33,811 results · Most relevant
Ryo Lu recommends that if you are building tools, you start from three small points: make important assumptions visible, make each edit small enough to feel safe, and make handed-off work retrievable when needed.
The narrator advises targeting keywords that people actually search for and that you can realistically rank for, rather than just publishing pages.
The narrator recommends starting with search, AI search, and video as your foundation, because once a page or video ranks it will keep sending free traffic for years. Get email going from day one to keep visitors coming back from all the other channels. Once that foundation is in place, add short-form video and paid ads on top as accelerators — you don't want to build everything on top of them, but they're great for reach and speed.
Adam Nahari advises working hand-in-hand with the CFO to calculate whether the cost of the initiative at scale — including compute, tokens, and the individuals involved — is significantly lower than the impact it would generate. Many pilots die at the unit-economics stage because they don't pass this test.
Adam Nahari looks for five things in a good AI use case: it moves a line the CFO recognizes; it is measurable with a baseline agreed up front, not in retrospect; the tech and data infrastructure is ready and the company owns the data; a named user — a person, not a department — wants it and can champion the initiative; and it is one of the CEO's top five priorities.
Adam Nahari's rule: before putting any money at risk, ask where the CEO is — whether they have committed publicly that the initiative is a company priority. If the CEO has not publicly committed, he does not fund the initiative; the correlation with success is that strong.
Adam Nahari recommends working hand-in-hand with the CFO to calculate whether an AI initiative's cost at scale — including compute, tokens, and personnel — is significantly lower than the impact it would generate.
Adam Nahari's pilot discipline: before starting a pilot, write down the kill metrics that determine whether it continues or stops, and let the CFO keep score. Use a time gate — give each stage 90 days, and if no value is seen in those 90 days, stop the initiative. Use a metric gate — A/B testing at each stage relative to a baseline decided in advance. Use an adoption gate — were over 50% of crew schedulers using the tool.
Adam Nahari advises: pick one or two problems, not 40, when driving an AI transformation. Start with the P&L and work your way backwards when choosing AI problems, and reimagine the entire process end-to-end rather than building on processes that may already be obsolete.
Measure time to first audio, not time to first byte, because a byte does not equal audio: popular TTS providers ship up to three quarters of a second of silence after the first byte, so a dashboard showing 210ms time-to-first-byte can still ship up to 750ms of silence. Similarly, measure time to first sentence for LLMs, not time to first token, because a TTS model needs an entire sentence before it can start speaking. Two seconds of silence after a caller asks 'did my payment go through?' is not reassuring, so measure the silence gap the caller actually hears.
Jesse Hall recommends using a turn detector that listens to audio as well as the transcript, so it can hear vocal inflection and wait when the caller is not actually done — for example when a payment amount is stated with a pause mid-sentence.
Jesse Hall advises setting the latency budget the caller hears before anything else, and measuring time to first audio and time to first sentence — not time to first token — as the key metrics, including measuring the silence gap the caller actually hears.
The narrator advises remembering that the default tab gets more attention than other tabs because users don't have to do anything to select and view it, and ensuring non-default tabs don't contain critical information for frequent workflows. Putting critical information in non-default tabs causes users to miss it.
The narrator advises using tabs only for content that has clear groupings, only when there are few groupings so users can see all options at a glance, and only for content that doesn't need to be compared. Tabs help chunk long content into scannable sections, which reduces cognitive overload.
The episode's overall method: have a pull hypothesis, fill out the template, schedule cold outbound with a surprising message that gets attention, do about five calls, and assess whether there's pull. If there's no pull, hopefully you've asked the Hail Mary question — 'Is there anything that we can do right now that is a problem on your to-do list?' — and learned something worth double-clicking into. The Hail Mary generates new information you can use to do another set of five calls; if it yields nothing, go back, get first-hand knowledge, and try again. Throw out Hail Mary questions, maybe enter goblin mode, come up with another pull hypothesis, and do five more calls, repeating until you find pull — and once you find pull, you have a business.
host's synthesis of the guest's process; the five-call heuristic is described as rough
Rob Snyder recommends setting a goal of two to three conversations per day with potential customers, noting that to get there you first have to get your first conversation. He suggests warm intrusions as one way to get early conversations.
None stated.
Rob Snyder recommends founders start by reaching out to potential customers on LinkedIn or via email. He discourages template messages ('hey, do you have XYZ pain points? Would you like to buy our game changing solution?') — no template you get online will get responses except 'leave me alone.' Instead, use 'founder magic' outreach: the message is basically 'you're interesting, I'm interesting, let's chat and see.' Keep it to a sentence or two, say something interesting to cause them to meet with you, and don't craft a whole value pitch or problem identification — the pull exists on the prospect's to-do list, and the message only exists to get them to talk to you. Every founder has their own version of an interesting message; it doesn't have to be super magical.
None stated.
Aparna Dhinakaran lays out a method for building always-on agents. First figure out what triggers the agents to wake up: an ongoing cadence (e.g., every three or six hours, pull the last X traces and identify problems), events such as Slack messages when a user complains, or metric-based triggers. When a run starts, pull in all context — traces plus third-party data like APM logs and metrics — and give the agent repo access and keys, all inside a sandbox that starts up each time the agent wakes, with a single proxy as the only way information gets in and out. Trace everything inside the sandbox, save every generated issue and output to memory each run so the next run doesn't start from ground zero, and save learned skills to the repo. After the run, tear down the sandbox, removecredentials
None stated.
Aparna Dhinakaran describes the AI engineer workflow: build the agent to be always-on, run evals before shipping it into production, then ship it. Once in production, the goal is to trace everything — not just the agents but also the coding agents behind the agents that built it, such as Claude Code, Cursor, and Codex, so you can understand what happened in the factory of agents. All agent traces can then be evaluated online, and production telemetry and evals become new data pulled back into agents to read the telemetry, the evals, and surface problems inside the traces. The system becomes: collect data, eval that data, run agents on top of it to find problems and figure out where the gaps are to improve your systems.
None stated.
Aparna Dhinakaran says an AI engineer really needs to learn how to write evals and should probably spend a lot of their time doing so, including looking at traces with their agents to identify common errors. Evals need to be aligned with human judgment, and defining what is good becomes a key part of the job. A common team question is how to trust the evaluator — whether it will catch issues and whether it is as good as human judgment. Her approach is to evaluate the evaluators: run evals through datasets, verify them against human output, and align so the judge agrees with the human before shipping a new evaluator, spending decent time building golden datasets. Evals can be used both offline and online, and can be as simple as a dataset or a simple prompt with outputs tested for regressi
None stated.
Choose sources → to get a library of your own and a daily brief.