Good morning. It’s Monday, August 10th.

You read. We listen. Let us know what you think by replying to this email.

OpenAI appears to be running a classic two-track play. On the compute frontier, rumors indicate a hard split between Astra, a long-horizon, multi-agent model family, and Doug, a massive pre-training run targeting base capabilities through raw scaling rather than post-training tricks.

As base capabilities jump, so do the tail risks. OpenAI flagged that Astra hit critical cybersecurity risk thresholds after autonomously generating zero-day exploits, forcing the lab to lock model weights and restrict development to sandboxes. To establish formal safety guarantees before capability overhang gets worse, Fields Medalist Jacob Tsimerman joined the safety team, bringing pure mathematics to a field long dominated by empirical hand-waving.

Meanwhile, at the user interface boundary, OpenAI is rapidly shortening the gap between inference and execution. GPT-Live now accepts files and manages persistent project state, turning voice chat into a contextual workspace.

At the same time, OpenAI acquired presentation startup NextSlide to turn raw model outputs into polished slide decks. It is an aggressive play to build a full office suite inside ChatGPT, absorbing human workflows before models take over completely.

Related video:

Toronto math genius on why he's joining OpenAI

Anthropic is turning Claude Code ‘Auto mode’ on by default for Pro, Max, and Team tiers starting August 14. The update replaces step-by-step confirmation prompts with an automated classifier that blocks destructive, irreversible, or out-of-bounds actions.
In testing across 1,053 paid users, Auto mode caught 89% of dangerous commands, compared to a meager 13.6% caught by human reviewers suffering from approval fatigue. Teams using the feature shipped 25 percent more pull requests.

Crucially, this operational shift relies on recent safety advances. Head of Claude Code Boris Cherny reports that Anthropic has trained newer models to resist indirect prompt injection, where hostile web content tricks agents into exfiltrating SSH keys or passwords.

Across 720 indirect prompt injection attacks evaluated by Trajectory Labs on Claude Fable 5, Opus 5, and Sonnet 5 in Auto Mode, zero succeeded.

Google is shifting strategy away from pure frontier-model dominance toward infrastructure scale and market distribution. The financial payoff is already visible: Google Cloud revenue hit $24.8 billion last quarter, up 82 percent year over year, with TPU sales projected to hit $120 billion by 2027 as Google supplies competitors like Anthropic.

DeepMind's DiffusionGemma illustrates this push for specialized compute efficiency. Converted from Gemma 4 26B-A4B using under 10 percent of its training budget, it replaces sequential generation with parallel 256-token block denoising, reaching 1,500 tokens per second on an Nvidia H100.

At the product layer, Google is monetizing custom workflows. An unconfirmed feature flag indicates free Gemini Gems will retire on October 20, forcing manual migration to Skills. Running on Gemini Spark, Skills require paid Pro or Ultra subscriptions while excluding work or school accounts, consolidating lightweight customization behind a paywall.

Ankon AI renders narrated whiteboard explainer videos from text prompts, scripts, PDFs, or reference images.

Omniwork operates as a desktop agentic OS, running specialized AI models to automate complex creative pipelines.

Prompt Golf is a competitive puzzle game where players craft minimal prompts to trigger target AI outputs.

DocsAlot CLI enables AI agents to build, maintain, and publish documentation sites using natural language.

Soloop is an approval-first Agent OS that deploys autonomous AI executive teams to build solo companies.

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.