bn
Good morning. It’s Monday, March 30th.
An AI Tool that I’ve been obsessing over lately is Google Stitch. It’s an AI design tool for applications. You just describe the vibe, look, and functionality you want for an app, dashboard, or website, and it spits out beautiful, unique, editable UI screens in seconds. Then you can export them and drop straight into Claude to build the whole thing. It’s free right now on Google Labs!
Google Stitch + Claude Code is an absolute app factory. Build something with it, send it to me, and I’ll share it in the newsletter!
-Jeff
AI Breakfast
P.S. - Our sponsor today, Peec.ai, shows you how to get your website discoverable on LLMs. Think SEO for AI model output. Googling something is a thing of the past. This is the future! Check them out below.
You read. We listen. Let us know what you think by replying to this email.
AI search is the fastest-growing discovery channel.
Your customers ask ChatGPT, Perplexity, Claude, and Gemini for recommendations daily. Is your brand the answer?
Peec AI shows you exactly where you stand:
Track your visibility and sentiment across all major LLMs
Benchmark against competitors to see your share of voice
Get step-by-step guidance on improving your AI visibility
Turn AI search from a black box into a measurable growth channel with clear metrics and a scalable strategy.
Google’s ‘agent smith’ AI explodes in popularity but access gets restricted as employees rush to use it
Agent Smith, Google’s internal AI built on Antigravity, handles autonomous coding, workflow execution, and document access via mobile, with demand forcing restricted use. Sundar Pichai and Sergey Brin are tying adoption to performance reviews, while Project EAT helps standardize AI use across teams.
NotebookLM now supports off-page generation with mobile notifications, letting users pick up work asynchronously. Gemini for Business will integrate NotebookLM notebooks directly, with agents summarizing research, plus a new Skills framework, including pre-made tools and a “Skill Architect” for custom AI assistants, and an AI Coding hub unifying Antigravity and Gemini CLI access. Read more.
Anthropic confirms ‘Mythos’ leak, slams OpenAI’s “tobacco industry” AI approach
Anthropic confirmed testing Claude Mythos (aka Capybara), calling it a major “step change” in reasoning, coding, and cybersecurity. The company further positioned itself as an ethical alternative to OpenAI criticizing the latter’s massive general-purpose models as a “tobacco industry” approach. A CMS misconfiguration exposed nearly 3,000 internal documents, including drafts and an invite-only European CEO summit, though customer data and core systems remained safe.
Claude now enforces session limits faster during peak hours. Former DeepMind researcher Nicolas Carlini said Claude outperformed him at finding zero-day vulnerabilities. Anthropic is also trialing Claude Operon, a biology-focused desktop mode with multi-session workflows and local file access powered by Mythos. Read more.
Meta tests self-improving hyperagents and Avocado AI
Meta’s hyperagents, built on the Darwin Gödel Machine, combine task agents with editable meta-agents that rewrite themselves, generate variants, track performance, and build persistent memory. The DGM-H system shows strong gains across coding, paper review, robotics, and Olympiad math, with improvement strategies transferring across domains. Experiments remain sandboxed with oversight.
Meta is also testing Avocado variants including 9B, Mango (multimodal agent), TOMM, Thinking 5.6, and Paricado. Some requests are routed through Google’s Gemini to fill capability gaps. Avocado is expected to be proprietary, a shift away from open-source Llama models. Read more.
MetaClaw lets AI agents learn during downtime
MetaClaw is designed to upgrade AI agents while they’re in use. When an agent fails, a separate model extracts a compact rule from the error and injects it into the system prompt, improving future behavior instantly.
In parallel, reinforcement learning with LoRA fine-tuning updates model weights during idle windows detected via calendar, inactivity, and sleep signals. Tests show large gains, with weaker models nearly matching stronger ones, though results are based on simulations rather than real-world use. Read more.
Agentation converts UI annotations into structured context AI coding agents can parse navigate and act on precisely.
SUN generates on demand AI audio courses with interactive playback for continuous hands free learning in real time.
Requestly uses AI to manage cookie consent for functionality analytics user experience and advertising with configurable user preferences.
Crossnode turns AI agents into paid products with built in billing access control and no backend setup required.
CrabTalk is an 8 MB open source daemon that streams every AI agent event in real time with zero hidden layers.
Thank you for reading today’s edition.
Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.
Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!
Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.
