X / Twitter
Swyx, AI Builder
Swyx shared how he currently divides work among models on “big projects”: Sol Ultra plans, Fable 5 critiques, Sonnet 5, Terra Ultra, and SWE 1.7 handle ultracode or slop-cannon work, and Devin Review uses Kakuna for review. The value lies not in the model names but in assigning different cognitive roles rather than seeking one strongest model. He almost always uses Matt Pocock’s grill-me or a variant of Thariq’s interview-me to force decisions into the open before execution. This shows workflows shifting from immediately putting agents to work toward using prompt structures to extract requirements, constraints, and judging criteria first. Another complaint about a Resend API key reflects real toolchain friction: repeatedly requesting the same service key quickly becomes infrastructure noise for users.
https://x.com/swyx/status/2076811977918484795
Thariq, Claude Code
Thariq focused on Claude artifacts’ growing expressive capabilities and their combination with local Claude Code sessions. His example was creating a project dashboard in Claude Tag that others and a local Claude Code session can both edit. This advances artifacts from one-off displays into project interfaces for collaboration and ongoing agent maintenance. For builders, the dashboard becomes a shared object representing project status, actions, and agent collaboration rather than simply an AI-generated page. If editing and synchronization remain reliable, Claude Code could expand from command-line tasks toward a team workspace.
https://x.com/trq212/status/2076790799011131735
Amjad Masad, Replit CEO
Amjad Masad demonstrated receiving live progress updates during model training, comparing it to early vibe coding except that the object is now “making a personal model” rather than software. This suggests training is being productized into a more immediate, responsive builder workflow. Traditionally, training has been offline, slow-feedback work associated with research or platform engineering; his emphasis is real-time visibility in personalized model training. The change for developers is not whether training is possible, but whether it can become as observable, iterative, and interruptible as working with a coding agent.
https://x.com/amasad/status/2076776737074184661
Guillermo Rauch, Vercel CEO
Guillermo Rauch said the two most popular features of https://v0.app are ease of use, the filesystem API, and observability, and Vercel will double down on these areas. This suggests v0’s appeal extends beyond generating interfaces to placing results in familiar filesystems and observable workflows. In another post, he connected feature flags with agents that can configure and tune experiments, making them building blocks for autonomous, self-optimizing websites and applications. The direction is clear: agents participate in runtime experimentation and optimization loops, not only code generation. He also cited open-weight models accounting for 29% of gateway tokens, up substantially from 11% in April. Overall, he is focused on the full chain from generation and deployment through observability to self-optimization.
https://x.com/rauchg/status/2076817174073880957
Aaron Levie, Box CEO
Aaron Levie offered a long-term view of the AI stack: frontier intelligence will keep advancing the industry, with leading labs buying the most data, building the most compute, and remaining at the forefront of training breakthroughs. Meanwhile, open weights will quickly absorb those breakthroughs, offer cheaper intelligence, and permit workflow- and domain-specific post-training. The applied AI layer can combine frontier intelligence, open models, and inexpensive closed models to orchestrate industry workflows and optimize performance against cost. Enterprises should focus not on everyone training a model but on continually connecting business context, data, and workflows to AI systems. He also expects frontier models acting as managers and cheaper models handling workhorse tasks to become the routing template. On enterprise model training, he cautioned that the most valuable information is often sensitive and constantly changing, so it cannot simply be put into a model; permissions and security layers especially must not live inside models or agents. His conclusion is that AI is not a zero-sum stack: frontier models, open weights, applied AI, and enterprise context all offer substantial opportunities.
https://x.com/levie/status/2076882332821373381
Ryo Lu, Cursor Designer
Ryo Lu used Cursor to build custom e-reader firmware, bringing AI coding into hardware hacking. It targets beautiful Latin and CJK typography, supporting vertical text, prohibited-line-break handling, large character sets, book and reading-progress synchronization with ryOS, and faster rendering and caching. The substantive point is hands-on work on real reading details—typography, character sets, caching, and synchronization—rather than a web demo. Such projects show AI coding’s value to designers extending beyond faster frontend development into firmware, hardware experiences, and personal operating-system experiments. He also said Jenny would join and lead the team, restoring his focus and mental space for bigger ideas. For builders, this is a familiar signal: when leadership capacity is added, a founder-like designer can return to longer-horizon product imagination.
https://x.com/ryolu_/status/2076713331113734641
Zara Zhang, Builder
Zara Zhang proposed three levels of organizational AI adoption and placed most companies at level 2. The material does not define the levels, but the judgment suggests most organizations may be beyond initial trials without yet reaching deeper organizational redesign. She also shared a 45-minute outdoor conversation with Ashe from March, covering building in public, growing an X audience without creating slop, and her views on vibe coding. Notably, she places audience building and AI-builder practice in one framework rather than treating content growth as separate marketing work. For independent builders, this can be more useful than advice to post more, because it addresses the tension between real output and public storytelling.
https://x.com/zarazhangrui/status/2076862290985730481
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari released the Ramp-Autofill skill, built with the tryramp CLI and Claude Fable, and open-sourced its repository. It addresses specific reimbursement problems: finding receipts in iMessage and Gmail, using Playwright to convert linked webpages into PDFs, and attaching them as receipts. It also fills in meeting participants from Google Calendar events, learns how users previously wrote memos and categories, and automatically categorizes transactions with missing classifications. It verifies its work, flags discrepancies, and can run as a scheduled job so users need not repeatedly handle expenses. He added that one version was generated by Fable in one shot from a voice prompt while traveling by car to SF, followed by steering and edits. This exemplifies productizing an agent skill: connecting CLI, email, messages, calendars, browsers, and enterprise expenses into reusable automation rather than showcasing technical tricks.
https://x.com/nikunj/status/2076775924650107151
Peter Steinberger, OpenClaw and OpenAI
Peter Steinberger said he moved his maintainer agents to the cloud, where they had begun “fighting.” Though joking, this points to a real issue: when multiple maintenance agents run continuously in the cloud, coordination, conflict resolution, and state consistency become more important than in isolated tasks. Another post called “stress test” a good prompt, consistent with agent-engineering practice. Compared with merely requesting task completion, it pushes models to seek boundary conditions, failure paths, and hidden assumptions. Together, the posts concern maintaining quality through stronger prompts and testing habits while moving from local single-agent workflows to cloud-based multi-agent maintenance.
https://x.com/steipete/status/2076923300593422560
Official Blogs
Building intelligent apps for Apple platforms with Claude in the Foundation Models framework
Anthropic released a Claude Swift package for Apple’s Foundation Models framework, allowing Apple developers to hand Claude more complex requests in Swift. The framework suits fast local tasks such as summarization and extraction, returning typed Swift values through guided generation and @Generable annotations. Claude lets one application first use Apple’s on-device models for lightweight work, then hand off when multistep reasoning, code generation, web search, or data analysis is needed. The blog emphasizes that typed outputs produce cleaner Claude API inputs rather than sending raw user text directly to a cloud model. Examples include journaling apps generating daily prompts locally before Claude finds themes across months of entries, and study apps explaining terms locally before Claude answers why they relate to other material. Access uses an Anthropic API key, with streaming, tool calls, and structured responses integrated into the SwiftUI view. Supported platforms include iOS 27, iPadOS 27, macOS 27, visionOS 27, and watch OS 27.
https://claude.com/blog/claude-for-foundation-models
Podcasts
The MAD Podcast with Matt Turck — Inside Nemotron & NVIDIA’s AI Lab | Bryan Catanzaro
Key takeaway: Further AI progress requires not just more compute but greater efficiency to obtain more intelligence as compute, power, and cost approach their limits.
Bryan Catanzaro leads NVIDIA’s Nemotron open foundation models. His background spans early GPU-for-AI research, NVIDIA research, and work with Andrew Ng, Dario Amodei, and others at Baidu’s Silicon Valley AI Lab. His perspective has value on two levels: he views training through NVIDIA’s accelerated-computing lens and has firsthand experience of differences and mutual learning between US and Chinese AI research communities.
His first judgment on open-source AI is that open technology matters not because it must win a contest against closed source, but because AI, like the internet, will be applied very differently across industries. Retail, healthcare, and manufacturing use the internet differently, and AI will be similar. Open technologies provide room to customize models deeply integrated with enterprise data, customer understanding, guardrails, and workflows. He does not frame open and closed as simple opponents, emphasizing instead the entire community’s rapid progress over the past three months: specific gaps matter less than the overall speed of the field.
His position on Chinese models and distillation controversies is clear. He acknowledges mutual learning but rejects attributing other countries’ progress to a copycat mentality. Having worked at Baidu for two and a half years, he saw Chinese colleagues’ intelligence, diligence, and creativity firsthand and considers reducing their achievements to copying inaccurate. He further believes China has in some respects taken the lead in open-AI community collaboration, and other regions need to understand the benefits of building open technologies together.
He described Nemotron’s three tiers: Nano with 30B total and 3B active parameters, Super with 120B total and 12B active, and Ultra with 550B total and 55B active, addressing small, medium, and large deployments. Super is popular for balancing cost and intelligence. NVIDIA’s foremost LLM interest is agents solving problems for users day and night, and Nemotron aims to be fast and capable enough for agentic workflows.
Efficiency is his central technical theme. Nemotron Ultra and Super use MVFP4 for four-bit pretraining, requiring algorithms that converge to strong results despite coarse arithmetic rather than simply compressing models. Four bits provide only 16 values, but enable higher throughput on Blackwell Ultra and substantial energy savings. He put it sharply: “If you accept that we will operate at the limits, the way to get more intelligence is to become more efficient.” This also informs his description of AI as an “external brain”: people have always built external organs to solve problems—a kitchen is an external stomach—and AI may be an external brain whose impact no one yet understands.
On safety, he offered a somewhat counterintuitive framing: open technologies are generally safer because they receive more sunlight. More people can evaluate, challenge, and contribute improvements, which is more robust than a few people deciding everyone’s boundaries. He also sees AI as fundamentally exploring ideas, making diversity safer than monoculture. Real diversity is not effortless agreement but allowing deep disagreements to exist and be tested. For builders, future competition concerns combining open models, enterprise secrets, agent workflows, efficiency engineering, and safety evaluations into sustainable systems, not just building the largest model.