AI Builders
Digest

Swyx from Cognition Labs highlights METR_Evals' FrontierCode benchmark, showing over half of SWE-Bench results are unmergeable slop. FrontierCode features 3000+ rubrics for maintainable code quality, with Opus 4.8 scoring just 13.8% on Diamond tier. It marks the shift to 'Maintainable Code' era after autocomplete (HumanEval) and test-passing (SWE-Bench). Rapid progress in late 2025 shows models saturating easier tasks, enabling advanced agentic coding loops. The benchmark will evolve annually as AI capabilities grow.

OpenAI CEO Sam Altman posted the company's current plan, outlining their strategic direction amid rapid AI advancements. This comes as frontier models continue evolving and new coding capabilities emerge across the industry.

The Takeaway: As autonomous AI agents proliferate in enterprises, specialized oversight agents trained on historical behavior are essential to mitigate exponentially growing risks without stifling adoption. Onyx Security CEO Maxim Bar Kogan, co-founder of the Israel-based startup, explains how his team builds AI control planes using small specialized models for efficient monitoring that flag issues for deeper review, inspired by AutoGPT's vision of capable agents. Enterprises categorize agents into low-code automations, first-party builds, and dominant autonomous coding tools like Claude Code, with the latter over 50% and growing fastest, often lacking controls. Existing security tools fall short due to agents' need for broad permissions and unpredictable actions. 'We're allowed to look at a lot of historical data of how these agents have behaved, but enterprises today are not willing to have Anthropic or OpenAI give that historical data.' Long-term, Onyx aims beyond enterprise security to help control advanced AI through mechanistic interpretability and independent oversight, leveraging Israel's cyber-AI talent.

Google VP Josh Woodward announces NotebookLM's killer feature: expanding search beyond your own source files. The latest update supports new output formats including PDFs, DOCX, XLSX, PPTX, charts, and more to enhance research capabilities.

Claude Blog: Building intelligent apps for Apple platforms with Claude in the Foundation Models framework. Apple developers can now use a new Swift package to hand off complex multi-step reasoning, code generation, and web search from on-device models to Claude, receiving typed Swift values and streaming responses. This enables seamless hybrid experiences in journaling, study, and document apps. Available on iOS 27, iPadOS 27, macOS 27 and others. https://claude.com/blog/claude-for-foundation-models

Anthropic's Boris Cherny discusses Claude Code evolution: auto mode usage, routines fixing bugs proactively, and coding from phone. OpenAI's Thibault Sottiaux showcases nested loops in Codex and advanced controller interfaces. Peter Yang notes addiction to mobile Codex and questions on Google's competitor.

Box CEO Aaron Levie emphasizes that no amount of model intelligence replaces context. For general-purpose AI, instructions, domain knowledge, and proprietary data in the context window remain essential for differentiated value, explaining varied AI gains across users and why applied AI layers add significant value.

Replit CEO Amjad Masad highlights making games for Tesla on your Tesla. Vercel CEO Guillermo Rauch notes DeepSeek entering the chat amid competitive AI model developments.

Amanda Askell on future Claude versions. Zara Zhang praises Markdown/HTML/SVG and resonates with 'home cook' programmer analogy. Nikunj Kothari on autonomous companies and VC theses. Dan Shipper reacts to notable content. Claude AI announces Tokyo event.