What AI I Use How and When
The following is a pretty detailed report of the various AI tools I use as an end-user. It contains little insight or entertainment value. It is fascinating to me to learn how people use AI, so if you publish something similar, please do share.
Philosophy and General Observations
Try everything new at least once. Revisit periodically to review newer versions and re-asses. Things are moving very fast, and I make adjustments regularly. Expect future revisions of this report to look somewhat different.
Adopt a stable set of tools and get to know them really well. There are significant advantages to being a power user.
Try to do everything with AI. Only give up and complete a task without AI if it's impossible to use AI effectively. It’s OK to invest a bit more time if the result is higher quality or if I learn something new. I still need to force myself sometimes.
Complete most tasks with multiple tools and models to compare, learn, and benefit from their different strengths.
Try to keep a record of prompts and processes that work well. Most things are re-usable, either because I repeat the same tasks or because they provide inspiration for future tasks. This is more aspirational than a consistent habit - I’m a lot worse at this than I wish to be.
Invest in customisation. Add and periodically revise custom instructions and tool settings.
Speaking and listening is a very efficient and engaging way to work with AI. I increasingly prefer that to reading and writing.
Daily Drivers
ChatGPT
The OG AI tool that started it all. And still the best and the one I turn to most often.
Trend: ⬇️ Proportionally declining, but only because it’s starting from a very high point and there are new tools I’m adopting in addition.
Where: Web, Phone, Tablet (very rarely I also use the desktop app)
How Much: Plus ($20/m). I get most of the good stuff in sufficient allowance for my use. I tried Pro ($200/m) for a month and liked it a lot, especially because of the availability of o1-pro, which is the best general reasoning model, but I don’t feel that I can justify the expense on a regular basis.
Highlights: GPT-4.5 - the beautiful model with personality, taste, and exquisite intuitive intelligence - the best for writing or for having an intelligent conversation with; o1 - strong general intelligence with reasoning; o3-mini for coding and quant tasks; Tooling - Canvas for iterative editing and vibe coding, Code Interpreter for data analysis and processing, Web Search; Live voice mode - great for chatting with AI on the go or getting it to interview me and get thoughts out of my head and into text form; Deep Research - the best research product, slow but amazingly thorough!
Lowlights: Slow - the web app is slow and clunky and the models can at times respond very slowly and it makes everything a bit frustrating; So many models and tools to choose from, it can be confusing.
Perplexity
Search (and research) for the AI era. Also just incredibly versatile general AI tool.
Trend: ↔️ Stable. Preplexity has become one of my workhorses and I use it all the time.
Where: Web, Phone, Tablet
How Much: Pro. I get it as a freebie from my digital wallet subscription, but I would pay anyway ($20/m), you really get a lot of goodies with very generous limits.
Highlights: Superb search; Document-oriented output - it’s the wikipedia that writes itself; Fast and mostly good enough Deep Research; Spaces - where you can group multiple files and threads under one customised roof - the most useful implementation of this concept and the one I use regularly; Access to not only Perplexity’s fast Sonar model but also the best models from OpenAI, Anthropic, and DeepSeek (uncensored and US-hosted).
Lowlights: UI is a bit cluttered and keeps changing.
Raycast
Raycast is a desktop launcher for Mac (and soon Windows and iOS) that was already cool before it hopped on the AI speed-train. It now has the best AI desktop with a quick Ask AI and AI Chat, AI Commands that let you package prompts with templates that can dynamically insert content like the clipboard or the current open web page, and very recently AI Extensions that allow you to chat with the tools installed in your Raycast - this can include things on your desktop but also many cloud services - from the chat window or packaged commands.
But even without its built-in AI functionality Raycast is a must-have tool if you work a lot with AI. Its Clipboard History functionality is insanely useful when working with multiple AI tools and lugging chunks of text from between apps. Snippets are the perfect repository for prompt templates at your fingertips. Notes are an easy and quick way to keep text around and get back to it quickly. And of course, keyboard shortcuts for all the many AI tools I use.
Trend: ⬆️ Increasing exponentially. I only warmed up to Raycast recently and since then I discover new ways I can use it every day.
Where: Desktop (an iOS version is planned, and I eagerly wait for it)
How Much: Pro ($10/m). In addition to various non-AI features this includes access to AI using the cheaper models. There’s an “Advanced AI” bolt-on for double the price, if you want the best and latest models, which I don’t find the need for because the basic subscription includes Gemini 2.0 Flash (with or without thinking) and Claude Haiku and various other models that are good enough for casual use.
Highlights: Fast and good looking - this is one of the best tools I get to use for anything, including AI; Great integration with my desktop environment and many additional tools via extensions; Infinitely versatile and configurable, yet intuitive to use.
Lowlights: Only on desktop (at least until the iOS version becomes available). And it will never replace the other chatbots for me without a web version, which I’m not aware of.
Reading / Writing / Research / Learning
Notion
I got back to Notion as my “2nd brain” repository for everything not only but primarily because of its excellent AI integration. Pretty much everything you’d expect from an AI “copilot” for your knowledgebase, and more.
Trend: ↔️ Stable. I use Notion all the time, my use of its AI features is a bit random, sometimes a lot, other times not so much.
Where: Web, Phone, Tablet
How Much: Plus + AI ($20/m)
Highlights: Excellent AI integration that is a pleasure to use; Best way to search and mine the information in my vast Notion repositoy; AI data fields that are being written automagically by AI.
Lowlights: Not “agentic” - it mostly won’t go and do things for you at your request; Can have AI magic in database fields, but not in the main text of documents.
Readwise Reader
More than just read-it-later, this is the main app I use for reading internet articles, eBooks, X threads, PDFs, … The main AI feature I use is the automatic summary, which helps me get a quick overview of an article and in many cases … decide to delete it without reading further. What a great timeserver! Also pretty good text-to-speech for listening to articles. Many other AI features which I use very rarely.
Trend: ↔️ Stable. I’ve been using Reader since it became available and while I only use the most basic AI capabilities, they are integral to my reading habits.
Where: Web, Phone, Tablet
How Much: Full ($8/m). Some AI is built-in, and it lets me add an OpenAI API key for more sophisticated use (my bill comes up to ~$1/m).
Highlights: Automatic summaries with my own template; Decent text-to-speech for listening to articles.
Lowlights: Somehow the UX of the other AI features feels too complicated so I don’t use them much.
NotebookLM
Google’s “experimental” tool for chatting with large collections of large documents. Also has a cool podcast generator.
Trend: ⬇️ Declining. I was very excited about it initially, then again when the audio overview became available, but I now use it much less.
Where: Web (sometimes on the Phone for listening to a deep-dive podcast)
How Much: Free (a Gemini subscription includes some benefits like a larger number of notebooks, but I don’t use this enough to justify the additional expense).
Highlights: Deep-dive podcasts - great way to turn a bunch of documents to something engaging you can listen to; Research over large documents and even larger collections of documents.
Lowlights: The UI is grating, and development is stagnant (rumours that the team quit Google to go work on startups).
CardCraft
My own home-grown solution for automagically creating flashcards for spaced repetition from materials I learn like books, articles, videos, and in many cases output from other AIs.
Trend: ↔️ Stable. This is now integrated into my learning regime, and I use it a few times a week as I ingest new knowledge.
Where: Web
How Much: I’m the developer. I’m offering it for free to testers during early preview. I have vague plans to activate some kind of pay-per-use scheme later on to cover the AI costs.
Highlights: Saves me hours every month by creating flashcards that I would otherwise have to manually write myself. This is huge.
Lowlights: Not yet as creative with the format and structure of card decks as I or another human would be; I’m the developer, so it only improves if I find the time and patience to work on it.
Hacking
Visual Studio Code (Insiders edition) + GitHub Copilot
A few years ago I switched from Emacs to VSCode and haven’t looked back. It’s the one true IDE. And GitHub Copilot is the first AI coding tool and it keeps getting better.
Trend: ↔️ I don’t do a lot of coding but when I do it’s almost always with this setup.
Where: Desktop (sometimes web with GitHub Codepsaces)
How Much: GitHub Copilot ($10/m)
Highlights: Code completion; Suggested edits; Contextual chat; The new Agent mode.
Lowlights: A bit slow; UI a bit confusing; Agent doesn’t work autonomously - I have to approve every change manually.
v0
v0 is a front-end “vibe coding” tool from Vercel. I fell in love with it for its elegance and the playful joy it brings to creating UIs and sharing them.
Trend: ⬇️ Declining slowly. I was super excited about v0 when I first discovered it, but now that the novelty wore off, I just don’t have as much appetite for front-end vibing.
Where: Web (very rarely also on the tablet)
How Much: Pro ($20/m)
Highlights: Gorgeous; Fun and easy to use; Has such great taste in aesthetics - this is the front-end designer/developer buddy I never had.
Lowlights: No Git or VSCode integration, so the only way to bring things back into a normal project is by copy-and-pasting or retrieving blocks using the shadcn CLI.
Warp
Post-modern terminal app. Lots of cool features. And also AI.
Trend: ↔️ Stable. It’s a terminal. I use it every day. Not a lot.
Where: Desktop (terminal)
How Much: I use the free version. I think it has some limits on how much AI you can use but I’m not using it enough to notice.
Highlights: Really nice terminal and the AI integration is handy.
Lowlights: AI integration is not that intuitive - I don’t use it that much because I’m never exactly sure what I need to press and what it is going to do. Probably my fault.
llm
Simon Wilison’s command-line tool for working with LLMs. Incredibly versatile with lots of useful features and plugins for all cloud and local LLMs. Pipe stuff in and out of it (pbcopy and pbpaste, or files, or the output of other programs) or go digging in the SQLite database it keeps with a record of all prompts and responses.
Trend: ⬇️ Declining. I loved this tool when the LLM Cambrian explosion happened, and I wanted to experiment with every new model. I still use it regularly but now only when it really makes sense to do something on the command line.
Where: Desktop (terminal)
How Much: Open Source (needs API keys for cloud or local models - I mostly use cloud models from OpenAI, Anthropic, and Google and pay ~$3/m)
Highlights: Lots of great features covering pretty much anything you can do with an LLM; Plugins with sophisticated functionality; Access to all cloud and local LLMs; Saves all interactions in an SQLite database, so very easy to go back and pick things up.
Lowlights: Honestly not sure there are any. For a command line tool this is pretty much the gold standard.
Audio
MacWhisper
goodsnooze.gumroad.com/l/macwhisper
Great Mac app for using OpenAI’s open-source Whisper models for text transcription, with lots of features and great integration with the OS. Great for dictation - I use it all day long to talk to my computer - as well as for transcribing recordings.
Trend: ↔️ Every day all day. I don’t see how it would be possible for me to use it more, and I can’t imagine using it less.
Where: Desktop (an iOS version is planned, and I can’t wait to start using it)
How Much: Pro version was a one-off payment of $50.
Highlights: No recurring subscription - usable free version and the fully featured pro version is a one-off purchase; Support for lots of models of different sizes, as well as some cloud options; Easy to set-up shortcuts for dictation - I use this all day long to talk to the computer; Easy to set up LLM transformation of the transcribed text.
Lowlights: This is a power-user tool with lots of configuration - can be confusing at times; UI is a bit clunky.
ElevenReader
Cool app from text-to-speech leaders Eleven Labs for listening to documents. Fun and easy to use and the quality of the voices is excellent.
Trend: ⬆️ Only started using it recently and I’m finding new opportunities frequently.
Where: Phone
How Much: Free
Highlights: Excellent Text-to-Speech for documents and web pages - very realistic and pleasant voices.
Lowlights: There’s a podcast feature but it’s not as powerful as NotebookLM’s.
More Tools / On Probation
Grok
The AI from Elon’s xAI. Second only to OpenAI in quality. And improving rapidly.
Trend: ⬆️ Since Grok 3 came out I’m using it a lot. Good chances it will be in the “daily drivers” section the next time I report.
Where: Web, Phone, Tablet
How Much: I get generous allowance of all features with my X Premium subscription. There’s a $30/m SuperGrok subscription option for users who don’t want to pay for X.
Highlights: Grok 3 - excellent big model, second only to gpt-4.5; Grok 3 thinking - powerful and fast reasoning intelligence, almost o1-pro quality, but much faster and more afforable; Absolutely no censorship - this AI does and says whatever you want; Integration with X; It’s FAST!
Lowlights: The app (both web and phone) feels very beta quality; The voices of the live voice mode are annoying so I don’t use it so much, even though it’s very powerful; The model is highly intelligent but probably needs a bit more fine-tuning (RLHF) and can be unpredictable - you never know what you’ll get, sometimes it’s brilliant, other times insane - a bit like Elon.
Google AI Studio
Google is funny - they create some of the best AI capabilities but never fail to disappoint with their end-user apps and integrations. AI Studio, one of the best deals in AI, is their developer playground, and its a great way to get all the AI goodness from Google, for free. It has all their models and features, and they even let you install it as a web app on your desktop or phone.
Trend: ↔️ It’s been around for a while. I use it when I need something specific or when there’s a new feature or model to test.
Where: Web (very rarely also on phone and tablet)
How Much: Free with extremely generous daily rate limits.
Highlights: All the models and features from Google DeepMind, including the latest experimental releases, for free; Huge context window (1 million tokens) - this is what I use if I need to process very large documents; Excellent multi-modal (vision, audio, image generation) capabilities, including live audio/video.
Lowlights: It’s not supposed to be a consumer tool, and it wouldn’t pass as one.
Goose
New open-source tool from Block that puts an agent, connected to lots of tools via MCP (Model Context Protocol) on your desktop (or in your terminal). Not quite a polished consumer app, but a great way to experiment with MCP.
Trend: ⬆️ This is very new. I picked it up because I wanted to play with MCP servers and I’m finding interesting things to do, but mostly it’s still in probation and not integrated into my workflow.
Where: Desktop
How Much: Free and Open Source (needs API keys for cloud models; I use Claude and Gemini)
Highlights: Free and easy way to experiment with MCP.
Lowlights: Alpha quality software.
What I don't use
Claude, Gemini, DeepSeek, Mistral Le Chat, MS Copilot, …
So many chatbots, and they all have their strengths, but none of these seem to offer anything interesting enough for me to adopt, learn, and pay for yet another tool.
Cursor, Windsurf, Replit, Loveable, Bolt, Claude Code, …
There’s recently a proliferation of AI programming tools, and it’s interesting to test them, but programming is one area where intimate knowledge of your toolbox pays off, so the switching costs from good old VSCode and the lovely v0 are too high.
Sora, Midjourney, Suno, …
When artistic AI first became available with image generators I was pretty excited about it, but with time I realised that I don’t have the mental bandwidth or the time for getting really interesting results. Nowadays I dedicate my creative energies to banging on drums and I admire the AI creations of others.
Character, Replika, …
I find the idea of developing a complex relationship with an AI entity enticing and creepy in equal measures. Ultimately, whenever I give it a try, I get bored quite quickly. Maybe I’m too old-fashioned, or maybe it’s just that we haven’t crossed the proverbial uncanny valley yet.
Apple Intelligence
Tried it. It’s unusable.
