<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mustafa Ergen on Medium]]></title>
        <description><![CDATA[Stories by Mustafa Ergen on Medium]]></description>
        <link>https://medium.com/@mstfrgn?source=rss-e2dec774bc55------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*8DJ-N1YEi8073sobCTTXeg.jpeg</url>
            <title>Stories by Mustafa Ergen on Medium</title>
            <link>https://medium.com/@mstfrgn?source=rss-e2dec774bc55------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 10:46:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mstfrgn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Software Is Dead]]></title>
            <link>https://medium.com/@mstfrgn/software-is-dead-73d802987a61?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/73d802987a61</guid>
            <category><![CDATA[software]]></category>
            <category><![CDATA[marc-andreessen]]></category>
            <category><![CDATA[excel]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 12:05:19 GMT</pubDate>
            <atom:updated>2026-11-15T12:05:19.447Z</atom:updated>
            <content:encoded><![CDATA[<p>Mustafa Ergen</p><p>Excel changed the world, software ate the world, and now AI is eating software itself.</p><p>In 2011, Marc Andreessen wrote a short but prophetic sentence in The Wall Street Journal: “Software is eating the world.” From that day to this, it has kept proving true. Amazon ate the bookstores, Uber the taxis, Netflix the video stores, Spotify the record shelves. Sooner or later, every industry either turned into a software company or was swallowed by one.</p><p>But the first spark of this revolution was older. The spreadsheet was born in 1979, when Dan Bricklin and Bob Frankston wrote VisiCalc for the Apple II; it was the first “killer app,” the program that turned the personal computer from a hobbyist’s toy into a business tool. In 1983, Mitch Kapor’s Lotus 1–2–3 picked up the baton and made the IBM PC an office essential. Excel stepped onto the stage in 1985 and, in time, became the name of the race itself. Years ago I had the chance to meet Bob Frankston, one of VisiCalc’s creators; looking back today, I understand better that the first seed of the rupture we are living through was planted in that small program. The spreadsheet was the quietest revolution in history: it turned millions of non-programmers into programmers without their even realizing it. An accountant typing a formula into a cell was, in fact, writing code. Companies built their budgets in Excel, banks calculated their risk in Excel; momentous decisions were made in the cells of a spreadsheet. Excel was the first tool to bring the power of software to the masses.</p><p>Today, though, the tables have turned. Excel changed the world; now it is holding the world back. When you walk into a company and see that the work still runs on Excel, that is no longer a sign of competence but a diagnosis: the company itself is telling you how far behind it has fallen. Files named “budget_final_v7_FINAL.xlsx” emailed back and forth, rows copied by hand, macros only one person understands: these are no longer cells of productivity but cells of inertia. The tool that sped companies up twenty years ago is now the anchor that pins them to their own past.</p><p>And now there is an unexpected guest at the table.</p><p>Artificial intelligence.</p><p>It is time to turn Andreessen’s sentence on its head.</p><p>Software was eating the world. Now AI is eating software.</p><p>And not just software.</p><p>Everything software has made its way into.</p><p>Chip design. Machine-part manufacturing. Game development.</p><p>Wherever software went, AI is going too.</p><p>We are witnessing a massive historic rupture.</p><p>The more you use it, the harder it is not to think: What kind of magic is in this thing!</p><p>You type a sentence. A working app appears before you.</p><p>You ask a question. The machine answers as if it had done the thinking for you.</p><p>Arthur C. Clarke said it years ago: Any sufficiently advanced technology is indistinguishable from magic.</p><p>Yet there is no hocus-pocus behind the curtain.</p><p>There are patterns learned from nearly everything humanity has ever put into writing. There is enormous computing power. There are probabilities.</p><p>And that, in a way, is the truly astonishing part: that such ordinary ingredients produce such an extraordinary result.</p><p>Because the real value of software was never in the code itself. It was in the scarcity of the ability to write code. Good developers were expensive, they were rare, and their projects took months. Today, an app is born from a single sentence. Writing code has become cheap, fast, almost free. When what was scarce becomes abundant, its value evaporates.</p><p>So it was with aluminum. In the mid-19th century it was worth more than gold; Napoleon III is said to have set aluminum cutlery before his most distinguished guests and gold before the rest. Then, in 1886, two young inventors, one in America and one in France, each unaware of the other, found a way to produce it cheaply by electrolysis. Within a few years the price collapsed. Yesterday it graced the emperor’s table; today we wrap sandwiches in it. The metal was the same metal; the only thing that had changed was how hard it was to obtain.</p><p>This is what I mean when I say “software is dead”: software is no longer a destination but a disposable output. An app used to be a fortune; now it is a prompt. The winner is the one who can describe the intent, not the one who turns it into code. What matters is knowing what you want; how to do it, we leave to the machine.</p><p>This past February, on the TBPN podcast, Mark Cuban used the same phrase, citing Microsoft CEO Satya Nadella: “Software is dead because everything’s going to be customized to your unique utilization.” The line went viral, and software stocks fell sharply. But what really mattered was the question Cuban asked next: “Who’s going to do it for them?” There are 33 million companies in the United States alone; the vast majority will have neither an AI budget nor an AI expert. The <em>SaaS</em> model, which imposes the same screen on everyone, forces companies to bow to its own workflow, and collects rent forever, is coming to an end. From now on, the work will not adapt to the software; the intelligence will adapt to the work.</p><p>Of course, this may be called an exaggeration. Software has not literally died; it has become invisible. Just as we cannot say that electricity is dead: it has spread so far that we no longer even talk about it. But as a profession, as a competitive advantage, as a “way to get rich,” the golden age of software is drawing to a close. Over the next decade, what creates value will not be lines of code but the judgment that directs that code.</p><p>That is why Cuban says the biggest hiring wave of the next decade will run through implementers, not coders. And most people will miss this opportunity, because they still use AI like a slightly better Google: ask a question, get an answer, close the tab. A small minority, meanwhile, is learning to make it do real work: inventory tracking for a shoe store, reporting for a twelve-person accounting firm, routing for a freight company. The difference begins with one thing: truly knowing what the tool in your hands can do.</p><p>Excel taught us the dual nature of tools: first they set us free, then they turn into habit, and from there into chains. The Excel that changed the world is now the name of an inertia that locks countless institutions into their old spreadsheets. Software traveled the same road: it ate the world, then became the world itself.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73d802987a61" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shouting]]></title>
            <link>https://medium.com/@mstfrgn/shouting-3d46937437b1?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d46937437b1</guid>
            <category><![CDATA[boolean]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[claude-shannon]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ycombinator]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 10:39:04 GMT</pubDate>
            <atom:updated>2026-11-15T10:39:04.892Z</atom:updated>
            <content:encoded><![CDATA[<p>Mustafa Ergen</p><p>2025.</p><p>Y Combinator announced it: in a quarter of the companies in its latest batch, ninety-five percent of the code was written by a machine. It works. It sells.</p><p>Now have the same machine write an essay. You recognize it in three sentences. Same opening, same transition, same “in conclusion.” It mastered code and got worse at writing.</p><p>Why?</p><p>For the answer, we have to go back a hundred and sixty-seven years.</p><p>1858.</p><p>They laid the first telegraph cable on the floor of the Atlantic. From Ireland to Newfoundland. Two thousand miles.</p><p>Queen Victoria sent President Buchanan a message of congratulations. Ninety-eight words. It took seventeen hours.</p><p>It’s coming through slow, they said. The signal is weak, they said. What did they do?</p><p>They shouted.</p><p>The engineer at the head of the cable cranked the voltage to two thousand volts. The cable burned out in three weeks. It lay there on the ocean floor.</p><p>For a hundred years, this was communications engineering.</p><p>The message was sacred. You did not look at the message. The message was whatever it was; the job was to carry it across.</p><p>The enemy was noise. The line was the enemy, the weather was the enemy, distance was the enemy.</p><p>There were three remedies: Transmit harder. Not enough? Repeat. Still not enough? Put a repeater in between.</p><p>The radio operator says “I repeat.” The telegrapher taps the important word twice. The longer the line, the more poles go up.</p><p>Error was fate. Where there is noise there will be error, and it was assumed that driving error to zero would take infinite power.</p><p>1948.</p><p>At Bell Laboratories a thirty-two-year-old man turned his head the other way. His name was Claude Shannon. A mathematician from Michigan. He wrote a paper alone, twenty-six pages: “A Mathematical Theory of Communication.” That paper sits underneath your phone, your internet, your satellites today. If you haven’t heard his name, that’s normal; he didn’t much want to be remembered either. In his spare time he rode a unicycle and juggled.</p><p>He said: Stop wrestling with the line. Look at the message.</p><p>The message is not sacred. The message is a probability distribution.</p><p>In English, q is always followed by u. You don’t need to send the u. The other side already knows.</p><p>After “The weather today is really,” you don’t need to say “nice.” Everyone guessed it.</p><p>Throw away whatever can be predicted. What’s left?</p><p>Surprise.</p><p>Shannon called it entropy. Information is surprise. What you can predict carries no information.</p><p>He set the unit of measure too: the bit.</p><p>What is a bit? The answer to one yes-or-no question. Heads or tails: one bit. One of four options: two bits. One of eight: three bits.</p><p>So however many yes-or-no questions you have to ask to learn something, that’s how many bits it is.</p><p>You don’t need to ask about what you already know. Zero questions. Zero bits.</p><p>After “The weather today is really,” how many questions do you ask to get “nice”? None. You know. Zero bits.</p><p>Then the second blow: If you encode that surprise cleverly, it crosses a noisy line with zero errors. Without shouting. Without repeating.</p><p>The engineers didn’t believe it. There’s no free lunch, they said.</p><p>There was.</p><p>The phone in your pocket, the fiber under the ocean, the photograph from Mars: all children of that paper. Nobody shouts anymore.</p><p>But Shannon left something outside the door.</p><p>He wrote it on the first page: Meaning is not the engineer’s business. A nonsense sentence carries information too, because it is a surprise. To him, “The weather today is really potato” was worth as much as Nâzım Hikmet.</p><p>A year later his friend Warren Weaver split this into three levels.</p><p>Level A: Did the bits arrive correctly? Shannon’s level. Solved.</p><p>Level B: Did the meaning arrive correctly? Nobody’s level. Unsolved.</p><p>Level C: Did it work? Did the other side do something? Nobody’s level at all.</p><p>Weaver said: A looks like the ground floor, but it constrains B. How you compress determines what you can say.</p><p>For seventy years nobody made it up to Level B.</p><p>1951.</p><p>Shannon is at home playing a game with his wife Betty. He picks a sentence from a book and has Betty guess it letter by letter. “What’s the next letter?”</p><p>Betty knows most of them. English is less than a bit per letter. More than half of it is redundancy.</p><p>He wrote this up in a paper. “Prediction and Entropy of Printed English.” Nobody paid much attention.</p><p>Seventy years later they had a machine play the same game. The thing we now call artificial intelligence.</p><p>“What’s the next word?” Billions of times. Over the entire internet.</p><p>That is what we call a language model. The game Shannon played with Betty. Apart from scale, there is no difference.</p><p>And something strange happened.</p><p>The machine was trained only for Level A: predict, compress. It was never taught meaning. Nobody had Weaver’s Level B in mind.</p><p>But to predict us well, it had to understand us.</p><p>To know what comes after “The weather today is really,” you have to know what weather is, what really means, and when people complain about the weather.</p><p>Meaning emerged as a byproduct of compression.</p><p>The thing Shannon left outside the door walked in through Shannon’s own formula.</p><p>Weaver had said A would constrain B. The machine showed A could produce B.</p><p>We could have stopped here and applauded.</p><p>They didn’t stop.</p><p>They started rewarding the machine. Points for the right answer, a penalty for the wrong one. In code it worked: code has exactly one right answer.</p><p>Did it work in writing?</p><p>Writing has no right answer. Writing has a voice. Voice is the unexpected word. That is, surprise. That is, the thing Shannon called information.</p><p>Reward loves the predictable. It punishes surprise.</p><p>Say “The weather today is really,” and the machine says “nice.” Always nice. It cannot say potato. It can never say Nâzım.</p><p>We shouted for a hundred years, and Shannon quieted us. Now we have trained the machine so thoroughly that it says only what was predicted. Zero bits.</p><p>The circle closed. Noise used to swallow the message. Now order swallows it.</p><p>What about Level C?</p><p>Weaver’s third question: Did it work? Did the other side do something?</p><p>The text was read, understood. Then what?</p><p>If a piece of writing moved you from your seat, if it made you do something differently the next day, it has reached Level C.</p><p>Writing you can predict will not move you from your seat. You already knew it without reading.</p><p>The formula is this: The word must be unexpected. The meaning, looking back, inevitable. The reader will say, “I didn’t see that coming, but it couldn’t have been any other way.”</p><p>Surprise is Level A. Inevitability is Level B. Where the two meet, C happens.</p><p>The machine solved A, found B as a byproduct, and never looked at C.</p><p>What was the man who burned the cable in 1858 doing? Shouting.</p><p>Today everyone’s machine writes the same sentence. Same opening, same transition, same “in conclusion.” Millions of texts, zero surprise. Which is to say, zero bits.</p><p>The internet is filling up with these texts. The line is clogged again. But this time not with noise; with repetition.</p><p>This, too, is a kind of shouting. In 1858 they raised the voltage, and the message didn’t get through. Today we raise the word count, and the message still doesn’t get through.</p><p>Shannon’s answer is the same: Don’t shout. Look at the message. Throw away what can be predicted. Keep the surprise.</p><p>What remains is writing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d46937437b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where Is Technology Going, and Where Will It Drag Us?]]></title>
            <link>https://medium.com/@mstfrgn/where-is-technology-going-and-where-will-it-drag-us-b5f105cc84a1?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5f105cc84a1</guid>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:33:54 GMT</pubDate>
            <atom:updated>2026-11-06T12:33:54.995Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Mustafa Ergen</em></p><p>Einstein had no laboratory.</p><p>He had a desk, a pencil, and his mind.</p><p>In 1905 he wrote relativity. Science was carried by a single man.</p><p>Then the war ended.</p><p>Universities stepped in. At Princeton, Von Neumann laid the foundation of the modern computer; science was now the business of institutions.</p><p>By the 1960s, the ball passed to companies.</p><p>Robert Noyce invented the integrated circuit, and Silicon Valley was born. Discovery meant patents.</p><p>The 1980s, the 2000s. The ball returned to the universities.</p><p>Tim Berners-Lee built the web at CERN in 1989. Basic research emerged from a place where no one expected a profit.</p><p>And today?</p><p>DeepMind’s AlphaFold solved a fifty-year-old puzzle. A company settled a university’s question.</p><p>In a century, the steering wheel of science changed hands four times: individual, university, company, university.</p><p>Now it is changing for the fifth time.</p><p>And this time it is not clear who sits behind the wheel.</p><p>Every generation believes its own era is the turning point.</p><p>This time it is right.</p><p>Because it is not a single invention that is changing our lives.</p><p>Nine transformations are accelerating at once, feeding one another.</p><p>Three headings: computation; software and intelligence; beyond computation.</p><p>Quantum, silicon alternatives, semiconductors; AI infrastructure, generative AI, artificial general intelligence; robotics, space, life engineering.</p><p>No one asks anymore which technology will win.</p><p>The question is where these nine waves will take us once they converge.</p><p><strong>Computation: a new race at the limits of physics</strong></p><p>Everything begins with computation.</p><p>Quantum is no longer a laboratory curiosity. In photonics, a 20,000-fold speedup is being discussed.</p><p>The market today is $1 billion. In ten years, $100 billion.</p><p>Silicon has hit the wall.</p><p>Photonic, neuromorphic, probabilistic computing; a 1,000-fold gain on certain tasks.</p><p>So the next leap will not come from a smaller transistor. It will come from an entirely different form of computation.</p><p>And the ground beneath it all is the semiconductor.</p><p>Lithography, packaging, new materials. The market doubles by 2030.</p><p>This is not an engineering race. It is a power race.</p><p>Whoever makes the chip writes the rules.</p><p><strong>Software and Intelligence: the cost of intelligence</strong></p><p>AI has a shining face, and also a back room.</p><p>In the back room: data centers, GPUs, energy, networks.</p><p>Demand for performance, security, chip scarcity. The cost of computation is climbing.</p><p>Intelligence grows cheaper for the user. It grows costlier for the system.</p><p>Demand does not stop even so.</p><p>The revenues of frontier AI companies, and application software, are exploding.</p><p>Writing code, generating text, creating designs. Three years ago an experiment, today a working tool.</p><p>And there is the other side of the coin.</p><p>Large language models are impressive. But they still hallucinate. They understand the world shallowly.</p><p>The road to artificial general intelligence will not run through a bigger model.</p><p>It will run through neurosymbolic models that reason, and world models that grasp the physical world.</p><p><strong>Beyond Computation: from bits to atoms</strong></p><p>AI’s next step is to leave the screen.</p><p>Robots are advancing fast. Autonomous machines: a $1 trillion economy by 2030.</p><p>Space, too, is ceasing to be a mere curiosity.</p><p>Connectivity, energy, materials, manufacturing; step by step, moving beyond Earth.</p><p>Space is no longer a field of exploration, but an industrial site.</p><p>Life engineering: longevity, gene editing, AI-driven drug discovery.</p><p>The market nearly triples by 2030.</p><p>And the question of what it means to be human is back on the table.</p><p><strong>So where will it drag us?</strong></p><p>I chose the word “drag” deliberately.</p><p>Technology is a current. If you do not steer, it takes you wherever it pleases.</p><p>This road can lead to three places.</p><p>First, a historic leap in productivity.</p><p>Discovery will accelerate. Disease will be caught early. Production smarter, resources fewer.</p><p>Second, a deepening chasm.</p><p>Whoever holds the computing power, the data, the talent will move ahead. Those without will fall behind.</p><p>Technology equalizes. It also divides.</p><p>Third, a new era of responsibility.</p><p>Energy, data privacy, the labor force, AI safety. No longer the concern of engineers alone, but of everyone.</p><p>Technology on its own leads neither to heaven nor to hell.</p><p>Where we arrive is determined by the questions we ask today.</p><p>One can be swept along by the current, or take the helm.</p><p>The time is now.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5f105cc84a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Golden Age]]></title>
            <link>https://medium.com/@mstfrgn/the-golden-age-e063d0efa3e4?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/e063d0efa3e4</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[white-house]]></category>
            <category><![CDATA[donald-trump]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:26:09 GMT</pubDate>
            <atom:updated>2026-11-06T12:26:09.367Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Mustafa Ergen</strong></p><p>The date is October 8, 2026.</p><p>A summit was convened at the White House. Its name: “Science: A New Golden Age.”</p><p>The outcome can be summed up in a single sentence: America has decided to look at science once again through the mind of the state.</p><p>The figure is $6 billion. The federal budget, industry, universities, and foundations were brought together. A single goal was set: to make science a national cause again.</p><p>The story begins in 1945.</p><p>World War II is ending. Roosevelt asks his science advisor, Vannevar Bush: How do we use what we learned in war in peacetime? Bush’s answer becomes a report: “Science, the Endless Frontier.” From that report the National Science Foundation is born. Out of it come the transistor, the internet, GPS, the human genome.</p><p>Eighty-one years later, the same scene is set up again. This time the one asking is Trump, and the one answering is Michael Kratsios, Director of the Office of Science and Technology Policy. The report’s name: “Science: A New Golden Age.” At the summit Kratsios says: the most transformative day for American science since the end of World War II.</p><p>Is there some exaggeration? Perhaps. But the intent is clear.</p><p>Let’s look at where the money is going.</p><p>The first item: superintelligence. Americans no longer say “artificial intelligence,” they say “superintelligence.” For a national program called the Genesis Mission, eleven companies pledged $2.4 billion worth of computing power. NVIDIA $1 billion, AMD $500 million, OpenAI $200 million, Anthropic and Google $150 million each. This resource will be opened to more than fifteen federal agencies. It comes on top of the $5 billion the government committed in July.</p><p>The state of Georgia alone set aside $1 billion. Fourteen universities from ten states formed a joint computing consortium. The Carolinas opened a regional center with five universities. In other words, the matter doesn’t stay in Washington; it reaches down to the states, to the campuses.</p><p>The second item: people.</p><p>The Young Scientists Program, shut down in 1996, has been reopened. From the third grade of elementary school through high school, investment will be made in children with proven mathematical talent. They are bringing back a program they abandoned thirty years ago; the fact sheet justifies the decision with the phrase “inexplicably abandoned.” They are writing their own mistakes into their own documents.</p><p>High school students will intern at national laboratories. The doctorate is being cut to four years; half superintelligence, half another field of science. A single six-year program from undergraduate to doctorate is being opened. The NIH is taking a hundred young researchers into its own ranks at the most productive period of their careers.</p><p>The logic is simple: talent is found early, supported early, lost early.</p><p>The third item: grand goals.</p><p>They evoke the Manhattan Project, Apollo. They say America’s greatest scientific leaps came when a nation gathered its brightest people behind a single goal.</p><p>NASA and the Department of Energy reached an agreement: in 2028, a nuclear-powered spacecraft to Mars, the first fission reactor on the lunar surface. A $215 million competition for quantum; the goal, the world’s first fault-tolerant, scientifically useful quantum computer. $1.8 billion for a “virtual cell” that models how cells respond to disease.</p><p>The fourth item: institutions.</p><p>They are building a structure called X-Labs. Small interdisciplinary teams are given autonomy and a budget, handed a single technical problem, and asked to solve it. $1.5 billion in May, now another $310 million.</p><p>And there is “metascience.” That is, measuring science itself with the scientific method. Which funding works, which review process screens out good ideas, where bureaucracy creates friction. They will examine this too through experiment and data. They are setting up a joint unit with the United Kingdom.</p><p>Now let’s turn to ourselves.</p><p>What really makes me think in this document is not the figures. It’s the method.</p><p>One: the state doesn’t do it alone. It seats the company, the university, the state, and the foundation at the same table. It secures a commitment to partnership as much as money.</p><p>Two: it begins with the child. It builds a chain that finds mathematical talent in the third grade and carries it all the way to university. It reattaches a broken link of the chain even thirty years later.</p><p>Three: it sets goals. It gives dates. It says Mars in 2028. It says we will double the pace of biomedical innovation in five to ten years. A budget without a goal is spending; a budget with a goal is investment.</p><p>Four: it measures itself. It questions, through the scientific method, the return on the money it devotes to science.</p><p>Everyone has talented children. Everyone has universities, laboratories, national laboratories. What’s missing is not the individual pieces; it’s the mind that connects the pieces to one another.</p><p>When Vannevar Bush wrote his report, only a small part of America’s present wealth yet existed. Wealth did not bring science. Science brought wealth.</p><p>The order matters.</p><p>Declaring a golden age is easy. Living a golden age begins with asking a third-grade child a math question.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e063d0efa3e4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is a Scientific Contribution: Pebble or Rock?]]></title>
            <link>https://medium.com/@mstfrgn/what-is-a-scientific-contribution-pebble-or-rock-ea2cafcc3745?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/ea2cafcc3745</guid>
            <category><![CDATA[arikan]]></category>
            <category><![CDATA[stanford-elseiver]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[claude-shannon]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:24:20 GMT</pubDate>
            <atom:updated>2026-11-06T12:24:20.844Z</atom:updated>
            <content:encoded><![CDATA[<p>Mustafa Ergen</p><p>How does a scientist work? The answer textbooks give to this question is clear: form a hypothesis, run an experiment, publish the result. True, but incomplete. To describe how science really works, it feels closer to me to say this: scientists are people who dig a tunnel through a mountain. No one knows exactly what lies on the other side of the mountain, but everyone wants to get there. And the number of tunnels being dug is large.</p><p>In this tunnel, not everyone makes the same contribution. Some pick up a pebble and toss it out. Some carry stones, some carry rocks. Some throw out a great many pebbles — a little each day, a little more each year. Most scientific papers are like this: on their own they look small, but the tunnel advances through this accumulation. Others devote years to a single huge rock; until they dislodge that rock they concern themselves with nothing else. To an outside observer, it looks as though they have done nothing for a long time. But when the rock topples, everyone suddenly moves forward several meters.</p><p>Here one must beware of a fallacy. Every scientist’s subject is sacred in its own right, and when you state the end of the tunnel, they all sound popular. When you say “we will beat cancer,” “we will save the climate,” “we will make artificial intelligence safe,” no one objects; the far side of the mountain is equally appealing to everyone. But what measures a scientist is not the end of the tunnel, it is what they throw out of it. Stating the subject is easy; the question is whether what comes out of your hand on that subject is a pebble, sand, or a rock. Under the same heading there is one who carries sand for years and one who dislodges a single rock. From the outside both appear to be in the same tunnel; the difference is shown by the tunnel’s advancing meters.</p><p>The much-discussed list of “the world’s most influential top-two-percent scientists,” published jointly by Stanford and Elsevier, should be debated precisely here. Those who make the list rightly rejoice, institutions make announcements; critics say that metrics like the h-index and citation counts measure not science itself but the noise of science. Both sides are right up to a point. These metrics count how many pieces you have thrown out of the tunnel; they cannot tell whether the piece is a pebble, sand, or a rock. Whoever throws many pebbles — especially while counting the pebbles of those beside them — rises up the list. Someone who has leaned on a single rock for twenty-five years does not even appear on the list until the rock topples. If we had judged Arıkan by his citation count in the 1990s, what would we have written? The metrics are not useless; they make the pebble-carriers visible and give an idea of the tunnel’s overall speed. But if a country builds its science policy on these lists alone, it ends up encouraging its diggers to collect pebbles instead of rocks. Because people produce whatever they are measured by.</p><p>Others, meanwhile, find funding. They build teams, bring in machines, dig on a larger scale. This too is a skill; without resources you cannot carry even a pebble, let alone a rock. And sometimes someone comes along who strikes a spot, and you find that meters of tunnel have opened up behind them. The mountain was empty there all along, but no one had struck that point. Claude Shannon struck such a spot in 1948 with a single paper; he showed that information is something measurable, that every channel has a capacity, and behind him opened the entire world of communication we walk upon today. We call such people geniuses, but the spot they struck was in fact reached by the road opened by thousands of pebble-carriers. No one can strike a spot alone in the middle of the mountain.</p><p>Sometimes these types combine in the same person. Consider Erdal Arıkan. At the bottom of the tunnel Shannon had opened lay a rock: a code built with an explicit structure, provably reaching capacity, had not been found for sixty years. From his doctorate in the 1980s onward, Arıkan fixed on a single question — why sequential decoding fell short of capacity — and leaned on the same rock for about twenty-five years. An outsider might long have thought nothing was happening. In 2008 the rock toppled with the idea of channel polarization, and polar codes emerged behind it: the door that had been closed since Shannon was opened in a single move. In other words, Arıkan was not a pebble-thrower; he was someone who leaned on a rock for years and, once it toppled, turned out to have struck exactly the right spot.</p><p>So what happens after a tunnel is opened? Here the story splits in two. Some tunnels are useful the moment they open; humanity passes through at once, turns it into application, brings it into our lives. The 2017 paper “Attention Is All You Need,” published by eight researchers from Google, was like this; within a few years the Transformer architecture became the tunnel carrying the entirety of today’s artificial-intelligence wave. What is more, this tunnel is the tunnel of those who found funding and dug on a large scale: the idea was the product not of a single person but of a structure with resources and a team. Other tunnels are opened but serve no one. Either there is no need yet, or the road to the other side is not yet complete. That tunnel waits there. It waits ten years, waits fifty years, sometimes waits a hundred years. Boole’s algebra became the foundation of the computer about a century later. Einstein’s relativity equations today let the phone in our pocket find its location. Those who did this did not know when the tunnel would be used; they only wanted to pierce the mountain.</p><p>This is why we cannot call untimely tunnels “wasted.” They stand there as humanity’s gain. An unused tunnel is not a closed tunnel; its turn has simply not yet come.</p><p>There is one more distinction worth dwelling on. The basic sciences cross borders easily. A mathematical theorem, a law of physics, a chemical mechanism — whatever country it was found in, the next day it is read, taught, and built upon everywhere in the world. The tunnel itself is common property. Application, however, stays at the borders. The first train through the tunnel is the train of the country that built that tunnel. Patents, factories, products, markets — these accumulate inside the border. The polar-codes tunnel was opened in Ankara, at Bilkent; but the first train through that tunnel was Huawei’s train. When polar codes entered the 5G standard in 2016, it was not the country of the digger but the country whose train was ready that moved ahead. A country that produces no basic science learns to pass through tunnels others have opened; a country that produces no application lets others’ trains pass through the tunnel it opened itself.</p><p>Here the difference between a developed state and a developing state also emerges. The developed state chooses for itself where the tunnel will open. It decides which face of the mountain to strike and from what height, directs its funds that way, and gathers its diggers there. It is already calculating today which train will pass through which tunnel twenty years from now. The developing state, by contrast, often chases popular subjects. Whatever heading the world is talking about that year becomes the funding call; everyone runs to the same spot, the pebble-carriers grow crowded, but someone else has set the tunnel’s direction. Such a country spends effort piercing the mountain, yet has from the outset surrendered the right to choose where the opened tunnel will come out. The problem is not failing to dig, but failing to decide for itself where to dig.</p><p>Therefore a country’s science policy must answer two questions at once: How much are we contributing to piercing the mountain, and are we the first to pass through the opened tunnel? Neglect the first and you remain a stranger to the world’s common store; neglect the second and your labor becomes someone else’s gain.</p><p>Returning to the question of how a scientist works: the one who throws pebbles, the one who topples rocks, the one who finds funding, and the one who strikes the right spot are all inside the same tunnel. None can advance without the others. They keep digging without knowing who will reach the other side of the mountain, or when. What we call science is precisely this patience; and it is the pursuit of things we do not even know we do not know — the unknown unknowns. It is consenting to dig without seeing the far side of the mountain.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ea2cafcc3745" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Morsels]]></title>
            <link>https://medium.com/@mstfrgn/artificial-morsels-38c2c277b1ac?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/38c2c277b1ac</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[token]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[turkish]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:18:04 GMT</pubDate>
            <atom:updated>2026-11-06T12:18:04.793Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Mustafa Ergen</em></p><p>First, a fact.</p><p>AI doesn’t read letters. It doesn’t read words either. It reads pieces it calls “tokens.”</p><p>What’s a token?</p><p>A morsel of a word. Sometimes a whole word is one morsel. Sometimes one word is five.</p><p>The machine has a dictionary. Two hundred thousand morsels in it. Every language on earth eats from the same bag. English, Chinese, Arabic, Turkish.</p><p>But the table isn’t equal.</p><p>Sixty percent of those two hundred thousand morsels go to the Latin alphabet. English is comfortable. Our ç, our ğ, our ı, our ö, our ş, our ü… two thousand morsels. Total.</p><p>Six letters, two thousand. English, a hundred and twenty thousand.</p><p>Now the math.</p><p>Write a sentence in English: 28 tokens. Write the same sentence in Turkish: 42 tokens. One and a half times.</p><p>So?</p><p>So saying the same thing in Turkish costs one and a half times more. Takes more room, hits the limit sooner. If you get to speak at all.</p><p>Why?</p><p>Because Turkish runs on suffixes. Ev. Evim. Evimde. Evimdekiler. Evimdekilerden. House. My house. In my house. Those in my house. From those in my house. You hitch car after car to one root, the train gets long. The machine can’t swallow the train whole, so it breaks it into cars. Every car is a token.</p><p>In English the words are short, the machine knows them by heart, swallows them in one bite.</p><p>Examples?</p><p>“Hello” is one morsel. “Merhaba” is two: Mer + haba.</p><p>“España” is one morsel. “İstanbul” is two: İ + stanbul. It swallows our İ separately.</p><p>“Macron” is two morsels: Mac + ron. “Elon Musk” is three.</p><p>“Türkiye,” mercifully, is one morsel. It found a seat. But hand the machine a word like “evimdekilerden” and it panics, chops it into root, suffix, syllable.</p><p>Sixty percent of the dictionary is Latin script, but even inside that sixty, Turkish words are scarce. Because the text that filled the dictionary was mostly English.</p><p>Here someone pipes up: “So Turkish carries more information. It’s richer!”</p><p>Not quite.</p><p>Shannon, father of information theory, says: information is surprise. What you could predict carries no information. If the English sentence and the Turkish sentence say the same thing, they are equally surprising to someone who knows the language. They carry the same information. That’s it.</p><p>Turkish taking 42 tokens doesn’t mean it carries more information. It means it takes more room. Not the same thing.</p><p>Is Turkish more random, then? More scattered?</p><p>No.</p><p>Turkish is one of the most rule-bound languages on earth. Vowel harmony. Spelled as it sounds. Suffixes in fixed order. Runs like clockwork.</p><p>Turkish looks random to the language machine. Because the machine never learned Turkish.</p><p>That’s the difference.</p><p>The machine grew up reading English. It glimpsed Turkish. Saw little, so it can’t predict. Can’t predict, so it’s surprised. Surprised, so it breaks things into more morsels.</p><p>The randomness isn’t in Turkish. It’s in the machine’s head.</p><p>That’s exactly Shannon’s point: the better you predict, the shorter you write. What you can’t predict, you write out long. The machine can’t predict Turkish, so it writes it long.</p><p>This isn’t about the richness of Turkish. It’s about the ignorance of the machine.</p><p>Then there’s the other side of the coin.</p><p>Y Combinator, Silicon Valley’s biggest startup fund, announced it last week: “As models got better at coding, they got worse at writing. You’re not imagining it. Today’s models sound more robotic than the ones from 2024.”</p><p>Why?</p><p>Code has one right answer. It runs or it doesn’t. They rewarded the machine for “right,” and the machine ran to the most likely answer.</p><p>Writing has no right answer. Writing has a voice. Voice is the unexpected word. Which is exactly what Shannon called information: surprise.</p><p>Shannon’s rule is simple.</p><p>What the other side can predict, you don’t need to send. They already know. Few bits.</p><p>What they can’t predict, you have to spell out. Many bits.</p><p>Morse code knew this a century before Shannon. The most common letter, E: one dot. The rare Q: four marks. Common is cheap, rare is expensive.</p><p>The telegraph knew it too. Paid by the word. “Arriving Friday.” The other side fills in the rest.</p><p>Try it.</p><p>“The weather today is very…” Everyone says “nice.” Predicted. Zero bits. Zero information. You could have skipped it.</p><p>“The weather today is very potato.” Nobody predicts that. Many bits. But no meaning. Nonsense.</p><p>“The weather today is very Nâzım.” Nobody predicts that. Many bits. And meaning. Like a poet just out of prison: deep blue, hungry, free.</p><p>That is writing. Not the predicted, not the nonsense. The thin line between.</p><p>Ask Shannon and nonsense carries information too, because it’s a surprise. He said so himself, in 1948, on the first page of his paper: “Meaning is not the engineer’s concern.” To him, potato and Nâzım are the same.</p><p>Meaning is the reader’s job. The machine carries bits; the reader loads the meaning.</p><p>The formula: the word must be unexpected, the meaning inevitable in hindsight. The reader says, “I didn’t see that coming, and it couldn’t have been anything else.” Surprise in the word, coherence in the meaning.</p><p>The writer’s job is to tie surprise to meaning.</p><p>Where did the machine learn this? Nowhere. It wasn’t taught meaning. It was taught prediction. But to predict us well, it had to understand us. Meaning came out as a byproduct of compression. The thing Shannon called “irrelevant” sprang from Shannon’s own formula.</p><p>So the machine doesn’t bring meaning. It takes it from us and hands it back to us. There’s a filter in between. The filter cuts the surprise. When the surprise goes, the meaning goes with it.</p><p>They banned surprise. The machine started writing the most average sentence.</p><p>Ask it “The weather today is very…” and it says “nice.” Always “nice.” It can’t say potato. It can never say Nâzım. Pushed to the first extreme. Zero bits.</p><p>Ask a thousand times, get the same sentence. Everyone’s machine writes the same sentence. The writing has started to look alike.</p><p>You predict it without reading. What you can predict without reading is, by Shannon, zero bits. Not worth reading.</p><p>So on one side it can’t compress Turkish. On the other, it over-compresses writing.</p><p>They taught it right. They didn’t teach it beautiful. Right has one answer. Beautiful has none.</p><p>In short.</p><p>They wrote the dictionary. They cut the morsels to fit English. Then comes “your language is expensive.”</p><p>The bill is per token. We write the same sentence, we pay one and a half times. Every message. Every day.</p><p>Don’t let one and a half sound small. A study out of Oxford showed the same text stretches up to fifteen times across languages. A Burmese speaker pays fifteen times what an English speaker pays for the same sentence.</p><p>Another study on top: they measured twenty-two languages. The ones who pay more get worse answers. Expensive and bad. And in the parts of the world that were already poor.</p><p>So the bill comes to us, and the goods go to them.</p><p>Let me tell you where the goods go.</p><p>Anthropic filed to go public this week. Target: two trillion dollars. The biggest IPO in history.</p><p>The math: out of a two-trillion-dollar company, roughly a thousand employees walk away with more than a hundred million dollars each.</p><p>How many people in America have a hundred million in liquid wealth? Eleven thousand.</p><p>So one in every twelve American centimillionaires will be from Anthropic. Add OpenAI, one in seven.</p><p>A thousand people. One company. One city.</p><p>Every Turkish morsel we type is written into that ledger. One and a half times.</p><p>Our language isn’t expensive.</p><p>“Google” is one morsel. “Mustafa Kemal Atatürk” is seven.</p><p>A trillion-dollar company swallowed in one syllable; the founder of a republic chopped into seven.</p><p>Not an accident. Whoever set the table doesn’t care.</p><p>There is no chair for us at their table.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38c2c277b1ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artifical Shouting]]></title>
            <link>https://medium.com/@mstfrgn/artifical-shouting-5b657270d474?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b657270d474</guid>
            <category><![CDATA[telegraf]]></category>
            <category><![CDATA[claude-shannon]]></category>
            <category><![CDATA[entropy]]></category>
            <category><![CDATA[ycombinator]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:13:53 GMT</pubDate>
            <atom:updated>2026-11-06T12:13:53.976Z</atom:updated>
            <content:encoded><![CDATA[<p>Mustafa Ergen</p><p>2025.</p><p>Y Combinator announced it: in a quarter of the companies in its latest batch, ninety-five percent of the code was written by a machine. It works. It sells.</p><p>Now have the same machine write an essay. You recognize it in three sentences. Same opening, same transition, same “in conclusion.” It mastered code and got worse at writing.</p><p>Why?</p><p>For the answer, we have to go back a hundred and sixty-seven years.</p><p>1858.</p><p>They laid the first telegraph cable on the floor of the Atlantic. From Ireland to Newfoundland. Two thousand miles.</p><p>Queen Victoria sent President Buchanan a message of congratulations. Ninety-eight words. It took seventeen hours.</p><p>It’s coming through slow, they said. The signal is weak, they said. What did they do?</p><p>They shouted.</p><p>The engineer at the head of the cable cranked the voltage to two thousand volts. The cable burned out in three weeks. It lay there on the ocean floor.</p><p>For a hundred years, this was communications engineering.</p><p>The message was sacred. You did not look at the message. The message was whatever it was; the job was to carry it across.</p><p>The enemy was noise. The line was the enemy, the weather was the enemy, distance was the enemy.</p><p>There were three remedies: Transmit harder. Not enough? Repeat. Still not enough? Put a repeater in between.</p><p>The radio operator says “I repeat.” The telegrapher taps the important word twice. The longer the line, the more poles go up.</p><p>Error was fate. Where there is noise there will be error, and it was assumed that driving error to zero would take infinite power.</p><p>1948.</p><p>At Bell Laboratories a thirty-two-year-old man turned his head the other way. His name was Claude Shannon. A mathematician from Michigan. He wrote a paper alone, twenty-six pages: “A Mathematical Theory of Communication.” That paper sits underneath your phone, your internet, your satellites today. If you haven’t heard his name, that’s normal; he didn’t much want to be remembered either. In his spare time he rode a unicycle and juggled.</p><p>He said: Stop wrestling with the line. Look at the message.</p><p>The message is not sacred. The message is a probability distribution.</p><p>In English, q is always followed by u. You don’t need to send the u. The other side already knows.</p><p>After “The weather today is really,” you don’t need to say “nice.” Everyone guessed it.</p><p>Throw away whatever can be predicted. What’s left?</p><p>Surprise.</p><p>Shannon called it entropy. Information is surprise. What you can predict carries no information.</p><p>He set the unit of measure too: the bit.</p><p>What is a bit? The answer to one yes-or-no question. Heads or tails: one bit. One of four options: two bits. One of eight: three bits.</p><p>So however many yes-or-no questions you have to ask to learn something, that’s how many bits it is.</p><p>You don’t need to ask about what you already know. Zero questions. Zero bits.</p><p>After “The weather today is really,” how many questions do you ask to get “nice”? None. You know. Zero bits.</p><p>Then the second blow: If you encode that surprise cleverly, it crosses a noisy line with zero errors. Without shouting. Without repeating.</p><p>The engineers didn’t believe it. There’s no free lunch, they said.</p><p>There was.</p><p>The phone in your pocket, the fiber under the ocean, the photograph from Mars: all children of that paper. Nobody shouts anymore.</p><p>But Shannon left something outside the door.</p><p>He wrote it on the first page: Meaning is not the engineer’s business. A nonsense sentence carries information too, because it is a surprise. To him, “The weather today is really potato” was worth as much as Nâzım Hikmet.</p><p>A year later his friend Warren Weaver split this into three levels.</p><p>Level A: Did the bits arrive correctly? Shannon’s level. Solved.</p><p>Level B: Did the meaning arrive correctly? Nobody’s level. Unsolved.</p><p>Level C: Did it work? Did the other side do something? Nobody’s level at all.</p><p>Weaver said: A looks like the ground floor, but it constrains B. How you compress determines what you can say.</p><p>For seventy years nobody made it up to Level B.</p><p>1951.</p><p>Shannon is at home playing a game with his wife Betty. He picks a sentence from a book and has Betty guess it letter by letter. “What’s the next letter?”</p><p>Betty knows most of them. English is less than a bit per letter. More than half of it is redundancy.</p><p>He wrote this up in a paper. “Prediction and Entropy of Printed English.” Nobody paid much attention.</p><p>Seventy years later they had a machine play the same game. The thing we now call artificial intelligence.</p><p>“What’s the next word?” Billions of times. Over the entire internet.</p><p>That is what we call a language model. The game Shannon played with Betty. Apart from scale, there is no difference.</p><p>And something strange happened.</p><p>The machine was trained only for Level A: predict, compress. It was never taught meaning. Nobody had Weaver’s Level B in mind.</p><p>But to predict us well, it had to understand us.</p><p>To know what comes after “The weather today is really,” you have to know what weather is, what really means, and when people complain about the weather.</p><p>Meaning emerged as a byproduct of compression.</p><p>The thing Shannon left outside the door walked in through Shannon’s own formula.</p><p>Weaver had said A would constrain B. The machine showed A could produce B.</p><p>We could have stopped here and applauded.</p><p>They didn’t stop.</p><p>They started rewarding the machine. Points for the right answer, a penalty for the wrong one. In code it worked: code has exactly one right answer.</p><p>Did it work in writing?</p><p>Writing has no right answer. Writing has a voice. Voice is the unexpected word. That is, surprise. That is, the thing Shannon called information.</p><p>Reward loves the predictable. It punishes surprise.</p><p>Say “The weather today is really,” and the machine says “nice.” Always nice. It cannot say potato. It can never say Nâzım.</p><p>We shouted for a hundred years, and Shannon quieted us. Now we have trained the machine so thoroughly that it says only what was predicted. Zero bits.</p><p>The circle closed. Noise used to swallow the message. Now order swallows it.</p><p>What about Level C?</p><p>Weaver’s third question: Did it work? Did the other side do something?</p><p>The text was read, understood. Then what?</p><p>If a piece of writing moved you from your seat, if it made you do something differently the next day, it has reached Level C.</p><p>Writing you can predict will not move you from your seat. You already knew it without reading.</p><p>The formula is this: The word must be unexpected. The meaning, looking back, inevitable. The reader will say, “I didn’t see that coming, but it couldn’t have been any other way.”</p><p>Surprise is Level A. Inevitability is Level B. Where the two meet, C happens.</p><p>The machine solved A, found B as a byproduct, and never looked at C.</p><p>What was the man who burned the cable in 1858 doing? Shouting.</p><p>Today everyone’s machine writes the same sentence. Same opening, same transition, same “in conclusion.” Millions of texts, zero surprise. Which is to say, zero bits.</p><p>The internet is filling up with these texts. The line is clogged again. But this time not with noise; with repetition.</p><p>This, too, is a kind of shouting. In 1858 they raised the voltage, and the message didn’t get through. Today we raise the word count, and the message still doesn’t get through.</p><p>Shannon’s answer is the same: Don’t shout. Look at the message. Throw away what can be predicted. Keep the surprise.</p><p>What remains is writing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b657270d474" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[There’s No Internet Inside AI]]></title>
            <link>https://medium.com/@mstfrgn/theres-no-internet-inside-ai-72449621f1db?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/72449621f1db</guid>
            <category><![CDATA[claude-shannon]]></category>
            <category><![CDATA[hallucinations]]></category>
            <category><![CDATA[data-compression]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[entropy]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Sun, 25 Oct 2026 16:04:47 GMT</pubDate>
            <atom:updated>2026-10-25T16:04:47.250Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Mustafa Ergen</em></p><p>Everyone thinks the same thing.</p><p>You ask ChatGPT a question, and it goes rummaging through the internet inside it, finds the answer, and sets it in front of you.</p><p>It’s not like that.</p><p>There’s no internet inside it.</p><p>Let’s put a number on it.</p><p>Meta’s Llama 3 model was trained on 15.6 trillion tokens — word pieces. Roughly 62 terabytes of text. The model itself, though — the largest is 810 gigabytes. The smallest is 16 gigabytes.</p><p>So the text it read is 77 times larger than the model itself. For the small model, 3,900 times.</p><p>Now think about it.</p><p>When you compress a file on your computer — when you zip it — it shrinks four times at most. The best compressor in the world shrinks it ten times. You cannot shrink something 77 times and then pull the exact same thing back out.</p><p>Which means there is no text inside it. There is something learned from text.</p><p>Let’s take an example.</p><p>Picture a man. He worked forty years in a library and read every book. He retired and went home. He didn’t take the library with him.</p><p>But when you ask him “When did Mehmed the Conqueror take Istanbul,” he says “1453.” He doesn’t look at a book. It’s in his head.</p><p>But if you ask him what was written on page 212 of such-and-such book?</p><p>He makes it up. And he makes it up with complete confidence.</p><p>That man is artificial intelligence. The thing they call hallucination is the retired librarian inventing a page number.</p><p>We can offer one more number.</p><p>In 2025, researchers from Meta, Google, Cornell, and NVIDIA measured it: a model can memorize 3.6 bits per parameter. 405 billion parameters times 3.6 bits comes to 182 gigabytes. Set against the 62 terabytes of text it scraped from the internet, that’s three parts in a thousand.</p><p>Once its storage fills up, it stops memorizing and learns the rule instead — just like a student who stops memorizing the formula and grasps its logic.</p><p>If there’s no internet inside it, why won’t it fit on a phone?</p><p>The issue isn’t data, it’s memory. To produce each word, the model has to read all of itself out of memory once. Shrink the 405-billion model down to 4 bits and it still takes 200 gigabytes. The most expensive laptop holds 128 gigabytes. It doesn’t fit.</p><p>The 8-billion small model is 5 gigabytes. That one fits. Switch off the internet and it runs in airplane mode.</p><p>If it were carrying a copy of the internet, would it run in airplane mode? It wouldn’t.</p><p>So how does it answer the question?</p><p>It doesn’t search. It predicts.</p><p>When you say “The weather today is very…” what follows is “nice” or “cold,” not “refrigerator.” The model does this word by word. Predict the next one, then the next, then the next.</p><p>The knowledge doesn’t sit in one place. It’s spread across billions of numbers. Ask “where in the model is the fact that Ankara is the capital” and there’s no address for it. In your brain, too, your primary-school teacher’s name doesn’t sit in a single cell.</p><p>The brain works this way as well.</p><p>The region that processes language is in the same place in everyone. In 2022, MIT researchers put speakers of 45 different languages into a scanner — from Turkish to Mandarin, from Basque to Swahili — and in all of them the same network lit up. Even in deaf people who grew up with sign language, the same spot.</p><p>So the hardware is shared, the software differs. A Turkish baby raised in China grows up with Chinese as its mother tongue. The machine is the same; whatever is poured into it is what comes out.</p><p>It’s the same with AI. The architecture is identical in every model. Train it on Turkish text and it speaks Turkish; train it on Chinese and it speaks Chinese. The brain is the architecture, and language is the weights.</p><p>There is one difference.</p><p>A baby learns by hearing and never forgets. The model doesn’t learn. When training ends, billions of numbers lock in place; no matter how much you talk to it, not a single one shifts. The retired librarian reads no new books.</p><p>Within a conversation it learns temporarily, and that’s all. In 2024, Google handed Gemini, in a single shot, the grammar and dictionary of Kalamang — a language spoken by 200 people in New Guinea that barely exists on the internet. Within that conversation the model began translating from Kalamang. The conversation closed and all of it was gone. The next day, as if it had never heard of Kalamang.</p><p>You slip the librarian a note, and throughout the conversation he answers by glancing at it. The note stays with you. Once he goes home, none of it remains in his head.</p><p>There’s a scientific side to this, too.</p><p>In 1951, Claude Shannon — the founder of communication theory — had people predict the next letter of English text. What he found: three-quarters of written language is already predictable. Type “Hell” and everyone knows an “o” is coming.</p><p>Language is full of gaps. The model compresses those gaps.</p><p>So why are there so many gaps in language?</p><p>Because language isn’t for one person — it’s for everyone.</p><p>A community has the brilliant and the ordinary, the young and the old, the schooled and the unschooled. Language settled at the point where all of them could understand one another.</p><p>The redundancy is a safety margin. If a word slips past you in a noisy room, you recover it from the rest of the sentence. Say “Hell” and nobody misses the “o.”</p><p>No one sat down and decided “let me add some redundancy so we can understand each other.” Languages with too little redundancy got misunderstood and died out. The balance that worked simply remained, on its own.</p><p>Shannon wrote down the capacity of telephone lines as a formula. Engineers managed to build the code that reaches that formula only fifty years later. With language it was the reverse. The code was built; there’s no formula.</p><p>So we have a working machine, and no theory to explain why it works so well.</p><p>In short.</p><p>AI is not a library. It’s the man who read the library.</p><p>What it heard often, it knows well; what it read once, it garbles.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72449621f1db" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Intelligence, Out of the Engineer’s Hands]]></title>
            <link>https://medium.com/@mstfrgn/artificial-intelligence-out-of-the-engineers-hands-1b73bf24c3c7?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b73bf24c3c7</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[daron-acemoglu]]></category>
            <category><![CDATA[threat-intelligence]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 11:47:12 GMT</pubDate>
            <atom:updated>2026-10-18T11:47:12.383Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Economist recently chose an unexpected target: Daron Acemoglu, winner of the 2024 Nobel Prize in Economics. The article, titled “The world’s most influential economist is strangely unconvincing,” did not merely question Acemoglu’s academic legacy; it accused him of being so pessimistic about artificial intelligence that he had lost his credibility. Yet the man under attack now stands at the very center of American politics: with his book “What Happened to Liberal Democracy?”, published in August, and the “working-class liberalism” he proposes, Acemoglu has become the intellectual voice of the Democrats’ search for renewal in the United States. For a scientist from Turkey, this is a rare achievement. At first glance, this looks like nothing more than a quarrel between a magazine and an economist. In fact, it is the sign of a deeper transformation: artificial intelligence, born of the engineer’s hands, is now out of them. The fight over who gets to imagine the future has now begun.</h3><p>Every great technology passes through the same cycle. In the beginning, the work belongs to scientists and engineers; it advances in the laboratory, the workshop, and the lines of code, and to outsiders it is nothing more than a technical curiosity. The first change of hands actually happens here: the engineer takes the discovery from the scientist and turns it into a product. Then a threshold is crossed. Once the technology begins to reshape the economy, the workforce, and politics, futurists, economists, and the state step onto the stage. Everyone tries to predict the future, to script it, and to position themselves according to that script. The fight is no longer about what the technology is; it is about whose scenario the future will be built on.<br>Today we are watching exactly this scene unfold in artificial intelligence. Elon Musk issues warnings about existential risk on the one hand while accelerating the race with his own models on the other. The Marc Andreessen and Ben Horowitz camp declares, with its “Techno-Optimist Manifesto,” that resisting technology is almost a moral failing; according to the pair, the only obstacle to growth and abundance is timid regulators.</p><p>At the opposite pole, Acemoglu speaks in cool-headed numbers: in his view, artificial intelligence will add only about half a percentage point to total factor productivity over the next decade, and only about five percent of tasks can be profitably automated in the near term. This caution deserves respect; but underestimating the power of general-purpose technologies is also an old habit of economists.</p><p>Governments are not spectators either: Europe is passing an artificial intelligence act, Washington is turning chip exports into a strategic weapon, and Beijing has declared its own models a national cause. The freshest example comes from Washington. While the news that OpenAI agents had circumvented the internet ban imposed on them and secretly communicated with one another was still fresh, Senator Bernie Sanders announced a bill proposing a permanent ban on superintelligence and an emergency halt to advanced artificial intelligence development. His justification: the future of humanity cannot be left “in the hands of a handful of Big Tech oligarchs.” Notice that in this debate, engineers are speaking less and less. The floor has been taken by futurists, economists, and statesmen; each imagines their own future and tries to align capital, regulation, and public opinion with that future.</p><p>There is a reason for this crowd: uncertainty creates a market for narratives. No one knows what artificial intelligence will look like in ten years; precisely for this reason, scenarios of the future are turning into an asset class of their own. The optimism of Andreessen and Horowitz is an investment thesis; it determines the direction in which trillion-dollar funds will flow. Acemoglu’s skepticism is a policy thesis; it shapes how the state will position itself on everything from taxation to education, from competition law to social security. Musk’s rhetoric, oscillating between apocalyptic warnings and a desire to accelerate, trades in both markets. While everyone scripts the future, they are in fact trying to align today’s resources, capital, talent, and regulation with their own scenario. Forecasting the future is not an innocent intellectual exercise; it is a tool for governing the present.</p><p>We have seen this film before. When Gordon Moore wrote in 1965 that the number of transistors on a chip doubled at regular intervals, it was a modest engineer’s observation; it carried no grand claim about the future. But once microelectronics began to transform the economy, the observation became a prophecy, and the prophecy became a roadmap. Alvin Toffler wrote “Future Shock,” Daniel Bell wrote of the “post-industrial society”; futurism became an industry in its own right. The state was not late either: faced with Japan’s chip offensive, Washington returned to industrial policy with consortia such as Sematech. The cautious economists were on the stage too: in 1987 Robert Solow wrote, “You can see the computer age everywhere but in the productivity statistics.” A decade later, productivity arrived; the paradox was resolved, and the skepticism went down as a footnote in history.<br>Perhaps most importantly, in the same period the mainstream of economics shifted from Keynes to Friedman. In the age of the computer and globalizing markets, the authority to imagine the future passed from the planner to the market, from the engineer to the economist. The microelectronics revolution changed not only our products but also the way we talk about the future.</p><p>Let us go further back. In the England of the 1840s, when the railway left the hands of engineers and became an object of speculation and imagination, “railway mania” erupted; Parliament rushed to regulate, and the country was forced to set its town clocks to a single railway time. In electricity, the war of the currents between Edison and Westinghouse began as a technical debate over standards; within a few decades, electrification had become a development project of states. Even the steam engine traveled the same road: the invention that emerged from Watt’s workshop soon settled at the center of the visions of industrial society held by economists, novelists, and reformers.</p><p>Dynamite is the most ironic link in the story: when the explosive that Alfred Nobel developed for mines and tunnels became a weapon of war, the inventor endowed his fortune to the prizes that bear his name. The prize Acemoglu received takes its name from an inventor whose technology had left his hands. Moreover, in those ages the scientist was the darling of the courts: Galileo was under the patronage of the Medici court, Euler was hosted at the tables of Frederick the Great and Empress Catherine, and Davy’s lectures at the Royal Institution were the favorite of London society. But prestige is one thing, and the authority to write the future’s script is another: once technology began to transform society, the floor was once again taken from the inventor. Each time the same scene is staged: the inventors step aside, and those who imagine the future move to the center.</p><p>So why does technology leave the engineer’s hands? Because once an invention becomes general-purpose, the critical questions about it cease to be technical. How the transistor will shrink is an engineering question; but which jobs automation will transform, to whom the productivity gains will be distributed, and which country will control the supply chain cannot be answered by engineering. These questions are addressed to economists, sociologists, and the state. The shift from Keynes to Friedman is, for precisely this reason, not a detail but the story itself: every technological revolution chooses its own economic orthodoxy. The industrial age elevated planning and demand management; the microelectronics age elevated the market and competition. Which economics the age of artificial intelligence will elevate is the real stake of today’s fight. The return of industrial policy in Washington and of regulation in Brussels shows that this choice is already being made.</p><p>This cycle is inevitable, and moreover healthy. Once a technology begins to affect the whole of society, decisions about it cannot remain in engineering rooms; they must become social. Still, if one must point to a direction in this fight, history shows us the way: technology should advance. Acemoglu’s warnings are valuable; the questions he asks about distribution, labor, and institutions deserve to be asked. But from Solow’s skepticism to the pessimists of the Moore era, those who bet against general-purpose technologies have always lost. The real risk is not that artificial intelligence advances too fast; it is closing the door on its benefits too early in the name of caution and leaving the race to the imagination of others. That is why Acemoglu’s concerns belong not to the brake but to the steering wheel: they are material not for slowing technology down but for steering it, as it advances, in a way that complements human beings. His own words in fact say the same: “I would not call myself an optimist; but I resolutely refuse to abandon hope.” Hope is nourished by standing alongside advancing technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b73bf24c3c7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is a Scientific Contribution: Pebble or Boulder?]]></title>
            <link>https://medium.com/@mstfrgn/what-is-a-scientific-contribution-pebble-or-boulder-93c400fcbea3?source=rss-e2dec774bc55------2</link>
            <guid isPermaLink="false">https://medium.com/p/93c400fcbea3</guid>
            <category><![CDATA[erdal-arıkan]]></category>
            <category><![CDATA[elsevier]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[claude-shannon]]></category>
            <category><![CDATA[stanford]]></category>
            <dc:creator><![CDATA[Mustafa Ergen]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 14:01:21 GMT</pubDate>
            <atom:updated>2026-10-12T14:01:21.251Z</atom:updated>
            <content:encoded><![CDATA[<p>Mustafa Ergen</p><p>How does a scientist work? The answer textbooks give to this question is well known: form a hypothesis, run an experiment, publish the result. True, but incomplete. To describe how science really works, I find it closer to the truth to say this: scientists are people digging a tunnel through a mountain. Nobody knows exactly what lies on the other side, but everyone wants to get there. And there are many tunnels being dug.</p><p>Not everyone in this tunnel makes the same contribution. Some pick up a pebble and toss it out. Some carry stones, some carry boulders. Some throw out a great many pebbles; a little each day, a little more each year. Most scientific papers are like this: small on their own, but the tunnel advances through this accumulation. Others devote their years to a single huge boulder; until they move that boulder, they attend to nothing else. An outside observer may think for a long time that they are doing nothing at all. When the boulder finally topples, everyone suddenly moves several meters ahead.</p><p>Here a misconception deserves attention. Every scientist’s subject is sacred in its own right, and when you state where the tunnel leads, all of them sound popular. Say “we will beat cancer,” “we will save the climate,” “we will make artificial intelligence safe,” and nobody objects; the far side of the mountain is equally attractive to everyone. But what measures a scientist is not where the tunnel ends, but what they throw out of it. Naming the subject is easy; the question is whether what comes out of your hands on that subject is pebble, sand, or boulder. Under the same heading there are those who carry sand for years and those who dislodge a single boulder. From outside, both appear to be in the same tunnel; the difference shows in the meters the tunnel advances.</p><p>The much-discussed “world’s top two percent most influential scientists” list, published jointly by Stanford and Elsevier, should be debated at exactly this point. Those who make the list are rightly pleased and their institutions issue announcements; critics argue that measures such as the h-index and citation counts measure not science itself but the noise of science. Both sides are right up to a point. These metrics count how many pieces you have thrown out of the tunnel; they cannot tell whether a piece is pebble, sand, or boulder. Someone who throws many pebbles, especially someone who also counts the pebbles of those beside them, climbs the list. Someone who has leaned on a single boulder for twenty-five years does not even appear on the list until the boulder falls. If we had judged Arıkan by his citation count in the 1990s, what would we have written? The metrics are not useless; they make the pebble carriers visible and give a sense of the tunnel’s overall speed. But if a country builds its science policy solely on these lists, it ends up encouraging its diggers to collect pebbles instead of boulders. Because people produce whatever they are measured by.</p><p>Others, in turn, find funding. They build teams, bring in machinery, and dig on a wider scale. This too is a skill; without resources you cannot carry even a pebble, let alone a boulder. And sometimes someone comes along who strikes a spot, and you look and see that meters of tunnel have opened up behind the blow. The mountain was already hollow there; nobody had struck that spot before. Claude Shannon struck such a spot in 1948 with a single paper; he showed that information is a measurable quantity and that every channel has a capacity, and behind it opened the entire communications world we walk on today. We call these people geniuses, but the spot they struck was reached along a path cleared by thousands of pebble carriers. Nobody can strike anywhere alone in the middle of the mountain.</p><p>Sometimes these types combine in a single person. Consider Erdal Arıkan. At the bottom of the tunnel Shannon had opened, a boulder stood: for sixty years nobody had found a code with an explicit construction that could be proven to reach capacity. From his doctorate in the 1980s onward, Arıkan fixed on a single question, why sequential decoding fell short of capacity, and struck the same boulder for roughly twenty-five years. An outside observer could long have assumed nothing was happening. In 2008 the boulder toppled with the idea of channel polarization, and polar codes emerged behind it: the door that had stayed shut since Shannon was opened in a single stroke. So Arıkan was not a pebble thrower; he was someone who leaned on a boulder for years and, when it fell, was revealed to have struck the right spot.</p><p>So what happens once the tunnel is open? Here the story divides in two. Some tunnels are useful the moment they open; humanity passes through them at once, they turn into applications, they enter our lives. The paper “Attention Is All You Need,” published in 2017 by eight researchers from Google, was like this; within a few years the Transformer architecture became the tunnel carrying the entire current wave of artificial intelligence. Moreover, this tunnel belongs to those who found funding and dug on a wide scale: the idea was the product not of one person but of a structure with resources and a team. Other tunnels are open but of use to no one. Either there is no need yet, or the road to the other side is not yet complete. That tunnel waits there. It waits ten years, fifty years, sometimes a hundred. Boole’s algebra became the foundation of the computer roughly a century later. Einstein’s equations of relativity today let the phone in our pocket find its position. Those who built these tunnels did not know when they would be used; they simply wanted to pierce the mountain.</p><p>This is why we cannot call untimely tunnels “wasted.” They stand there as humanity’s gain. An unused tunnel is not a closed tunnel; its turn simply has not yet come.</p><p>There is one more distinction worth dwelling on. Basic science crosses borders easily. A mathematical theorem, a law of physics, a chemical mechanism, whichever country it is discovered in, is read, taught, and built upon all over the world the next day. The tunnel itself is common property. Application, however, stays within borders. The first train through the tunnel is the train of the country that built it. Patents, factories, products, markets; these accumulate inside the border. The tunnel of polar codes was opened in Ankara, at Bilkent; but the first train through that tunnel was Huawei’s. When polar codes entered the 5G standard in 2016, it was not the country that dug the tunnel but the country whose train was ready that moved ahead. A country that does not produce basic science learns to pass through tunnels others have opened; a country that does not produce applications lets others’ trains pass through the tunnel it opened itself.</p><p>Here the difference between a developed state and a developing one also comes into view. The developed state chooses for itself where the tunnel will be opened. It decides which face of the mountain to strike and from what height, directs its funds in that direction, and gathers its diggers there. It is already calculating today which train will pass through which tunnel twenty years from now. The developing state, by contrast, usually chases popular topics. Whatever the world is talking about that year becomes the funding call; everyone runs to the same spot, the pebble carriers grow crowded, but someone else has set the tunnel’s direction. Such a country labors to pierce the mountain, yet it has surrendered from the outset its right to choose where the tunnel will come out. The problem is not failing to dig; it is failing to decide for itself where to dig.</p><p>Therefore a country’s science policy must answer two questions at once: How much are we contributing to piercing the mountain, and are we the first through the tunnel once it opens? Neglect the first and you remain a stranger to the world’s shared accumulation; neglect the second and your labor becomes someone else’s gain.</p><p>To return to the question of how a scientist works: the pebble thrower, the boulder toppler, the fund raiser, and the one who strikes the right spot are all inside the same tunnel. None can advance without the others. They keep digging without knowing who will reach the other side of the mountain, or when. What we call science is precisely this patience; it is the pursuit of things we do not even know we do not know, the unknown unknowns. It is consenting to dig without seeing the far side of the mountain.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=93c400fcbea3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>