Daily deep dives

🧬 Google DeepMind's AlphaEvolve combines Gemini language models with evolutionary computation to autonomously discover and optimize algorithms across computing tasks. The system has already recovered 0.7% of Google's worldwide compute resources—a significant achievement given the scale of the company's operations—by enhancing data center scheduling and hardware design. Beyond practical applications, AlphaEvolve has made contributions to theoretical mathematics, including the kissing number problem, showcasing AI's potential to advance both industrial efficiency and scientific discovery. DeepMind plans to extend access to academic researchers, potentially accelerating innovations in material science, drug discovery, and sustainability.

🛒 Perplexity AI's partnership with PayPal will enable direct purchasing within its chat interface, allowing users to buy products, book travel, and purchase tickets without leaving the conversation. Set to launch this summer in the US, the integration will reach PayPal's massive base of 430 million active accounts across nearly 200 markets. The collaboration represents a significant advancement in agentic commerce, potentially transforming how consumers discover and purchase products online. Meanwhile, Perplexity continues to attract major investment interest, reportedly seeking $500 million in funding at a $14 billion valuation with backing from tech heavyweights like Jeff Bezos, Nvidia, and SoftBank.

⚙️ ServiceNow's entry into the CRM market challenges traditional approaches by reimagining customer relationship management as a cross-functional "nervous system" that orchestrates service across organizational silos. The AI-powered platform integrates technologies like AI Agent Fabric and Workflow Data Fabric to streamline customer interactions beyond simple record management. With $1.4 billion in annual contract value and 30% year-over-year growth, ServiceNow is positioning itself as a serious disruptor in the space, demonstrated by pilot customer Pure Storage's impressive results—40% cost reduction and an 82% Net Promoter Score. This shift represents a broader industry trend toward workflow-based CRM approaches that address the increasing complexity of digital customer journeys while incorporating responsible AI governance.

🎨 Figma and Bolt's new partnership enables non-technical entrepreneurs to transform design mockups into functional applications through "vibe coding," where users describe desired functionality in plain language. The integration supports four common app types—those that perform actions, store information, document processes, or sell products—making app development accessible to a wider audience without coding knowledge. While this collaboration significantly lowers barriers to digital entrepreneurship, the resulting applications have security and complexity limitations, positioning them more as prototyping tools than production-ready solutions. Success still requires strong business fundamentals and domain expertise, with entrepreneurs needing to approach the process as an iterative learning journey rather than a one-and-done solution.

⚖️ Harvey AI's new $250+ million funding round values the legal tech startup at $5 billion, marking a striking increase from its $3 billion valuation just months ago amid impressive revenue growth to $75 million in annualized run rate. Led by Kleiner Perkins and Coatue, with increased participation from Sequoia Capital, this investment highlights the accelerating adoption of AI in the legal sector, where Goldman Sachs estimates 44% of legal work could eventually be automated. Harvey's technology leverages models from OpenAI, Anthropic, and Google through partnerships with consulting firms like PwC and direct sales to corporations, positioning it at the forefront of a legal tech sector that attracted $2.1 billion in global investments in 2024.

🔍 Patronus AI's newly launched Percival represents a significant advancement in AI oversight, offering an automated monitoring platform that detects and fixes failures in enterprise AI agent systems. Using an innovative "episodic memory" architecture, Percival learns from past errors to identify over 20 different failure modes across four categories, adapting to specific workflows while suggesting optimizations for improved performance. Early adopters report dramatic reductions in agent workflow debugging time, with enterprise AI innovators like Emergence AI and Nova already implementing the technology. This solution addresses a critical challenge in the AI ecosystem—the compounding error risks inherent in complex, mission-critical agent applications—and integrates seamlessly with multiple frameworks including Hugging Face Smolagents and OpenAI Agent SDK.

Read our other letters