In The Loop

Jev is a new kind of AI model that can't write anything. Instead it does the most important but most expensive thing the very best models do: make's decisionsIt reads an email, a lead or a document, picks from answers you've written down, and does it in under a second for a fraction of the cost than your Claude & GPT subscription. Nine days after it launched, the startup behind it, TypeSafe AI, was reported to be in talks to raise over a billion dollars, and OpenAI and Amazon and are just about to release their own versions.In this episode of In The Loop, I'm walking through what the Jev model is, how it works, and the six ways people are already using it.I also cover how to try it yourself in TypeSafe's playground and what a model this cheap does to the companies that charge by the token. The open question is how much of it lasts now the big labs have copied it.⏭️ Episode highlights(01:05) – What Jev is and how it works(05:00) – Seven hundred adverts broken down for nine cents(07:35) – Sorting email and leads as they arrive(09:00) – Checking every paragraph against your style guide(11:25) – Setting it up in TypeSafe's playground(14:05) – What this means for the big AI labs🔗 Links & resourcesTypeSafe playground - https://console.typesafe.ai/playgroundTypeSafe launch post, "Introducing System One Models & Jev" - https://typesafe.ai/blog/introducing-system-one-models-and-jevTypeSafe add-on for Claude Code and other coding assistants - https://docs.typesafe.ai/agent-skillRaphaëlle d'Ornano on Jev and AI economics, Decoding Discontinuity - https://www.decodingdiscontinuity.com/p/typesafe-jev-decision-models-ai-economicsEpisode transcript with more resources on the Mindset AI blogIf you enjoyed this episode, rate, follow, and share. It helps others stay ahead of the latest AI trends.

Anthropic's Dario Amodei published an essay saying the whole industry has to slow down, and Musk, Altman and Hassabis all backed it inside a day. In this episode of In The Loop, I'm going through what Amodei is actually proposing and whether it could work⏭️ Episode highlights(01:00) – OpenAI paused, then said nothing for seven weeks(02:40) – The safety test that caused the break-out(04:30) – Jacob Coxon resigns and gives up his equity(06:20) – Amodei's two reasons for slowing down(08:30) – The proposal, in plain language(10:30) – The speed limit nobody can measure(13:00) – The case against: liability, rivals and an IPO(16:00) – The suppliers nobody has named🔗 Links & resourcesDario Amodei, "We Must Pace the Frontier" - https://darioamodei.com/post/we-must-pace-the-frontier

Within his first two weeks as COO at Turtl, Dave Martin and his team built a churn prediction model. Tested against historic customer data, it predicted churn with 98% accuracy.Work like that normally takes six months or more, plus an agency and external partners. Most companies still make that call on gut feel, or by putting some basic data into Claude or ChatGPT.In this episode of In The Loop, I'm joined by Dave to walk through exactly how he and his team did it, using a method he calls pattern of life analysis.We cover how they paired customers who renewed with similar ones who left, got 12,500 minutes of call recordings through Claude without blowing the context window, the landmines you'll hit if you try this yourself, and what it takes to replace the loudest voice in the room with evidence.⏭️ Episode highlights(08:15) – Why a scale-up can't wait six months(09:40) – Testing the model: 98% on historic data(11:10) – Rebuilding each customer journey, call by call(24:25) – Why Claude's first answer was "absolute garbage"(42:05) – The "rudimentary" test that proved three ideas(45:55) – Why loud opinions make terrible decisions🔗 Links & resourcesDave Martin on LinkedIn - ⁠https://www.linkedin.com/in/mrdavemartin/⁠Turtl - ⁠https://turtl.co⁠Jack Houghton LinkedIn - ⁠https://www.linkedin.com/in/jack-houghton1/⁠Mindset AI website - ⁠https://bit.ly/40lJr6B⁠

Anthropic went from $9bn of annualised revenue in December to $65bn by the end of July, and it's expected to list in October at close to $2 trillion - the highest price any company has ever carried into a stock market debut. If you work backwards from that valuation, it has to be earning about $1.2 trillion a year within a decade. The entire world spends $1.5 trillion on software. So the Anthropic IPO isn't a bet on software at all. It's a bet on wages.In this episode of In The Loop, I'm exploring the six things that all have to be true for a $2 trillion Anthropic valuation to make sense - using their IPO as a way of understanding the strategies of the biggest AI companies in the world.⏭️ Episode highlights(01:00) – Why nobody in AI talks about software any more (02:00) – $9bn to $65bn in seven months, and the caveat (03:10) – Anthropic's own written date for powerful AI (04:20) – One in five firms, 78% of the workforce (05:20) – The 60 pence in every pound that goes back out (06:20) – Memory costs more than the processor (07:30) – Three of the four hyperscalers are burning cash (08:40) – The one line to find in the prospectus

In this episode of In The Loop, I'm going through four tactics engineers any many others use every day that almost nobody outside engineering has heard of: Fan out, adversarial review, ChatGPT's computer history, and the browser Claude got of its own this month. ⏭️ Episode highlights(01:00) – Why engineers are years ahead of everyone else(01:45) – Fan out: several Claudes, one job(02:45) – Why long chats get worse the further down your list they go(04:20) – Adversarial review: the blank chat that has no stake in your work(05:40) – The four sentences that do all the work(06:35) – ChatGPT watching your screen(08:20) – Claude's own browser