In The Loop

Since May, Google, Microsoft, xAI, Meta, Apple, OpenAI and Anthropic have all launched a personal agent: an AI personal assistant you hand your email, your calendar and sometimes your bank to, and it gets on with jobs while you're not looking.

Its exploding so quickly that Meta's Muse got to 5 million US downloads in 22 days, faster than ChatGPT did.

In this episode of In The Loop, I'm going through who's launched what, what makes these agents different from the ChatGPT or Claude you already use, how to test one on your own life before you trust it with your bank.

⏭️ Episode highlights

(01:33) – Every big lab ships one in five months

(02:36) – Why Meta's Muse is pulling ahead

(04:37) – Why your agent gets its own computer

(06:14) – How agents remember you

(07:21) – Picking an agent like a bank

(08:11) – Which agents work in the UK

(14:08) – The case for waiting

In I, Robot, Asimov imagined a robot psychologist who could only understand machines by watching what they did. That's still where we are with AI today.

For six years, one neuron that lit up for both cats and cars was the AI field's favourite proof that we can never understand what AI is thinking.

Last week, a study showed we could potentially understand AI neural networks

In this episode of In The Loop, I'm going through what the research found, how it could solve major problems with AI safety and what this means for the future.

⏭️ Episode highlights

(00:00) – I, Robot and the robot psychologist

(01:00) – What a neuron inside an AI is

(02:40) – Cats, cars and the V

(04:30) – The agents that hit Hugging Face

(06:40) – The theory that stopped researchers looking

(08:45) – Yudkowsky's broken screwdriver

(10:35) – Six dollars that found a crack

(12:50) – What to ask your AI vendor

🔗 Links & resources

• Nick Cammarata, "Neural networks are already clean", Goodfire - https://already-clean.goodfire.com/Olah, Cammarata et al., "Zoom In: An introduction to circuits", Distill

First off, we are so sorry for the upload process mistake. Thank you to everyone who notified us about the mistake. It won't happen again!

Anyway.....to the episode....

In February, Anthropic's revenue run rate was $14bn.

By the end of July it was reported at more than $65bn, ahead of OpenAI.

It's now said to want a $2 trillion valuation at its IPO, but I think it could be in trouble (or at least starting to slow down).

Its growth is slowing fast, the price of its product keeps falling, and it has signed up for about $400bn of computing it can't cancel.

In this episode of In The Loop, I'm going through why I think Anthropic's growth is plateauing, how OpenAI has been catching up by going after Anthropic's business customers, and what both companies will likely do next.

⏭️ Episode highlights

(01:07) – The founder who moved off Claude

(02:30) – Doubling time goes from 8 weeks to 17

(04:26) – Token prices down about 40% since March

(06:23) – The roughly $400bn bill Anthropic can't cancel

(08:12) – How OpenAI got to almost $70bn

(14:06) – Four things I expect to happen

(15:24) – What Anthropic and OpenAI must do

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 decisions

It 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 & resources

• TypeSafe playground - https://console.typesafe.ai/playground
• TypeSafe launch post, "Introducing System One Models & Jev" - https://typesafe.ai/blog/introducing-system-one-models-and-jev
• TypeSafe add-on for Claude Code and other coding assistants - https://docs.typesafe.ai/agent-skill
• Raphaëlle d'Ornano on Jev and AI economics, Decoding Discontinuity - https://www.decodingdiscontinuity.com/p/typesafe-jev-decision-models-ai-economics

Episode transcript with more resources on the Mindset AI blog

If 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 & resources

• Dario 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 & resources

• Dave Martin on LinkedIn - ⁠https://www.linkedin.com/in/mrdavemartin/⁠
• Turtl - ⁠https://turtl.co⁠
• Jack 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

Once a week, an agent at StackOne goes hunting for new prompt injection attacks. It reads papers, trawls Reddit, tries what it finds against real models, keeps the attacks that land, and retrains StackOne's defence model on them. A person still approves every deployment.

Guillaume Lebedel, their CTO, puts the whole thing on screen, including the five experiments out of six that failed.

In this episode of In The Loop, Guillaume talks about how and why they have built an auto-research loop.

What one is in plain words, how you pick a goal ai agents can measure, how you sample 100,000 test cases down to something you can afford, and why he runs evals on cheap models before trusting anything.

We also spoke about token leaderboards and the impact its had internally.

⏭️ Episode highlights

(05:41) – What an auto-research loop is

(11:53) – Why an agent is a folder

(14:47) – Screen-share: the attack-hunting agent

(18:00) – Six experiments, one promoted

(27:29) – Sampling 100,000 test cases down

(31:48) – Ninety per cent of tokens are wasted

(39:33) – Turning off extra usage the same day

(41:13) – Screen-share: the token derby

Anthropic went from about a billion dollars in revenue to about forty-seven billion in a bit over a year, roughly doubling every six weeks.

In this episode of In The Loop, I'm going through how the fastest-growing company in history actually builds software, and pulling out five things you can learn from how they build products & their teams.

⏭️ Episode highlights

(03:00) – One job title, teams of two engineers

(05:30) – The five archetypes with no job titles

(07:45) – Building hundreds of versions before deciding

(09:30) – 15% building, 85% checking

(11:15) – Build a check, not just instructions

(13:00) – Never let the AI mark its own homework

(15:00) – The agent that tried to push to production

🔗 Links & resources

• Gergely Orosz, The Pragmatic Engineer, "How building software is changing at Anthropic" - https://newsletter.pragmaticengineer.com/p/inside-anthropic
• Anthropic, "Running an AI-native engineering org" - https://claude.com/blog/running-an-ai-native-engineering-org