AI

I wanted to write an article about humanoid robots. Robots are getting impressive. They can walk, run, jump, do acrobatics, move everything around, dance, and sometimes look like they are one software update away from replacing half of humanity. But I thought there was one underrated blocker: hands.

Human hands are ridiculous pieces of biological engineering. We do not just grab things. We feel them. Pressure, texture, slip, vibration, temperature, pain, softness, sharp edges, tiny surface changes. Blind people can read Braille with their fingertips. Our fingertips have one of the highest densities of sensory receptors on the body.

So my original thesis was: Humanoid robots will not become universal workers soon, because I thought we could not replicate the sensor density of human hands - 2,500/cm². You cannot take a robot like KUKA, built for manufacturing, change the software, and ask it to make an omelette. It will break the eggs for sure, but not in the pan.

Then I did what I often do now. I asked AI to argue with me. This is one of the most useful ways to use AI. Not to agree with you and write for you. But to attack your idea before you embarrass yourself with it in public.

And in this case, it worked very well. AI pulled examples of modern tactile sensors.

Meta and GelSight’s Digit 360 artificial fingertip has around 8.3 million taxels, responds to omnidirectional touch, captures multimodal signals, and uses on-device AI to process the data in real time.  

Then there is F-TAC Hand, published in Nature Machine Intelligence, which reports 0.1 mm spatial resolution across 70% of the robotic hand surface.

Wow. Yes, I was wrong. Modern technology is already way beyond density of sensors in human hands. What is missing is packaging the whole thing, including microcontrollers and servomotors, into something cheap, durable, fast, waterproof, and power-efficient. We've done it with smartphones, and then with flexible touchscreens. Human-like robot hands are next.

The real lesson was this: AI helped me invalidate my own idea before I invested too much ego into it. AI is very useful as an argument opponent. A good idea should survive being attacked. A weak idea should die early, preferably in private, before you share it with people who may laugh at you.

Here are five ways to use AI for validating ideas:

1. Ask AI to attack the idea

Use a reasoning model. Tell it to disprove the idea, not improve it. Ask for hidden assumptions, weak logic, missing evidence, counterexamples, and what would make the idea false. This works for science, legal arguments, essays, books, strategy, or technical design. OpenAI recommends reasoning models for complex multi-step tasks (source).

2. Check current facts, not model memory

For any idea depending on facts, law, science, technology, history, or market reality, use web search or deep research with citations. Ask AI to separate facts from interpretation. OpenAI’s web search and deep research tools are designed for up-to-date, sourced research, including legal and scientific research (source).  

3. Ask AI to find stronger opposing arguments

Do not only ask “is this right?” Ask: “What would the smartest critic say?” For legal arguments, ask for the other side’s best case. For science, ask for alternative explanations. For a book, ask why readers may reject the premise. Claude’s docs recommend clear role/task framing and structured prompts (source).

4. Test the structure, not just the conclusion

Ask AI to turn the idea into claims, assumptions, evidence, and gaps. Then test each link. This is useful for essays, theories, legal reasoning, research proposals, architecture decisions, and books. Use XML sections or clear headings for context, claim, evidence, objections, and verdict. Anthropic recommends structured prompting and XML-style separation (source).

5. Use a scoring rubric

Make AI judge the idea against fixed criteria: clarity, evidence, falsifiability, novelty, internal consistency, external contradiction, practical consequences, and weakest point. Structured outputs are especially useful if comparing many ideas, because they force consistent fields instead of free-form vibes. (source).

People understand Artificial General Intelligence (AGI) wrong. AI will absolutely solve problems humans could not solve. It already started. Anthropic’s Mythos reportedly found zero-day vulnerabilities in major operating systems and browsers, including bugs that survived ten or twenty years of human review, automated testing, and pentesting. One was a now-patched 27-year-old bug in OpenBSD — an operating system known for security.

But it still does not mean AI is a human mind. Look at the caption image generated by AI. You will notice immediately that something is off with the “superhuman” shoulder and the banner. Your brain spots it naturally because it reconstructs 3D structure from a flat 2D image. But for AI noticing an "AI slop" shoulder in a picture may be almost as hard as finding a rare bug in OpenBSD for you.

Evolution optimises for survival, and in humans that did not mean simply being stronger or faster. It meant self-awareness, critical reasoning, long-term planning, reading other people’s intentions, detecting lies, building trust, forming groups, and surviving inside those groups. Not good at it? You die, with your whole group. That is how survival of the fittest works. Human cooperation and social cognition are our AGI, and they are much more complicated than finding bugs in old software. We never evolved to be coders. Our brains are bad at it, and GitHub is the evidence.

Take chess. There is no human alive who can beat a top chess engine. But this does not mean chess engines are “smarter” in the human sense. It means humans were not evolved for chess.

Imagine a dystopian world where people were only allowed to have children if they were in the top 20% of chess players. Keep this show running for 1000 generations. All kids would become chess masters at kindergarden. This is not magic. This is selection. We did similar things with domestic animals. Wolves did not start as natural sheep managers, but selective breeding produced dogs that can herd, guard, retrieve, track, and read human signals in ways wolves do not.

But we never bred humans for chess. Neither for finding bugs, coding, doing materials science, drug discovery, engineering design... That is why we are naturally bad at it.

Modern AI is loosely inspired by biological neural systems. Very loosely. Like saying an aeroplane is inspired by a bird. True, but please do not expect it to lay eggs.

Human brains learn continuously. Our neural network changes all the time. That is neuroplasticity. Long-term memory is not stored as notes attached to a prompt. It is embedded into a living, changing system. LLMs are different. A trained model is mostly frozen. It does not learn from every conversation the way a human does. “Memory” in current AI systems is usually external: conversation history, saved notes, markdown files, vector databases, retrieval systems, tool logs, user profiles, or some other data pulled back into the prompt.

Imagine driving this way. Every time you approach a turn, someone gives you the road rules, your car manual, a map, your old driving notes since high school, and a few memories of accidents you avoided (or not). Then you infer what to do next - turn or not.

And that is just one example of how the human brain and AI are different.

Will AI be smarter than humans in specific domains? Yes. In many places, it already is. Will that automatically become AGI? No.

A machine can be superhuman at chess, hacking, coding, scientific research, while still being fragile with memory, context, grounding, and judgement.

The human brain is not just neural paths in the cortex. It is a significantly more complicated organ. There are some processes we still don't fully understand. This matters because LLMs were inspired by pieces of biological intelligence, and that worked incredibly well. So the most likely path to AGI may be more of the same: keep copying, abstracting, and engineering more functions from the brain and nervous system. We just started.

For three decades, startup folklore has repeated the same advice: build the MVP in a way that can grow into the real product. That was sensible when software was expensive to produce, expensive to change, and expensive to rewrite. But AI coding has changed the economics. When even Y Combinator says a quarter of its Winter 2025 startups had codebases that were 95% AI-generated, we are not looking at a minor tooling improvement. We are looking at a shift in how cheap first-pass software has become.

That does not mean architecture is dead. It means the old definition of MVP is dying. The MVP used to be a tiny version of the final machine. Now it is increasingly a fast instrument for finding the right workflow, the right UX, and the right customer pain. In other words, the value is moving from code reuse to learning reuse. Thoughtworks has already published one of the clearest signals of this change under a title many engineers would once have dismissed as heresy: “Partner with the AI, throw away the code.”

That is why I think the middle row in the classic MVP diagram deserves a comeback.

Pre-AI, shipping a skateboard when your end goal was a car was often bad advice because you were creating a dead-end product. You still had to rebuild everything later, and rebuilds were painful. In the AI-agentic era, that dead end is far less dangerous. If you can get a working product in front of users today instead of spending the next month trying to make it “scalability-shaped,” the smarter move is often to launch the ugly thing, learn fast, and replace it once the market proves you deserve a second version.

So the right way is not “scale from day one.” The right way is build for replacement. That means you can be relaxed about future scale and still be serious about today’s feedback loop. You do not need bank-grade compliance to test with a small business. You do not need perfect multi-tenant elegance to validate a painful workflow for few early adopters. You need something users can touch, judge, complain about, and maybe pay for (I wish, haha). In the AI era, UX clarity and time-to-learning beat technical dignity far more often than most architects want to admit.

Of course, there is a catch. Cheap code also makes cheap mess. GitClear’s 2025 research, based on 211 million changed lines of code from 2020 to 2024, found growing signs of eroding code quality, including a sharp rise in copy-pasted and duplicated code. But... AI is evolving fast!

That is the distinction too many articles miss. The prototype can be disposable. The thinking cannot. The first version should be built quickly, but with enough discipline that you know what was validated, what users loved, what they ignored, and what must survive the rewrite. So yes: “Build the MVP so you won’t need to rewrite it” is becoming bad advice. Build it so you can test the market immediately. Build it so you can throw it away without mourning. And if the market says yes, do not scale the prototype. Rewrite the winner.

If this resonates with you, ask me how to implement it in your organisation. Happy to help!

One day I had an idea: why don’t we use vacuum clothes dryers? Water evaporates faster at low pressure, you need a lower temperature, and heating requires a lot of electricity. Best of all, vacuum drying should be gentler on fabric.

I am not a mechanical engineer, so of course I asked ChatGPT for help. I described the idea and asked it to help me calculate the efficiency. As always with AI, the process quickly became entertaining. The idea turned into a complex thermo-vacuum hybrid solution, and the comparison with existing dryers ended up in favour of my design.

But one detail caught my attention:

I asked ChatGPT about it, and it admitted the error and recalculated. The new number looked more realistic, so I asked ChatGPT to draw a schematic blueprint of the device we had “invented”. Here was the result:

I don’t need a mechanical engineering degree to understand that something is wrong with it. It looks nice, but a lot of things just don’t make any sense. I pointed out my concerns, ChatGPT admitted there were flaws, and redrew another one, even more confusing. Then one more. And another one. After that “improvement”, I decided to stop:

The problem is that, without knowing the right technical jargon or understanding the design principles of complex thermodynamic systems, I was unable to prompt correctly. And ChatGPT itself confirmed it:

We are in a bad loop:

  • the schematic is not grounded in a proper
    engineering drawing process,
  • the image model is good at making plausible-looking
    diagrams,
  • and each revision is based on partly wrong prior
    visuals plus natural-language corrections.

So instead of converging, we are accumulating inconsistencies.

In software, I can usually catch the drift. In thermodynamics and pump topology, I can’t. So the interaction turned into a dangerous loop: plausible output, weak verification, accumulating errors.

This is the part too many people miss in the “vibe coding” conversation. AI is excellent at generating options. It is much worse at reliably telling you which option is physically, architecturally, or operationally correct. That final step still belongs to someone who knows what good looks like.

AI can help you build faster. It cannot tell you when you are confidently building a trainwreck.