On Friday evening, I had 50% left of my weekly limit in my Claude Max subscription (ironically, because I spent 2 days OOO last week at an AI conference!).
Like most weekends recently, I realise that I have a lot of super intelligent agents' time to do things for me, that will expire by Sunday evening.
Here is what I built with just some prompts dictated to my computer (ie I talk to my computer or record voice notes):
→ 𝗡𝗼𝘁𝗲𝗲: Markdown editor, replacing VS Code as my note-taker after 6 years of daily heavy usage.
→ 𝗙𝗶𝗹𝗲𝗲: macOS file manager, replacing Path Finder (used daily since 2022) + Find My File + Hazel + scripts.
→ 𝗘𝘃𝗲𝗻𝘁𝗲𝗲: a custom event app to help me manage pre-event, during event with a dedicated iPad app, and post-event follow-ups. Used a pre-version last week at the AI conference (best event app ever!), now more polished/improved after field experience.
→ 𝗛𝗶𝗳𝗲𝗲: my personal music app, replacing Audirvana + Apple Music
→ 𝗗𝗲𝗺𝗼𝗲𝗲 (I know I need a better name): to help me create demo videos, replacing OBS + ScreenFlow + Elgato Prompter
→ 𝗠𝗶𝗱𝗻𝗶𝗴𝗵𝘁 𝗖𝗹𝗼𝘀𝗲𝘁: a custom iPhone app for my daughter, which went from "can we create an iPhone app, because I don't like the ones I tested or they just want your data?" (#prouddad).. to fully functioning app in a few hours & with only 2 voice notes (her first voice note was 40mns, the 2nd was 10mns, to answer open questions).
Plus a lot of different work tasks, and improvements on previous apps I built.
These are all native macOS, iPad or iPhone applications (since I joined the Apple Developer Program a couple weeks ago - great investment!). All just for my personal use or my family's use.
What is mind-blowing to me is that each app was 100% functioning and much better than the original, 𝐈𝐍 𝐎𝐍𝐄 𝐒𝐇𝐎𝐓, just on the back of 𝐓𝐀𝐋𝐊𝐈𝐍𝐆 𝐓𝐎 𝐌𝐘 𝐂𝐎𝐌𝐏𝐔𝐓𝐄𝐑 (or rather rambling).
They use a fraction of the application memory of the original apps, while being 100% tailored to my needs, fully integrated in my automated workflows (CLI for each), with only local data, and improved easily and quickly, as I use them, also just with voice commands.
This is months or even a year's worth of work done over a weekend, giving vocal instructions to my computer every now and then while doing other things (no, I didn't spend my entire week-end vibe-coding apps!).
I'm deep into AI but I remain mind blown every week by the progress and the capabilities 🤯
🚀
PS: by Monday 6am I had used 100% of my weekly limit, so max ROI 😁
I’m at @WorldSummitAI in Amsterdam today, if anyone wants to meet to geek out about (conversational) AI.
Built my own custom iPad event app with AI for it too 😁
Roughly 60% of the jobs people hold today are in job types that didn't exist in 1940.
Solar photovoltaic electrician. Circuit layout designer. Software applications engineer. None of those could have gone on a careers form in advance.
David Autor & his team at MIT put the number on it, tracking 80 years of new job titles in the US census: about 6 out of 10 of us are in work that's genuinely new.
When I worry about what's left for the next generation, this is one of the facts I come back to.
The economy keeps inventing whole new categories of work, not just dealing us a smaller hand from the same deck.
And it's not only blue-collar churn: over 70% of professional jobs are in roles that didn't exist in 1940 either.
The catch: nobody can name the next categories in advance. That's genuinely uncomfortable when it's your kid asking what to study.
So I've stopped trying to pick the safe job for them, and started backing the safe habits: get good at learning fast, building things, working with people, asking better questions... and learn to leverage AI.
If a 15-year-old asked you what to study for a job that doesn't exist yet, what would you tell them?
#AI
#FutureOfWork
250 years of automation, and record employment
US payrolls held 30 million jobs in 1939.
Today they sit near 159 million (BLS).
Five times the jobs, after a century that automated farms, factories & a good chunk of the office.
The loom, the steam engine, the assembly line, the computer.. each one was meant to be the one that finally left us idle.
Each one was wrong.
Worldwide it's the same picture. Around 3.5 billion of us in work, the most ever, with global unemployment near 5% in 2024 (ILO) - close to its pre-pandemic low.
So after 250 years of building machines to do our jobs for us, more people are working than at any point in human history (says the guy who sells AI for a living..).
The caveat: "more jobs overall" is cold comfort to the person whose specific role just went, and I don't want to dismiss that, and the pain caused at individual level.
But overall, every technology has "increased the pie" instead of just replacing slices in the existing one.
The lump-of-labour fallacy:
I see posts saying AI is about to end the accountants.
Same line ran in the 80s about spreadsheets. The US now employs ~1.5 million accountants & auditors, up by more than half a million since.
That gap, between the confident prediction and what happened, has a name worth knowing before the headline runs again.
Economists call it the "lump of labour" - the idea that there's a fixed amount of work in the world, so if a machine takes a slice, there's less left for the rest of us. An Englishman, David Schloss, named that assumption in 1891 (yes, the panic is that old).
When VisiCalc & then Excel made financial maths almost free, the safe bet was that accountants were finished.
Two things happened instead.
The pure arithmetic got automated, and a different role, the bookkeeping clerks, took the hit. That part of the prediction came true.
But cheap calculation also meant you could run far more of it.. budgeting, forecasting, scenarios nobody bothered modelling when each one cost a week. A whole job barely existed before: the FP&A analyst who lives in those models. The judgment work grew faster than the arithmetic shrank, because the appetite for analysis turned out near-bottomless.
That's the pattern I keep coming back to with AI. It takes the dull tasks, and the work moves up to the parts that need judgment.
The same move the spreadsheet made, one rung higher.
Two honest caveats, because I'm not selling the easy version. That move up isn't automatic for the person displaced.. the clerk who lost the arithmetic is rarely the one who landed the forecasting seat. And maybe this time AI climbs the judgment rung too, with nowhere left above it.
I don't think so. But that's the real argument worth having.
#AI
#FutureOfWork
Every tech wave was going to take the jobs.
In 2013 an Oxford study gave the fear a number: 47% of US jobs at "high risk" from automation. The figure went global in a week.
9 years on, the opposite: +16 million jobs, unemployment at a 50-year low.
Here's the part the panic skipped: that 47% was never a forecast of jobs disappearing.
It measured how many jobs were technically automatable over 10-20 years (susceptibility, not a body count), and the authors have spent years saying so.
The headline invented the apocalypse.
Same fear, every wave, loud every time.. and every time the work grew.
The disruption was real for the people in the way of it, and I don't wave that off: the weaver whose trade ended rarely got the new job that replaced it (the gains landed on the next generation, not on him).
But the machines didn't kill the trade, they moved it (more cloth, cheaper, so more hands than before.. different hands, different work, but more of them).
AI is the newest verse of a very old song!
#AI
When will @Microsoft update Powerpoint and Excel file formats to be more friendly for programmatic workflows? Always messy when using 3rd party Python libraries..
I prefer working with AI than humans to get things done. Having an “army of interns” (agents) is so liberating and exhilarating 🤓
This thread might be the trigger for one of the best political decisions, applicable across countries.
Nic retweeted
Pourquoi Demis Hassabis (CEO de DeepMind ET prix Nobel de chimie 2024) est une personne que vous devez connaître !
1/ Pendant que la Silicon Valley brûle des dizaines de milliards de dollars dans des chatbots réversibles, Demis Hassabis accumule discrètement de la connaissance scientifique pure et définitive.
2/ Hassabis refuse la course au gigantisme bête et méchant. Il ne croit pas que l'empilement infini de puissance de calcul suffira à créer de la vraie intelligence. Son obsession ? Les Root Nodes (les nœuds fondamentaux) : des verrous scientifiques qui, une fois brisés, débloquent des pans entiers de l’économie mondiale.
3/ Contrairement à un ChatGPT (un produit commercial qui peut devenir obsolète en 6 mois face à un concurrent open-source), les découvertes de DeepMind sont irréversibles. Une fois qu'AlphaFold cartographie 200 millions de protéines, la science ne peut plus "désapprendre" cette avancée. Elle s'inscrit pour toujours dans l'histoire humaine.
4/ L'impact industriel est déjà là avec sa seconde entreprise, Isomorphic Labs. Son moteur de conception de médicaments par IA (ISO DDI) simule en quelques secondes des interactions moléculaires complexes.
Révolution majeure : le tout premier médicament 100% conçu par ordinateur entrera en phase clinique humaine d'ici fin 2026 !
Le secteur pharma traditionnel, habitué à des cycles de 10 ans, est en train de muter radicalement.
5/ Mais Hassabis ne s'arrête pas à la biologie. DeepMind attaque simultanément 5 autres fronts majeurs :
La fusion nucléaire : Optimisation du contrôle du plasma pour décrocher l'énergie propre et illimitée.
La science des matériaux : Via l'outil Gnome, qui prédit la stabilité de 2,2 millions de nouveaux cristaux (800 ans de recherche humaine condensés) pour les batteries et supraconducteurs de demain.
Mais Demis HASSABIS attaque aussi les sujets de : La météo (GraphCast) et les mathématiques fondamentales (AlphaProof) et enfin la génétique (AlphaGenome) pour décoder les 98 % non codants de notre ADN.
6/ Paradoxe ultime : Hassabis est un pur scientifique fatigué et tiraillé. Il confie qu'il aurait préféré garder l'IA dans les laboratoires plus longtemps, sur un modèle collaboratif lent et méthodique type CERN. Mais la sortie de ChatGPT a forcé la course commerciale. Aujourd'hui, il développe Gemini (le chatbot grand public réversible) uniquement pour financer ses véritables cathédrales scientifiques irréversibles.
En résumé : ne jugez pas la guerre de l'IA aux valorisations boursières éphémères. Demis Hassabis n'a pas le sens du spectacle de Musk ou Altman, mais il possède un prix Nobel, des millions de structures physiques validées et les clés de l'infrastructure scientifique des 50 prochaines années. Il n'écrit pas un business plan, il écrit l'Histoire.
I started sharing some of my AI-assisted work with "(A)I", like in "(A)I pre-filled the answers of that questionnaire.".. or "(A)I wrote this post" 😜
The parentheses matter. It's me - with AI helping. Not AI pretending to be me.
Meaning I drove it, AI did the heavy lifting, I reviewed and signed off.
The current binary of "AI-generated" vs "human-written" is useless.
Almost nothing is purely one or the other anymore (or will be very soon).
"(A)I" is my shorthand for the middle ground.
Small, honest, and it doesn't slow anything down.
Thought I'd share..
Just found out about the /rename command in Claude.. feeling dumb in hindsight that I didn't look for it a long time ago! 🙄 Makes managing many sessions in parallel much easier! I like the clear blue label..
Finally managed to find time over the long week-end to resume work on a dormant hobbie project: a site to index all brands across various (passion) areas - indexAll.
When I rabbithole in any given topic, I like to explore all the brands (or apps) that exist.
Especially things like “what is the oldest French watch brand?” or “what are Hi-Fi amplifier brands made in Germany or Austria?” 🤷🏻♂️
This is the site I have always wanted to have - perhaps it helps others.
Lots of work still though.. 🤗
indexall.io/
Claude Cowork seems unable to just proceed in trusted folders and always asks for permission to CRUD files. Tried now for weeks via Always allow button, all settings to "allow", Global instructions, system prompt, prompt, settings.json... still gettting a ton of these. How can I get it to just proceed? @claudeai
Nic retweeted
We built our launch video in Claude Code using HyperFrames.
Now it's yours.
Open source, agent-native framework. HTML to MP4.
$ npx skills add heygen-com/hyperframes
RT + Comment "HyperFrames" to get the full source code of this launch video (must follow)