34,920 results · Most relevant
Having an expert human in the loop is very important when building AI agents in specialized domains. A subject matter expert can check the assumptions, inputs, and outputs, and make the right corrections. Even when most checks are hard-coded, a human check with professional judgment is still needed. You can also codify the domain expert's checks directly into the model to make it more autonomous.
An analyst working in finance should never show a client or PM a number they can't explicitly validate or trace. Revenue figures presented to clients must be actual, correct numbers.
When the only viable choice is bad, hesitation guarantees the feared outcome. Trust your eyes, do the necessary thing immediately, and don't hesitate — the consequences of acting are never as bad as leaving the problem in place.
AI models often write startup sites like investor pitch decks, pitching the company rather than speaking to users — a common pattern in YC applications. Copy about who the company targets ('Enterprise. Middle market. Private equity.') is about the company, not the user, and doesn't help users solve their problem. Startups should instead speak to users: the product, who should use it, why it's valuable, and what problem it solves. The user cares about having a problem solved, and making visitors figure out which category they belong to before understanding the product is too much work.
Isabelle Williams emphasizes that when real corporate money is on the line, you need to be able to audit agent actions, limit what an agent can do, and revoke it when needed. An agent that can change its own budget or approval rules is an agent that has gone rogue; agents must be prevented from changing budgets or approval rules even when the human user can change them. When an agent hits a permission ceiling, route the escalation to the human user who retains full permissions with their own user token.
The most important lesson over the years centers around talent: acquire the best talent at all costs. Hiring is very hard, and Chris attributes a lot of his success to finding A players. Most of the time you have to recruit A players — you do it through relationships, learning their characters and characteristics. A players are not cruising Craigslist and cruising Indeed.
Chris Dreyer says to always want to be the most expensive and create enough value to be the most expensive. In his current city there's one steak joint with a much higher price, better experience, and significantly less competition — an example of premium positioning. Most people race to the bottom by lowering their price and creating less value — instead, think the opposite: how can I create more value and have less competition? By focusing on adding more value, he stood out in a crowded market and established himself as a high-worth digital marketer.
Starting the Personal Injury Mastermind podcast was one of the most valuable things Chris Dreyer has done as an agency owner. A podcast is leverage for distribution and marketing, builds authority, and creates one-to-one relationships with any guests you have on. The most important thing about hosting a podcast — that no one talks about — is that it is a feedback loop for learning: ask guests about the issues they actually have, since they're the best source on their own problems (e.g., a personal injury attorney on their own issues).
CJ Hess phrases it to his team: you can't automate a process that does not exist. When 40 people all roughly do the same function but each does it a little differently, it is way harder to build an agent that proactively does that work in the background. Without some form of standardization, it's impossible for an agent to know whether it's doing the work right, to build evals around the system, and to deliver the value an agent doing that function would have.
Rob Walling advises that if you have an existing audience or reputation in a space, you should factor it in and choose startup ideas that leverage that unfair advantage.
Rob Walling offers four rules of thumb for launching a SaaS product: serve businesses rather than consumers, solve a problem businesses will pay to fix, charge higher prices to reduce churn, and raise prices sooner than you think.
Rob Walling stresses that any strategic approach to finding startup ideas works best when pointed at a real problem you've actually observed, not one invented in the abstract.
Use experiments to test what happens when you change the model, system prompt, or agent tools/data, and prototype changes before promoting to production. Pushing changes because 'it looks kind of good' — running on vibes — is the common but wrong approach, and merging into main or prod without passing QA is not a best practice.
Pamela Fox observes that people keep being told 'you got to run evals' but no one's running them, or they run the default evals like coherence and fluency — which she says to never run, since LLMs are great at being coherent and you never have to check that. Most of the time the built-in evaluators are not what you want: you want to build custom evaluators to really see if your agent is doing a good job, though you need help bootstrapping those custom evaluators. She does sometimes use the groundedness and relevance evaluators for her RAG things.
Pamela Fox notes that a lot of the time you realize your evaluator itself is incorrect, so with LLM-as-judge you have to figure out whether the problem is your app or your judge. You also want to calibrate your judges with human input, and it helps to work with a domain expert who can label your data and check whether they agree with the LLM judge on what was a pass or a fail.
Commercial context: The creator describes this as their own product.
Nick Nisi connects the verification approach back to basic unit-testing practice: bring deterministic checks to the models in a way that they can self-heal and self-improve. This includes moving manual developer verification habits — like flexing the app in Chrome while debugging late at night — into code and a deterministic CLI so agents get the same experience, and giving agents visibility into all aspects of performance such as networking, traffic, and rendering.
Nick Nisi argues that the real value of local verification is the rapid development loop: give the agent a task, let it go, and verify locally without waiting on CI — CI is a feature of it, but not the actual tool. The faster feedback loop allows running faster and parallelizing oneself across multiple agents; he reports running three agents on a low day (at a conference) and normally 10 to 15. He also notes that when delegating, you tell the agent where you want to end up — e.g., add an email to the waitlist, persisted and saved — and how to verify it, and that telling the agent exactly what you want gets there without generating a bunch of additional useless code.
Every eval has holes but the holes are not in the same place; stack as many evals as you can on top of each other so it's hard for an error to get through all of them. Code evals catch formatting problems but miss meaning; LLM judges catch reasoning gaps but miss subtle hallucinations; humans catch the subtle stuff but can't look at every trace. Cramming every failure into one giant LLM-as-judge eval is an unreliable, complicated way of not solving the problem — build several small evals instead.
When presenting analysis or doing financial analysis on an investment, you need to show where your numbers came from and attribute your decision to the body of evidence you had. 'Trust me, I ran the numbers' without a trace should not be trusted.
In data and finance you need a separate point of reference, a counterfactual you can check against, without redoing your plan entirely. Have something review and give you a rough answer or challenge if something doesn't look right.
Use saved training trajectories and a controlled intervention to explain a failure, separating equilibrium theory from observed empirical stability.
Demonstrate all 3 repaired bugs and defend why a passed gradient check is necessary but not sufficient for generalization.
Use the tractable fixture and posterior predictive results to separate approximation, sampling and model errors; justify one revised modeling assumption.
Choose sources → to get a library of your own and a daily brief.
Map each implemented component to the original paper and explain 2 reduced-model failure modes without claiming full-scale reproduction.
Use calibration and shift measurements to justify a monitoring or abstention threshold, and give an explicit confounding counterexample to causal feature-attribution claims.
Defend a causal claim or refusal using the DAG, numerical estimates and a counterexample showing why predictive model fit cannot establish identification.
