35,848 results · Most relevant
Bostrom argues that a short delay (weeks to months) at the moment a key AI breakthrough lands can be highly valuable: a few extra weeks to test everything properly and crank capability increases up gradually could produce a large reduction in risk.
attributed
To get compute access, you fill out a request form and then receive an API key for the cluster.
Jesse Hall encourages the audience to run their own agents through the same benchmark used in the talk to catch surprises and get their scores, and to let LiveKit know how their agents perform.
The air bike is a big favorite for conditioning because it allows pushing specific conditioning with absolutely zero technique. You're in fixed fulcrums with fixed feet and hands, so you can't do it wrong — all you can do is not go as fast or go too hard. Any athlete, even the speaker's parents in their late 60s, can go to town on it straight away. Unlike 400s or hill sprints, you don't need to be good at sprinting to get conditioning benefit.
The creator notes that the biggest challenge with compliance agents is that they always start in the repo where the agent lives, so they may lack permissions in other repositories; the fix is to actually start the agent in the other repo so it has permissions there.
The creator recommends the Orchestrate command in the Copilot app as horizontal scaling for cross-repo work: set GitHub issues as goals, spin up sub-agents in the repository they're supposed to be working in, and keep pre-canned agents on task in individual repositories.
The creator recommends setting up an alert so that every time an agent exceeds its P90 for expected cost to run, it goes to investigate why it exceeded its threshold and reports debugging information for further evaluation or improvement.
The creator recommends that when a tool call is failing over and over, you A-B test different front matter or different models to see if they produce the right output, so you can pick a cheaper model.
The creator recommends pulling pre-built pipeline agents off the shelf from the GitHub Next organization's Agentix project — around 30 different agents including code deduplication and various reviewer agents (QA reviewer, joke-making reviewer) — and dropping them into your pipeline rather than building from scratch.
You can ask Codex to read the repo and dive really deep into how the harness and app server work behind the scenes, look at the actual system prompt of Codex in the repo, and fork the open-source harness and app server to make it your own.
Dominik Kundel recommends using dynamic tools to expose additional functionality to the agent, such as game actions like granting armor, health, or weapons.
On a per-thread basis you can modify any other configuration for your agent, such as setting a different model provider, choosing a different model, or fine-grained sandbox control over network connections and which files it can modify.
Dominik Kundel notes you can connect the Codex harness to any other model provider with a Responses API compatible API, and use Codex from within Claude Code via a plugin that uses the same app server and harness behind the scenes.
With the app server protocol, apps can offer sign-in with ChatGPT so users bring their own ChatGPT subscription and usage goes off their Codex usage limit. If you've used something like T3 code or similar, you were already using this through the app server protocol.
Pedro Lopez advises against MCP elicitation for connecting tools and credentials. Elicitation is a spec feature for requesting information from the user; form mode is not applicable for sensitive data like API credentials, and URL mode — recently available and similar to Airbyte's link-based credential flow — has the problem of inconsistent client support (Claude Desktop doesn't support it), forcing Airbyte to do its own thing.
When shipping a product, measure the reliability of the product before giving it to customers, and aim for around 92% reliability before expecting customers to fully hand over workflows to your agent.
Asking for a whole not-fully-thought-out feature set back in one shot means asking without really understanding what you're asking for. Gabriel Martinez advises not to throw a whole not-fully-thought-out feature set back in one shot without really understanding what it is you're asking for.
Gelsinger advises against committing capital to new datacenter/GPU projects without first securing the energy to power them — energy will be an upper bound, and more project defaults (like the Oracle/Comet case) are coming.
When she subtly implies she's in a rush (food, dinner plans), recognize you only have a short window — plan for roughly a five-minute interaction instead of trying to bounce to another location.
If she starts to leave mid-interaction (e.g., to get her food), position yourself slightly off to the side and reopen when she returns. If she completely leaves and avoids you, accept it gracefully — 'that's all good.'
Choose sources → to get a library of your own and a daily brief.