<?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"><channel><title><![CDATA[The Tech Journal]]></title><description><![CDATA[The Tech Journal]]></description><link>https://thetechjournalcs.hashnode.dev</link><image><url>https://cdn.hashnode.com/uploads/logos/6af5ac90c2444806424f294b/4b705fa7-300f-48ad-af6c-3e3dafe8998d.png</url><title>The Tech Journal</title><link>https://thetechjournalcs.hashnode.dev</link></image><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 20:12:33 GMT</lastBuildDate><atom:link href="https://thetechjournalcs.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[What Large Language Models Actually Learn — And What They Don’t]]></title><description><![CDATA[When you ask Google a question, it doesn't write the answer. It just looks at your keywords, searches a massive index of the web, and hands you a list of…]]></description><link>https://thetechjournalcs.hashnode.dev/what-large-language-models-actually-learn-and-what-they-don-t</link><guid isPermaLink="true">https://thetechjournalcs.hashnode.dev/what-large-language-models-actually-learn-and-what-they-don-t</guid><category><![CDATA[Artificial Intelligence]]></category><category><![CDATA[Computer Science]]></category><dc:creator><![CDATA[Fatemeh Sadat Khalesi Nejad]]></dc:creator><pubDate>Thu, 12 Nov 2026 13:40:13 GMT</pubDate><content:encoded><![CDATA[<p><a href="https://youtu.be/5sLYAQS9sWQ?is=mdMkFmF-2z_umxfJ">https://youtu.be/5sLYAQS9sWQ?is=mdMkFmF-2z_umxfJ</a></p>
<h2><strong>Introduction</strong></h2>
<p>When you ask Google a question, it doesn't write the answer. It just looks at your keywords, searches a massive index of the web, and hands you a list of links. Think of it as a really fast librarian. Large Language Models (LLMs) like ChatGPT or Gemini are a completely different animal. They don’t have a database of facts they can just look up. Instead of retrieving information, they generate it from scratch. When you hit enter on a prompt, the system isn't fetching an answer. It's calculating one.</p>
<h2><strong>The Core Concept</strong></h2>
<p>How exactly do you calculate an answer? It all comes down to next-token prediction. An LLM essentially does one thing: it guesses the next word in a sequence. It’s a lot like the autocomplete feature on your phone, just scaled up to an absurd degree. If you text, "I'm going to the," your phone might suggest "store" because that's what you usually type.</p>
<p>An LLM does the same thing, but it bases its guesses on the billions of pages of text it processed during training. It predicts one chunk of a word (a token), sticks it onto the end of your sentence, and then runs the math all over again for the next one. It strings these words together so fast that it tricks us into feeling like we're having a real conversation.</p>
<p><strong>What They Actually Learn</strong></p>
<p>Because they sound so human, it’s easy to assume they understand what we're talking about. They don't. When an LLM digests the internet, it isn't learning facts or logic. It's learning statistical patterns.</p>
<p>It maps out how words connect in specific contexts. It knows that the word "Paris" often sits right next to "France" and "Eiffel Tower." It learns the syntax of a Python function and the polite, slightly boring tone of a corporate email. But it hasn't seen the Eiffel Tower, and it doesn't know what a Python script actually executes. It just knows the shape of the language we use to describe those things. It masters syntax, not meaning.</p>
<p><strong>What They Don't Learn</strong></p>
<p>Since they only care about syntax, LLMs have a huge blind spot. They don't know what is true and what is false.</p>
<p>Because they are just generating the next most likely word based on statistical patterns, they can confidently write sentences that make perfect grammatical sense but are completely wrong. We call this a "hallucination." An LLM might invent a historical event or write a chunk of code that looks flawless until you actually try to compile it.</p>
<p>It isn't lying on purpose. It just doesn't have a concept of reality. It plays the probability game to predict a word that fits the linguistic pattern, without stopping to check if that word matches the real world.</p>
<p><strong>My Perspective</strong></p>
<p>As a CS student, looking under the hood of these systems changes how you see them. When I started learning C# and grinding through CS50x, everything I wrote was strictly deterministic. You write a function, pass in variables, and get a guaranteed output based on strict logic. A <code>for</code> loop does exactly what you tell it to do, every single time.</p>
<p>Studying algorithms and data structures makes me look at LLMs differently. I don't see them as "thinking" artificial brains. I see them as incredibly complex probability engines. They are just giant matrices of weights and biases crunching floating-point numbers. It’s brilliant engineering and fascinating math, but it’s still just math. There is no ghost in the machine.</p>
<p><strong>Conclusion</strong></p>
<p>At the end of the day, LLMs are just tools. They are fantastic for brainstorming, formatting messy data, or getting past the blank page when coding. But they make terrible oracles. If you treat an AI like an all-knowing source of truth, you're going to get burned. Use them to write and code faster, but always verify the output yourself. They know how to talk. It's up to us to make sure what they say is actually right.</p>
<p><strong>Further Reading</strong></p>
<p>If you want to dig deeper into how these systems actually work without getting lost in the hype, these are great places to start:</p>
<ul>
<li><strong>Stanford Encyclopedia of Philosophy:</strong> The entry on "Artificial Intelligence" (<a href="https://plato.stanford.edu/entries/artificial-intelligence/">https://plato.stanford.edu/entries/artificial-intelligence/</a>) breaks down the actual logic and philosophy behind machine learning.</li>
<li><strong>MIT Technology Review:</strong> Search their AI topic index (<a href="http://technologyreview.com">technologyreview.com</a>) for articles specifically explaining neural networks and model architecture for non-researchers.</li>
<li><strong>Google Research Blog:</strong> (<a href="https://research.google/blog/">https://research.google/blog/</a>) A highly reliable place to read how modern model architectures are actually trained, straight from the engineers building them.</li>
</ul>
]]></content:encoded></item></channel></rss>