Tuesday, May 19, 2026
The Agentic Era has arrived.
May 19 · 12 videos
Stripe Sessions 2026 launched today.
Patrick Collison says money is data.
Agents are now autonomous economic actors.
Sam Altman calls OpenAI a token factory.
Stripe volume hit $1.9 trillion.
Defense is the next trillion dollar category.
“In Lisp, the core idea is that code is data. Well, the basic idea in Stripe is that money is data.”
AI Dev 26 x SF: Emma McGrattan: Engineering the Context Layer
Emma McGrattan · DeepLearningAI · 15 min
Watch on YouTube →Emma McGrattan argues the context layer is the most critical component of enterprise AI. It grounds LLMs in business reality through hybrid architectures.
- Gartner reports the average enterprise manages over 400 data sources.
- Cloud latency ranges from 20 to 200 milliseconds.
- Regulatory mandates like HIPAA often require on premises or edge deployments.
- Egress costs are a major financial burden for scaling AI applications.
- Multimodal retrieval is expected to arrive within 12 to 18 months.
- Silent failures like hallucinations are more dangerous than system crashes.
Stripe Sessions 2026 | Keynote
Patrick Collison · Stripe · 87 min
Watch on YouTube →Patrick Collison and the Stripe team announce the Machine Payment Protocol for agentic commerce. Stripe now processes 1.6 percent of global GDP.
- Stripe payment volume reached $1.9 trillion with 99.9996 percent API availability.
- The Machine Payment Protocol enables HTTP based agent transactions.
- Tempo blockchain facilitates real time streaming payments via stablecoins.
- Tokens Paid as Burned model eliminates unpaid inference bill risks.
- Platform Growth Studio provides benchmarking against the entire Stripe network.
- Link Wallet for Agents provides secure authorization for autonomous actors.
Sam Altman in conversation with Patrick Collison
Sam Altman · Stripe · 56 min
Watch on YouTube →Sam Altman discusses OpenAI's transition into a mega scale token factory. He emphasizes user empathy over technical skills for future founders.
- GPT 5.5 and Codeex are becoming the primary interfaces for human computer interaction.
- OpenAI aims for a low margin infrastructure model similar to a utility.
- Material science is identified as the most underrated domain for AI transformation.
- Successful enterprise adoption must be led by CEOs automating their own jobs.
- GPT 4 was held internally for 8 months after training before its release.
- The era of technical only founders is shifting toward those with deep user empathy.
Nat Friedman and Daniel Gross in conversation with John and Patrick Collison
Nat Friedman · Stripe · 55 min
Watch on YouTube →Nat Friedman and Daniel Gross discuss the Golden Age of Tinkering. They argue the primary goal of AI labs is automating the research process itself.
- Cycle time from idea to user feedback is the most important organizational metric.
- Global compute capex is approaching 1 percent of global GDP.
- Frontier models like Claude Code allow individuals to reverse engineer hardware.
- Process heavy environments are described as an indignity to top tier engineers.
- Agents are predicted to leapfrog legacy banking for stablecoin based economies.
- Beauty and aesthetics in data center design are essential for public acceptance.
Stripe Sessions 2026 | Indexing the economy
John Collison · Stripe · 31 min
Watch on YouTube →John Collison analyzes the 2026 global economy through Stripe's data. He explores how AI reduces coordination costs and enables solopreneurship.
- The software sector lost $1 trillion in market value in January 2026.
- Five million Americans now run one person businesses generating over $100,000.
- The Solo Paradox suggests productivity gains may lag technology by 30 years.
- Proprietary data becomes exponentially more valuable when agents can reason over it.
- Revenue growth for the Stripe Atlas class of 2026 is 5x higher than 2024.
- Software is becoming more abundant and substitutable, shifting rewards to profitability.
AI Dev 26 x SF | Marc Brooker: It's Time to Be Right
Marc Brooker · DeepLearningAI · 14 min
Watch on YouTube →Marc Brooker explains why knowledge work agents are limited by defect rates. He advocates for combining neural models with symbolic frameworks.
- The industry must shift from flashy demos to low rates of low consequence defects.
- AWS is investing in Hydro for distributed systems and Cedar for policy authorization.
- Auto formalization turns ambiguous natural language into mathematically precise specifications.
- AI safety should be treated as operational reliability including durability and cost.
- Operational metrics should replace superficial benchmarks like Pass@10.
- The agent loop allows building great systems on top of faulty components.
Burnham, Streeting & The Fight to Be Prime Minister
Rory Stewart · The Rest Is Politics · 46 min
Watch on YouTube →Rory Stewart and Alastair Campbell analyze the leadership threats to Keir Starmer. They explore the rise of populism in Germany despite high quality public services.
- The AfD party has reached 41.4 percent support in regions like Saxony Anhalt.
- Reform voters are characterized as 85 percent former Tory voters with low university degree rates.
- Public service quality is a tangible metric for political popularity in regional mayoralties.
- Anti corruption campaigns are cited as the most effective way to combat right wing populists.
- Mainstream parties can survive on 28 percent of the vote in a five party system.
- Economic prosperity does not insulate a country from identity based populism.
What Breaks When You Build AI Under Sovereignty Constraints - Bilge Yücel, deepset GmbH
Bilge Yücel · AI Engineer · 19 min
Watch on YouTube →Bilge Yücel defines the four pillars of AI sovereignty for organizations. She explains how moving away from vendor dependence breaks existing architectures.
- Technical sovereignty requires control over data flow, model choice, and infrastructure.
- Sovereignty is a spectrum ranging from SaaS to air gapped environments.
- Haystack provides an orchestration layer for model swappability and auditable version control.
- European providers may offer regulatory advantages for EU based organizations.
- True sovereignty means responding to incidents without third party vendor support.
- Retrofitting sovereignty often requires a full re evaluation of prompt engineering.
Don't Build Slop (4 Levels of AI Agent Maturity) - Ara Khan, Cline
Ara Khan · AI Engineer · 18 min
Watch on YouTube →Ara Khan warns against building slop and introduces a four level maturity framework for agents. He argues for a Kanban style UX for long running tasks.
- Every addition to an agent prompt risks making the performance worse.
- GPT 5.3 system prompts are roughly one third the size of GPT 5 prompts.
- Humans are transitioning from individual contributors to engineering managers of AI agents.
- Agents are often inference bound and run for 8 to 10 minutes per loop.
- Cloud native fleets allow for scaling agentic workflows to thousands of concurrent operations.
- Frameworks like LangChain are excellent for finding product market fit but may lack production modularity.
Is Defense the Next Trillion-Dollar Category? | a16z American Dynamism Summit
Dino Mavrookas · a16z · 12 min
Watch on YouTube →Dino Mavrookas and Michael Duffey discuss the shift to autonomous first defense manufacturing. They advocate for a commercial first acquisition strategy.
- Saronic autonomous vessels require 50,000 labor hours compared to 7 million for a destroyer.
- The Pentagon is moving toward private capital investment to expand production capacity.
- Fragility in defense stems from sole source suppliers that are not profitable.
- Manufacturing instructions should be simplified like IKEA manuals for quick workforce training.
- Founders are urged to build for a 100 year generational timeframe.
- Commercial markets should provide wartime production capacity during peacetime.
Zepto: How Two 17-Year-Olds Built India's Largest Seller Of Fruits and Vegetables
Aadit Palicha · Y Combinator · 28 min
Watch on YouTube →Aadit Palicha and Kaivalya Vohra explain how Zepto became India's largest quick commerce platform. They emphasize controlling dark store infrastructure.
- Zepto processes millions of deliveries daily and employs over 200,000 people.
- Controlling the end to end experience resulted in 10x volume improvements over intermediaries.
- The company scaled a high margin advertising business to hundreds of millions in ARR.
- AI is used for supply chain forecasting to reduce external software spend to nearly zero.
- Extreme customer delight is the precursor to lowering customer acquisition costs.
- The naivety of being a young founder is an advantage for challenging impossible constraints.
Personalization in the Era of LLMs - Shivam Verma, Spotify
Shivam Verma · AI Engineer · 20 min
Watch on YouTube →Shivam Verma details Spotify's transition to a unified generative framework for personalization. The system treats the music catalog as a vocabulary for LLMs.
- Spotify manages 750 million monthly active users and over 100 million tracks.
- Semantic IDs compress complex content vectors into hierarchical tokens for the LLM.
- Interaction history is treated as a prompt for foundational user modeling.
- Users gain steerability through natural language interfaces like the AI DJ.
- Domain adaptation of open weight models is preferred over building from scratch.
- Soft tokenization projects user embeddings directly into the LLM latent space.
References
PeoplePatrick Collison (x.com/patrickc) · Sam Altman · Nat Friedman (x.com/natfriedman) · Daniel Gross (x.com/danielgross) · John Collison (x.com/collision) · Marc Brooker · Bilge Yücel (x.com/bilgeycl) · Ara Khan (x.com/arafatkatze) · Dino Mavrookas (x.com/MavrookasD) · Michael Duffey (x.com/USDASDuffey) · Aadit Palicha · Shivam Verma (x.com/kaffeinated) · Andrej Karpathy (x.com/karpathy) · Erin Price-Wright (x.com/espricewright) · Erik Torenberg (x.com/eriktorenberg)
ToolsStripe · OpenAI · Claude Code · Haystack · Hydro · Cedar · Lean · Zepto · Spotify · Llama · Qwen