35,595 results · Most relevant
A variant of the job hopper list should include people who wanted to purchase but couldn't — they know you, want to work with you, and may eventually change jobs. Track lost-deal buying team members and check whether their new role lands in your territory.
The host calls 'posted on LinkedIn in last 30 days' the most important Sales Navigator filter, because active posters are more responsive to outreach, and recommends always adding it to any filter set.
Pamela Fox recommends using A-Cert to write judge evaluators based on a spec, describing the context, judge model, and evaluation dimensions in the A-Cert YAML, and stratifying eval dimensions (e.g., budget pressure: ample, tight, invisible) either manually or by letting the tool generate a stratification.
Stakeholders you don't know exist — like a VP of finance who viewed your champion's recap twice but was never introduced to you — can kill the deal at the finish line; multithread to them before it's too late.
Seinfeld's production week: writers rewrote constantly (including overnight rewrites on day three), NBC and Castle Rock watched a full run-through at the end of day three and gave notes — after which, per Jason Alexander, Larry David would quit, the notes would be diminished, and Larry would return. Day four was camera block day where cameramen learned their shots and rehearsed with actors as a unit ahead of the live audience. Day five was morning rehearsal, then filming with the live audience in the afternoon, running the entire show in chronological order. The next day the team came back for a table read, making the process non-stop.
A common technique: pass a large amount of random text through the model and record the intermediate numbers for every example. Then look for patterns — numbers or series of numbers that occur only when the model processes English text, French text, text about love, text about eyes, etc. This decoding into a guess of what results mean is the first phase; researchers then run experiments such as asking harder questions and labeling all the intermediate states to see what the model is doing. An immense amount of effort is dedicated to decoding what the models' intermediate results mean, and this recording and pattern analysis reveals intermediate results representing all the concepts you expect.
The speakers describe running LLM output and then using unit tests around it to verify in different ways, catch edge cases, and verify that the code is a real interpretation or real representation of what they're expecting to get back from a system-one model — using that to prove they're on the right track.
Adaptive stress testing requires a definition of failure and a measure of likelihood; for language models, use perplexity as the likelihood measure and search for low log perplexity with high toxicity. If the most likely failure is really likely, study it and put mitigations in place. Note: regular adaptive stress testing typically zeros in on only one failure mode, missing others.
An always-on AI system monitoring traffic and data can flag anomalous behavior, such as keys being generated and shipped to a personal Gmail account.
Coach Kyle recommends responding to what she shares with a little personal input about your own experience on that topic (e.g., 'I was actually in all black yesterday, but today I had to switch it up') instead of only asking questions.
Coach Kyle recommends adding a single relevant adjective (e.g., about how she's dressed or what she's holding) on top of the basic opener so it feels slightly unique to the specific woman, which he says increases the likelihood of a positive response.
Coach Kyle recommends starting with a simple, direct opener that essentially says 'I saw you walking by and I thought you looked really cute.'
Coach Kyle recommends that after she shares information, instead of asking the next predictable question, you look at her and make an observation and assumption about her (e.g., 'you probably do something artsy because you don't have a very professional outfit on').
If you're spending less than $10k a month on AI, you're better off finding strong use cases within your organization than optimizing the token bill — a strong use case can bring way more than $10k a month in value. Also apply AI models selectively rather than to everything, since many models are expensive.
When a working solution (a 'twin') already exists, leverage or fork it rather than spending effort building it again. The speaker is confident solutions for your problem already exist.
Eval questions come from real user-testing rounds with actual members; they define behavioral guidelines (no advice, stay neutral about competitors, say clearly when something isn't in the rulebooks) and run about 86 end-to-end tests with promptful: an LLM judge against the rubric, a check that the right source document was cited, plus the original retrieval-only evaluation.
Paden Gayle relays that the security team asks which environment will host the system, and warns that deploying AI in an unconfirmed hosting environment creates security and compliance risk.
Press T in Figma to select the text tool and type text. In Figma, each text style (display, H1, H2, H3) has a style applied, visible when selecting the text. Four dots in Figma signify that a style is applied to text. Click apply styles to see any styles local to the document and apply them to text. You can break (detach) a style if you need to make a small adjustment.
In Figma, select the frame, come to the menu, go down to appearance, and click on the appearance panel to see modes associated with the frame. Change the appearance mode to swap between brands (e.g., a sans serif brand with a different layout, or brand three entirely different) while still retaining a typographic scale. Appearance modes can change brand by brand and also by responsiveness for different sizes. Make composite styles for each brand's typography properties.
Figma has accessibility options baked into the interface: find them via Command K and 'accessibility settings'; adapt content for screen readers, underline links, simplify focus navigation, and use enhanced contrast mode; use the color contrast checker to test and fix contrast, and motion variables for reduced motion.
Choose sources → to get a library of your own and a daily brief.