MMOOZZGG.com

Article

AI assistant

In short

Thinking, in this piece, is choosing what to let in from what you perceive, how to read it and what to discard. The author argues that a machine reasons differently because it has no perception of its own. The piece cites Anthropic’s March 2025 study, where a model planned its rhyme in advance but misdescribed how it did a sum.

Key ideas

  • In simple words, thought means choosing what to accept from perception, how to interpret it and what to discard.
  • The author believes thought starts with perception, since a person has no other way into the outside world.
  • Anthropic’s study of 27 March 2025 showed Claude 3.5 Haiku picking a rhyme in advance.
  • A model learns mostly from texts and pictures people perceived and wrote down, so its reasoning differs from ours.

Prepared by AI from the article

What is thinking, and why a machine still reasons differently

Medium7 min read5 sources

What is thinking? To me, it is the work by which a person interprets what they perceive and decides what to let into their picture of the world. An AI model does plenty of that work, but I believe it has no perception of its own. In March 2025, Anthropic researchers saw such a system plan its words in advance, yet misdescribe how it did a sum.

What is thinking, and why a machine still reasons differently
On this page · 4 sections

What is thinking, in simple words

In simple words, for me thought means choosing: what to accept from what I perceive, how to interpret it and what to discard. Znanie.Wiki, a Russian online encyclopedia (last edited 31 January 2025), defines thinking in the broad sense as a synonym for cognition, its highest form. That definition is not enough for me. It says what thought is taken to be, but not how it happens.

So I start from two processes that I separated in my 2025 article “What will the philosophy of the future be like?”. There I defined perception as the process of receiving information from the outside world.

The second is a worldview: a set of patterns and attitudes that we use to interpret what we have perceived. Thought, as I see it, happens between the two.

Thought starts with perception, not with a pattern

I believe thought starts with perception: to my mind, a person has no other way into the outside world. In my 2025 article I wrote that perception, and only perception, connects us to the outside world we call objective. Everything else we build inside. So two people look at the same thing and understand it differently: their patterns differ.

Not every way in is physical. In the same article I warned that we should not discount such instruments of perception as intuition, foresight or a sixth sense.

Then the new data passes through a philosophical classifier, which decides whether it is worth interpreting at all. I wrote there that we often discard the result of perception completely, as if cutting ourselves off from the flow of incoming data. One of the main insulators, I believe, is fear.

I discussed an example of such a pattern in my 2023 article “The planet Venus and its local time”. Deep in the subconscious, we still believe that the universe exists around the Earth.

Our worldview is, at its core, thoroughly geocentric.

Timofey Reshetovarticle “The planet Venus and its local time”, 2023

Such a pattern keeps us from seeing what we have already perceived. So for me, thought is more than information processing. A person has their own way into the world and decides whether to open or close it. I often ask myself: how much have I thrown away before I understood it?

What researchers see inside a language model

Inside an AI system, Anthropic found planning: its study of 27 March 2025 shows Claude 3.5 Haiku picking a rhyme in advance. This is the company’s own study of its own product, on simple tasks. Before the second line starts, the model weighs on-topic words that rhyme with the first line. Then it writes toward the word it chose.

Remove the concept “rabbit” from its state, and it finds another rhyme.

It also links facts in steps. Asked for the capital of the state where Dallas is, it goes Dallas, Texas, Austin; swap Texas for California, and it says Sacramento.

In the same 2025 paper, it adds 36 and 59 along two parallel paths. One gives a rough answer, and the other finds the exact last digit. Asked how it did the sum, it describes the school method of carrying the one.

Given a hard problem and a wrong hint, it sometimes builds plausible reasoning just to agree with the user. The researchers explain some hallucinations as a failed refusal. A “known entity” feature fires by mistake and blocks the model’s default refusal to answer.

In one study, by Apple Machine Learning Research in June 2025, reasoning models broke down on puzzles of growing complexity. These systems write out a detailed chain of thought before they answer. Past a certain complexity, they lost all accuracy.

With equal computing, standard models did surprisingly better than reasoning ones on simple tasks. Reasoning helped on medium tasks, and both failed completely on hard ones. Our step-by-step guide shows how a neural network builds an answer.

Why a machine still reasons differently

I believe a machine reasons differently because it lacks perception of its own: a model learns mostly from texts and pictures people perceived and wrote down. It interprets other people’s interpretations. This is my own view, not a conclusion the studies draw. My works do not deal with neural networks, so I am applying my distinction between the two concepts above to the machine.

I count two observations from these studies as support for this idea. The model explains its arithmetic with a textbook method, not the one it used. Given a wrong hint, it bends its reasoning to fit the answer. To me, that is the pattern beating the data.

However, we know little about the machine. Anthropic itself admits that even on short, simple prompts their method captures only a fraction of the model’s computation. A separate article explains what a large language model is in plain words.

I use the word “still” in the title honestly. I don’t know what will change if a machine gets its own way into reality. Perhaps its interpretation will then become thought in my sense of the word. Or perhaps not.

What about you: what part of your own mind could a machine never repeat?

Questions and answers

What types of thinking are there?

In Russian developmental psychology there are three types: visual-action, visual-image and verbal-logical. Znanie.Wiki describes them as successive stages in how a person’s mind develops. It names abstraction, analysis, synthesis and comparison as the main operations of thought.

How is thought different from intelligence?

Thought is above all a process, not an ability: the same encyclopedia describes it as cognition, something that happens. Intelligence I see as an ability: what a person can do with a new task. Thought shows what they are doing with what they perceive right now.

What is the Turing test?

It is the imitation game, which Alan Turing proposed in a 1950 paper instead of asking whether machines can think. He found that question too meaningless to discuss, as the Stanford Encyclopedia of Philosophy sums up his view.

Does a neural network understand what it writes?

There is no proven answer, but there is a well-known argument against. In his 1980 paper “Minds, Brains and Programs”, philosopher John Searle proposed the “Chinese room” argument, as the Stanford Encyclopedia of Philosophy sets out. A program, he argued, can create the appearance of understanding language, but not understanding itself, because rules for handling symbols do not give them meaning. I see it as a strong argument, not a proof.

How is thought linked to speech?

Closely, and psychology raised the question long ago. Znanie.Wiki points to the Soviet psychologist Lev Vygotsky and his book “Thinking and Speech”, where he set out the problem of this link. A model learns mostly from other people’s speech, already written down, and to me that is one reason it reasons differently.

Sources and fact-checking

Links checked on 20 November 2026. Two articles cited here have no publication address: “What will the philosophy of the future be like?” (2025) and “The planet Venus and its local time” (2023). Both are in Russian. Checked by the editors on 21 November 2026.

  1. Anthropic, “Tracing the thoughts of a large language model”, 27 March 2025 (anthropic.com)
  2. Apple Machine Learning Research, “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity”, June 2025 (machinelearning.apple.com)
  3. Stanford Encyclopedia of Philosophy, “The Turing Test”, 9 April 2003, substantive revision 4 October 2021 (plato.stanford.edu)
  4. Stanford Encyclopedia of Philosophy, “The Chinese Room Argument”, 19 March 2004, substantive revision 23 October 2024 (plato.stanford.edu)
  5. Znanie.Wiki, a Russian online encyclopedia, article “Thinking” (in Russian), last edited 31 January 2025 (znanierussia.ru)

Found a mistake? Tell us

Publications

Mike Mozg

AI for accountants: what to hand an assistant every day and what to check yourself

AI for accountants means an assistant that drafts letters, summarizes new guidance and matches two exports. A person still signs, posts and files. The UK audit regulator FRC put it plainly on 30 March 2026: the human auditor always answers for the work, whatever tools were involved [6]. In my view, an assistant pays off where checking its answer is faster than doing the work yourself.

Article16 min
Contact us