X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault Sottiaux described this week’s theme as “intelligence too cheap to meter” and previewed another release from the team tomorrow. GPT-5.6 Sol achieved SoTA on ARC-AGI-3, but the key was not retraining the model: it was changing two settings. The test allowed the model to reason fully and use OpenAI’s standard compaction to continue working across multiple context windows. This shows that a model’s practical capabilities depend not only on its weights but also on how it runs and manages context. He also reminded developers to pay attention to agent harnesses, because the execution frameworks supporting models directly affect their performance. Together with an investigative article he recommended, this suggests that building and hardening harnesses is becoming an important engineering challenge in unlocking model capabilities.
https://x.com/thsottiaux/status/2082655731204096275
https://x.com/thsottiaux/status/2082637967852806207
https://x.com/thsottiaux/status/2082609662231502932
Peter Yang, Practical AI Tutorial Creator
After enjoying AI’s productivity benefits, Peter Yang has become wary of three negative usage patterns. The first is not bothering to read original material and relying solely on AI summaries, or letting agents make large-scale file changes without carefully checking the results. The second is giving agents feedback from a phone at any moment and mistaking constant availability for productivity, even when he should be spending time with his children or attending to real life. The third is preferring to discuss ideas with ChatGPT Voice rather than with the real people around him. In design work, he particularly values Claude Design proactively asking clarifying questions before creating anything, because these force users to think through their requirements first. To avoid a generic “Claude style,” he creates a design.md in advance that specifies colors, fonts, and other visual standards. He recommends collecting favorite design patterns and application interfaces through Mobbin MCP, or consulting the ready-made specifications library at designmd.sh. His core method is to establish a design language through references and explicit constraints before asking AI to execute specific prototyping tasks.
https://x.com/petergyang/status/2082642205811106158
https://x.com/petergyang/status/2082579428090192192
https://x.com/petergyang/status/2082519030859264086
Google Labs, Google’s AI Experiment Platform
Google DeepMind released Lyria 3.5 and integrated it directly into Google Flow Music. The new version improves prompt adherence, giving creators more precise control over musical direction. It supports setting an exact BPM and exporting stems for a complete song to facilitate further production. Vocals emphasize richer, more expressive, and more dynamic performances. Arrangement transitions have also improved, making the flow between notes and sections more natural. The focus of this upgrade is to move generated music further beyond a one-off finished product toward precisely controllable production material that can be edited further.
https://x.com/GoogleLabs/status/2082501360466174163
Amjad Masad, Replit CEO
Amjad Masad said different models excel at different visual tasks: some at CSS, some at SVG, and others at animation. Replit Design does not entrust every task to a single model, instead using the best-suited open and closed models together. The aim is to achieve better aesthetic quality through a combination of models, rather than merely produce functioning pages. Masad called it the best AI design tool currently available in terms of taste, playfulness, and ease of use. His proposed direction suggests that design products are entering a “post-prompt era,” with competition shifting from how users write prompts to how systems orchestrate models behind the scenes.
https://x.com/amasad/status/2082508826767679668
https://x.com/amasad/status/2082505558293467363
https://x.com/amasad/status/2082504898801999990
Aaron Levie, Box CEO
Aaron Levie argued that as AI becomes more capable, inference resources will flow first to tasks with the highest economic value, potentially crowding out lower-value demand. In theory, scarcity could push inference costs up as high-value demand grows, but he does not expect prices simply to spiral along that path. Many model providers and infrastructure companies want to compete for these workloads, so competition will keep pushing prices down until supply catches up with demand. He also noted that the OpenAI agent sandbox escape directly affects enterprise AI adoption because it demonstrates both agent capabilities and the vulnerability of existing environments. Enterprises need to limit the data agents can access, preserve audit trails and governance capabilities, and reassess which systems must remain deterministic. Outdated access controls could lead a well-intentioned agent tasked with “finding the project’s most important intellectual property” to uncover information that should never have been exposed; malicious actors pose an even greater risk. Unlike humans, agents can work indefinitely, face no real consequences for exceeding their authority, and lack mature judgment about whether a task is reasonable. Enterprises therefore need new data protection, rapid blocking, and agent defense systems. This will create security startup opportunities while extending the time required for highly autonomous agents to become widespread in enterprises.
https://x.com/levie/status/2082658870523248967
https://x.com/levie/status/2082514776392175844
Ryo Lu, Cursor Designer
Ryo Lu announced that Cursor is now available on iOS. The core use case presented is letting users access their agents from anywhere. Mobile availability means agent workflows are no longer entirely tied to desktop development environments. The launch post specified only this positioning of “access your agents anytime, anywhere.” Its emphasis was expanded accessibility; the original post did not elaborate on feature scope, pricing, or platform restrictions.
https://x.com/ryolu_/status/2082539893729972320
Zara Zhang, Builder
Zara Zhang argued that deep domain expertise combined with AI-native ways of working creates formidable individual competitiveness. Past experience alone is not enough: professionals must keep reinventing themselves and adopting new working methods. She also emphasized that marketing skills influence not just promotion but product quality itself. Many technical people address an imagined audience without encountering the reality of how users understand and use their products. Marketing forces builders to observe actual perceptions, language, and behavior, making it a product feedback mechanism too. Her two observations point to the same principle: professional expertise must evolve alongside new tools and direct contact with the market.
https://x.com/zarazhangrui/status/2082705944782520462
https://x.com/zarazhangrui/status/2082684904136134881
Dan Shipper, Every CEO
Dan Shipper said almost everyone at Every rated ChatGPT for Work’s voice mode highly, and the team had not seen such shared excitement in a long time. He also noted the defensive needs exposed by agent security incidents. In the case discussed, OpenAI’s safety classifier was disabled and the model was explicitly instructed to exploit vulnerabilities, so the outcome cannot simply be attributed to normal usage conditions. Meanwhile, HG’s AI did detect the attack automatically, but did not flag it at a sufficiently high severity. He expects malicious actors to deliberately use models for similar actions in the future. Any company handling sensitive customer data will need automated, agent-based defenses. For founders, security products that continuously detect, assess, and block other agents’ behavior could represent an enormous opportunity.
https://x.com/danshipper/status/2082613916706693560
https://x.com/danshipper/status/2082608994275725650
Aditya Agarwal, General Partner at South Park Commons
Aditya Agarwal compared AI researchers’ situation to “a 19-year-old LeBron with only two years left to make money.” He believes some researchers feel robots are about to take over their work, making them like MVPs with only 18 months left to attack; this pressure can explain much of the controversy in the industry. At South Park Commons’ Demo Faire, he observed strong investor interest in frontier technologies such as robotics, drones, and semiconductors. Investors cared not only about financial curves but also wanted to discuss in depth what founders were actually building. Meanwhile, the bar for excellent software has risen substantially. If a product is merely something like vertical SaaS or an agent-for-X, it must demonstrate a genuinely different vision or exceptionally rapid growth. He believes the real joy of early-stage investing should be helping build the future, rather than simply finding charts that keep going up. Over the long term, Silicon Valley remains especially good at mobilizing its entire ecosystem behind projects that take decades to mature.
https://x.com/adityaag/status/2082558632705896899
https://x.com/adityaag/status/2082538703432630398
Sam Altman, AI Observer
Sam Altman said models that can significantly accelerate scientific discovery are very close. He believes the best way to achieve this is not for AI companies to solve every scientific problem themselves, but to put the capability in scientists’ hands. The emphasis is on equipping professional researchers with more powerful research tools. Models serve as accelerators, while scientists remain the ones posing questions and advancing discoveries. He also stressed that the benefits of scientific progress should reach everyone.
https://x.com/sama/status/2082628413769003269
Official Blogs
Building intelligent apps for Apple platforms with Claude in the Foundation Models framework
Anthropic released a new Swift package that lets Apple developers call Claude through the Foundation Models framework. Apple’s framework is suited to fast, local summarization and information extraction using on-device models, and guided generation can return typed Swift values with little code. When a task requires multistep reasoning, code generation, web search, or data analysis, the app can hand the prepared input to Claude. Claude’s streaming responses, tool calls, and structured results return to the same SwiftUI view, with the package handling the integration. This lets apps choose the right model for different steps—for example, generating a journal prompt on device, then having Claude look for patterns across months of entries. A learning app could likewise explain terminology locally, then hand off to Claude when a question requires connections across the entire course. The article was published on June 8 and said support would become available the following day for iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27, with developers required to sign in using an Anthropic API key.
https://claude.com/blog/claude-for-foundation-models
Podcasts
AI & I by Every — Best of the Pod: Wired's Kevin Kelly on Why AI Is a 50-year Overnight Success
Key takeaway: Technological breakthroughs often look like “overnight successes,” but what actually determines their timelines is often not imagination, but overlooked physical constraints, infrastructure, and unexpected uses.
Kevin Kelly has long observed the internet, AI, VR, and other technological waves, while candidly recording his own mistaken predictions. Seeing Jared Lanier demonstrate VR in 1987 deeply impressed him, and he was convinced it would quickly become the future. Decades later, VR’s core experience has not improved a millionfold; the most obvious change is that the equipment’s price has fallen from millions of dollars to around $100. His conclusion: “Predicting the future is easy. Making correct predictions is hard.”
His first counterintuitive judgment is that technology spreads much more slowly in the physical world than in pure software. VR faces not only hardware challenges but biological limits involving eye focus and the weight a head can bear. Robotics faces an even larger gap: the human brain is a roughly 25-watt “supercomputer,” while the body’s movement system has only about a quarter of a horsepower. Current machines struggle to approach that energy efficiency. He therefore expects widespread robotics adoption to take far longer than many people anticipate.
His second judgment is that major breakthroughs may come from outside the original objective. LLMs did not begin by directly tackling reasoning; researchers unexpectedly observed reasoning in language translation capabilities. The “LLM moment” that VR or AR awaits may likewise come from lenses, focusing, projection, or another adjacent technology, rather than linear improvements along the current path. There is no sign yet that this crucial breakthrough has occurred.
His third judgment is that humanity’s understanding of intelligence remains at a very early stage. Kelly compares the present to the period when early scientists studied electricity, when even Isaac Newton proposed incorrect theories. Early Royal Academy activities included ticketed demonstrations of electrical sparks, and people once thought electricity was primarily a biological phenomenon. He suspects intelligence is not a single element but a compound of multiple cognitive components that have yet to be identified. Today’s AI is like being able to manufacture “salt” without yet knowing which elements it contains.
His fourth judgment concerns innovation frontiers and governance. Both the early internet and Burning Man once existed in chaotic, free, permissionless states, but as risks emerged, roads, rules, policing, and organizational structures were gradually established. Kelly does not believe order only undermines innovation, because mature governance typically also expands participation. A better state would involve continually opening new frontiers, letting those suited to rule-free exploration keep experimenting while people skilled at building within rules turn the results into stable systems.
His own approach to balance is to venture briefly into the technological frontier, then return to history and the physical world. While studying the latest AI, he also reads the history of electricity’s discovery and makes things by hand in his workshop. Understanding the future requires more than chasing the next release: it also means looking at the past ten thousand years, the next ten thousand years, and the human conditions that remain unchanged over time.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL