Good morning. It’s Monday, August 3rd.

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

Remember when the news was about what happened, not how to feel about it? 1440's Daily Digest is bringing that back. Every morning, they sift through 100+ sources to deliver a concise, unbiased briefing — no pundits, no paywalls, no politics. Just the facts, all in five minutes. For free.

Read the newsletter trusted by 4.5 million fact-seekers.

Anthropic researcher Levent Alpoge revealed that Claude Fable autonomously replicated 5 of OpenAI Astra's 10 headline mathematical proofs in just 24 hours. Operating without internet access or custom prompting, Fable matched Astra's reasoning leaps on identical math problems. The result shows that major mathematical breakthroughs can be reproduced across model families using generic prompting setups.

Separately, Andrej Karpathy tested Claude Opus 5 by giving it a 1-million-token budget (~$10) and the first paragraph of The Lord of the Rings. Over two hours, Opus 5 procedurally generated a 3D JavaScript world using 5,500 lines of code, placing and animating (x,y,z) polygon assets entirely on its own.

Karpathy noted the experiment points toward a future of instant, hyper-custom interactive media, while highlighting a clear hardware and software gap in native visual perception.

To check its work, Opus 5 had to slowly take and inspect static screenshots, showing that models still lack native, real-time video processing to debug complex visual outputs.

OpenAI quietly teased its next major model family, Astra, inside a technical post about mathematics. An internal version of the multi-agent system solved 10 long-standing open problems across high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, group theory, arithmetic circuit complexity, and extremal combinatorics.

The results include non-sofic group construction, disproving Connes's rigidity conjecture, tighter sphere-packing bounds, and resolving Erdős conjectures. Astra coordinated tasks across agents, generating formal Lean 4 certificates for machine-checkable proofs at roughly $2,000 in inference under Sol API pricing.

Reports suggest Astra could launch as GPT-5.7, GPT-6, or under a new brand, with more capable versions potentially released under restricted access.

While touting Astra's capabilities, OpenAI CEO Sam Altman proposed using ChatGPT Work to turn family calendars and news into personalized morning audio briefings. But the idea drew immediate pushback over automating parent-child relationships.

Google pulled Nano Banana 2 from Google Earth just two days after launch. The tool allowed users to edit real satellite imagery with text prompts, but users quickly generated and exported fake scenes of refugee movements and damaged hospitals. Google paused the integration to build stronger safeguards against exported screenshots spreading misinformation.

In a new paper, Google DeepMind introduced SkillSmith, a framework that enables LLMs to synthesize custom AI skills by reading textual knowledge and prefix weights at the same time. Using "instruction-steered parametric synthesis," SkillSmith ingests prefix parameters and text descriptions to output new prefix weights tailored to a target task, outperforming standard text-only prompting and weight-merging techniques.

Meanwhile, Google cancelled its standalone Android and iOS AI Studio apps despite nearly 800,000 pre-orders, opting to integrate its "vibe coding" tools into Gemini while preserving the AI Studio web platform.

Google is also updating the Gemini desktop app toward feature parity with the web experience by adding dedicated image and video creation tabs, a direct camera capture tool, and support for custom Model Context Protocol (MCP) servers in Gemini Spark.

Additionally, Gemini for macOS introduced "Speak to Window," allowing users to hold the Fn key to dictate, format, and edit content across desktop applications using optional on-screen context reasoning.

MiniMax H3 generates 2K video with stereo sound from multimodal inputs for motion design and branding.

AgentMicro tracks live parallel Codex Desktop and CLI tasks directly from your macOS menu bar.

Port22 streams Mac AI coding agent prompts to your phone for remote monitoring and instant approvals.

DepthData unifies company AI tool spend and adoption tracking into one audit-ready system of record.

Pally connects to iMessage and WhatsApp inboxes to monitor texts and automatically send contextual replies.

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.