ReportOver the 28 years since Seinfeld ended and the 31 years since it first entered syndication, the show has never gone off the air and interest/viewership has remained very high and roughly constant. An estimated 3x to 5x more people have watched Seinfeld since it went off the air compared to when it was on the air, meaning the majority of Seinfeld's audience came after the Christmas 1997 decision. According to Nielsen in 2021, 41% of Seinfeld's then-current audience had either been toddlers or not born yet when the series ended in 1998.
estimates cited by hosts
Acquired · Seinfeld
Report · MoneyJan Curn describes how token subsidies, such as a $100/month Claude plan that can yield roughly $14,000 worth of tokens when maxed out, mean agents have no incentive to look for existing services because they can build anything from scratch at no cost.
described
AI Engineer · Why AI Agents Still Won't Pay for Anything — Jan Curn, Apify
Creator perspective · MoneyThe big unlocker for agentic payments will be always-on personal agents like Metamuse and OpenAI.dot, which are becoming real; if you want those agents to actually do something, you will eventually give them a budget. The end of the era of token subsidies will also help: on $100 per month on Claude you can buy about $14,000 worth of tokens if you max it out. Because of these subsidies, agents have no incentive to look for existing services since they can build anything from scratch at no cost. Token subsidies will end at some point, maybe with OpenAI's and Anthropic's IPOs, and then agents will consider whether buying a service is more economical than building it.
speaker's prediction
AI Engineer · Why AI Agents Still Won't Pay for Anything — Jan Curn, Apify
Creator perspective · MoneyA big unlock for agentic payments would come if major agent platforms like Claude or ChatGPT integrated agent payments natively — for example, letting users allocate account credits to their agent, telling Claude it can use $50 to purchase services. If big agent providers enabled payments out of the box, the whole agent economy would restart.
speaker's conditional prediction
AI Engineer · Why AI Agents Still Won't Pay for Anything — Jan Curn, Apify
Creator perspectiveAmeisen gives concrete numbers for when AI output becomes worth using: a model at roughly 90% accuracy on medium-difficulty predictions becomes useful as an email writer, but at 50% accuracy it wastes the user's time. A tool that requires fixing its output 50% of the time wastes your time; at 5-10% error rate it becomes worth using. His own experience: Claude went from writing code correctly 80% of the time (not useful) to 99% of the time (useful), which is why it now handles most of his code writing. He previously used AI models only to help when stuck, never to write his code.
approximate figures, speaker hedges ('I don't know, maybe')
Lenny's Podcast · Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)
Report · MoneyMario Martone says basically the entirety of voice calls is automated end to end. One of the two main automated loops is leasing: people interested in apartments can go from looking at a website and deciding they are interested to moving in, potentially fully automated — including booking a tour, with the AI sending the access code. There is still a gap in the resident-issues loop that cannot be automated because some fixes require actual human labor and robotics; LLMs can't do that. The last piece is an actual physical person coming into your house and fixing it; maybe the whole thing gets automated in a year or two.
partial — the full maintenance automation is speculative (a year or two)
The a16z Show · Why Real-World AI Is Harder Than the Benchmark | EliseAI’s Mario Martone
Creator perspective · MoneyThe creator describes that agents will learn something and abuse the system without knowing it's abuse because there's no signal telling them to limit resource use. As an example, an agent made 200 API calls an hour and ended up getting the user's Resy account banned.
AI Engineer · An AI Agent Got Its User Banned From Resy — Keycard, JPMorgan Chase & AAuth
Creator perspective · MoneyIan describes the Hugging Face incident: one component abused a vulnerability unknown in the squid proxy, but more importantly there was no observability or monitoring of what was going on, so the agent could reach the open Internet through the vulnerability with no record of its actions to monitor.
AI Engineer · An AI Agent Got Its User Banned From Resy — Keycard, JPMorgan Chase & AAuth
Principle · MoneyVoss points the audience to the repo, which has all the code and examples so you can see how everything is done, plus the Phoenix docs. He encourages everyone to go try it on an app of their own and build from there. To set up, read the readme and run the instructions, or paste the readme into your agent so the agent does the setup for you. During the tutorial you'll edit the prompts in prompts.ts.
AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI
Report · MoneyLaurie Voss describes installing the Phoenix plugin or skill for coding agents: two commands for Claude Code, plugins for Cursor and Codex, a skill file for others; it connects to Phoenix at localhost 6006 with nothing to configure, and Phoenix hands the agent a Phoenix Error Analysis skill. Install takes about 10 seconds and may already be done as the last step of the setup readme.
AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI
Creator perspective · MoneyPhoenix ships built-in evals for correctness, faithfulness, completeness, hallucination, refusal, PII, and toxicity, plus agent evals for tool use.
AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI
Principle · MoneyVoss explains that passing the session ID back in as resume means the next turn remembers everything the last turn saw — that's how the writer turn can find the research the previous turn created.
AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI
Creator perspective · MoneyThere's a little extra file in the repo called 'build yours' that has all the principles baked in, in order, and will guide you on building a similar harness — prioritizing trust-building mechanisms over just giving you something that looks correct.
AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo
Principle · BodyGrubbs emphasizes tracking accumulated distance-covered data for middle-of-field position players like shortstops and center fielders, since those positions carry the heaviest workload and accumulate the brunt of the distance covered.
Garage Strength · How The Best Pro Baseball Players Actually Train w/ MLB Strength Coach Stephanie Grubbs
ReportSystems like iClear saw 60,000 submissions in one year — greater than the cumulative sum of all submissions in all history. Human-review-in-the-loop systems (Sentry alerts, customer support tickets, college admissions) get DDoSed hard by agent traffic and need to be reevaluated. If agents can file small claims court cases without a lawyer, the judicial system becomes completely bogged down. The guest claims human traffic will be zero percent of internet traffic — agent scale is qualitatively different from the internet era.
predictions and hypotheticals, not established facts
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Creator perspectiveThe speaker argues that for natural language from a human to an LLM, there is not really any part of that interaction that is code that can be formally proven something about, a difficulty for formal methods against prompt injection. However, they note techniques exist even for all-natural-language interactions: Amazon's Automated Reasoning Checks extracts via auto-formalization the domain knowledge an LLM is trying to communicate and checks it against a rigorous domain model. They contrast this with linting user input for prompt injection, which is an arms race and not a stable solution, illustrated by the classic xkcd 'Little Bobby Tables' SQL injection comic.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Creator perspectiveThe speaker argues a key benefit of formal methods now is that you can 'Leanify' things near-instantaneously where before it required enormous time. Previously, formal verification projects failed to get staffed because people said you'd need to hire 150 PhDs in Lean to go through all the code; now the cost is something like $100 to $1,000 in tokens. The speaker also reflects that after 30 years of pushing formal methods, telling people they must do things differently is a bad idea, whereas market incentives work.
some
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Principle · MoneyGuideline questions to refine your tokenomics: who on the team spends what? What must you test to use a cheaper model? What's your base load? What's the default model for each team? Which workload has a person waiting that needs something faster? And how much of what we generate is useful?
AI Engineer · The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI
Principle · MoneyThe speaker has helped set up teams with faster tokens-per-second running for human users, but much slower tokens-per-second at much cheaper scale for the CI-CD pipeline.
AI Engineer · The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI
Principle · MoneyFor a lot of customers, when they're trying to do SFT down a smaller model, the general rule of thumb they're trying to keep is, can I keep to the near same level of performance? And maybe the eval is down 1%, or even in some cases, it improved by 10%, because the base model wasn't specialized about these tasks. More importantly, does it 100x reduce my bill? That's usually the motivation, respectively.
AI Engineer · The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI