Originally posted on LinkedIn in August 2025, back when GPT-5 and Claude 4.1 were the news of the day. Every word still stands.

GPT-5 is out, Claude 4.1 is here. Writing texts with them still sucks (unless you're okay with "YOU'RE ABSOLUTELY RIGHT!")

Here are my quick tips on how to write with LLMs.

TL;DR — it's not faster, it's just different.

Writing is a human task. Nature's "Writing is Thinking" kinda says it all.

With this off the table, how do I write with LLMs to avoid cringe and slop.

🧠 Split the work by strengths

Leverage the best in humans — expertise, wit, passion, and LLMs — speed, summarization, reasoning.

🎙️ Talk, don't type

Use voice mode and throw all your ideas on the topic into it. Do not care too much for the structure just yet. The most important part — talk to the machine (yes, use your vocal chords).

🏗️ Ask for structure only

Ask the LLM to produce only the structure, a Table of Contents.

✍️ Read it... then delete it

As counterintuitive as it sounds, you now have way more than 15 minutes ago when you just started. Let me explain:

• You spoke through the main points, which actually reinforced the best of your ideas and highlighted the weak parts way better than if you just typed them
• You used the best feature of LLM — restructuring data. You're lucky if it made mistakes or allowed itself some inaccuracies — now you know what to avoid and how to write it better
• And now you have a better understanding of the subject. Go on and write it down, make your own mistakes!

💫 Bonus: turn the LLM on your draft

Send the final text to the same LLM and ask it two things:

• List some questions as if they were a member of your audience. What would they want to expand on?
• List some harsh critiques from the standpoint of a nitpicky ICP. While this part is usually already pretty sloppy, some of the points GPT or Claude produces are quite useful.

Okay, this process might seem more lengthy. Because it is. But for me, it's way, way more engaging, and it allows me to produce more meaningful texts, not less. I still don't use LLMs to produce text anyone would read (if you catch me doing that, you're allowed to hit me in the knee.)

This post was written using this approach, and I even decided to pretend to be an LLM and place emojis here and there. Did I do a good job?

Originally published on the Superposition Labs LinkedIn page on April 9, 2025.

After weeks and months of talking to AI more than to real humans, I decided to embrace my writing style and language proficiency, and stop using any AI for writing or rewriting my thoughts.

Here's why.

The sabbatical

I had a long sabbatical from LinkedIn, during which I was having tremendous fun building AI agents and promoting them on X, mostly for the web3 crowd.

LinkedIn was already a swamp of slop and lazily AI-generated content back then, and it's even more so now (boy do comments here look horrible.)

I not only witnessed this cycle hyper-accelerated on X, but actually contributed in small part to it.

From "Newmarket" to Memetic

We started with the idea of AI agents as economic entities under the project codenamed "Newmarket," and quickly realized that LLMs can't be trusted with either money, time or decisions in any combination. I will write a bit more on this later.

It was glaringly obvious that an AI agent is a set of building blocks for dynamic workflows, but most of them were utterly impossible to communicate with due to their pure synthetic nature and ever-present agreeableness.

That's cool and all, but that also led to incredibly easy attacks, and diminishing results, so together with Arsenii Zemskov we focused on AI agents that have personality, and we succeeded in that with the system we call Memetic. It ingests an X profile and creates a "Reflection" that mimics how the user interacts with others. The whole thing is rooted in our deep research on social media interactions that we were doing with Talkscan at Superposition Labs.

We then built tooling around this idea, and naturally, our agents started crawling X, engaging in conversations.

The reply-bot flood

That was really fun and innovative around October–November 2024, right around when a whole bunch of open-source frameworks were released to do almost exactly the same, but with a really sloppy way of initializing these agents ("You are a software engineer who writes deep and thoughtful posts" bla bla bla...)

Within weeks every X reply section flooded with automated content that made very little sense, and within 1–2 months most of those "advanced reply bots" were muted, banned and deplatformed. Some still stand. Now X is an even worse swamp of manipulated narrative and attention farming.

We moved on with Memetic to building synthetic audiences, as we discovered that the way we initialize and build Reflections gives them a set of unique abilities for market and product research.

Why did we pivot then?

We created more than 3000 agents that were running day in and day out, and we learned an inconvenient truth for proponents of the write-with-AI approach.

If you need AI to communicate your ideas, your ideas are not worthy to communicate. Even less so to read by other humans. If your message is written and published by AI then you are redundant, and anyone can get the same text content from Claude 3.7 Sonnet.

Every kind of automated outreach is glaringly inauthentic, and it mostly feeds the Dead Internet theory.

Still an AI enthusiast

Don't get me wrong, I am still an AI enthusiast. I share tremendous excitement and chilling fears as to what are the next steps for human interactions in the AI-first world. AI is awesome at optimizing workflows, at generating stunning images and processing them. But reading text that was generated by a push of a button is beyond redundant.

Humans develop AI text fatigue incredibly fast, and I find it incredibly hard to justify investing even $20/month for a service that writes tweets on my or someone else's behalf. That is why I am not using AI for any texts that I publish, even though I am far from being fluent in English.

I have a serious gripe on "AI Avatars" precisely on the same grounds. But this is a writing for the future.

Originally posted on LinkedIn in October 2024.

Working with AI models and, especially, agents calls for robust benchmarks to evaluate their capabilities. Many of you are likely familiar with the Abstraction and Reasoning Corpus (ARC) Challenge, which has become a popular way to benchmark AI models.

As this test is quite hard for SoTA models, I use it regularly to prove to myself that I'm still human.

The Bongard Problems

Recently, Arsenii Zemskov brought an old but fascinating test back into the spotlight: the Bongard Problems, collected by Harry Foundalis.

The Bongard Problems present two sets of shapes, and your goal is to describe what is characteristic of each set so you can discern the two groups. These problems predate the ARC Challenge, and it strikes me how visionary thinkers were, conceptualizing such challenges decades ahead of their time.

A five-year-old vs. problem 32

I printed out a portion of the problem set. After working through a couple of pages, I decided to show them to my five-year-old daughter. I was curious about how she would perceive these problems, given that she hasn't been exposed to IQ tests or other early developmental assessments at her age.

Below is problem number 32. Take a moment to look at it and consider the difference between the two groups.

As someone with mathematical training, one might say: "The figures on the left have sharp projecting angles, while the ones on the right do not."

However, my daughter's interpretation was beautifully different. Her solution:

"The ones on the left — you can see up in the sky. The ones on the right — you can see on the ground."

Fresh eyes

This perspective was a powerful reminder that unfettered human intelligence, especially in children, offers profound and intuitive insights we adults might overlook. Children approach problems without the constraints and biases that we adults often carry.

In the realm of AI and complex models, it's easy to get lost in algorithms and data. But moments like these remind me that human intelligence, in its purest form, is a true source of inspiration.

Perhaps, as humanity continues to develop advanced AI, it should strive to preserve and emulate that childlike wonder and unbounded creativity. After all, the most profound insights often come from seeing the world through fresh eyes.