Explore

Make important elements painfully obvious: prominent CTAs, readable text, and visible 'free'

If something is important to communicate, it should be painfully obvious regardless of the angle the user views it from. The primary CTA button should be large and prominent — not so small it's barely visible at 100% zoom — and a 'download for free' CTA should be immediately visible and legible. Important text like links should be highly readable; hard-to-read links are a failure. There's a reason 'free' is displayed giant on pricing pages: make it prominent.

Y Combinator · How To Stand Out In A World Of AI Designs

Assume users will miss content behind other tabs

With any tabbed UI, you have to expect that people will not see anything behind the other tabs — if a user scrolls through, they will miss the tabs not currently shown, and you have to assume they'll only see one tab's content. Related: on some sites the auto-advancing tab timer starts before the user is actually seeing it, so users only see snippets of each tab; users may even start a tab's timer by scrolling past without reading the content, as happened when 'voice drafts' was highlighted without being read. Paying attention to whether auto-advancing timers start when the user actually sees the content is a good thing to check.

Y Combinator · How To Stand Out In A World Of AI Designs

Your next user is an AI agent

Isabel argues that developers should consider that their next user is an AI agent: users are increasingly turning over research, product discovery, and getting started with a product to their personal agents, so applications need to let agents do product discovery and onboarding, and connect those agents to real customers.

Commercial context: The creator presents this as their own offering and directs viewers to get it.

AI Engineer · How to Let AI Agents Sign Up for Your App Safely — Isabelle Williams, WorkOS

Second- and third-order market connections require a human or a recommendation system

Anekha Sokhal argues that to surface second- and third-order market connections — for example, an oil trader needing to know right now when an energy equities company does something that indirectly impacts oil — a news stream is not enough. The only ways to connect two simultaneous information points are a human analyst making the connection or a recommendation system mimicking a human.

Presented as the speaker's thesis; the 'only way' framing is her claim, not an established fact.

Latent Space · You are the Team — Anekha Sokhal, Moshi

Wells Fargo built its first foundation model, the Wells Fargo Foundation Model

Wells Fargo built its first foundation model, the Wells Fargo Foundation Model. Freddy Lecue explains why they did it, how they did it, and what they accomplished.

AI Engineer · A Bank Trained Its Own Foundation Model to Catch Fraud — Freddy Lecue, Wells Fargo

Wells Fargo built a transaction foundation model as a shared backbone

Wells Fargo decided to build a single transaction foundation model backbone that can be fine-tuned for downstream tasks like credit scoring, fraud detection, anti-money laundering, and marketing, rather than maintaining many siloed models.

AI Engineer · A Bank Trained Its Own Foundation Model to Catch Fraud — Freddy Lecue, Wells Fargo

Evaluations should falsify, not confirm; eval systems should be failure-driven

Attributing the proposal to Peter Watson, Vishal Srivastava endorses the principle that any evaluation should be designed to falsify, not to confirm, and applies it by proposing that eval systems need to be failure-driven.

quoted/endorsed principle attributed to Wason via the host

AI Engineer · 1,368 AI Failures and What They Teach About Evals — Vishal Srivastava & Tanmay Sah, New York Life

Deployment setup determines whether errors cause harm

The deployment setup matters: how much authority the system has, whether someone reviews its output, and whether monitoring can catch errors before they cause harm. Exposing the system to external customers means any small issue can propagate fast, unlike internal employee exposure where errors can be contained. For high-risk systems, take care of deployment controls first, as this gives assurance that failures can be handled in advance, in contrast to looking at technical failure alone.

AI Engineer · 1,368 AI Failures and What They Teach About Evals — Vishal Srivastava & Tanmay Sah, New York Life

Design evals to falsify, not to confirm

Any evaluation that needs to be done has to falsify, not to confirm — following Peter Wason's research methodology. Design eval systems to be failure-driven rather than confirmation-driven, starting from the system claim and actively trying to break it. Confirmation-biased humans designing evals that confirm rather than falsify will produce evals that miss real failures.

AI Engineer · 1,368 AI Failures and What They Teach About Evals — Vishal Srivastava & Tanmay Sah, New York Life

Agentic evals add complexity, and the evaluator itself must be checked and governed

Tanmay Sah explains that agentic deployment systems have many possible inputs, tool responses, and action sequences (trajectories or rollouts), adding new complexity; sometimes another AI agent or AI judge generates tests or judges results. Because that judge can make mistakes and tests can miss important failures, he recommends checking whether the evaluator is reliable, what the evidence covers, and whether it applies to the system actually being deployed. He argues the evaluator is the part of the system we govern.

guest's argument

AI Engineer · 1,368 AI Failures and What They Teach About Evals — Vishal Srivastava & Tanmay Sah, New York Life

Longer chats carry forward history and increase credit usage; start a new chat for new tasks

The narrator explains that continuing to chat with the agent in one conversation means the agent carries forward all that conversation's history, and as a chat gets longer it increases credit usage even for simple prompts. When switching to a new task or direction, start a new chat instead of continuing the old one.

Figma · 6 efficiency tips for working with the Figma agent

Defenders hold structural advantages: knowledge of their systems, architecture, and code

Kurtz argues the defender knows their own systems, has the architectural diagram, and has the code, and that leveraging these advantages can dramatically increase the probability of outsmarting offensive AI.

No Priors: Artificial Intelligence | Technology | Startups · The New Apex Predator: CrowdStrike CEO George Kurtz on Adversaries Armed With AI Agents

Labeling: name the concern, then stop talking

Before defending your position in a design review, identify the other person's concern out loud (e.g., 'It sounds like you're worried that we're going to push the launch back'), then stop talking — the tension drops and the conversation shifts to solving the real problem. Nobody can hear your argument until they feel heard.

Design Better · Tom Foley: How to pair typefaces and why the best brand fonts go unnoticed

AI enables paying down security tech debt in previously unthinkable ways

Brockman claims AI lets organizations pay down their tech debt in a way that was previously unthinkable. As one example, he highlights formal verification of software — proving security properties of software — noting the technology has existed for a long time but was so painful, niche, and expensive that no one did it for anything significant. AI models are good at proving things, making this approach newly practical.

No Priors: Artificial Intelligence | Technology | Startups · The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman

Data bottlenecks: math and physics are data-rich, biology is data-starved

Greg Brockman explains that you can get lots of data in mathematics and theoretical physics, but getting lots of data in biology is very difficult and data-bottlenecked. He notes OpenAI has been working on techniques to work with less data, and that in areas where fundamental data does not exist, innovations on wet labs and how simulations are done will become incredibly important.

No Priors: Artificial Intelligence | Technology | Startups · The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman

If AI outcomes disappoint, the problem is likely you, not the model

Arcolano's central thesis: when teams fail to get desired outcomes from AI, the cause is usually the team or organization rather than the model itself. He frames this with a golf analogy — an instructor telling a struggling golfer they don't need better clubs, they need better fundamentals.

AI Engineer · Better AI Models Won't Fix Your Engineering Team — Nicholas Arcolano, Jellyfish

Tool choice explains <1% of PR throughput variance; skill use ~3%; consumption ~9%; company/team ~12%; the individual engineer ~75%

From a curated dataset of ~50,000 engineers at 447 companies (Cursor and Claude Code users, this year), the statistical decomposition of PR throughput change: the tool (harness + model) explains less than 1% of variance; how well you use it (skills) about 3%; how much you use it (raw token spend) about 9% — roughly 10x the tool's contribution; where you work (company and team) 12%; and the individual engineer is the biggest factor at about three quarters of the model.

AI Engineer · Better AI Models Won't Fix Your Engineering Team — Nicholas Arcolano, Jellyfish

Open with broad questions to find business problems and root causes

Sam McKenna explains that the purpose of opening with a broad question is to find business problems rather than product or situational issues, and that prospects who state a specific problem usually have a deeper root cause you need to uncover to win the deal.

30 Minutes to President's Club | No-Nonsense Sales · #619: Active Listening to Loosen up Your Discovery Calls and First Dates | Sam McKenna | 30MPC Hall of Fame

Treat your creator program like a door-to-door sales team

The guest frames creator-driven growth as a sales-team problem: teach people how to sell your product the right way and give them the opportunity to succeed, and the business blows up. He argues scaling this is not a money problem — billion-dollar brands want to throw money at it but lack the relationships and education.

Foundr · The NEW Creator Playbook that Took Me from $80K Debt to $250K/Day | Trybe

A successful brand needs traffic, a great offer, and a great product

The host presents a three-part principle for a successful brand: getting customers to the website, having a great offer, and having a great product. They add that even the most amazing creative engine doesn't matter if your product is bad and LTV is terrible.

endorsed

Foundr · The NEW Creator Playbook that Took Me from $80K Debt to $250K/Day | Trybe