35,595 results · Most relevant
After learning about Duchamp, Chase got into Virgil Abloh. He defines readymade as dumbing everything down to its simplest form, and says that idea is where everything began for him.
: none
Dru Knox explains that raw, unassisted harness engineering is hard for reasons that come down to human psychology: it is fundamentally unplanned work (you cannot anticipate whether agents will fail, how, or how long fixes take); the work competes with shipping and slows feature delivery; teams become AI researchers just to keep up, with best practices going out of date within weeks and becoming anti-patterns; and there is a trap — skip the work and the agent never improves, do the work and miss the feature deadline so no one ever does it. He advises planning upfront how you will make time and space to keep up with knowledge that changes every week.
Dru Knox states the sweet spot for starting harness engineering and a software factory is when coding agents frequently do the right thing on easy and medium complexity tasks; if you aren't already using coding agents, this is probably not where you want to start.
described
Eugene Teo notes that people have a negative bias, so fear-based titles like 'don't do this' or 'avoid this' perform better, but overusing them creates a fear-monger perception of the brand. He cites his own 'Why I Stopped Training With Barbells and You Should Too' video: despite balanced content, he still gets criticized for demonizing barbells by people who didn't watch it, and he blames a slightly too negative title for lasting brand damage. He now tries to take a slightly more neutral-to-negative stance on titles when possible.
Eugene's experience and judgment; includes his own anecdote
The brain is the ultimate model of hardware-software co-design: hardware and software are one and the same and have co-evolved. When taking pieces of brain-inspired design onto other computing substrates or types of physics, one has to be very careful, because whatever is native to that form factor is probably the right way to improve it.
Consequential creator advice requires independent verification before Today can promote it.
The speaker explains that hiring an agency for outreach led to weeks of waiting, wrong people being targeted, and emails written by people who didn't understand the business. He decided to build a fully autonomous AI agent pipeline that prospects, vets, and writes personalized outreach emails instead of continuing to pay thousands for outreach that barely works.
Hormozi advises choosing games of your own design, and checking when you win that you are happy with the W and that the price paid was one you were willing to pay.
Hormozi told his friend that she was not the one defining the physique that would win her competition — meaning she was playing someone else's game. He also asked whether, even if the winning physique were clearly defined, she would want it.
Hormozi argues that just because you know how to win doesn't mean you should win, and it doesn't mean the game is worth playing — there might be a different, parallel, or superior game that would result in what you actually want.
Sam argues that the model is the brain of an AI system, and when building something that makes decisions on your behalf, the brain matters more than anything else. Agents make decisions and handle complex scenarios and branching paths, so if the AI is a yes man that guesses and never asks clarifying questions, it builds agents that make bad decisions. The pushback, clarifying questions, and 'are you sure' before it acts are essential when building systems that think and act on their own.
Anish Acharya argues that humans are a critical ingredient and that most work in an organization cannot be done fully autonomously. Assuming it can leads to over-reliance on AI. He notes an AI-native company still needs humans for sales, support, strategy, and exceptions, and someone to say 'here's the thing I think we should make' and be right about it. He also warns that expecting an agent to generate a good business idea without setting it in the right direction leads to failure.
When building agent products, make the right choices obvious to users because customers will blame the vendor even when the customer made the mistake — and the vendor should accept that responsibility. Build agents with guardrails so they make fewer mistakes than humans would in the same product.
Prioritize building company-wide context — tacit knowledge, idiosyncrasies, and unstructured internal information — as the foundation for AI to perform work rather than just retrieve information. The better a company establishes that context, the more of the perform piece AI can do instead of the human.
Common predictable questions come up in every interaction, so prepare flavorful answers. Stale, flat answers make the conversation boring and unattractive, while cool answers to the same questions create far more intrigue and attraction. The way you answer questions and talk about your passions, life, mission, goals, and who you are as a man is what draws her in and gets her excited and intrigued.
For common questions (work, where you live, what you do for fun), have a flavorful answer with a fun backstory instead of a straightforward, boring, logical reply; the goal is to display cool things about yourself while slightly dragging out the answer.
Gabriel Chua warns that failing to run periodic goal audits lets long-running agents silently deviate from original goals, and not checking in on blockers like missing API keys or authentication issues leaves the agent stalled without you knowing.
Charlie Guo advises evolving your Codex workflow from asking in individual turns to building loops, thinking about what the loop needs to check at each turn; staying in individual-turn mode caps your leverage.
Charlie Guo explains subagents: they get used out of the box by default, and the model is good at figuring out on its own when it should hand off into subagents, though you can also tell it to think about delegating. Subagents keep context separate, with models communicating with each other, which helps keep context windows efficient, and Gabriel Chua adds they are particularly helpful in long-running tasks because they keep context well separated and let agents focus. Use subagents when the work primarily needs to be visible to the model, not to the user — like delegating to a subcontractor whose conversations you don't care about, only the returned work. Use thread-to-thread handoff instead when you still want to be in the loop and see everything happening (released roughly two weeks
endorsed
Charlie Guo shares a series of escalating questions he repeatedly asks himself to get from an individual ask to something significantly bigger: 'Have I tried asking Codex to do it?'; as the agent progresses, 'What blocked Codex from getting to the answer?' — checking whether it needed context or permissions it didn't have; building needed context into repeatable tools like skills or plugins; scaling from one agent to many by prompting Codex to implicitly delegate structure or orchestrating across multiple threads; and 'Why does this process need a human at all?' — taking the human out of every step that doesn't need one (e.g., a plugin can fetch an API key and set it up locally), while keeping humans in the loop only where they need to be. The overall goal is to figure out how to build the
Kent argues that product engineers should take ownership and accountability over what gets built rather than treating implementation as their only job. He frames this as the core of the workshop: technical expertise married to understanding of the user and their actual problems significantly shapes the product's direction. He also warns that engineers who only listen to someone tell them what to do and turn it into code are easily replaceable, especially as AI agents become competent.
Choose sources → to get a library of your own and a daily brief.