Brex CEO Pedro Franceschi argues that AI represents a platform shift as significant as electricity, and most founders still underestimate its impact. He advocates for 'token maxing'—personally pushing AI to its limits to understand its capabilities—and insists that CEOs must become chief AI officers to rebuild their companies around intelligence on demand. Franceschi describes how Brex is transforming from a fintech company into an AI-native organization, using AI teammates rather than chatbots to create customer world models and ultimately build a form of company AGI.
Latest insights
Insights come from up to date trends.
Gwynne Shotwell is the key figure behind SpaceX, saving the company by negotiating a $1.6 billion contract with NASA and stabilizing client relationships after the Falcon 9 explosion. She translates Musk's ambitious vision into executable business plans, driving the global expansion of Starlink. As SpaceX approaches its IPO, her leadership story offers a model of operations and strategy for the tech industry.
The next major bottleneck for AI expansion may not be chips or software, but copper and other critical minerals. Dan Dreyfus argues that the US faces a looming supply-demand shock as China restricts mineral exports, while AI data centers and electrification will require as much copper in the next 18 years as was mined in the last 10,000. Combined with a dying grid, craft labor crisis, and $140 trillion in debt, this sets the stage for a commodity supercycle that investors should watch closely.
Anthropic has released Claude Fable 5, its most capable model from the previously secret Mythos series, now available to all users. The model achieves 80.3% on SWE-bench Pro, helped Stripe migrate 50 million lines of code in one day, and accelerates protein design by 10x. However, its high per-token cost (double Opus) and tendency to perform more work per round cause subscription credits to drain quickly, especially after the free window ends on June 23. This makes it a powerful but expensive tool for developers tackling complex, long-horizon tasks.
Anthropic has proposed a controversial idea: temporarily pausing all AI development to assess risks from recursive self-improvement. The suggestion highlights growing unease about AI systems potentially improving themselves beyond human control. This sparks debate on whether a pause is feasible or just a panic response.
Anthropic has released Claude Fable 5, a Mythos-class model that balances advanced capabilities with robust safety safeguards, enabling broad deployment without compromising on security. It excels at sustained, complex tasks across coding, research, finance, and law, handling problems that require days of focused work. This release signals a shift toward more capable and trustworthy AI assistants for professional use.
Bill Maris, founder of Google Ventures, shares four critical lessons from his career, including how Google could leverage its data and compute to crush AI competitors. He argues that small VC funds often outperform larger ones due to better incentives and focus, and compares current AI development to an 'Atari stage' where foundational breakthroughs are still emerging. The discussion also covers OpenAI's valuation challenges in an AI price war and the broken incentives in venture capital that hinder deep tech innovation.
Recursive self-improvement (RSI) proposes using AI to optimize AI, creating a feedback loop where stronger models discover new knowledge and insights. The ultimate goal is to input compute and output scientific discoveries, accelerating humanity's understanding of the world. However, as AI gains the ability to modify code, design experiments, and make unexpected discoveries, safety and interpretability become critical. Interpretability is not just a safety mechanism but also a tool to improve R&D efficiency, helping humans judge whether the AI is discovering something new or crossing the line into uncontrollable territory.
CC Switch is a desktop tool that lets developers switch Claude Code's underlying model to alternatives like DeepSeek with a single click, eliminating the need to manually edit environment variables. The tool supports hot-swapping, so no restart of Claude Code is required. This simplifies experimenting with different AI models for coding tasks, making it easier to compare performance and choose the best fit.
A desktop tool called CC Switch lets developers instantly swap Claude Code's underlying model to DeepSeek without manually editing environment variables or restarting the terminal. The tool supports hot-switching, making it seamless to test different AI models during coding sessions. This is especially useful for indie developers who want to compare model outputs or leverage DeepSeek's strengths while keeping Claude Code's interface.
Y Combinator is reviving its Startup School as a two-day in-person event in San Francisco, July 25–26. The program features talks from world-class speakers, hands-on demos, and small group sessions with YC partners, targeting ambitious technical builders globally. This signals a return to community-driven startup education beyond the typical online format.
Palo Alto Networks CEO Nikesh Arora reveals that an AI system (Claude) found years' worth of security vulnerabilities in their code in just six weeks, demonstrating AI's transformative potential in cybersecurity. The discussion explores whether cyber defenders are losing the race against AI attackers, and argues that analytical SaaS is dead—only AI-native products will survive. Arora also weighs in on AI commoditization, predicting that while models become utilities, value will shift to applications and data moats.
Two Claude Code skills form a pipeline: claude-session-manager compresses raw conversation files into structured Markdown (reducing size ~17x), and mining-session-skills evaluates whether extracted patterns are worth turning into reusable skills. The key insight is that automatic skill generation without quality gates produces noise, so a judgment step is critical. The video also demonstrates how the second skill was built using Claude's own methodology, making it a meta-example of skill authoring.
Emergent is a no-code platform that lets anyone build, ship, and monetize software by chatting with an AI agent, reaching $100M ARR in just nine months. The company's success stems from a multi-agent architecture that outperforms traditional coding benchmarks like SWE-bench, and a strategy of tinkering and iterating rapidly. For indie devs and AI builders, the story highlights the power of second-mover advantage and building globally from India.
This week's GitHub highlights feature five notable projects: Open Design offers an open-source alternative to Claude Design for UI generation; Headroom compresses AI context to reduce token costs; Taste Skill trains AI models in aesthetic judgment; English-level-up-tips provides an unconventional English learning guide; and ai-engineering-from-scratch is a comprehensive course on building AI systems from scratch. These tools cater to developers seeking cost-effective, open-source solutions and skill-building resources.
Paxel is a free tool from Y Combinator that analyzes how developers build with AI coding agents like Claude, Codex, and Cursor. It generates a builder profile covering planning, steering, execution, and patterns across sessions, helping developers understand their workflow. The tool integrates with Startup School applications, offering a way to showcase building style without risk.
Legora, a legal AI startup from YC W24, reached $100M ARR in just 18 months by targeting the conservative legal industry with AI agents that automate M&A and other legal workflows. CEO Max Junestrand shares how the company moved fast post-YC, convinced law firms and corporate legal departments to adopt AI, and built a moat beyond foundation models. The discussion covers fundraising, scaling culture, and Legora's ambition to expand beyond legal tech into broader enterprise automation.
Tianyuan Dong's newly founded lab, Recursive Superintelligence (RSI), is valued at $4.65 billion, betting on recursive self-improvement—where AI systems autonomously design and enhance their own capabilities without human intervention. The conversation explores the technical and organizational challenges of this path, including Coconut reasoning pathways and the role of synthetic data. It also analyzes how new labs like RSI compete with big companies through organizational efficiency and talent acquisition, and what the AI era means for individual careers.
OpenClaw enables content creators to build a fully automated AI assistant that curates, writes, and generates cover images for a daily AI newsletter—all without writing any code. The video walks through deploying on MyClaw, connecting Telegram, configuring data sources and skill packs, using GPT image for covers, and setting up scheduled tasks. This workflow saves creators hours each morning, letting them focus on reviewing and refining the output.
Qualcomm is betting big on edge AI as the next battleground, with on-device intelligence transforming phones, cars, glasses, and robots. The video explores how agent-based interactions are replacing traditional apps, and why 6G is being architected specifically for AI workloads. Physical AI—where AI interacts with the real world through robotics and autonomous systems—is highlighted as the ultimate opportunity, with Qualcomm's integrated chip-connectivity-AI strategy positioning it as a key enabler.