consistently becomes as important as their capability ceiling. Efficiency also directly affects cost and response speed at deployment scale. When choosing models, enterprises will look beyond individual demonstrations to long-term operational performance.
https://x.com/thsottiaux/status/2091581575108653374
Peter Yang, Practical AI Tutorial Creator
Peter Yang divided AI evals into top-down and bottom-up categories. Top-down evals start from a task description and consider which capabilities should be checked without existing examples; Claude is well suited to helping generate these evaluations. Bottom-up evals come from observing large volumes of actual output and require converting human intuitive feedback into explicit evaluation criteria. Citing Shreya, he noted that Claude is not good at independently discovering these issues grounded in real examples, so this work remains highly dependent on human judgment. Peter also recommended Shreya and Hamel's AI evals course, which has a 4.7 rating and 900 reviews on Maven, with the next cohort starting September 5. He additionally shared his experience of working for six months with his human assistant Char, who uses Claude Code and Codex for podcast postproduction, show notes, and short-video editing. Char also copies and continually adapts Peter's AI skills to fit Char's own workflows, illustrating that an AI-native assistant needs not only to operate tools but also to turn reusable processes into a personal working system.
https://x.com/petergyang/status/2091586298779955512
https://x.com/petergyang/status/2091529714104168870
https://x.com/petergyang/status/2091631590799737306
Madhu Guru, Senior Director at Meta AI
Madhu Guru proposed a Goldilocks principle for building evals: their granularity should be neither too coarse nor too fine, but aligned with distinct jobs to be done. For a financial analysis agent, checking only which stock it ultimately recommends obscures quality issues across several critical preceding stages. The full process includes understanding the customer's portfolio, risk tolerance, investment horizon, goals, and constraints, then gathering evidence about stocks, industries, the macroeconomic environment, Fed policy, and news events. It must then analyze revenue growth, valuation guidance, and growth projections, narrow the candidate stocks, and finally provide tickers, buy and sell prices, and time frames. Each stage produces intermediate outputs, making it possible to establish separate evals and arrive at a diagnosis such as 92% for customer understanding, 92% for evidence extraction, 70% for data analysis, and 75% for the final recommendation. These layered results quickly identify data analysis as the concentration of problems, rather than merely labeling the final answer wrong. If a stage remains too complex, it can be broken down further, but the degree of decomposition should depend on whether it enables diagnosis and action. The core principle is to make evals just granular enough to locate problems and guide improvement.
https://x.com/realmadhuguru/status/2091684812012875981
Guillermo Rauch, Vercel CEO
Guillermo Rauch observed that following OpenAI Sol's price reduction and discounts from Vercel AI Gateway, Sol had become the Gateway's fastest-growing frontier model. He inferred that demand for intelligence is highly elastic, with lower inference costs quickly driving usage growth. Price fluctuations also give gateways clearer commercial value: they can help applications take timely advantage of price changes, lower operating costs, and improve margins. Rauch therefore sees the growing interest in routers as no coincidence and believes gateways will ultimately become inevitable infrastructure. He also described his philosophy for extending fx, emphasizing open protocols such as MCP, Skills, and Plugins. The deeper reference remains Unix: small programs should each do one thing well, then compose functionality by invoking other programs. Combined with libfx, which can be embedded in complex programs, developers can build their own CLIs, background agents, or software factories and choose to run them locally or in the cloud.
https://x.com/rauchg/status/2091671326897713424
https://x.com/rauchg/status/2091583525661384813
Garry Tan, Y Combinator CEO
Garry Tan predicted that systems of record must gradually become AI harnesses or risk being replaced by agents. Traditional systems of record derive their core value from storing and organizing business data, but agents place more active execution demands on software. In his judgment, merely serving as a static data source will no longer be enough to sustain a product's position. Such systems need to provide an environment suited to AI invocation, control, and task execution. The future competitive focus may therefore move from who stores the records to who enables agents to use those records safely and effectively to get work done.
https://x.com/garrytan/status/2091742825042030681
Official Blog
Anthropic released a new Swift package allowing Apple developers to call Claude through the Foundation Models framework for complex workflows such as multi-step reasoning and code generation. Apple's framework lets developers use models directly from Swift and return typed Swift values with a small amount of code through guided generation, making it suitable for quick, local tasks such as summarization and information extraction. When a request exceeds what an on-device model is suited to handle, the application can pass structured input to Claude and stream results back into the same SwiftUI view. Claude can also search the web for current information and execute code for data analysis. Official examples included a journaling app generating daily prompts on-device before asking Claude to find connections across months of entries, and a learning app explaining terms locally before handing follow-up questions about broader conceptual connections to Claude. The package handles streaming, tool calls, and structured responses, letting developers choose different models for different steps within one user experience. The company said support would become available the following day for iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27, and integration requires an Anthropic API key.