In his lecture "Achilles' Heel: Where Do the Limits of Artificial Intelligence Lie?", published on 16 October 2026, Eng. Abu Bakr Soliman, an NLP and LLM specialist, makes one clear argument: today's AI is a masterpiece built from matchsticks, not the Great Pyramid. It learned everything it knows from people, it can collapse at a single weak point without noticing, and much of what we are told it can "remember", "reason" or "learn" is engineering wrapped around a frozen model. Human intelligence, with its ability to sense that something has gone wrong and to adapt to situations it has never seen, is still the reference point. Below is a faithful summary of the lecture in its original order, with links to the primary sources behind each fact and a short section on what it means for companies.
Who is Abu Bakr Soliman and what is this lecture about?
Abu Bakr Soliman (MSc) works on applied AI solutions and strategy, with a focus on natural language processing and large language models. He shares his work on his YouTube channel and on LinkedIn. The lecture is a 71-minute live session that asks a simple question: we all know what AI can do today, but do we know where its abilities end?
He opens with a game borrowed from an old Spacetoon show called "The Strange Link": find the story that connects seven unrelated things. His list was the Egyptian philosopher Zaki Naguib Mahmoud, the Great Pyramid, a fish market, South Korea, the novelist Ahmed Khaled Tawfik, tea with mint and the comedian Samir Ghanem. By the end of the lecture each one has a place in the story.
Why does the lecture start by defining "perception"?
The first story is familiar to anyone in Egypt. A group of friends sit at a local coffee shop and one of them says "why don't we start a project together?" Everyone gets excited, roles are shared out, money is collected. Then reality arrives: tasks are harder than expected, partners drop out, details nobody thought about appear. At some point you realise that something is wrong.
That moment of realisation is what Soliman calls perception. He links it to Zaki Naguib Mahmoud's point that many arguments are really arguments about undefined terms. Before debating whether AI "perceives", "thinks" or is "AGI", we should agree on what those words mean. His own working definition of perception is the ability to notice, in the middle of a situation, that something strange is happening. Keep this definition in mind, because the whole lecture comes back to it.
How does a neural network learn, in five minutes?
To make the rest of the lecture accessible, Soliman explains neural networks with a translation example. You prepare pairs of inputs and correct outputs, such as an Arabic word and its English translation. The network, which you can picture as a table of numbers that are multiplied and added together, takes an input and produces a prediction. The prediction is compared with the real answer, and the gap between them is called the loss. The network then adjusts its numbers so that the loss is smaller next time.
It works like a teacher testing a student: ask a question, measure how far the answer is from the right one, correct the student and ask again, until the mistakes shrink.
What did Move 37 and Move 78 reveal about AI?
For decades the dream of AI researchers was to beat humans at their own games. IBM's Deep Blue beat world chess champion Garry Kasparov in their 1997 rematch. Chess, Soliman says, eventually became "boring" for researchers because every piece has a limited set of moves. The next challenge was Go: the rules are simple, but a player may place a stone on almost any point of a 19 by 19 board, so the number of possible games is enormous.
In March 2016 DeepMind's AlphaGo played Lee Sedol, a winner of 18 world titles, in Seoul. According to DeepMind, AlphaGo first learned from thousands of expert games and then improved by playing against versions of itself. It won the match 4 to 1. In game 2 it played the famous Move 37, a move that commentators did not understand at first and that, as Demis Hassabis wrote after the match, had a 1 in 10,000 chance of being played by a human. Hassabis later shared the 2024 Nobel Prize in Chemistry for AlphaFold.
This is the story everyone tells. Soliman wants us to remember a different one. Lee Sedol lost the first three games, and with the match already decided he played game 4 without pressure. On Move 78 he played a move that AlphaGo did not expect, which Hassabis described as another 1 in 10,000 move. AlphaGo's reply on move 79 was a mistake, and as NPR reported, Hassabis said the system did not recognise its error until move 87. It never recovered and resigned the game.
Soliman's reading is simple: AlphaGo had learned from strong players and from itself, so a move from outside that world broke it, and it kept playing weak moves without realising anything was wrong. "Everyone talks about Move 37," he says, "nobody tells you about Move 78", the moment a human found where the machine breaks.
What is AI's "Achilles' heel"?
In Greek myth, Achilles was dipped into a river that made him invulnerable, except for the heel his mother held him by. An Achilles' heel is not an ordinary weakness. It is the single point where the whole system collapses.
Soliman's example from computer vision is the one-pixel attack. Researchers Su, Vargas and Sakurai showed that changing a single pixel was enough to push 67.97% of the CIFAR-10 test images, and 16.04% of the ImageNet test images they tried, into a wrong class. No human would see the difference between the two images, yet the model sees a different object. Worse, the model has no sense that it is failing, so it carries on with full confidence. Research teams now spend real effort probing models for exactly these breaking points.
How are language models trained, from next-word prediction to ChatGPT?
Most of today's excitement comes from large language models that write code, emails and study notes. Soliman explains their training in three steps.
- Pretraining on raw text. Take sentences from books, articles and web pages and turn them into examples: given these words, what word comes next? The result is a base model that completes text, like a very powerful keyboard autocomplete. Ask it a question and it may reply with more questions. That, Soliman says, is why GPT-3, released in 2020, did not excite the wider public.
- Instruction data written by people. The missing piece was examples of tasks and good answers. Companies hired people to ask questions and experts to answer them, then retrained the base model on that data. According to OpenAI, ChatGPT, launched on 30 November 2022, was trained with human AI trainers who wrote conversations playing both the user and the assistant.
- Synthetic data. Paying people is expensive, so the next idea was to let models write both the questions and the answers. Hugging Face's open Cosmopedia dataset, for example, contains over 30 million files and 25 billion tokens of textbooks, stories and articles generated by a model. "Machines teach machines" became the new slogan.
The same idea explains a current controversy. In February 2026 Anthropic reported that DeepSeek, Moonshot and MiniMax used about 24,000 fraudulent accounts to run over 16 million exchanges with Claude, in order to train their own models on its answers. This practice is called distillation: training a weaker model on the outputs of a stronger one.
Can the student ever surpass the teacher?
Here Soliman offers a thought experiment. Take 1,000 teachers and 50,000 students, about 50 students per teacher, and let the teachers teach one course for three months with no outside sources. Then give the same exam to both groups. His rule is that, if the experiment is sound, the students' average score cannot rise above the teachers' average.
For AI, the teachers were people: their code, articles, books and stories. Models improved quickly from a low starting point, which impressed us, but in his view their average performance has not reached the average of the humans they learned from, and they are hitting a wall. He separates this from speed. A model that builds a landing page in ten minutes while a person needs a week is faster, not more intelligent, and a skilled designer's page can still be better.
Would an AI trained in 2000 have invented Git or React?
He asks developers in the audience to imagine a coding assistant trained on everything written by programmers up to the year 2000. Would it later propose single-page applications and frameworks like React, invent Git with branches and merges, design CI/CD pipelines, suggest moving JavaScript to the server, or call for new languages like Go and Rust? His answer is no: a system that only learned from the developer of 2000 stays inside that world.
He adds that AI output is still easy to recognise, pointing to the purple gradients that appear on so many generated landing pages. In 2000 some people said there was nothing new left to invent in software, and some say the same in 2026. If we hand the future to systems trained on the past, we stop producing the next Git. His advice: keep ideas and experiments competing, and do not surrender to the current scope of AI.
Déjà vu or jamais vu: what can humans do that AI cannot?
Soliman recalls a short novel from Ahmed Khaled Tawfik's Safari series, whose hero Dr. Alaa Abdel Azim suddenly realises that his life has changed: he is now a husband and a father with new responsibilities. Tawfik writes about déjà vu, the feeling of having seen something before, and its lesser known opposite, jamais vu, literally "never seen".
Humans constantly face situations they have never seen, and perception lets them adapt and look for a new solution. A system without perception, Soliman argues, cannot reach a genuine "never seen before" idea on its own.
Why does a trained model stop learning?
AI goes through two phases. In training, the model is active: it receives data and updates itself. In inference, the phase we use every day, the model is frozen. It only generates answers. Whether one person or a billion people talk to it, the model itself does not change.
Improving a model also has a cost. Make it better at translation and coding gets worse; improve its Python and its C slips. Soliman quotes Samir Ghanem's famous line about fixing the jacket only for the trousers to go wrong. Researchers call this catastrophic forgetting, and DeepMind has called it one of the fundamental limitations of neural networks. Fine-tune a model to analyse legal documents and you should expect it to lose some of its ability to translate or write stories. People do not learn this way. We pick up a new skill without forgetting everything else.
What are "fabricated capabilities"?
This is the part of the lecture that will stay with many developers. Soliman argues that the industry borrowed human words for things that work very differently under the hood.
| Capability | In humans | In today's AI, as the lecture describes it |
|---|---|---|
| Memory | The brain recalls relevant context naturally | The software around the model writes notes to files ("the user's name is Abu Bakr") and inserts them into each new prompt |
| Reasoning | An internal process that happens before we speak or act | More generated text that is part of the output itself |
| New skills | Learning something new that changes how we work | A file of written instructions, often in Markdown, loaded into the prompt |
Yann LeCun makes the same point about reasoning in his talk at ETH Zurich: "We reason internally, we don't reason in token space." Soliman's warning is that these names are partly marketing. When a model "remembers" you, it is often reading a summary file, and when that file grows too long, it gets trimmed and the "memory" is gone.
What does Yann LeCun say today's AI is missing?
Soliman recommends LeCun's June 2026 talk at ETH Zurich, "World Models: Enabling the next AI revolution", and draws three problems from it and from LeCun's wider argument.
- Errors snowball. A person who says a wrong word in the middle of a sentence corrects it and moves on. A language model builds on its own mistake. Researchers have measured this effect and named it hallucination snowballing: models defend an early wrong answer with further false claims, even though, asked separately, ChatGPT recognised 67% of those claims as wrong and GPT-4 recognised 87%.
- Language is only one part of intelligence. LeCun says in the talk that "language is easy by comparison" and that the real world is messy. Intelligence needs a model of the world that can predict, plan and ignore irrelevant detail. When a car speeds through a pile of leaves, your brain does not track every leaf. It notices that something disturbed them.
- Learning is far too expensive. In the same talk LeCun says a large language model is trained on about 20 trillion words, roughly 30 trillion tokens, which would take a human about 400,000 years to read. A four-year-old has been awake for about 16,000 hours and has taken in a comparable amount of data through vision alone. And while self-driving companies have millions of hours of training data, LeCun describes intelligence as the ability to learn to drive in about 20 hours.
The conclusion Soliman draws is that the way current AI learns has to change before we can compare it to human intelligence.
Why should you distrust people who can explain everything?
Soliman turns to Karl Popper, the philosopher of science who left Marxism as a young man because it seemed to explain every outcome: if the revolution succeeded there was a reason, and if it failed there was a reason too. Popper admired Einstein's relativity because it made claims that could be proven wrong. A theory that explains everything, he concluded, is not science.
The lecture applies this to business commentary about AI. The same voices explain why investment pours into AI and why it fails to pay back, why AI will send programmers home and why it will create more programming jobs. Soliman's request is modest: admit that this is a jamais vu moment that nobody has seen before.
He also challenges the famous line "if I had asked people what they wanted, they would have said faster horses". It is usually attributed to Henry Ford, although there is no solid evidence he said it. Soliman's point is different: Ford did not give his cars away for free, while AI companies hand out free usage to everyone, and a lot of money is burned on content with no real value.
Why shouldn't you ask the fish seller?
If your mother sends you to the fish market to find the best mullet, you would not ask a fish seller whether his fish is the best. He has an interest in the answer. Soliman applies the same logic to AI: do not take your view of the future from the companies selling it. Every new model is "the best", just as every new detergent removes "the toughest stains".
In his view the market is shaped less by AI itself than by the enormous amount of money poured into it. He quotes Dr. Ibrahim Sobh, who argues that in moments like this the people without a stake should be heard early: researchers, educators, teachers and social scientists, to ask whether spending so much of humanity's money and energy is justified.
Matchsticks or the Great Pyramid: what is the conclusion?
Imagine someone builds a pyramid out of matchsticks, one stick at a time. It is a real masterpiece and it deserves respect. But if he then tells you it is the Great Pyramid, you have to say no. That, in Soliman's words, is where AI stands: an impressive achievement, still at the matchstick stage, while human intelligence is the pyramid we are trying to reach. AI as a system has to improve and correct itself before it gets there.
Until a new kind of AI appears, his advice is to stay balanced: do not chase every announcement, do not take predictions from people with something to sell, and remember that today's AI has serious breaking points. Nobody knows what tomorrow holds, and the change may well work in your favour.
What did he advise in the Q&A?
- AI will become like a calculator. Accountants still exist after calculators and Excel. People who use the tool well remain essential.
- Transitions take longer than they feel. More than a hundred years after the car was invented, donkeys and carts still have a place in Egyptian markets. The shift for developers has been fast, but it will settle.
- The tech market is short of money, not killed by AI. He expects a new wave of startups once funding conditions improve, run by founders who learned from the money burned in this cycle.
- Aim to "be born senior". System design, software architecture and deep software engineering should be part of undergraduate study, not something juniors wait four or five years to learn.
- Get used to boredom. Short videos have weakened our ability to focus. Finish the long book, the full film, the whole game. Boredom, he says, was a cure we did not notice.
- Build your network. Many of his own opportunities came through communities and people, not only through skills. He ended the session by asking attendees to share their LinkedIn accounts and connect with each other.
What does this mean for companies using AI?
At HBS Group we build AI into real business systems, and this lecture matches what we see in projects. A few practical lessons follow from it.
- Test for your Move 78. Before launch, feed the system the unusual cases: odd customer requests, messy data, inputs that are almost right. That is where it breaks, not in the demo.
- Keep a person where perception matters. AI does not notice when it is failing. Put human review on decisions that affect money, customers or reputation.
- Look under the hood of "memory" and "reasoning". Ask how a tool stores context, what happens when it grows and where your data goes.
- Measure results, not announcements. Judge AI by what it saves or earns in your own workflows. Our guide on how to measure AI ROI shows which metrics to track, and our comparison of AI agents, chatbots and automation helps you pick the right tool for the job.
- Invest in your people. The value comes from people who understand both the work and the tool, not from the tool alone.
If you want to bring AI into your company in a way that respects these limits, connected to your own systems and with people in control, read about Amrak, our service for bringing AI into your company, or talk to our team.
You can watch the full lecture on Abu Bakr Soliman's YouTube channel. Thanks to Esmael Yahya, who prepared the video chapters.
Editor's note: this article summarises the lecture as presented. We checked the facts it mentions against primary sources and linked them above. Where a detail in the talk differed from the record, we used the source: Move 37 was played in game 2, ChatGPT launched in November 2022, and we left out a claim about the fifth game that we could not verify.
