Good morning. It’s Friday, June 5th.
I spent the day yesterday at Oumi’s Seattle office to see what they were building. Their thesis is that most businesses don’t need to use high-end expensive models for simple tasks, and cheap (or free open-weight) models with a bit of reenforcement learning perform actually better than frontier models on narrow tasks.
Their speciality is walking the non-technical people through the process of training a small model on various simple tasks like customer email classification, risk analysis, and database retrieval. Very cool concept! Thanks for the invite to the offices! Definitely a company to keep an eye on.
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
In partnership with AhaCreator
Influencer marketing works — until you actually try to run it at scale.
Finding the right creators, chasing replies, reviewing drafts, managing contracts, tracking deadlines, handling payments… it quickly turns into a full-time operations mess.
AhaCreator is an AI-native influencer marketing platform built to automate that entire workflow.
Its AI helps brands discover creators that actually fit their audience, detect fake engagement, manage outreach and follow-ups, review creator drafts against the brief, and keep campaigns moving from first contact to final delivery.
AhaCreator also handles the trust layer: standardized contracts, escrow protection, global payouts, and refund protection if a creator doesn’t deliver.
For startups, agencies, and brands trying to scale creator campaigns without adding more manual work, AhaCreator is basically an AI operations team for influencer marketing.
Thank you for supporting our sponsors!
Anthropic stops testing new Mythos model after unreleased 'Oceanus' leak
Anthropic is raising the alarm about recursive self-improvement, warning that accelerated updates could quickly spiral past human control. Anthropic’s internal metrics back it up. Claude now writes over 80% of the company's production code, triggering a massive 8x explosion in developer output.
On one end, cybersecurity is fundamentally breaking. Anthropic’s Logan Graham says AI is now finding flaws faster than organizations can patch them, making remediation the key constraint. He also predicts today’s leading cyber models could look “dumb” within 6–12 months.
This raw power is already being weaponized: Financial Times has reported that NSA has embedded half a dozen forward-deployed Anthropic engineers to help the agency deploy Claude Mythos for offensive cyber operations, using its advanced exploit-detection capabilities to plan strategic strikes despite past friction between the lab and Washington.
However, even as Anthropic calls for a global pause to rethink safety, the commercial machine keeps churning. The company just launched a massive Partner Hub to help consulting giants like Accenture and Deloitte train hundreds of thousands of workers on Claude.
Anthropic also dropped the new ant CLI for the Claude Platform, a developer tool that makes every API endpoint runnable from the terminal. The CLI supports interactive OAuth logins and lets developers version-control Claude Managed Agents as YAML files in Git, giving Claude Code agents direct terminal-level execution skills.
Meanwhile, a secret, unreleased version of their next Mythos model, codenamed Oceanus, was just leaked and sold via Chinese API proxies, forcing Anthropic to panic-pause its pre-launch safety testing.
OpenAI deploys 5x more compute-efficient Dreaming memory architecture to free ChatGPT tier
OpenAI is rolling out a new background memory system called "Dreaming." Instead of forcing you to manually save notes, Dreaming quietly tracks your multi-chat history to build structured, narrative dossiers of your work, hobbies, and travel. It automatically adjusts when your real-world context changes, cuts backend compute costs by 5x, and pushes fact-retrieval accuracy to 82.8%. Best of all for everyday users, the tech is coming to free tiers soon.
At the same time, OpenAI pushed a massive update to GPT-Rosalind, their dedicated life sciences model. It now handles complex genomics analysis and drug discovery workflows using 31% fewer tokens than standard models, making deep scientific reasoning much cheaper to run.
To smooth things over in Washington while this commercial rollout explodes, CEO Sam Altman is leading a political charm offensive. He just endorsed a light-touch regulatory blueprint that invites tech labs to voluntarily share models with the government 30 days before public release. To keep OpenAI out of the political crosshairs, Altman explicitly stated he is not financing any campaigns in the 2026 US midterms, cutting ties with the Greg Brockman-backed Leading the Future PAC.
Gemma 4 12B model is designed to run on any laptop with 16GB of RAM
Alphabet is pushing a massive local hardware play while backing it up with historic capital investment. The company closed a record-setting $85 billion equity sale, anchored by a $10 billion investment from Berkshire Hathaway, to bankroll an intense $180–190 billion infrastructure buildout.
At the same time, Google DeepMind just launched Gemma 4 12B, a compact model running under an Apache 2.0 license that crams multimodal text, vision, and raw audio processing onto standard 16GB laptops. By building an encoder-free setup that feeds visual patches and audio directly into the decoder backbone, Google completely removes traditional processing lag, allowing local laptops to run native tool calling and a massive 256K context window.
This local-first strategy is taking over Google’s creative tools, too. The Magenta team dropped Magenta RealTime 2, an open-weights live music model that runs right on your MacBook using a custom C++ engine built for Apple Silicon's MLX.
Meanwhile, Google Labs is launching Dreambeans for iOS and Android, an experimental, illustrated journal built explicitly for paid Google AI Ultra subscribers. Its Personal Intelligence layer combs through your connected apps, distilling everything into 10 to 14 short, customized daily stories. The app drops practical, real-world recommendations into your morning feed, all wrapped up with custom illustrations generated by the Nano Banana 2 image model.
Frontier models and product moves
Agents and the agentic stack
Business, labor, and institutions
Security and surveillance
Hardware and infrastructure
Research and science
Robotics
Watch
Factory Router selects the best model for every task automatically, maintaining top performance while cutting costs by 25%.
Mailwarm 2.0 premium warmup and deliverability system uses automated engagement and expert guidance to keep emails out of spam.
Coral allows agents to query data across APIs, databases, and file systems using standard SQL.
Empromptu AI captures live app usage and human corrections to train custom, accurate models you own.
Astra Autonomous Pentesting deploys AI agents to continuously uncover, validate, and fix complex chained software vulnerabilities.
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.
