Good morning. It’s Monday, May 25th.

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

Attio is the AI CRM that keeps you ten steps ahead.

Ask Attio anything. Where should I focus? What deals are at risk? Search, update, and create across your customer data.

Ask more from CRM. Ask Attio.

Try Attio for free

OpenAI is preparing for the reality of recursive self-improvement by opening a safety position within its Preparedness team, offering a salary of up to $445,000. Seeking technical executors labeled as "tasteful and strategic," the role requires modeling abstract, long-horizon security risks that do not yet exist.

The researcher will mitigate data poisoning, create interpretability tools to audit model activations, and establish internal metrics to track the automation of technical staff. This initiative feeds into aggressive internal developer targets across the industry, such as Sam Altman’s benchmark for an automated AI research intern by September 2026 and predictions from Demis Hassabis and Jack Clark pointing toward autonomous R&D by 2028.

Concurrently, Gartner named OpenAI a Leader in its Magic Quadrant for Enterprise AI Coding Agents, marking a distinct shift toward autonomous engineering. Driven by the release of GPT-5.5, Codex now commands 4 million weekly users across major corporate setups like Datadog and Dell. The platform pairs deep OS-level sandboxing with granular enterprise governance rules, allowing agents to execute complex tasks safely. This architecture enabled Cisco to compress the development of its AI Defense platform down to weeks. Read more.

Google DeepMind’s AlphaProof Nexus has autonomously solved 9 of 353 open Erdős mathematics problems and proved 44 of 492 open OEIS conjectures. Operating through agentic loops, the system pairs LLMs with the Lean formal proof assistant. The LLM independent subagents propose proof paths while Lean acts as a formal verifier, checking and rejecting logical steps to eliminate hallucinations.

This hybrid process tackles deep research problems rather than textbook exercises, with some breakthroughs costing only a few hundred dollars per problem. Because formal verification provides concrete logical guarantees, this shift from assistant to autonomous research engine has direct implications for cryptography, smart contract auditing, and zero-knowledge proof systems.

This milestone drops right into a widening philosophical rift over the trajectory of machine reasoning. DeepMind CEO Demis Hassabis suggests humanity is "in the foothills of the singularity," predicting human-level AI within five years that could spark a transformation ten times faster than the industrial revolution.

Conversely, Yann LeCun maintains current LLMs lack true intelligence, arguing that genuine reasoning requires solving novel problems without prior training, not just scaling learned patterns. DeepMind's Oriol Vinyals takes a middle view, noting that while today’s models are undeniably potent in structured fields like coding and mathematics, they still cannot learn from real experience or produce foundational discoveries on their own. Read more.

Jeff Bezos has pushed back against rumors that his secretive, $38 billion startup Project Prometheus is building humanoid robots or factory automation. In a CNBC interview, Bezos clarified that the company has nothing to do with robotics. Instead, it is building an "artificial general engineer" designed to completely automate how physical objects are created. He described the core technology as a highly advanced, next-generation take on computer-aided design (CAD) software.

The venture operating out of San Francisco, London, and Zurich launched in late 2025 with $6.2 billion from Bezos and co-CEO Vik Bajaj, a former Google X executive. It recently closed a $10 billion funding round backed by JPMorgan and BlackRock. The startup has poached around 120 engineers from top labs like OpenAI, DeepMind, Meta, and xAI. Instead of training models on human text, Prometheus focuses on the laws of physics to run billions of cloud simulations, creating stress-tested manufacturing blueprints. Read more.

Freu AI learns directly from user actions to automate complex cross-software workflows without configuration or APIs.

Memdex automatically saves your scattered AI conversations locally to detect and inject past context into any prompt.

TestSprite 3.0 deploys parallel AI agents to autonomously generate, execute, and auto-heal backend and frontend end-to-end tests.

Forsy captures real-time agent workflows across traces and tasks to build high-fidelity data infrastructure for the agent economy.

VenturusAI automatically runs frameworks like SWOT and PESTEL on your business idea to evaluate market viability.

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.