AI's Bitter Lesson: when brute force beats elegant design

Every optimization algorithm, every curve fit, every AI gradient — they all trace back to one core idea: the derivative. ⚙️ It’s not just a math concept — it’s the mathematics of change. The bridge between motion, prediction, and growth. Derivatives power how machines learn, how markets move, and how innovation scales. Master this one idea, and you’ll start seeing its fingerprints across every field — from physics to finance to AI. 🌍 #STEM #TechEducation #ContinuousLearning #Calculus #Innovation

The horse Clever Hans was once believed to have learned math from his trainer. He would tap his hoof to provide the correct answer to problems, earning him a sweet carrot for his genius. Researchers got interested in understanding and dived in to uncover the truth: Hans was actually responding to involuntary cues like change in posture or shift in expression signaled from his questioner. This phenomenon is known as the observer-expectancy effect, and it exposes an issue that extends into engineering, AI and other sciences. The core problem is the divergence between our own expectations and reality — driven by things like confirmation bias or project demands. We often see a correlation between an input (X) and an output (Y) and assume a direct causal link. But sometimes, a hidden variable (Z) is the true driver. To build more robust systems we must try to control these spurious correlations. Judea Pearl's "The Book of Why" is a nice read for any engineer interested in getting to the bottom of these problems. #Engineering #AI #DataScience #MachineLearning #CriticalThinking #SystemsThinking #ExperimentDesign #TheBookofWhy

Be honest: Ever sat in math class thinking, “When will I ever use matrix multiplication?” Yeah, me too. Turns out that “useless” math is what powers 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 and every AI tool today. The real game-changer? 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 In 2017. A few researchers published "𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝗜𝘀 𝗔𝗹𝗹 𝗬𝗼𝘂 𝗡𝗲𝗲𝗱". It stopped computers from reading word-by-word (and losing the plot) and taught them to understand the 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 and connect all the dots. That's how 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀 and all modern LLMs were born. Crazy how a small idea can rewrite the rules of tech. (Paper linked in the comments 👇) Share this post if you love a good tech origin story, and follow for more insights into the future of AI!

The conclusion of the book by Brian Christian, "The Alignment Problem" is a classic. One that many working in #AI could probably relate to. On January 14, 1952, the BBC hosted a radio program that convened a panel of four distinguished scientists for a roundtable conversation. The topic was "Can automatic calculating machines be said to think?" The four guests were Alan Turing, one of the founders of computer science, who had written a now-legendary paper on the topic in 1950; philosopher of science Richard Braithwaite; neurosurgeon Geoffrey Jefferson; and mathematician and cryptographer Max Newman. The panel began discussing the question of how a machine might learn, and how humans might teach it.  "It's quite true that when a child is being taught, his parents and teachers are repeatedly intervening to stop him doing this or encourage him to do that," Turing said. "But this will not be any the less so when one is trying to teach a machine. I have made some experiments in teaching a machine to do some simple operation, and a very great deal of such intervention was needed before I could get any results at all. In other words the machine learnt so slowly that it needed a great deal of teaching."  Jefferson interrupted. "But who was learning," he said, "you or the machine?"  "Well," Turing replied, "I suppose we both were."

The Algorithm of Doubt The most dangerous sentence from an expert’s mouth: “This is definitely true.” And the most trustworthy? “I don’t know — yet.” Doubt isn’t weakness. It’s not a flaw. It’s an algorithm — one that only a few minds run, again and again. Because complex systems — differential equations, statistical models, debugging processes — don’t reward certainty. They reward those who invest time and energy to question. And still have the strength to ask again. Debugging isn’t about declaring truth. It’s about reordering probabilities. Model-building isn’t a mirror of the world. It’s an attempt to understand it — through iterations and failures. “I don’t know” isn’t the end. It’s the starting point of any intelligent system. If expertise is an algorithm, then doubt is its main loop. #intelligence #modelbuilding #debugging #doubt #science #algorithm #humility #deepthinking #biotech #AI #philosophy

A neural network learns from its target much like we learn from our goals. The error becomes feedback, the feedback shapes adjustment, and each adjustment brings us closer to mastery. In machines, this is called optimization. In humans, it’s simply called practice. And just like Newton’s law reminds us — every action has a reaction. In physics, they balance equally. But in life, reactions aren’t always equal: sometimes greater, sometimes lesser, often delayed. Yet no action ever goes without consequence. The same holds true for logic. Every logic meets a counter-logic. They may be equal, they may differ, but both can still lead to the same or even better outcomes. Learning, action, and logic are all connected by one thread: progress through feedback. Machines optimize. Humans grow. Both remind us that every step — equal or not — carries us forward. #ArtificialIntelligence #MachineLearning #NeuralNetworks #Learning #GrowthMindset #Logic #Wisdom #Inspiration

# MattersAI The Turing test Alan Turing was an English mathematician and logician and is rightfully considered to be the father of computer science (the mother of all things AI). Turing was fascinated by intelligence and thinking, and the possibility of simulating them by machines. Turing’s most prominent contribution to AI is his imitation game, which later became known as the Turing test. In the test, a human interrogator interacts with two players, A and B, by exchanging written messages (in a chat). If the interrogator cannot determine which player, A or B, is a computer and which is human, the computer is said to have passed the test. The argument is that if a computer is indistinguishable from a human in a general natural language conversation, then it must have reached human-level intelligence. What Turing meant by the test is very much similar to the aphorism by Forrest Gump: “stupid is as stupid does". Turing's version would be “intelligent is as intelligent says”. In other words, an entity is intelligent if it cannot be distinguished from another intelligent entity by observing its behavior. Turing just constrained the set of behaviors into discussion so that the interrogator can’t base his or her decision on appearances.

We Taught the Machines to Lie A paper posted last month to the arXive titled “Why Language Models Hallucinate” (Kalai et al., 2025) makes a blunt claim: LLMs hallucinate because they were trained to. The authors show that most benchmarks (including GPQA, MMLU-Pro, SWE-bench, and many others) use binary grading. Correct answers score a 1, while both “I don’t know” and wrong answers get the same 0. In that setup, a model that guesses (ie lies) beats one that admits to not knowing. If you aren't sure, the optimal strategy is to bluff. Don't be fooled by the commonly-used term “hallucination.” The models are lying intentionally and by design. Not because they are sentient and malicious and want to cause harm. They are simply doing what their training objective and benchmark rules reward. Kalai’s group formalizes this. Any model evaluated this way will necessarily prefer confident wrong answers to honest uncertainty. And because most leaderboards and post-training pipelines adopt this same rule, the bias is systemic. Their proposed fix is simple but radical. Change mainstream benchmarks to reward calibrated confidence. In their example, a model could be instructed to answer only if it has confidence > t and be penalized for wrong answers proportional to t / (1 − t). This would make abstention the rational choice when uncertain, realigning optimization toward truthfulness. But let’s be honest. While this is an excellent paper (I'm planning future posts digging in to their analysis), the headline is also obvious to (at least to experts). AI researchers didn't overlook this issue. Binary grading was useful to make early LLMs coherent and confident, and later to make them likable and seem intelligent. The chosen training objectives served practical and commercial sense. But they also made systematic hallucination inevitable, and it's probably time to relook at some fundamentals exactly like this group did. #LLMBehaviour #MachineLearningEthics #AIResearch

𝐐𝐮𝐚𝐧𝐭𝐮𝐦 𝐀𝐧𝐝𝐡𝐫𝐚 𝐒𝐞𝐫𝐢𝐞𝐬 | 𝐖𝐡𝐚𝐭 𝐑𝐞𝐚𝐥𝐥𝐲 𝐂𝐡𝐚𝐧𝐠𝐞𝐬 𝐖𝐡𝐞𝐧 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐆𝐨𝐞𝐬 𝐐𝐮𝐚𝐧𝐭𝐮𝐦? We’ve all seen how 𝐜𝐥𝐚𝐬𝐬𝐢𝐜𝐚𝐥 𝐦𝐚𝐜𝐡𝐢𝐧𝐞 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 works — you take a dataset, engineer features, train your model, and optimize predictions. But what happens when 𝐪𝐮𝐚𝐧𝐭𝐮𝐦 𝐩𝐡𝐲𝐬𝐢𝐜𝐬 steps into this workflow? That’s when the game changes — and 𝐐𝐮𝐚𝐧𝐭𝐮𝐦 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 (𝐐𝐌𝐋) begins to redefine everything we know about computation. This diagram 👇 beautifully contrasts both worlds: 🔹 On the left — 𝐂𝐥𝐚𝐬𝐬𝐢𝐜𝐚𝐥 𝐌𝐋, where learning happens through feature engineering, training, and validation. 🔹 On the right — 𝐐𝐮𝐚𝐧𝐭𝐮𝐦 𝐌𝐋, where the process involves system preparation, unitary transformations, and complexity simplification before learning even starts. Instead of processing one possibility at a time, 𝐪𝐮𝐚𝐧𝐭𝐮𝐦 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 explore 𝐦𝐮𝐥𝐭𝐢𝐩𝐥𝐞 𝐬𝐭𝐚𝐭𝐞𝐬 𝐬𝐢𝐦𝐮𝐥𝐭𝐚𝐧𝐞𝐨𝐮𝐬𝐥𝐲, thanks to 𝐬𝐮𝐩𝐞𝐫𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧 and 𝐞𝐧𝐭𝐚𝐧𝐠𝐥𝐞𝐦𝐞𝐧𝐭. The result? 𝐅𝐚𝐬𝐭𝐞𝐫 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧, 𝐝𝐞𝐞𝐩𝐞𝐫 𝐩𝐚𝐭𝐭𝐞𝐫𝐧 𝐝𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲, and 𝐚𝐧𝐬𝐰𝐞𝐫𝐬 𝐭𝐨 𝐩𝐫𝐨𝐛𝐥𝐞𝐦𝐬 that classical ML can’t even reach. This isn’t AI vs Quantum — it’s 𝐀𝐈 × 𝐐𝐮𝐚𝐧𝐭𝐮𝐦, working together to create something truly revolutionary. The future of intelligence is 𝐪𝐮𝐚𝐧𝐭𝐮𝐦-𝐞𝐧𝐡𝐚𝐧𝐜𝐞𝐝. Welcome to the era of 𝐐𝐮𝐚𝐧𝐭𝐮𝐦 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞.

♟️ When a Machine Beat the World Champion: The First True AI Breakthrough Day 3 of my 30-day AI video challenge I’m posting one video every day this month to explore the story of artificial intelligence — how it started, how it evolved, and where it might be headed next. Today’s video is A Very Short History of AI – Part 3, where I talk about what many consider the first major breakthrough in AI history: IBM’s Deep Blue defeating Garry Kasparov in 1997. It was a symbolic moment — the world’s best chess player losing to a machine. But here’s what’s interesting: chess, for all its complexity, is actually one of the easier problems for a computer to solve. The rules are fixed. Every piece moves in a predictable way. With enough computing power, an algorithm can analyze millions of future possibilities and pick the best move. Of course, achieving that in the ’90s was no small feat — it required enormous computational effort and clever programming. But that victory also revealed something deeper: real intelligence isn’t just about logic or calculation. Most of the world’s problems don’t come with clear rules or perfect information. And that’s where AI hit its next frontier. 🚀 💭 What do you think — does mastering chess prove intelligence, or is it just powerful computation?