Recommendation · MoneyRob Snyder recommends that when things aren't working, you enter 'goblin mode': try a bunch of stuff, talk to as many people as possible, get on planes, go have 20 meetings, and just look for something. Just get on a plane, go to the office, go in person, do whatever it takes to find something.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyRob Snyder recommends not proactively offering pilots, proof of concepts, or design partners — go straight to selling and rolling out. If a prospect says the way they buy requires a pilot, that's okay — but don't proactively offer one.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyRob Snyder recommends that at a mature company with a new product, you put up a hypothesis, talk to five of your existing customers, look for that pull, see if they're going to actually buy it, then go through the same process again.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyRob Snyder recommends delivering firsthand teaches you a ton — keep asking 'Are you going to churn now?' and try to find the moment when they say 'OK, cool. I can't live without this.'
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyRob Snyder says you can tell a prospect 'you don't look like our case study, so it's probably going to be a lot of work and might not be a good fit.'
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyRob Snyder recommends that when talking to a prospect like SpaceX, ask: what is the case study we're going to write about them — how we got them, how we sold them, how we delivered it.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Recommendation · MoneyAparna Dhinakaran recommends designing agents to be durable: if the sandbox or harness fails mid-run, the agent should pick up from the step it failed at rather than starting from scratch, with state saved in memory outside the sandbox. If an agent fails, spin it back up quickly by launching a new sandbox and harness to pick up where it left off.
None stated.
AI Engineer · What Does an AI Engineer Do Now? - Aparna Dhinakaran, Arize AI
Recommendation · MoneyAparna Dhinakaran recommends that AI engineers set up systems and agents to do code reviews instead of reading diffs themselves, write evals to review non-deterministic tasks, and have agents read observability data and consume telemetry to find patterns instead of just setting up metrics and software telemetry manually.
None stated.
AI Engineer · What Does an AI Engineer Do Now? - Aparna Dhinakaran, Arize AI
Recommendation · MoneySallyAnn DeLucia recommends starting with a small scope (around 20 traces, one dimension, one judge), reading the judge explanations, iterating, and only then expanding to more evals.
AI Engineer · How to Write Your First AI Eval - SallyAnn DeLucia, Arize AI
Recommendation · MoneyWhen writing down eval ideas, record: the task, what success looks like, what failure looks like, what the eval judges, and three bullets on what pass and fail look like. The speaker's list is slightly garbled in the transcript, so the exact six items are ambiguous. The reason for bullets rather than a one-liner is that you'll use this to write your LLM-as-judge prompt, so specificity helps the LLM understand.
speaker's list is slightly garbled in transcript; exact six items ambiguous
AI Engineer · How to Write Your First AI Eval - SallyAnn DeLucia, Arize AI
Recommendation · MoneySallyAnn DeLucia advises saving root spans to the dataset when evaluating input versus output, and tool call spans when testing a specific tool call; choosing the wrong span type means the eval judges the wrong thing. Root spans can be filtered with parent ID equals null, and tool calls can be filtered by the tool's name.
AI Engineer · How to Write Your First AI Eval - SallyAnn DeLucia, Arize AI
Recommendation · MoneyRather than being rate-limited to one IDE, run many agents in the background and review their output when done — a concrete operating pattern for AI-first engineering.
Also discussed by Rania Khalaf on AI Engineer (unclassified source — not counted until classified). Not counted as independent sources.
The a16z Show · The New Startup Playbook | Box CEO Aaron Levie
Recommendation · MoneyLevie recommends writing very in-depth specifications and clear prompts so agents can execute correctly against your codebase — spec-writing skill is now extreme leverage.
The a16z Show · The New Startup Playbook | Box CEO Aaron Levie
Recommendation · MoneyPer Jared Friedman of Y Combinator: go do the job you want the agent to automate and become the expert in how it's done. Best case is you've done the job; second best is a co-founder who has; third best is shadowing someone doing it all day. If starting from scratch, pick an industry you know extremely well and reverse-engineer the work.
attributed advice (Jared Friedman) plus Levie's elaboration
The a16z Show · The New Startup Playbook | Box CEO Aaron Levie
Recommendation · MoneyVoss recommends writing down your high blast-radius code and making that the only thing you review manually, because automatic reviewers will handle the rest. He also recommends delegating everything that is cheap to verify to automated processes or agents.
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Recommendation · MoneyVoss recommends measuring how much stuff you are actually shipping rather than lines of code produced, measuring the gap between tests passing and a maintainer being willing to merge the code, and treating production as the last reviewer — watching what code does in production with another model grading it, which is an eval.
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Recommendation · MoneyVoss recommends spending significant time thinking about how to build an automatic way of telling whether a given thing is good or not, and then feeding that to your agent instead of trying to figure out the review process manually.
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Recommendation · MoneySet up GitHub Actions so each pull request runs through an experiment workflow against the evaluations your team has deemed important. Don't merge a PR until it meets all of the listed regression criteria; merging without regression criteria means pushing on vibes instead of with confidence. If a change is negative, go back, test, and iterate.
AI Engineer · Your Agent Fix Just Broke Something Else - Sofia Jakovcevic, Arize AI
Recommendation · MoneyFrancois recommends using paged attention — virtual memory paging with a buffer pool into DRAM — as a KV cache management technique. He notes it is mutually exclusive with radix attention.
None stated; presented as endorsed engineering advice.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyMonitor tool call pass rates, tool calling errors, and request shape over time without data retention to catch provider issues. Different providers may serve different output lengths, which indicates issues. DeepSeek/DLM serving issues can cause the model to get into an infinite loop on problems, detectable by watching request shape.
None stated; presented as endorsed monitoring practices.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club