35,848 results · Most relevant
Structure videos with 'beats' — a sequence of moments that keep the story moving — rather than relying on quick cuts. Sequence a demo video so each beat answers the question the viewer is about to ask (play song → lower volume → show earnings dashboard).
If your audio is bad, viewers will bounce from your video. A demo with low video quality (720p) but crisp audio and genuinely interesting content will retain viewers. Invest in a super cheap DJI mic (findable used online) or any mic that is not your MacBook mic or AirPods mic.
Joseph Kim encourages builders to ask what model capabilities they are underappreciating today because they aren't quite useful yet, consider where those capabilities might be a year from now, and think about what they would build if current capabilities crossed the threshold of usefulness and became reliable enough to delegate. Consider how the product's interface would change, decide what users would still need to control, review, or approve, and build the right product experiences now so users can take advantage of capabilities immediately when models cross the threshold. Building threshold-ready products helps you stay one step ahead of competitors. His closing question: what should you start building today to be ready for the thresholds models will cross tomorrow?
The hosts urge everyone to stop scroll jacking: when a site keeps scrolling after the user has stopped touching the mouse or trackpad, it breaks the natural precision of scrolling and feels frustrating and unnatural. The simple fix is to just let the scroll proceed naturally so viewers can see the content.
From a training perspective, the key question is just how many transactions you put in your context or history for doing the training. When post-training a foundation model for transaction fraud, the key training question is how many transactions to include in the context or history for training.
At inference time for fraud prediction, you put in the sequence of transactions (someone using the debit card) plus all historical information.
Evaluation is perplexity-based, plus an accuracy dashboard on downstream tasks. Evaluate the foundation model using perplexity-based evaluation plus an accuracy dashboard on downstream tasks.
As part of applying the build-break-fix loop, map out the uncertain and high-risk regions where the system's performance cannot be trusted, map the entire failure region, and route based on reliability and materiality. For high-risk rated systems, make sure the AI is only used in the region you are confident of.
endorsed recommendation
Kurtz's core defensive principle: the only way to fight AI today is with better AI. Defenders should exploit their unique advantages — knowledge of their own systems, architectural diagrams, and code — to dramatically increase the probability of outsmarting offensive AI. He drills home giving defenders capabilities equal to or better than attackers, since even a slight mathematical edge compounds into a defender advantage, and stresses rapidly tracking what the offense is doing to stay defense-dominant.
Evan Conrad argues that to solve these compute problems, you need to underwrite the value of the GPU cluster itself rather than underwriting the credit risk of the customer — and that solving them requires solving a bunch of different problems all at once. He notes this proposed mechanism is not guaranteed to work.
moderate
Greg Brockman argues that because software is becoming nearly free, its value now lies in reliability and security, so cyber/security spend should vastly outstrip software spend. He acknowledges he may not be right, and the host notes traditional security investment has been only low single digits percent of IT budget.
explicit
Look for startup opportunities in domains that were historically prohibitive for normal people or companies but have become straightforward with AI.
Sam McKenna says she never sets an agenda at the beginning of a discovery call even though she has one in the meeting invite, comparing agenda-setting to setting an agenda on a first date, and recommends starting the call as a conversation after rapport building and show-me-you-know-me.
Never open a call with 'Hey guys, where are you based?' — showing up unprepared burns credibility and executive presence. Instead, connect the research dots personally: 'Hey Armand, I see that you're based in San Francisco. Oh my gosh, I miss it so much. I'm sure things were crazy with Dreamforce.' — find something to connect with that also lets you share a bit about yourself.
For AI writing, Bouchard advises an extremely long and detailed prompt — it cannot be long enough, and the more the better. Include details on how you talk and write, tons of examples of your previous content, storytelling guidance, task directions, negative examples of things you don't like and words you don't want, and share your full newsletters or writing samples — it will just be better.
Coach Kyle recommends using the WeFrame in conversation to instantly make the interaction about the man and the woman hanging out together, calling failure to do this one of the biggest mistakes men make in conversation.
Set a deadline for when the usable report arrives, then choose a load that makes it. The first token may be reasoning, so you still need to measure when the usable report arrives.
When shaping the research workflow, fix the question on the source reports, use the same information cutoff, and agree on what makes a report acceptable before running the comparison.
Test with the number of users you expect and check whether reports arrive within the promised time. Check whether the result can be repeated as the third validation step. When checking the complete result, measure the complete implementation and run the same questions again to verify the answers still meet your checks.
Test retrieval by moving the important passage to different places in the input and asking the same question again. A long input is useful when the model can find the key fact.
Choose sources → to get a library of your own and a daily brief.