It is summer here, and I have analysis paralysis. I want to write about AI and I do have things to say, but I cannot start—there is too much I want to unpack right here, right now. I have two drafts about AI, agents, media buying, and creatives, but I do not want to post them yet. I will write something general. Here and below, I mean the hyped Gen AI, not the classic stuff like SVM, CatBoost, and the rest of ML. Hype is hype, whatever you write.
I work with AI a lot compared with the rest of my work communication. Right now I am skeptical about AI in its current form: it is a tool, no worse and no better than the others, with its own peculiarities. But generally I am an AI optimist.
For me personally and for my use cases, AI now looks like this:
- with supervision, a few attempts, and cherry-picking, it is easy to get a one-off wow result
- it is hard to keep getting a boring, autonomous result at stable quality
So almost every agentic solution, for example, is still more or less like that.
Thinking models
This is what I use most often: throw in (or dictate) more context and talk. Always with boundaries around the goals and a “be brutally honest.”
Why thinking? Because non-thinking models make a lot of things up.
Everything I have is in Obsidian. I keep all my notes in a way that makes them easy to use with AI. There are also cursor-based cases outside development (in Russian), but that is not what this post is about.
Then all sorts of things:
- Delegation: discuss development tasks, management or marketing situations, and processes.
- Storming: media-buying approaches, ways to solve engineering tasks, creatives, texts, that sort of thing.
- Talking to my own texts and asking them questions.
- Preparing for difficult meetings. For example, if I need to talk to lawyers, I first work through the situation with AI, then confirm it with the lawyers.
- Finding and comparing, getting a summary of solutions, frameworks, and approaches, then talking through the results.
- For the previous item I sometimes use deep research. I like OpenAI’s version. Google’s is okay too, but verbose and willful. I have a few funny cases where Google went off to research object1 when I wanted object2. It told me: I am now researching object1 (which the user calls object2).
- Technical tasks, something like: it would not start in the terminal, what am I doing wrong, halp.
- Talking to documentation: hello, context7.
AI does not prepare a concrete result for me. I never copy something out of a chat and use it. It is more like talking to a smart and slightly strange professional who gives me food for thought. The benefit is this:
- Talking to AI forces you to structure your thoughts well.
- Fast access to general knowledge and feedback based on it.
So far I see nothing better than o3-pro and o4-mini-high. Anthropic is close in quality, but OpenAI models are the easiest for me to work with. I dropped Pro a few times, then came back two or three days later.
NotebookLM
I started writing and realized this is a topic for a separate post. The short version: any article or pile of information turns into a decent podcast.
Code and technical tasks
I can write and read code, but I do it slowly. Claude Code and Cursor write one-off scripts and prototypes for me. Sometimes, if the task is well structured and the code will be compact and disposable, I use o4-mini-high.
That said, code is not all good news. Yes, vibe-coding one-off sandbox scripts works well if your hands are straight. For grown-up cases, see for yourself: tldr: take senior developers on a large codebase (exactly that)—development with AI takes 20% more time, even though the developers report a 20% speedup. There is a lot of this, and not just in code.
Replacing Google
When I need to find something and then follow a link, or get a short answer from search, I use Perplexity instead of Google. I do not like its Research / Lab: it hallucinates like hell and gets facts mixed up. The only place it is okay is consumer tasks—comparing headphones and reading reviews.
Other:
- non-thinking: translation, summaries, data parsing, ad texts, and so on
- everyday cases: draft a formal complaint, ask something while traveling
- local models when I do not want to share something with OpenAI