Principle · MoneyYi Chen recommends placing only one solid call-to-action at the bottom of the page, following Duolingo's pattern, and offering pre-written quick replies for messaging a seller, such as asking if the item is still available or negotiating the price.
Figma · What if Craigslist was Designed by Duolingo?
Principle · MoneyYi Chen recommends adding a progress bar at the top of a multi-step flow so users know how long it takes for the post to be ready on the platform, making it less likely for them to drop off.
Figma · What if Craigslist was Designed by Duolingo?
Principle · MoneyYi Chen argues that XP is not given just for sending messages because the goal is to reward users who are actually making the deal successful.
Figma · What if Craigslist was Designed by Duolingo?
Principle · MoneyYi Chen recommends that tiles follow the visual principle of rounded corner cards, highlighting things users care about like the price, and showing the item image in a dedicated section.
Figma · What if Craigslist was Designed by Duolingo?
Principle · MoneyYi Chen recommends showing a celebration screen after a user sends a message to reward progress on the deal, and showing a celebration screen with earned XP after the user hits post. Celebration screens are used to mark progress milestones, such as sending a message or posting an ad, to reinforce user progress.
Figma · What if Craigslist was Designed by Duolingo?
Principle · MoneyRob Snyder recommends starting outreach on LinkedIn because target customers have interesting profiles that do credibility work for you. He suggests adding 100–150 people per week (about 20 per day) and messaging them when they accept the connection request. Benchmarks: 30–40% should accept the request, 30–50% of accepted connections should reply, and the majority of repliers should say yes to a call. If acceptance is below 30–40%, the LinkedIn profile is likely the problem — it may look like you're about to pitch them.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Creator perspective · MoneyRob Snyder describes a talk track for when the prospect asks about price: use a case study — 'Hey, react to this kind of case study' — saying that everyone has priced differently, sometimes a small project for $80,000, other times up to a million, and asking 'You tell me, how are you thinking about this?' This makes it not 'me proposing pricing to you' but both parties looking at a semi-real example and letting the buyer react. He notes founders are classically afraid to ask for big numbers, and that if maximizing revenue for a demo day you throw out a number in a way that doesn't get you slapped.
guest's recommended script
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Principle · MoneyThere will be hundreds of companies with $1B, $5B, $10B market caps and beyond that take a job function in the economy and build an agent for that job function.
The a16z Show · The New Startup Playbook | Box CEO Aaron Levie
Creator perspective · MoneyTenant isolation is really important — it would be very unfortunate if Citadel Desk 1 saw Citadel Desk 2's research. Security must be built into the document level as well. Every hit is re-authorized in Postgres — checking the citation, the value, and the document across every computation. Every builder adds a new filter, building up complex filters over time to prevent leaking data inadvertently across different sources. Only auth itself can create a scoped token to create this information. This is how Kepler ensures alpha stays within the firm's ecosystem.
AI Engineer · Do Not Let Models Invent Numbers - Vinoo Ganesh, Kepler
Creator perspectiveThe grading standard requires that every fact in an investment memo points to the exact tool call or source it came from; facts without this traceability fail the reasoning-graph standard.
AI Engineer · Why Frontier Models Fail as Buy-Side Analysts - Janie Gu, Google DeepMind & Suyeol Yun, LinqAlpha
Report · MoneyStagehand is Browserbase's open source SDK that lets you control a browser, built way on top of Playwright, and your agent sits and uses Stagehand to control the browser. The reason Browserbase built Stagehand was to give agents the ability to navigate the web and control browsers.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Creator perspective · MoneyJay Sahnan recommends using Stagehand's extract method to pull data into a Zod schema, converting messy unstructured web data into easy-to-use, readable data for an agent.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Report · MoneyStagehand's extract converts messy unstructured web data into a Zod schema that is easy to use, readable, and can be sent to an agent.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Report · MoneySending an agent raw HTML with Playwright costs about 122k tokens for one page. Stagehand takes the DOM and accessibility tree of a website, passes it to an LLM/agent, and understands where to click, then clicks it. Stagehand shrinks the DOM to a core accessibility tree before hitting the agent, reducing the same example to about 4,000 tokens versus 122k tokens of raw HTML. This gives direct selectors, saving money on tokens and reducing context for quicker results.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Report · MoneyVoss reports that Cognition built a benchmark asking 'would you, as the maintainer of this repo, merge this PR?' instead of 'does it pass the tests?', and under that framing the pass rate falls from 87% to 54%.
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Report · MoneyVoss's thesis: when agents make code writing free and demand for software is infinite, the enduring human role is identifying what agents cannot do — finding out what people actually want. The job now is finding out what people need, writing down what good means, and making it something people like.
speaker's thesis
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Report · MoneyVoss argues that agents collapsed the cost of writing code, but the constraint did not disappear — it moved to everything after writing, currently code review. Evidence: 8x more code written converts to only ~30% faster shipping, and a 10,000-line diff would take a human about a week to review properly.
AI Engineer · AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
Principle · MoneyJeff recommends measuring what is the deviation in logprobs as the prompt gets longer, rather than testing at a single fixed length, to detect context-window degradation.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Report · MoneyIn the Hopper generation, overlap scheduling gives only about 1.1–1.5x throughput versus vanilla scheduling, but in the Blackwell generation it can reach about 2x in GP settings.
possible_numeric_error
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Report · MoneyDeepSeek reports it took about 2.8 million H800 hours to train DeepSeek v3, and it runs about 43,000 H800 hours per day of inference, implying roughly 6x more inference flops than training flops per year.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club