Saturday, May 9, 2026
Follow builders, not influencers
5 stories · 3 min read
The compute wars are evolving into platform wars. While companies fight over GPUs, the real question is who builds the infrastructure that developers actually want to use when those million-token models are ready for production.
01Google's Gemini architect leaves after building a comeback
Madhu Guru, one of the key architects behind Google's Gemini AI models, announced his departure from the company. Guru helped build Gemini from scratch over the past three years, transforming Google from an AI underdog to a frontier competitor. In his goodbye post, he reflected on building "the playbook for building AI models, the customer feedback flywheel, and the enterprise business" that culminated in Gemini 3.
Why it matters: Guru joined when OpenAI and Anthropic were dominating, then helped Google catch up. His departure signals either mission accomplished or a bet that the next phase of AI competition happens somewhere else entirely.
Source →
02YC's Garry Tan ships browser agents that dodge bot detection
Y Combinator CEO Garry Tan released GStack v1.28, adding download capabilities and anti-bot detection for AI agents running web browsers. The update lets agents run in "headed mode" on headless Linux containers and includes an llms.txt file so other AI systems can understand GStack's capabilities without guesswork.
Why it matters: When one of Silicon Valley's top startup accelerators is building tools to help AI agents evade website security measures, we've crossed a line. Every website's "are you a robot?" question just became a lot more complicated.
Source →
03Together AI tackles the million-token serving problem
Together AI published a deep dive on serving DeepSeek-V4's million-token context windows, calling it fundamentally "a serving-systems problem." The post covers compressed KV layouts, prefix caching, and kernel optimizations needed to make million-token inference actually usable in production on NVIDIA's new HGX B200 hardware.
Why it matters: DeepSeek-V4 can theoretically handle a million tokens, but Together AI is doing the unglamorous engineering work to make it actually fast enough to use. The gap between "technically possible" and "production ready" is where real businesses get built.
Source →
04Anthropic positioning for the infrastructure war
Every founder Dan Shipper recorded a podcast dispatch about Anthropic's evolution from AI model company to full developer platform. He argues Anthropic is turning their API into "full cloud infrastructure for developers" with managed agents and compute deals, directly competing with cloud providers rather than just model companies.
Why it matters: Yesterday we covered Anthropic's quality control tools for agents. Today, analysts are calling it a full platform play. When AI companies stop selling models and start selling infrastructure, Amazon and Microsoft should pay attention.
Source →
05OpenAI board member breaks down safety decisions
Investor Matt Turck interviewed Zico Kolter, OpenAI board member and Carnegie Mellon machine learning department head, on AI safety and frontier model governance. The conversation covered OpenAI's preparedness framework, model release reviews, and Kolter's view that "AI safety does not come from scale" — contradicting the common assumption that bigger models are inherently safer.
Why it matters: This is rare insight into how OpenAI's board actually evaluates model releases. When a board member says safety doesn't improve with scale, it suggests the safety work happens in post-training, not in making models bigger.
Source →
Get this in your inbox
3-5 stories. 3 minutes. Every weekday morning.