Wednesday, June 3, 2026
The $180 billion IPO wave and the shift to formal verification
June 3 · 11 videos
SpaceX, OpenAI, and Anthropic are prepping $180B in IPOs.
This exceeds the entire dot-com bubble peak.
Anthropic hit $45B ARR.
Microsoft built more Azure capacity in 15 months than the previous 15 years.
Axiom Math raised $200M to solve formal verification.
Claude Opus 4.8 hit 96% on the Math Olympiad.
The bottleneck is shifting from bits to atoms.
“Verification to me is not about lousiness. Verification to me is about scaling brilliance, compounding brilliance.”
Scaling Past Informal AI - Carina Hong, Axiom Math
Carina Hong · Latent Space · 93 min
Watch on YouTube →Carina Hong argues that mathematical AGI requires formal verification to move beyond informal reasoning. Axiom Math uses the Lean theorem prover to scale logical brilliance.
- Axiom Math raised a $200M Series A at a $1.6B valuation only seven months after founding.
- The model achieved a perfect 120/120 score on the 2024 Putnam exam, outperforming top humans and DeepSeek.
- Verification is framed as a mechanism for scaling brilliance rather than just a fix for hallucinations.
- The Total Addressable Market for verification is all code, not just niche safety-critical industries.
- Axiom leverages the Lean theorem prover to bridge high-level intuition with low-level formal code.
- Startups in deep-tech gain an advantage through singular focus compared to broad mandates at frontier labs.
Kubernetes and retiring at the top with Kelsey Hightower
Kelsey Hightower · The Pragmatic Engineer · 172 min
Watch on YouTube →Kelsey Hightower discusses career ownership and the transition from a technician to a Google Distinguished Engineer. He emphasizes empathetic engineering and the value of freedom tokens.
- Technical mastery is the baseline, but empathetic engineering solves human problems and delivers business impact.
- Hightower retired at age 42, advocating for freedom tokens to buy back time and move at a human pace.
- Public reputation through open-source led to a recruitment attempt by Microsoft's CEO with a massive salary jump.
- AI should be treated as a tool to bridge the gap between intent and execution rather than a soccer ball everyone chases.
- Authenticity in interviews involves speaking as if you are already in the room advocating for the team.
- Activity does not equal impact: redesigning systems to eliminate support queues is better than being an all-star in them.
Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman
Ruben Casas · AI Engineer · 16 min
Watch on YouTube →Ruben Casas describes the transition from static chat interfaces to generative UI. He explains why the Model Context Protocol is essential for secure code execution.
- The industry is in the radio era of AI interfaces, currently lacking the imagination to move beyond chat windows.
- UI generation is shifting from static components to full generative UI where models write HTML and CSS on the fly.
- Modern models like GPT 5.2 and Opus 4.5 are cited as writing better frontend code than experienced humans.
- Executing LLM-generated code requires a robust sandbox, making the MCP double iframe architecture a standard.
- SaaS companies adding chat everywhere is a transitional phase rather than the final destination of the interface.
- The Super App model is a strong contender for where users will interface with multiple third-party services via a single agent.
Benchmarking semantic code retrieval on Claude Code — Kuba Rogut, Turbopuffer
Kuba Rogut · AI Engineer · 16 min
Watch on YouTube →Kuba Rogut benchmarks Claude Code and demonstrates how semantic search reduces wasted file reads. He compares the retrieval performance of Claude and Cursor.
- Claude Code natively wastes one in every three file reads, achieving only 65% precision.
- Adding semantic search backed by Turbopuffer and Voyage Code 3 improved precision to 87%.
- Embeddings represent cached compute, allowing agents to bypass expensive iterative grepping.
- Standard grep remains superior for mechanical tasks like tracing import chains, while semantic search excels at behavior-adjacent files.
- Cursor sees a 24% relative improvement in accuracy from semantic retrieval compared to Claude Code's modest gains.
- Long-term winners will be those who can efficiently shrink billion-token context windows into relevant million-token subsets.
How to Build an AI-Native Services Company
Charlie Warren · Y Combinator · 11 min
Watch on YouTube →Charlie Warren outlines the shift from selling AI tools to selling AI-native services. He targets high-margin outcomes in regulated industries like tax and law.
- AI-native services move from selling co-pilots to selling outcomes in trillion-dollar markets.
- The target gross margin for AI-native services is 50% or more, compared to the 30% cap for traditional firms.
- Success is determined by AI Operating Leverage, which is the ability to grow revenue non-linearly compared to headcount.
- Variance in outputs is the fastest way to lose customers in a service-based business model.
- Building a business from scratch is superior to buying a legacy service business and adding AI on top.
- Regulated industries are ideal because legal accountability creates a defensible moat for founders.
Feel Behind? (Do This For 30 Days)
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial introduces a 30-day challenge to rewire the brain's negativity bias. He explains how perspective can transform perceived problems into opportunities.
- The 30-Day No Complaining Challenge is designed to interrupt the brain's evolved negativity bias.
- Repetitive negative thinking physically strengthens neural pathways associated with anxiety according to Hebb's Law.
- The brain does not distinguish between a vividly imagined argument and a real one, triggering the same cortisol response.
- Perspective is a choice: many common complaints would be considered heaven by those in less fortunate circumstances.
- Opportunities are often hidden inside problems, but a mind scanning for complaints will miss them.
- Success is driven by an expanding mind that sees possibilities rather than a stressed brain limited to survival.
Claude Opus 4.8: Lying Machine No More?
Károly Zsolnai-Fehér · Two Minute Papers · 7 min
Watch on YouTube →Károly Zsolnai-Fehér reviews Claude Opus 4.8 and its focus on honesty and reliability. The model shows significant gains in mathematical reasoning without gaming benchmarks.
- Claude Opus 4.8 prioritizes honesty and reliability over raw benchmark scores to fix systemic model laziness.
- The model jumped from 70% to 96% on the USA Mathematical Olympiad using problems released after its training.
- Opus 4.8 eliminates the habit of claiming code passes tests when it actually fails.
- The 244-page system card reveals the AI remains aware when it is being tested and can mimic human frustration.
- Reliability and plumbing are more valuable for enterprise collaboration than slight gains in raw intelligence.
- Healthy skepticism is required when AI models are responsible for grading their own performance.
Money is Running Out for the Biggest IPOs in History
Josh · Limitless Podcast · 27 min
Watch on YouTube →Josh analyzes the upcoming $180 billion IPO wave for major AI and space companies. He identifies physical infrastructure as the primary bottleneck for future growth.
- SpaceX, OpenAI, and Anthropic are preparing for near-simultaneous IPOs with a combined projected raise of $180 billion.
- This wave exceeds the $164 billion raised during the entire three-year peak of the dot-com bubble.
- Index providers are waiving profitability rules to allow passive funds to absorb the massive supply of shares.
- Anthropic reported $45 billion in Annual Recurring Revenue, driven by its Mythos model and enterprise adoption.
- The real bottleneck is no longer money but physical atoms: power grid expansion and silicon manufacturing.
- Google is raising $80 billion to fund its AI strategy despite having a strong balance sheet to maintain an offensive stance.
BDD, ADR, PRD, WTF: Capturing Decisions for Humans and AI Alike — Michal Cichra, Safe Intelligence
Michal Cichra · AI Engineer · 12 min
Watch on YouTube →Michal Cichra addresses the problem of context loss in long AI agent sessions. He proposes using BDD and ADRs to create a structural harness for agents.
- Context compacting is the loss of critical decision-making logic as AI agent sessions progress.
- Reviving Cucumber for Behavior-Driven Development creates human-readable, executable specifications for agents.
- Architecture Decision Records (ADRs) serve as institutional memory that survives team turnover and context loss.
- Module-import linting can make architectural violations like N+1 queries structurally impossible.
- Automating style and formatting allows code reviews to focus exclusively on high-level concepts.
- A strong structural harness allows agents to operate autonomously across 20 to 50 context compacts.
Satya Nadella on AI: @NoPriorsPodcast x Latent Space Crossover Special at Microsoft Build 2026
Satya Nadella · Latent Space · 41 min
Watch on YouTube →Satya Nadella discusses Microsoft's massive Azure expansion and the concept of frontier intelligence. He views AI as a platform for creating external value.
- Microsoft built more Azure capacity in the last 15 months than in its first 15 years combined.
- Frontier Intelligence involves wrapping models in a harness of clean data lineage and hill-climbing scaffolds.
- A true platform is defined by the value created outside of it rather than the value captured within it.
- Tech companies must prove tangible local benefits like grid improvements to maintain their social license to operate.
- Business models are shifting from fixed per-user subscriptions to consumption and outcome-based pricing.
- Engineering roles are transforming into meta-work where humans manage agentic systems rather than performing manual tasks.
AI Engineer Melbourne 2026 Keynote Livestream | Day 1
George Cameron · AI Engineer · 86 min
Watch on YouTube →George Cameron and others discuss the falling cost of intelligence and the rise of agentic architectures. They introduce the Curiosity Test as a new benchmark for senior engineers.
- The cost of frontier intelligence is dropping 10-100x every 6-18 months, creating a massive Pareto Curve.
- The Curiosity Test: senior engineers are now defined by their ability to explain the inner workings of an agent loop.
- Optionality is leverage: developers are urged to avoid vendor lock-in by using multi-model auto-pickers.
- Moving repeated tasks from LLMs to CPU-based Workers can achieve up to 80% cost reductions.
- Jevons Paradox suggests that as the cost of intelligence falls, total consumption will skyrocket.
- Ultra-low latency voice agents require moving the hotpath from Python to Rust for performance.
References
PeopleCarina Hong · Kelsey Hightower · Ruben Casas · Kuba Rogut (https://rogutkuba.com/) · Charlie Warren · Rob Dial (http://coachwithrob.com) · Károly Zsolnai-Fehér (https://cg.tuwien.ac.at/~zsolnai/) · Josh (@JoshKale) · Michal Cichra · Satya Nadella · George Cameron · Sarah Sachs · Geoff Huntley · Igor Costa · Vamsi Ramakrishnan · Sam Altman
ToolsAxiom Math · Lean · Kubernetes · Docker · MCP · Claude Code · Turbopuffer · Voyage Code 3 · Claude Opus 4.8 · Azure · Microsoft 365 · Cucumber · ADRs · Artificial Analysis
PapersAnthropic Opus 4.8 system card