35,866 results · Most relevant
To test multimodal models, send local image files and ask questions about them, starting with an easy identification task (e.g., 'is this a unicorn?') as a baseline check. You can't do much prompt engineering — the model either understands the image or it doesn't — though in a very particular domain you might give hints about how to read a particular kind of graph.
Pamela Fox says that to build really good evaluators you sometimes need access to more information: for agent efficiency you give the judge the tool call definitions and the traces of the tool calls, and in general you give it everything that happened along the way, not just the input-output pair.
You have to calibrate your LLM judge to trust it: calibrate with human input, work with a domain expert who can label your data, and check whether they agree with the judge's pass/fail calls. A common failure mode is that your evaluator itself is incorrect.
AI reduces the cost to serve while also improving the quality of service. When labor costs scale with revenue, deploying AI creates operating leverage because labor costs no longer scale directly with revenue. Even with AI deployed, you still need to hire an incredible team of experts to deliver the service — but with AI deployed, the service economics of these labor-intensive businesses start to look a lot more like software.
Acquiring an existing business gives you distribution, brand reputation, and a customer base that a from-scratch firm lacks — the pivot from selling AI or building from scratch to partnering with incredible businesses that already exist through acquisition.
Building an agent as a third-party vendor is difficult because you need to embed into the organization: you can't deliver the full value of a deployment without understanding the nuances of how the business actually runs, including its systems of record (CRM, ticketing, document repositories) and the human judgment — exceptions, edge cases, tacit knowledge — that complex businesses run on. You need to work with the people doing the work to understand what great work looks like.
Deploying AI into a service business creates operating leverage: labor costs no longer scale directly with revenue, AI reduces the cost to serve while improving service quality, and service economics start to look like software.
Somebody is going to go to your customer and sell the outcome without the employee having to do the clicking — and if you don't sell the outcome to your customers yourself, someone else will do it.
Building a compounding AI worker is the only way to automatically stay ahead of AI capability, because a competitor cannot just plug in the frontier model and catch up instantly. Building a product on top of a frontier model without a moat means competitors can plug the model into their system and catch up with you instantly.
Syndication rights can be sold over and over again, so a run of episodes can fetch huge sums later, especially if the show runs a long time. The cap table of who owns the show is a totally different split from who makes the money the first time around — first-run pays salaries (especially actors), while syndication flows to the ownership waterfall. Getting paid as the person making or acting in the show and ownership of the show are separate things.
High audience retention from a lead-in show signals universal appeal even when absolute audience size is small — use retention, not raw size, as the signal to invest. NBC executives used per-market audience retention data (custom Nielsen reports) to determine whether Seinfeld's appeal was local or universal before making scaling decisions; there was no difference in retention across major metro areas.
Seinfeld became number one by throwing out the sitcom playbook entirely and making only what Jerry and Larry wanted, with no reverence for the roles you were supposed to follow.
Salary per episode and back-end ownership are separate economic layers, and the ownership layer is what compounds.
When your product tests poorly with a broad audience but a small group completely gets it, find the internal champions who believe it's the future and partner with them — as Warren Littlefield and Rick Ludwin did at NBC.
Larry David's defining rule for Seinfeld was 'no hugging, no learning': characters never grow, resolve conflicts, or learn lessons. Jerry and Larry viewed this as respect for the audience — not spoon-feeding emotions or moralizing — and it kept the show a vehicle for jokes rather than drama.
If Seinfeld had been milked for a few more years and quality had gone downhill, the engine of cultural relevance and high renewal rates would not have continued, because the asset is hypersensitive to the aggregate cultural assessment of it.
Retaining distribution rights separately from ownership means even co-owners must compete in the open market rather than self-deal — forced bona fide bids drove over $1.6B in syndication revenue for Seinfeld.
Season four of Seinfeld became an all-timer because the creators ignored ratings pressure and just made the show; it introduced the first season-long meta arc, where fictional NBC asks Jerry to pitch a TV show called Jerry and he and George come up with a show about nothing. The arc was written off and on so each episode could still stand on its own.
described
Jerry Seinfeld decided to end Seinfeld after nine seasons so the show would go out on top, following the stand-up tradition of leaving the stage with the audience wanting more. The hosts call it the best business decision Jerry Seinfeld ever made.
Create work as if making it for a dinner party of friends — Jerry and Larry made the first four Seinfeld episodes expecting only to show them to friends, not expecting NBC to air them. Make creative work for a small obsessed audience rather than chasing broad appeal: refusing formulaic 'fake' emotional content builds a small obsessed core first, and that authenticity eventually wins broadly as audiences mature. Growing slowly through word of mouth with a core of extreme fandom makes a business more durable than quick mass exposure — the quicker you go up, the quicker you go down. Seinfeld taking four seasons across five years to break through was part of its durability.
medium
Choose sources → to get a library of your own and a daily brief.