35,595 results · Most relevant
GaryVee argues that anyone with a smartphone and internet connection can make content people care about, that distribution is commoditized with no gatekeeper since the internet is now distribution, and that creators should start immediately with whatever equipment they have — noting Wine Library TV had no mic and no lighting.
GaryVee argues that a small business owner should start their entire business empire from one location by building a personal brand through content, because the personal brand is the one true asset that doesn't get commoditized even in the AI world — the owner's IP and copyright is themselves.
The guest recommends getting to platforms early to extract underpriced attention, noting that being one of the few marketers on a new platform means you gather more attention. He references buying beachfront property in Malibu before anyone understands it's Malibu as an analogy, and cites his own track record of getting to platforms early.
Pamela Fox recommends building a good system for getting new user inputs into your eval cases once your app is in production, and adding those new cases to your evaluations to make sure that you don't drift.
Always start with power if you can — go up the food chain first, then work your way down. Don't try to personalize research-driven outreach at scale — it doesn't scale to 400 marketing managers; go to the C-level instead. Going to the CMO first also gives you a fallback: if the referred contact blocks you, you can go back to the exec and keep them abreast because they introduced you.
The speakers recommend treating everything your agent does — every run all day long — as very valuable data you can feed into LLMs to get fixes you need. They add that agent run data becomes even more valuable in the future as tooling improves and RSI loops get easier to set up, because it forms a dataset of exactly how your system performed, and recommend using the dataset of past runs to create comprehensive evals and have agents hill climb: loop, see their performance, and continue until they get better and better. With evidence from previous runs, agents are able to pass on their own.
future-facing portion is speculative
Laurie Voss advises writing your own rubric by asking what would make you act on a report and what you see in your own traces, then turning those observations into rubric criteria. Part five of putting together a rubric is constraining the answer — asking for things that are actionable or not actionable. An example of an actionable report has specific numbers, revenue growth, valuation compared to its sector, names a concrete risk, and ends with a recommendation you could actually follow.
Laurie Voss recommends using a judge model that is more capable than the agent itself, and even better, a model from a different provider entirely to avoid same-model bias. A model judging its own output tends to rate it higher (self-preference bias), which is why he judged Haiku with Sonnet; ideally you would judge Haiku with something from OpenAI.
Laurie Voss says you can't switch models without evals — running an experiment with a new model is what switching a model with an eval looks like. You can try the same rules with different models and the same judge with different rules. Upgrading the model improved faithfulness more than three times but it is still terrible — the agent is still making things up all the time. Whether a model upgrade is worth four times the cost is a question you need to make for yourself in production. He also notes a model can get more consistent just by getting more conservative — Haiku 5.5 said hold 41 out of 45 times, which is perfectly consistent between runs but probably worse. Tuning hyperparameters like temperature is the least useful thing you can pass to a model.
Laurie Voss advises running your evals multiple times (he ran his three times) because the agent is non-deterministic and each run of each example can mislead you in either direction. A prompt change you haven't measured is just a vibe — running on a couple of examples and eyeballing it puts you back to vibes-based evaluation.
Laurie Voss recommends bringing in a domain expert to label traces in Phoenix (10, 20, or 50 traces) to produce your golden dataset of whether outputs are good. For each golden dataset task, include an answer you know is good, which proves the task can be done. He warns that human labeler fatigue degrades golden dataset quality: after ~50 traces judges get bored and lazy, so 50 well-labeled examples beat 100 carelessly labeled ones.
Laurie Voss advises choosing an eval question that the judge can actually answer, and giving the faithfulness judge the full context including raw search results, not just the research notes. When his version of the eval only gave the judge the research notes, the judge flagged numbers as made up because it didn't know the agent searched the web and found the data. A correctness eval on an agent discussing recent events failed every report because the judge flagged fresh data as hallucination.
Laurie Voss advises building several small evals rather than one giant super eval — cramming every possible failure into one big LLM-as-judge is unreliable, complicated, and doesn't solve the problem.
Reverse engineer the full season schedule to find big stressor areas before building the training plan: identify the longest stretches of games without an off day (10, 13, 17 games), account for time zone changes, surface changes, and elevation changes, and recognize that MLB stadium field dimensions vary (short porches, big outfields, turf vs. grass) when planning load management.
When training baseball players, bucket them into three categories: bullpen guys, position players, and starters, since each has different game/usage patterns.
Don't change an athlete's entire training plan based on a single flagged metric. Change an athlete's plan only when three or four flags stack up across multiple data sources, and when they do, the performance team should respond with an action plan from their specific area (nutrition, supplements, sleep support like magnesium).
When running hefty agentic workloads scaled to hundreds of agents, prioritize scaling your safety and security posture to keep track of what's happening and ensure nothing bad happens on platform; if you had to get one thing right when building out your agents, it would be scaling safety and security.
Dr. Andy Galpin recommends selecting compound movements for power, speed, and strength, and thinking in terms of movement patterns rather than individual muscle groups.
Rick Rubin advises that if you keep thinking a project isn't the best it could be yet and keep changing it forever, you can do that indefinitely and never put anything out.
If your offer is perceived as a commodity, customers will always choose the cheaper competitor — and you will become a commodity if you use the exact same measured unit as competitors. Break the link between price and value by making your offer different on non-price factors: speed, risk, and ease.
Choose sources → to get a library of your own and a daily brief.