35,595 results · Most relevant
Harvey operates on three core values: Decisiveness (moving fast, preferring mistakes over overanalysis), Simplicity (avoiding overbuilding, simple solutions), and Jobs Not Finished (humility, openness to improvement, immediate pivot to next steps).
high
Roman Ugarte recommends viewing AI bots as colleagues and teammates, personifying them, and thinking about building AI products as building useful AI teammates rather than as a SaaS product. He frames the goal as a team of AI bots that help with both job and life, and suggests naming each bot specifically for its task or use case.
Karthic Rao recommends letting determinism govern control flow, evidence, and side effects in AI workflows, while letting intelligence (LLMs) handle interpretation and judgment.
Karthic Rao recommends asking where the boundary between determinism and intelligence is when designing AI system architecture, rather than asking whether the system should have more or less agency; he frames this as the one thing to remember, and warns that asking the agency question instead leads to a system that cannot reliably separate deterministic control from LLM judgment.
Karthic Rao recommends that when data is sensitive, you need a path to move workflows entirely to privately hosted LLMs, and to design the system accordingly.
Aparna Dhinakaran recommends running online evals on top of telemetry data as agents go into production. Skipping this means losing the ability to evaluate agent behavior on live data.
Aparna Dhinakaran recommends that AI engineers trace everything — not just the agents they build, but also the coding agents behind the agents that built the system. Failing to do so leaves you blind to what happened in the factory of agents that built your system.
Aparna Dhinakaran recommends building a system where agents can improve from the telemetry being collected, so that you're building the system to improve your product rather than just individual features. Only building individual features means missing the real AI engineering opportunity, and failing to pull telemetry and eval data back into agents means problems inside the traces go unsurfaced.
An eval is better than no eval at all. Start with out-of-the-box eval templates (QA correctness, hallucination, session coherence) from the eval hub, align them to your use case, and create more specific ones over time. As your application gets usage, ground evals in real production data and evolve them — don't set and forget, since evals become stale. Reusing basic evals across an org is great. The speaker's own example grew to 32 evals running on their agent.
Building a rigorous internal benchmark and evaluation suite allows a company to build faster by easily telling whether prompt or harness modifications broke something or improved the product. Publishing a benchmark externally encourages frontier labs to hill climb on your use case, pouring compute into improving base models for your specific use case so your product improves for users over time.
Sofia Jakovcevic advises running evaluation metrics to track agent performance and using that data to iterate and improve agents over time.
Sofia Jakovcevic recommends starting with observability: track log and tracing data for every production run so teams can see what their AI systems are actually doing at each step.
Continue the branding loop by associating with the things people want (money, looks, weight loss) and the people your audience finds cool or interesting — pairing over and over again.
Once you reach the top, you must defend the belt by fighting newer, hungrier challengers — the third step of the branding loop.
Farza argues that making the product and then magically making a hit video afterwards never works — you should know a launch is a hit before you even start working on it, because it just sounds good on paper.
Stephanie Jarmak recommends approaching agent evaluation the way a scientist would, relentlessly questioning assumptions about the system rather than trusting it.
Earning the right to the next meeting should also earn the right to a multi-threaded call — aim to talk to everyone in the decision-making circle as fast as possible, not just one contact. Treat a discovery call as a failure if you can't at least double the number of contacts or double their quality — 'double the quantity, double the quality.' If you've done active listening and created 'I want that' moments, the prospect should want to introduce you to everyone else in the organization — but expect that it isn't always that easy.
Jeff lays out a coherent method for the kickback: set the bench to about 30 degrees, lie face-down to position the arms correctly relative to the floor, extend the arm back to parallel with the floor, and work the forearm through a 90-degree arc from pointing straight down to straight back behind you.
Rob Walling advises validating a startup idea before writing any code.
GaryVee says his biggest blind spot was not being candid and not being as good a boss as he thought for 20 years. When he finally tricked himself into realizing everything not working in his life was based on his inability to be candorous — his kryptonite — he rebranded it as 'Kind Candor,' describing it as Mary Poppins: he can now deliver medicine with some sugar in it. He says six years ago when he owned it and became accountable, VaynerMedia did $100 million in revenue and is now doing $400 million, wildly because of that.
Choose sources → to get a library of your own and a daily brief.