Get tomorrow's brief in your inbox
Today: Cursor acknowledged building its Composer 2 coding model on top of China's open source Kimi, raising transparency questions. A compliance startup faces fraud accusations for generating fake evidence. Simon Willison showed how Claude can build custom skills to generate framework-specific code examples.
Cursor, the AI coding company valued at $29.3 billion and reportedly exceeding $2 billion in annual revenue, admitted this week that its new Composer 2 model was built on top of Moonshot AI's open source Kimi 2.5, a Chinese model backed by Alibaba. The company initially promoted Composer 2 as offering "frontier-level coding intelligence" without mentioning the Kimi foundation until an X user discovered code identifying Kimi as the base model.
Cursor's VP of developer education defended the decision, saying only about 25% of the compute came from the Kimi base while the rest came from their own training, producing "very different" benchmark performance. The company claims its use was authorized through a commercial partnership with Fireworks AI and consistent with Kimi's license terms. Read more
What this means for your business: Building on open source foundations is standard practice in AI development, but the initial lack of transparency matters. If you're evaluating AI tools for your business, look beyond marketing claims and dig into the actual technology stack. Ask vendors directly about their model origins, especially if you work in regulated industries where AI supply chain transparency matters. The broader takeaway is that the "built from scratch" narrative in AI is often marketing, not reality.
Claude Skills for Framework Documentation Developer Simon Willison demonstrated a clever workflow using Claude to create custom "skills" that help generate code for newly released frameworks. He had Claude clone the Starlette 1.0 web framework repository and build a comprehensive skill document with code examples, which then enabled Claude to generate a complete task management app using the latest API patterns. The skill system means you can have Claude study any codebase or documentation and create reusable knowledge for future coding tasks. Read more
If you're using AI coding assistants and frustrated by outdated examples, try this approach. Point your AI tool at current documentation and have it create reference materials before asking it to write code. This is free if you're already using Claude, and it dramatically improves code quality for frameworks released after the model's training cutoff.
Cloudflare DNS JSON API Willison also built a simple DNS lookup tool leveraging Cloudflare's CORS-enabled JSON API for its 1.1.1.1 DNS service (plus 1.1.1.2 for malware blocking and 1.1.1.3 for malware and adult content filtering). The tool demonstrates how modern APIs can enable simple diagnostic utilities without backend infrastructure. Read more
If you manage websites or troubleshoot connectivity issues, this kind of browser-based DNS checker is useful for quick diagnostics without installing command-line tools.
Vibe Coding for Non-Technical Founders Non-technical startup founders can now turn "vibe coding" into an execution advantage, according to a piece highlighting how AI tools have leveled the playing field for founders without engineering backgrounds. The article argues that founders who struggled with technical execution at previous companies now have practical options for building prototypes and MVPs using AI coding assistants. Read more
The practical application: If you've been waiting to build a software product because you can't code, the barrier is lower than ever. Tools like Cursor, Claude, and GitHub Copilot let you describe what you want in plain language and iterate quickly. Start with a simple internal tool or customer-facing prototype, not a full production app. Budget time for learning, because you'll still need to understand basic concepts like databases, APIs, and deployment.
AI Productivity Paradox AI productivity tools are raising expectations rather than reducing workload, according to recent analysis. Instead of freeing up time, AI tools are shifting the baseline for what's considered acceptable output quality and quantity, creating pressure to deliver more in the same timeframe. Read more
If you're implementing AI tools in your business, set explicit expectations about what "productivity gains" actually mean. Are you trying to do more work with the same team, improve quality with the same output, or genuinely reduce hours worked? Without that clarity, AI tools just become another source of burnout.
Delve Accused of Compliance Fraud Y Combinator-backed compliance startup Delve, which raised $32 million at a $300 million valuation last year, faces serious accusations of providing customers with "fake compliance" documentation. An anonymous Substack post from a former client alleges the company generated fabricated evidence of board meetings, tests, and processes that never happened, then routed all certifications through two India-based audit firms that rubber-stamped reports without independent review. The accusers claim this approach "inverts" normal compliance structure and creates criminal liability under HIPAA and fines under GDPR. Read more
Delve called the accusations "misleading" and said they contain inaccuracies, but the detailed allegations raise red flags. If you're using automated compliance tools, especially for HIPAA, SOC 2, or GDPR, verify that your auditor is actually independent and that evidence reflects real processes in your business. Compliance automation is valuable, but it should document what you actually do, not generate fictional evidence to pass certification. If a vendor promises drastically faster compliance than competitors, ask why.
Musk Announces Terafab Chip Plant Elon Musk announced plans for a "Terafab" chip manufacturing facility in Austin, Texas, to be jointly run by Tesla and SpaceX. The goal is to produce chips supporting 100 to 200 gigawatts of computing power per year on Earth, plus a terawatt in space, driven by concerns that semiconductor manufacturers can't keep pace with AI and robotics demand. No timeline was provided, and Musk has no background in semiconductor production. Read more Read more
For small businesses, this signals continued chip supply concerns at the high end. If you're planning AI infrastructure investments that depend on cutting-edge GPUs, assume constrained supply and rising prices for at least the next two years. Consider cloud-based inference services instead of buying hardware.
Use AI to Audit Web Performance Willison demonstrated using Claude to analyze the horrifying bloat on a PC Gamer article that triggered 431 network requests and 5.5 MB of transfer (ballooning to 200+ MB with autoplay videos), with 82% traced to ad tech and tracking scripts. The core content was just 10-15 KB of text. Read more
Try this on your own website: Use browser dev tools or an AI-assisted analysis to see how much of your page weight comes from ads, trackers, and third-party scripts versus actual content. If your mobile site takes more than 3 seconds to load on a 4G connection, you're losing customers. Simple fixes like lazy-loading images, removing unused tracking pixels, and switching to faster ad networks can cut load times in half.
The Cursor disclosure and Delve accusations both point to a transparency problem in the AI and automation space. Companies are building on foundations they're not disclosing upfront, or automating processes in ways that create legal risk for customers. For businesses adopting AI tools, the lesson is clear: ask more questions, verify claims, and understand what's actually happening under the hood. The "move fast and automate everything" mindset works until it creates compliance disasters or erodes customer trust.
Watch for more scrutiny on AI model provenance, especially as companies compete on claims of "proprietary" or "from-scratch" development that turn out to be fine-tuned versions of open source models.