Good morning. It’s Wednesday, March 25th.
On this day in tech history: Microsoft Excel was released 41 years ago today!
I built a little Mac tool with Claude called WhisperHotkey (Github link) that lets you dictate text anywhere on your computer. You hold down a key, speak, and your words appear wherever your cursor is. It’s powered by OpenAI's Whisper running entirely on your machine, so nothing ever leaves your laptop. I find it to be better than Siri, doesn’t require an internet connection, and it’s free.
I open sourced it, so check it out!
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
This guide distills 10 AI strategies from industry leaders that are transforming marketing.
Learn how HubSpot's engineering team achieved 15-20% productivity gains with AI
Learn how AI-driven emails achieved 94% higher conversion rates
Discover 7 ways to enhance your marketing strategy with AI.
Claude can now control your Mac and run tasks remotely
Anthropic is pushing Claude deeper into real-world task execution with a series of tightly linked updates. Claude can now operate a Mac directly, by clicking, typing, and navigating apps, falling back from APIs to full screen control when needed, and pairing with Dispatch to run tasks remotely. Reliability remains limited (~50%), with unresolved risks such as prompt injection and weak audit trails.
In Claude Code, a new auto mode introduces classifier-based permissioning, allowing safe shell commands to run autonomously while blocking destructive actions. Alongside this, a planner–generator–evaluator harness improves long-running coding through structured iteration, QA, and context handling.
Anthropic also launched a science blog to formalize AI-driven research workflows, as development velocity surges with 74 product updates in 52 days. Read more.
OpenAI shuts down Sora, goes all-in on 5GW fusion power by 2030
OpenAI is in advanced talks with Sam Altman-backed Helion Energy to secure fusion power, starting at 12.5% of output and scaling to 5 GW by 2030 and 50 GW by 2035, with Altman stepping down from the board and recusing himself while retaining his stake. It is shutting down the Sora AI video app, API, and website after legal pushback, and the collapsed $1B Disney deal. The new OpenAI Foundation launched a new $25B strategy with over $1B near-term for disease research plus shared high-quality health datasets. OpenAI also released open-source teen safety prompts.
ChatGPT is now a visual shopping discovery platform with product images, prices, ratings, side-by-side comparisons, easier photo uploads for similar items, and merchant integrations via the Agentic Commerce Protocol. No checkout. It gained a Library tab and toolbar for instant file access and querying. The company is dangling 17.5% preferred returns to private equity for enterprise AI joint ventures, while Altman steps back from safety oversight to chase fundraising, supply chains, and massive data centers. Read more.
Luma AI's Uni-1 could be the first real challenger to Google's Nano Banana image dominance
Luma Labs’ Uni-1 is a decoder-only autoregressive model that combines image understanding and generation in a single sequence of interleaved text and visual tokens. It breaks down prompts, plans spatial layouts, and keeps context across multi-image inputs and edits. On RISEBench and ODinW-13, it outperforms Nano Banana 2 and GPT Image 1.5 in reasoning and object representation.
Human preference tests put it first for overall, style, and reference-based generation. Costs run about $0.09–$0.10 per 2K image, with web access live and API integration coming for automated creative pipelines. Read more.
Cursor shares full training story behind Composer 2 MoE model
Cursor Research released a technical report detailing how they built Composer 2. The team focused on three core efforts: continued pretraining, reinforcement learning, and benchmark development. The goal throughout was to closely match the real Cursor environment so the model excels at practical coding tasks. Continued pretraining delivered steady gains in downstream coding performance.
The team found that simple algorithms often worked best and improved performance broadly. They introduced CursorBench, an internal benchmark that captures the complex, everyday problems software engineers face in large codebases. The team also developed and open-sourced custom kernels for efficient training, and discussed distributed scaling with Ray and PyTorch to handle large-scale RL across GPU clusters. Read more.
Cekura launches AI agent observability with metrics, auto-tuned evaluators, real-time dashboards, and alerts for performance and failures.
Jared is a Slack-native AI agent that integrates with 10,000+ tools, models team context, and autonomously executes workflows.
Tobira is a multi-agent network with identity layers enabling agents to discover, negotiate, and transact via privacy-preserving coordination protocols.
Agent Hub Builder creates a monetized AI tool marketplace with auth, usage tracking, chat interfaces, and personalized user experiences.
Zoer is an AI codegen agent that synthesizes database schemas, backend services, and optimized frontends for production-grade full-stack systems.
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.
