Daily Brief

A single token of KV cache on Mistral 7B costs 131 KB. Multiply that by 16,000 tokens of context and 80 concurrent users and the cache alone wants 42 GB of GPU memory, which is why requests start failing on a 24 GB card. Harshul Jain, a senior software engineer at Audible, and Tanmay Sah, an independent AI researcher, spend this workshop building that number up from first principles. They open on three symptoms every team hits: memory that climbs with context length, time to first token that degrades as prompts grow, and throughput that collapses because a naive server answers requests one aft

Full episode: https://youtu.be/X50zezLFWWI Me on twitter: https://x.com/dwarkesh_sp

Laurie Voss reran a year old benchmark and the models walked straight through its ceiling without noticing it was there. IFScale asks a model to write a business report containing a list of exact words, then counts how many actually appear. A year ago frontier models started dropping instructions somewhere around 200 to 300, which is a hard limit on how much you can put in a skills file. Voss, head of developer relations at Arize AI and a cofounder of npm, first replicated that result on the three models from the original paper still reachable by API, then pointed the same test at the current

Adding a nice interface made the product worse. With plain text results the model would run ten or fifteen job searches, filter them, and assemble a table. Once a rendering widget was attached, it called the tool once, saw results already on screen, and stopped exploring. Dustin Mihalik is a technical fellow at Indeed working on AI platform and guardrails, and this is a lessons from the trenches account of building MCP apps for Claude, ChatGPT, and Indeed's own job seeker agent. He starts with why a UI is worth having. A text response carries no branding, and getting a host chat app to link ou

Alex Hancock built one of the clients in this talk the night before he gave it, which is roughly the point. Hancock is a software engineer at Block, works on the open source agent harness Goose, now donated to the Linux Foundation, and maintains the Rust SDK for MCP. His argument is that the agentic stack already has a good standard for agents reaching outward to do things, which is MCP, and that the strongest thing about MCP is not its design but the fact that everyone uses it. What is missing is the other direction, a standard way for client software to tell a harness what to work on and to

Full episode: https://youtu.be/X50zezLFWWI Me on twitter: https://x.com/dwarkesh_sp

Why porting a massive codebase from Python to TypeScript was the right move. Mike Krieger, co-lead of Anthropic and founder of Instagram, breaks down the unconventional engineering decision behind this switch. He details how integrating Bun and a new deployment strategy justified the migration despite the scale of the project. Subscribe for more deep dives into complex engineering choices.

Rachit Kataria and Will Wang are the co-founders of Centralize (W24), an AI-powered platform that automatically builds org charts of enterprise accounts to help revenue teams map stakeholders, spot gaps, and close complex deals faster. The company recently closed a $15 million Series A led by NEA. In this Fireide, Rachit and Will sat down with Diana Hu, Managing Partner at YC, to talk about how Centralize's "trust graph" ingests emails, calls, and CRM data to surface human dynamics inside a prospect's org — and how one customer compressed a six-month sales cycle into weeks to close an eight-f

法律虽然无法强迫夫妻之间的彼此尊重，但是必须要捍卫底线的道德。

本篇为神奇组织之1977纽约大停电。