X / Twitter
Swyx (@Latent Space)
Swyx posted a chart comparing Anthropic and OpenAI's valuations and revenue: OpenAI at approximately $85 billion valuation and $30 billion ARR, Anthropic at approximately $90 billion valuation and $44 billion ARR. He noted different accounting methods: using OpenAI's method, Anthropic's actual ARR might be $8–10 billion lower. He also recommended standout talks from AI Dot Engineer.
https://x.com/swyx/status/2051440392722391180
https://x.com/swyx/status/2051329252344369626
Peter Yang, Roblox Product Lead
Yang announced he had finally downloaded Hermes and begun trying it, asking the community for real comparisons with OpenClaw. He emphasized that he wanted honest assessments, not sales pitches. The post received nearly 500 likes, reflecting strong interest in both agent products.
https://x.com/petergyang/status/2051129249348894754
Amjad Masad, Replit CEO
Masad demonstrated Replit's current agentic parallelism: 10 active agents running simultaneously, 198 in draft state, and more than 700 completed agent tasks. He considers it one of the internet's largest environments for parallel agents. He also shared an impressive case of someone building a multimodal learning platform for deaf, nonspeaking students using AI.
https://x.com/amasad/status/2051167532523074015
https://x.com/amasad/status/2051406536443035922
Garry Tan, Y Combinator CEO
Tan posted a series of reflections on "personal AI sovereignty." He sees personal AI's ultimate goal as enabling every AI-augmented individual to do meaningful work without capture by extractive institutions: freely writing their own prompts and owning their data. He calls this a central battleground of the new era and the point of open-sourcing GBrain: only by owning and running one's own prompts and data can one truly retain independent thought.
https://x.com/garrytan/status/2051099735176659256
https://x.com/garrytan/status/2051110206466302136
Builder Zara Zhang
Zhang described a fundamental AI-era change: before AI, making a small thing was unaffordable because software development was so expensive. You had to recruit a team, persuade others, and report to committees. Now it is just you and a coding agent. This changes the definition of what is worth building.
https://x.com/zarazhangrui/status/2051155065331941873
Nikunj Kothari, FPV Ventures Partner
Kothari called Gemini Flash's value for money astonishing: cheap, capable, supporting a million-token context and structured output, and his most-used production model. He also sharply criticized the 2023–2025 startup wave: too many teams focused on flashy launch videos and distribution while neglecting equally important product-market-fit validation. They obtained VC money but spent it in the wrong places.
https://x.com/nikunj/status/2051321911741972900
https://x.com/nikunj/status/2051349526171287930
Guillermo Rauch, Vercel CEO
Rauch announced open-sourcing `npx deepsec`, an agent orchestrator for deep security reviews. Vercel originally built it internally, validated its value by running it on several major open-source projects, and decided to release it publicly. This marks a new stage in coding agents' security applications. The post received more than 1,200 likes.
https://x.com/rauchg/status/2051386798899888539
Aaron Levie, Box CEO
Levie noted that Anthropic and OpenAI are both introducing initiatives to help enterprises deploy agents internally. He believes agents' expansion beyond coding into knowledge work will generate substantial demand for supporting upgrades. The trend is just beginning but will grow rapidly.
https://x.com/levie/status/2051344780328858040
Sam Altman (@OpenAI)
Altman announced a surprise benefit for everyone who applied to attend the GPT-5.5 launch event but could not get a seat.
https://x.com/sama/status/2051318922805436896
Podcasts
Training Data — Waymo's Dmitri Dolgov: 20 Million Rides and the Road to Full Autonomy
Training Data interviews Waymo co-CEO Dmitri Dolgov, a technology leader who has persisted in autonomous driving for more than twenty years. Born in the Soviet Union and raised in the United States, he returned to Moscow Institute of Physics and Technology, MIPT, to study mathematics and physics, later earned an AI doctorate, and found his lifelong direction at the 2005 DARPA Urban Challenge.
Dolgov recalled the 2009 launch of Waymo's predecessor, Google's self-driving car project. The initial team of approximately 12 had two goals: more than 100,000 fully autonomous miles and 10 intervention-free routes of 100 miles each in the San Francisco Bay Area. Working around the clock in shifts on code and hardware, they completed the goals in 18 months. He calls it the happiest period of his career.
Reflecting on repeated AV booms and downturns, Dolgov identified a pattern: breakthroughs such as convolutional networks, Transformers, and LLMs bring fast early progress and investment enthusiasm, but AV's fundamental challenge is the "last mile." The path from 80% to 100% is long and difficult. The reality that road accidents claim a life every 26 seconds motivates persistence. He emphasized that avoiding a "silver bullet" mentality and clearly understanding the problem's depth are key to resilience through dark periods.
On architecture, Dolgov introduced the Waymo Foundation Model, the AI system's core with three pillars: driver, simulator, and critic. It aligns closely with today's world-model concept, though Waymo has worked in this direction for years. The model must understand physical dynamics, what constitutes good driving, and driving's effects on other road users. Waymo has completed more than 20 million real rides, creating a substantial data-based competitive barrier.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8