35,595 results · Most relevant
Kyle's recommended open: 'Sorry, didn't mean to startle you,' followed by an intentful open like 'I was actually headed that way, but I just saw you walking and I thought you had a very cute, sporty look to you.'
When she has your undivided attention, take attention off the fact that she was leaving by making a new observation about her, a playful assumption, teasing her a little, or injecting personal input — instead of asking for a 'quick rundown,' which pressures her to summarize herself.
When a woman is walking quickly, jog up to close the distance so you hit the approach at the right time and location, rather than walking behind her and missing the window.
Dylan Cousin recommends using a vector search engine that embeds directly into your app: it opens a local store inside your process, writes embeddings with payloads, and queries offline in sub-millisecond time. He notes he works on such a piece (Qdrant) and that it's just one way to do it.
Implement two-factor authentication, managed authentication for SaaS, and granular file system permissions for AI agents. Disable auto-run features from removable media and require devices to update their operating system before allowing any other functionality upon first boot.
The host advises that the best way to develop range is not to try everything at once, but to focus on just one slightly challenging mode at a time. He challenges the audience to pick one characteristic (enthusiasm, silliness, shadow sharing, or selflessness) and make space to express it in their day.
Jeff Cavaliere recommends the hand-release push-up, also called the Superman push-up, to correct rounded shoulder posture. The exercise forces a full range of motion, engages the serratus anterior by extending fully at the top, and includes a Superman portion where you squeeze the glutes and lift the thighs off the ground. Beginners can perform it from their knees.
Jeff Cavaliere recommends performing a bridge with a reach-over to address weak core and weak glutes. The movement involves driving the hips up, reaching one arm across and pointing to the floor behind the opposite shoulder to create thoracic rotation, then lowering and lifting the legs into 'heels to the heavens'. He advises swiping hands on the floor to check if you are lifting high enough, moving the pelvis (not just the legs) to target the core, and doing as many reps as possible, aiming to fatigue the core before the glutes. If the glutes fatigue first, it indicates weak glutes that need attention.
The source endorses asking your champion to kick off procurement, security, and legal at the same time instead of one after the other the moment you are picked as vendor of choice — a move that alone can turn a six-week review into two weeks.
The host endorses using a redline deadline — a set of deadlines for when the customer's legal team delivers their first round of redlines, when the seller returns a response, and when the customer delivers their second round — calling it one of the most effective tools for project managing customers through the legal process. Setting it incentivizes the customer to hit those dates even without a give-get attached, and gives the champion a way to enforce moving through the review queue for their legal team.
The host endorses pre-scheduling the kickoff call before the contract is signed (deal cheat code #17). Pretty much everyone will say sure, let's go do that; a signed contract is required to hold the slot; and the pre-scheduled call gives the champion a reason to hold their legal and procurement teams accountable to the agreed timeline. The kickoff call covers data migration, integrations, and related implementation topics with the implementation team.
Ari Weinstein describes OpenAI's enterprise controls (part of Astra) that let businesses choose how much human-in-the-loop approval the agent requires, including a Guardian system where a smart model auto-reviews every agent action for safety and alignment with original intent.
James Sun describes enterprises as a natural fit for computer-use agents because they run legacy software around 15-20 years old with no MCP tools and confusing UIs that are hard for normal people to operate, and notes companies are building skills that teach the models to use that software better.
James Sun describes Astra as OpenAI's most aligned model, much better at resisting prompt injection and avoiding unintended actions, which he says makes enterprises feel safer using it for sensitive work like procurement or vendor communication.
Ari Weinstein sees promising enterprise automation use cases in two big categories, starting with automating tedious workflows: ERPs, CRMs, and every system of record; finance, where many financial software programs lack APIs for quickly pulling data and teams do manual month-end reconciliation across multiple ledgers; procurement, where teams fill out giant spreadsheets that ops manually enter into internal software before the procurement team contacts vendors; and sales, where system-of-record data entry is the tedious part agents can automate. Within company-approved software, he notes, automation is straightforward because companies want people to automate tedious work.
hedged
Daksh Gupta describes how Greptile disintermediates all context sources with a living wiki that acts as a world model of how the company works, incorporating codebase knowledge, production telemetry and logs, and roadmap information from documentation and Linear tickets.
Kevin Hou announces that agent teams are in public preview inside Antigravity: typing the slash command '/teamwork' enters a new mode that unleashes a swarm of agents onto the task at hand.
Kevin Hou reports that researchers were able to automate 90% of the side-by-side eval workflow by simply asking the agent about the evals in question using natural language; all the user had to do was load up a skills file and ask away.
Roland Gavrilescu lays out what he says you need to have a frontier agent: codified evals, a way to run evals, loops to continuously improve those evals, processed signals, the right tools for certain models, and different harness profiles for different models.
OptimizeAnything supports single-problem mode for one problem like a single matrix multiplication kernel, multitask search mode for n related problems where information transfer may help, and build-a-skill mode when optimizing on a set number of problems but deployment presents many new problems and generalization matters.
hedged
Choose sources → to get a library of your own and a daily brief.