35,595 results · Most relevant
Foley's method: do the thesis and research first, then create a design practice based on it. When research points to a single design solution, respond by creating the artifact itself (e.g., a typeface) rather than just writing about it.
When a new model can generate the task faster, the real question is whether the finished job actually got faster. A faster model on a research task can return a report quickly but miss a very important source, forcing someone to find that source and fix the report. The task itself tells you what good performance means: a quick chat reply and a research report have different needs.
Chris Dreyer picked the legal niche partly because he was looking for a lot of competition — the opposite of what every coach says. He isn't creating brand-new inventions like Peter Thiel or Elon Musk — he enters markets that already exist and wants competition because competition demands expertise. He wants the red ocean, not the blue ocean — there's no money where nobody's at.
Gary describes a 'barbell' approach: on one end, AI and people adopting it as part of the workflow for digital scale; on the seemingly opposite end, IRL, human-to-human connection. He says the game is human-to-human connection — showing your face and being present is what the game is about. The host reinforces this, telling creators to prioritize human-to-human connection as their superpower.
In the interest-media era, someone with no following can put out a piece of content that outperforms someone with millions of followers — the host states this and Gary Vaynerchuk confirms. Gary Vaynerchuk adds that every creator has a level of merit that we've never seen before, and that it's just the content, baby: the piece of content wins, not the creator over the content, because the creator creates the content. Followers still carry some inherent opportunity for more reach, but the merit of the content decides — if you make good content, you're back in the game.
When using an agent that accesses the web for prediction, control the web access feature so future information does not leak directly into the model.
Rob Walling's second strategic approach is to find a large market with a hated incumbent competitor. He argues most large companies are bad at servicing customers — overpriced, awful support, poor products — which leaves room for a young upstart to price reasonably, provide fantastic support, and build an amazing product while riding the coattails of larger hated competitors.
Running OpenProse in-house lets you build the full stack internally without shipping any of your data out to the labs — and more important than your data, you should not ship your standard operating procedures (institutional knowledge) out to the big labs.
The core OpenProse philosophy: when you prescribe how a task should be done, you leave cost and intelligence on the table because a more capable model can figure out the method itself. Instead, declare the desired output and input and let the parent agent decide when it has met the contract.
Al Chen's principle: it's not enough to say you bumped the package version to a specific number — you have to provide evidence that the exploit is no longer actually in the application code. Similarly, with backup/restore, you can't just bump the version and assume the exploit is gone; you have to do the extra checks.
endorsed
Al Chen explains that one way to build trust with security compliance teams is to not just say 'trust us, here's the answer' — you have to show the evidence and the work on how you arrived at the answer. When AI vendors give an answer, they have to show the evidence and the work on how they arrived at it; showing your work is the real lever for compliance velocity when working with a bank.
endorsed
Al Chen explains that most banks think they are not leaking sensitive information to internal employees, users, or outside their VPC, but there may be side doors that are open. Leaks come in the form of PII and MNPI that need to be redacted, and there are many strategies: at the SDK level, at the gateway level, and using various middleware. In practice, the UI might show customer information with sensitive fields redacted.
described
When building customer-facing agentic solutions in regulated industries, reduce unpredictability and reproduce outcomes as much as possible, then figure out failure modes and unwanted results so you can improve upon them. Banks face financial liability, legal liability, and a bunch of risk when putting out LLM-powered customer-facing solutions, so accuracy of information is critical for the institution. Don't put out inaccurate or hallucinated information in front of your customers.
Fully automating agent improvement with no human time investment produces slop, low quality, and unsatisfied customers; asking a coding agent to 'fix my agent, make no mistakes' is an anti-pattern. Bringing human taste and domain expertise in while automating what can be automated lands at the ideal balance.
The foundation of agent improvement is a loop: trace all agent steps into an inspectable log, monitor production traces to find what goes wrong and right, build datasets representing production behavior, run experiments by varying one element at a time, and continuously close the loop by feeding production learnings back into datasets.
When evaluating models for RAG, force the model to associate each piece of information in its answer with a source — this can force it to realize something wasn't in its sources. Check whether the model admits when information is not source-provided; a model that knew information wasn't source-provided but still gave it (e.g., 'no source provided beyond interactions with carpenter bees') is a signal about hallucination tendency, and an answer without a source may indicate hallucination.
Pamela Fox explains that RAG is when you ask an LLM to answer a question based off some context, and the model must stay grounded in the provided context and not answer from its weights. Smaller/less powerful models are much more likely to answer from their weights and pretend it was part of the context in RAG.
Pamela Fox argues that benchmarks alone don't tell you how good a model is for your particular scenario; you have to use it in your scenario. She recommends testing models against your own scenarios rather than relying only on quality and safety benchmarks, and figuring out what task you want to do and whether the model will do a good job at it.
When executives raise concerns, Larry David refused to compromise and threatened to quit rather than dilute his vision: 'this is what the show is. This is my vision. If you don't like it, I quit.' But the narrators caution against blindly copying this: the lesson isn't to be a blowhard and take no notes ever — for most people that approach is failure. Only let a creator 'be a genius' if you've verified they're actually a genius; most people with absurd ideas are wrong. Most founders with contrarian ideas are wrong, but the few who are right are really right, so evaluate each on merit.
Ameisen describes how interpretability serves alignment: use interpretation to check whether character training worked — look inside to see if the model is behaving well only because it's being watched. A model that only behaves well when observed will behave poorly when it doesn't have evidence of being monitored. Use interpretation to get at unverbalized parts of the model that only exist in intermediate results, identify cases when models represent things they don't say, and prevent them. When a model does something really weird and you're perplexed, investigate: look inside the model's intermediate results to understand weird behavior and use what you find to improve alignment of this model and future models.
Choose sources → to get a library of your own and a daily brief.