Does AI Actually Understand? YES!
In a Nutshell
The video argues that AI demonstrates genuine understanding by meeting the observable criteria of Bloom's Taxonomy, including explaining concepts, solving complex problems, and discovering new knowledge like improved matrix multiplication algorithms. It refutes the "just statistics" objection by showing that functional capability—rather than internal experience—defines understanding, consistent with both educational standards and Ashby's law of requisite variety. The core takeaway is that AI models build internal representations of their domains and exhibit measurable understanding through performance, not subjective awareness.
These notes were generated by AI and may contain inaccuracies.
A post by Gregory Kurtzer, CEO of a tech startup and co-founder of CentOS and founder of Rocky Linux, stated that large language models do not understand anything. He described them as statistical models that output symbols based on probabilities calculated through matrix operations and context, rather than through understanding of subject matter. The post received over a million views. Kurtzer challenged the creator of the video after being called blatantly wrong.
The debate centers on the word "understands." Bloom's Taxonomy was published in 1956 by a committee led by Benjamin Bloom. The committee rejected vague phrases like "truly understand" and instead defined learning through observable behaviors and outcomes. In the 2001 revision, understanding was defined as determining the meaning of what is taught, demonstrated through interpreting, summarizing, inferring, comparing, and explaining. The six levels arranged as a ladder from bottom to top are: remembering, understanding, applying, analyzing, evaluating, and innovating.
Remembering: The Humanity's Last Exam contains 2,500 difficult questions written by experts. When released in early 2025, the best models scored below 10%. By the time of the video, the best independent test result reached approximately 61%.
Understanding: At Harvard, physics students taught by an AI instructor learned more than twice as much as students in active learning classes. Kurtzer himself acknowledged that large language models can explain difficult concepts, which is one of Bloom's listed demonstrations of understanding.
Applying: In 2025, Google's DeepMind AI system won a gold medal at the International Mathematical Olympiad with a score of 35 out of 42 points, as evaluated by the Olympic coordinators themselves.
Analyzing: In DARPA's Cyber AI Competition, AI systems identified 54 out of 63 planted vulnerabilities in real-world software and additionally discovered 18 genuine unplanted vulnerabilities.
Evaluating: A 2023 study found that GPT-4 agreed with expert graders in more than 80% of cases, approaching the inter-expert agreement rate.
Innovating: In 2025, Google DeepMind's Alpha Evolve tool discovered a method to multiply 4x4 arrays of complex numbers using 48 multiplications instead of 49, breaking a record that had stood since 1969. A team of 16 Cloud customers wrote a 100,000-line C compiler capable of compiling the Linux kernel.
The argument that AI merely predicts the next symbol describes how the system works, not what it can do. Human brains operate through neurons firing signals, yet no one claims humans therefore lack understanding.
Cybernetics pioneer W. Ross Ashby formulated the law of requisite variety in 1956: only variety can destroy variety. The controller must be at least as diverse as the problems it addresses. In 1970, Ashby and Roger Conant published the theorem that every good regulator of a system must be a model of that system. Success in regulation requires building a sufficiently similar model, whether explicitly or through evolutionary improvement of the regulator. This describes the training process of AI models. In 2025, Google DeepMind researchers stated that any agent generalizing to new multi-step targets must have learned a predictive model of its environment.
As of May 2026, the Cloud model writes more than 80% of the code that Anthropic integrates into its software repository, up from low single digits before Cloud Code's release. An AI model folded laundry using robotic arms without prior training on those specific arms, matching the efficiency of expert human operators using the arms for the first time. A model trained solely on Othello game moves constructed an internal representation of the game board; when researchers altered this representation, the model's moves changed accordingly.
Kurtzer stated that for him, understanding requires conscious comprehension, while also acknowledging that large language models can produce exceptional results and demonstrate functional capabilities. This creates a categorical error by conflating understanding with awareness. Bloom's Taxonomy and Ashby's law evaluate understanding through performance rather than internal experience. Ray Kurzweil noted that true understanding can only be measured in a conscious person, rendering it immeasurable under that standard.
Alan Turing addressed this objection in 1950, stating that the only way to be certain a machine is thinking is to be the machine and feel oneself thinking. The same logic applies to other people. Turing observed that society follows a polite convention that everyone thinks rather than debating this endlessly. He also noted that oral examinations already exist to determine whether someone genuinely understands or has merely memorized like a parrot.
A 2002 Yale University study by two psychologists asked participants to rate their understanding of everyday objects such as a zipper, toilet flush, and lock. After participants attempted to explain the mechanisms step by step, their self-rated understanding decreased. This phenomenon is called the illusion of explanatory depth. Humans can subjectively feel they understand something while being incorrect. The internal feeling of understanding is unreliable and constitutes a self-referential loop that cannot be externally verified.
In 2025, Anthropic researchers examined a Cloud model and identified an internal "I know this" signal. By default, the model states it cannot answer. When the model recognizes a name such as Michael Jordan, this signal activates and overrides the default. When researchers manually activated the signal for an unknown name like Michael Patkin, the model began referencing sports including chess. Some hallucinations result from this signal failing to activate appropriately. Machines possess their own version of the "I know this" signal and fail in the same manner as humans.
John Searle's 1980 Chinese Room thought experiment describes a person who does not speak Chinese following a rulebook to produce fluent Chinese responses without understanding. Searle concluded that computers running programs similarly lack understanding. However, a simulated calculator still performs actual addition. When the simulated entity is information processing, the simulation produces real functional output. If the room could answer any Chinese query fluently, it would be employed.
The idealist standard measures AI against a Platonic ideal of human understanding and concludes it falls short because it is not human. This constitutes a circular argument. The realist standard evaluates AI through objective, verifiable tests. Under every objective standard available, the models demonstrate understanding. Whether they understand to the same degree or in the same manner as humans remains an unanswerable speculation.
AI understanding remains unequal to human understanding in specific domains. One model provided near-perfect driving directions in New York, but when researchers blocked just 1% of streets, accuracy fell to 67%. Models continue to struggle with long projects requiring days of human effort. Kurtzer expressed concern that treating AI as an independent conscious entity could shift responsibility away from accountable humans and companies. Accountability remains one of the final tasks humans will retain because machines cannot be prosecuted or imprisoned.
Declaring that AI understands nothing contradicts the standards used by every teacher and the foundations of mathematical control theory. The only standard supporting this claim is an internal subjective feeling. The models understand to varying degrees, in different ways, and sometimes poorly, but they understand in reality.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
