Catch up on Y Combinator videos in minutes

Roman Ugarte explains that his team incubated Grok Bot in seven weeks to bring autonomous engineering agent workflows to non-engineers. They designed the product around three principles: organizing work around bot identities rather than chats, giving bots their own computers and tools, and keeping them always on. Launching internally produced strong adoption across both AI power users and non-technical staff.

Roman Ugarte explains that his team incubated Grok Bot as an internal prototype in seven weeks to bring autonomous engineering agent workflows to non-engineers. They designed the product around three principles: organizing work around persistent bot identities rather than chats, giving bots their own computers and human-equivalent tools, and keeping them always on via proactive loops. Internal users across the technical spectrum quickly adopted the tool before its public launch.

Tyler Shultz recalls meeting Elizabeth Holmes at his grandfather George Shultz's Stanford home, where her charismatic persona and vision of running any lab test from a single drop of blood completely sold him. Convinced she was the next Steve Jobs, he switched his Stanford major to biology to work at Theranos.

In a live audience Turing test built by Francois, 33% of humans were mistaken for AI and agents fooled roughly 30% of interrogators. Shifting to biology, he argues that falling sequencing costs and foundation models like Evo 2 and AlphaFold enable recursive self-improvement loops across both in silico and wet-lab drug discovery.

Before launching full vibe coding, Replit generated about 50% to 60% of user code via autocomplete while non-programmers manually copied code and errors between ChatGPT and Replit. Watching users in the office inspired Amjad Masad to automate the loop until the program worked and provisioned its own databases.

Blake Scholl says outsourcing Boom's supersonic engine to Rolls-Royce was his worst decision, driven by industry pressure to partner with a major manufacturer. When Rolls-Royce's board fired its CEO in early 2022, internal succession politics stalled the contract and left Boom with only a temporary cover story.

Paul Copplestone reports that Supabase grew from roughly 3 million to 13 million signups and from 8 million to nearly 50 million managed databases in a year, with revenue projected to grow fourfold. Approximately 70% of new activity comes from AI agents rather than human developers clicking dashboards. This surge marks a fundamental shift from building software for people to building infrastructure that agents choose.

Marvin Schwaibold explains that while design remains highly iterative, his workflow has shifted from crafting static layouts in Figma to driving AI tools like Codex and Claude Code in the terminal. Figma now serves as a quick scratchpad before moving into live demos on a phone or computer. Because AI handles technical execution, the designer's primary job is acting as a selective editor who curates what actually looks good.

François explains that serving models is the primary bottleneck in AI and has created $220 billion in market cap over the last decade. Citing DeepSeek V3 figures of 2.8 million H800 training hours versus 43,000 daily inference hours, he calculates that a weight file consumes 6x more compute per year in inference than training, or 12x over a two-year lifespan.

Farza Majeed argues that startups no longer need TechCrunch press or $20,000 produced commercials because users now live on algorithmic platforms like X, LinkedIn, and YouTube. Instead of treating a launch as a one-time rocket event, his team works backwards from weekly feature demos and drafts the social caption and hook during development. He concludes that how a team communicates a feature is just as important as building it.

Alex Danilowicz explains that he and his co-founder entered Y Combinator with a revenue-generating text message analyzer before their YC partner advised them to build around their core expertise as front-end engineers. After nine months of pivoting through component editors and localhost sharing extensions without finding ten paying users who loved the product, they launched Magic Patterns in October 2023 alongside v0 as one of the earliest prompt-to-design tools.

Francois argues that public discussion of AI safety has drifted from falsifiable scientific rigor into unmeasurable hysteria. He notes that past technological leaps, from the atom bomb to Alan Turing's 1950s computing predictions and Bob Metcalfe's internet collapse warning, triggered similar catastrophic forecasts that failed to materialize.

Cisco is transforming into the essential picks-and-shovels infrastructure provider for the AI era by delivering secure, high-scale networking and data systems. Patel explains that the secular shift toward artificial intelligence requires unprecedented compute and data management. Cisco aims to anchor this ecosystem by providing the underlying safety, networking, and observability apparatus.

Modern artificial intelligence hardware has become bottlenecked by memory capacity and bandwidth rather than pure computational throughput. Standard GPUs have seen diminishing energy efficiency gains per gigaflop in recent generations, while transformer attention scaling exacerbates power consumption. Consequently, deep learning requires alternative paradigms that emulate biological efficiency.

Analyzing companies accepted over the past 12 to 18 months, Diana reports that hard tech startups rose from 8% to 20% of the Y Combinator batch while median monthly revenue by batch end jumped from $8,000 to $20,000. Physical categories like robotics, US manufacturing, defense, semiconductors, and power have tripled or quintupled.

Although critics dismiss agent harnesses as mere prompt engineering or wrappers, changing the harness yields an 18% performance jump and drives recent gains in how long agents can run autonomously. On the ARC-AGI fluid intelligence benchmark, Claude Opus scored 30% on the private holdout, whereas adding Prime Agent's harness pushed performance to 95% and Nvidia's AVO reached 100%. These results demonstrate that test-time scaffolding unlocks capabilities already latent in fixed model weights.

Jeffrey Morgan reports that Ollama Cloud token usage jumped 150x since the start of 2026, propelled first by coding agents and then by co-work tools like OpenClaw and Hermes. Enterprises such as AT&T have reportedly shifted 40% of their token consumption to open models to cut costs and customize workflows. Expanded context windows over one million tokens and capable out-of-the-box Chinese models turned open models into mainstream enterprise workhorses.

Max Junestrand states that Legora is an agentic operating system used by over 3% of the world's lawyers, competing against legacy legal software that looked built in the 1990s. Since launching into general availability in October 2024, Legora grew from $1 million to $100 million in ARR and expanded from three engineers in Sweden to over 750 employees globally. He concludes that starting a global software powerhouse from Europe adds a special point of pride to their rapid growth.

Francois Chaubard argues that inspecting training data matters far more than tweaking model architectures once expressive transformers are deployed in production. While academic PhDs spend 95% of their time on architecture, production teams flip that ratio to focus up to 97% on data curation. Real-world edge cases like fogged retail fridges cannot be solved by adding neural network layers.

Michael Kratsios explains that working for Peter Thiel for nearly seven years showed him how federal and local rules blocked portfolio companies like Lyft and SpaceX. He argues that government is the only arena where one can solve problems at national scale by unlocking commercial drones and autonomous vehicles. This realization drove his transition from venture capital into White House technology leadership.

While welding a life-size razorback hog in her garage, art history PhD Susan Kare was invited by high school friend Andy Hertzfeld to interview at Apple. She prepared bitmap sketches on graph paper and brought library typography books, securing the job and later seeing that notebook enter MoMA's collection.

Chelsea Finn argues that while software AI products like ad ranking, ChatGPT, and coding agents remain useful even when imperfect because human users filter their outputs, physical robots directly alter the physical world and must operate autonomously with far fewer mistakes. She points to Waymo surpassing 250,000 weekly autonomous rides as proof that machine learning can achieve trustworthy physical autonomy, setting a >90% reliability target for complex tasks like making espresso.

Frustrated that he could not check on his coding agents from his phone while away from his desk in November, Peter Steinberger prompted a model to build a Mac-to-WhatsApp relay in an hour. By stripping away terminal complexity and prompting the weights to act like a proactive friend, the tool felt like magic to him even when his Twitter followers initially ignored it.

Francois Chaubard argues that despite a decade of breakthroughs from AlphaGo and MuJoCo to ALOHA and Vision-Language-Action models, general robotics remains unsolved due to four fundamental walls. Video world models fail to respect real-world physics, action space representations remain inefficient, robots lack the dense full-body tactile sensing of human skin, and hardware suffers from embodiment drift as actuators corrode and batteries degrade.

Max Hodak introduces Science's retinal prosthesis, a sub-retinal solar-cell chip paired with camera and infrared laser glasses that bypasses dead rods and cones to restore vision. Following three clinical trials, one blind patient used the implant to read a 300-page novel. Hodak pivots from this clinical milestone to argue that operational logistics, rather than strategy alone, dictate startup outcomes.

Stephen Haney explains that Paper uses HTML and CSS as its actual rendering engine rather than a custom canvas engine. Because AI agents already have HTML and CSS in their training data, they can read and manipulate Paper files directly with lower token spend, faster execution, and higher accuracy. This architecture turns a human handoff improvement into a direct visual interface for agents.

Garry Tan recounts how 23-year-old Baruch Spinoza was excommunicated in Amsterdam in 1656, refused a 1,000-guilder bribe to stay quiet, survived a knife attack, and ground optical lenses by day while writing Ethics by night. He highlights Spinoza's concept of conatus, the innate drive in every living thing to persist and increase its power to act. Tan frames this independent striving as the core engine that modern startup tools are built to amplify.

While advising government space agencies at McKinsey, Philip Johnston visited SpaceX's Starbase in early 2023 and saw gigafactories built to produce three reusable Starships daily. Projecting 1,000 times more launch capacity at one-tenth the cost, he began evaluating which sci-fi concepts would become viable.

Dmitri Dolgov argues that building physical AI requires

Patrick Collison compares human knowledge to L1 processor cache, arguing that knowing facts in your own brain allows far faster reasoning round-trips than querying an AI agent. He adds that top companies and AI labs still place a massive premium on raw human cognitive ability.