Daily Brief

Once GPT 5.3 Codex Spark started serving a thousand tokens per second on Cerebras, inference stopped being the bottleneck and the network became it. The answer was websocket mode: a persistent connection replacing server sent events over HTTP, carrying stateful context so a turn ships back only the tool call result instead of resending every item. The same pressure shapes context construction, which fights size, flexibility and cachability at once. Tools can be marked deferred so they never enter the context window and surface through tool search when the model actually wants them, and the ava

A frontier scale checkpoint is around 500 GB, so shipping one to a rollout fleet in another region takes minutes to hours and kills any hope of weight updates landing in seconds. Nan Jiang's claim is that you can send roughly 500 MB instead and have the rollout engine reconstruct a bitwise identical weights version. Fewer than 1% of rollout visible weights actually change between consecutive versions, and the reason is not that gradients are sparse. Gradients are dense, about 99% of parameters get a nonzero gradient and the FP32 master update is dense too. It is just small. The mechanism is a

Leave agent adoption to individuals and the engineer shipping two PRs a day ends up reviewing the ten that the early adopter ships. They fall further behind, the code they are reading is worse, and they conclude the agents are the problem. Aditya Khandelwal's argument, from leading a team of ten through this, is that adoption is therefore a leadership problem and not an IC one, because the changes that actually work, restructuring a codebase for progressive disclosure and converging on a shared setup, are not changes one engineer can make alone. The symptoms of a bad setup are specific. Engin

Sequoia Capital partner Sonya Huang opens our Own Your Intelligence event with the case for why more companies are choosing to own their intelligence, down to the weights. She lays out the four forces driving the shift: cost, speed, performance, and controlling your own destiny, and explains why the race for the application layer is becoming the race for the intelligence layer itself. Sonya shares an opinionated framework for deciding what to own vs. rent when assembling your AI stack, and another for how to get going from zero to one. With today's open-weight models near the frontier, she ar

How does an application company compete with frontier labs that have more money, talent, compute, and data? At Sequoia Capital’s Own Your Intelligence event, Harvey co-founder and president Gabe Pereyra shares the playbook: leverage the frontier ecosystem instead of building everything yourself. Gabe walks through how Harvey built its research lab, starting with benchmarks like LegalAgentBench and its open-sourced diligence dataset, using domain experts to guide synthetic data generation, partnering with multiple neolabs for post-training, and building the serving and evaluation infrastructur

a16z's Joel De La Garza is joined by Emilio Escobar, Chief Information Security Officer at Datadog, to discuss what it takes to secure a company where nearly every employee is using AI and more than 4,000 engineers are working with coding agents. Rather than trying to block new tools, Emilio explains why Datadog chose to embrace AI early and build the security infrastructure needed to use it safely. They unpack how AI changes traditional assumptions around data permissions, credentials, developer access, and software supply chains. Emilio shares how Datadog uses role-based MCP servers and eph

Sonnet 4.5 developed what Anthropic's Applied AI team came to call context anxiety: approaching its context window limit, it would wrap work up early and stop with room to spare. They built context resets into the harness to compensate. Then Opus 4.5 shipped without the behavior, and the fix turned into pure overhead, adding latency and discarding cache it should have kept. That is the principle Gagan Bhat and Isabella Kai He build the whole session on: a harness encodes assumptions about what the model cannot do on its own, and those assumptions go stale as models improve. The architectural