Suppose you want to check whether a piece of data belongs to a larger set. Let&rsquo;s say you&rsquo;re Google and you want to check if certain url has been marked as spam, the dumb approach would be to iterate over every url marked as spam to see if you can find it. Sure, maybe you think, &ldquo;I&rsquo;ll save the spammy sites in a hashmap&rdquo;, but then, a hashmap of the gazillion sites that exists in the internet? There must be a way that uses less space.&#xA;Even if you index those urls you still will have a not so good &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/linux/the-big-o-notation-for-algorithm-analysis/">Big O</a>&#xA; performance, maybe O(log n) or O.&#xA;This is an interesting data structure that&rsquo;s probabilistic, it won&rsquo;t have a 100% chance of returning what you wanna know, but the tradeoffs are interesting enough.&#xA;<h2 id="ok-but-whats-a-bloom-filter">Ok but what&rsquo;s a Bloom Filter?</h2>&#xA;A Bloom filter is a data structure that helps you check if an item might be in a set. It gives <del>blazingly</del> fast answers with very little memory. The trade-off is that sometimes it can say an item is present when it is not.&#xA;This data structure can produce false positives. But the good news are that there are not false negatives.&#xA;Because of this, Bloom filters shine when you care more about speed and space than absolute accuracy.&#xA;<div id="app-bloom-filter"></div>&#xA;<script type="module" src="https://res.cloudinary.com/dwrscezd2/raw/upload/v1757180703/index-CC06Czmb_ncmurd.js"></script>&#xA;&#xA;&#xA;<h3 id="where-can-i-use-bloom-filters">Where can I use bloom filters?</h3>&#xA;Bloom filters show up in those systems where quick lookups matter (No, your Tinder for pets probably doesn&rsquo;t apply):&#xA;<ul>&#xA;<li>Databases: To check if a key might be in a table before doing a costly disk read. Of course we&rsquo;re taking about millions of records, not tiny databases. (Cassandra, HBase, and Redis)</li>&#xA;<li>Web caches: To test if a page or object might be cached.</li>&#xA;<li>Distributed systems: To cut down on network calls when asking if a node has certain data.</li>&#xA;<li>Security: To quickly filter known bad URLs or email addresses, like in the first example I gave you.</li>&#xA;<li>Content recommendation systems: Where you want to avoid recommending content already consumed.</li>&#xA;</ul>&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1757107240/coffee-bytes/bloom-filter-usage_fvyq87.png" aria-label="Bloom filter usage by reddit user">&#xA;&#xA; <img class="md-image" loading="lazy" src="https://res.cloudinary.com/dwrscezd2/image/upload/v1757107240/coffee-bytes/bloom-filter-usage_fvyq87.png" alt="Bloom filter usage by reddit user"/>&#xA;&#xA;</a>&#xA;Bloom filters help reduce time and resource usage at scale, if you&rsquo;re ok with some false positives.&#xA;<h2 id="how-does-a-bloom-filter-work-internally">How Does a Bloom Filter Work Internally?</h2>&#xA;A Bloom filter uses:&#xA;<ul>&#xA;<li>A bit array (all values start at 0).</li>&#xA;<li>A set of hash functions.</li>&#xA;</ul>&#xA;When you add an item, the filter runs it through each hash function. Each function gives you an index in the array. You set the bits at those positions to 1.&#xA;<h3 id="there-are-no-deletions-in-classic-bloom-filters">There are no deletions in classic bloom filters</h3>&#xA;Since you are not storing real data but the &ldquo;pattern&rdquo; produced by the hash functions, you have no way to tell which combination of data and hash function produced it, hence you cannot &ldquo;delete&rdquo; the item.&#xA;<h3 id="you-cannot-retrieve-members-from-the-set">You cannot retrieve members from the set</h3>&#xA;Once you construct the bloom filter, you cannot know which items produced that patter from the filter itself, you can only know if something might belong or definitely doesn&rsquo;t belong to the set.&#xA;<h3 id="bloom-filter-collisions">Bloom filter collisions</h3>&#xA;However when you start adding more and more elements the chances of getting a collision (false positive) increase.&#xA;You can always add more hash functions to diminish the collitions but then the complexity and memory required increase.&#xA;<h3 id="checking-if-an-item-exists">Checking if an item exists</h3>&#xA;When you check an item, the filter does the same. If all the positions are set to 1, the item might be in the set (this is why it&rsquo;s a probabilistic data structure). If any position is 0, the item is definitely not in the set.&#xA;<h2 id="basic-bloom-filter-flow">Basic Bloom filter Flow</h2>&#xA;Here’s a simple outline that you can blatanly use to inspire you in the dark arts of bloom filters:&#xA;<ol>&#xA;<li>&#xA;Decide the size of the bit array and initialize all elements to 0.&#xA;</li>&#xA;<li>&#xA;Pick a few independent hash functions.&#xA;</li>&#xA;<li>&#xA;To add an item:&#xA;<ul>&#xA;<li>Run it through each hash function.</li>&#xA;<li>For each function, mark the bit at the returned index as 1</li>&#xA;</ul>&#xA;</li>&#xA;<li>&#xA;To check an item:&#xA;<ul>&#xA;<li>Run it through each hash function.</li>&#xA;<li>If all bits are 1 → item might be in the set.</li>&#xA;<li>If any bit is 0 → item is not in the set.</li>&#xA;</ul>&#xA;</li>&#xA;</ol>&#xA;<h3 id="bloom-filter-pseudocode">Bloom filter pseudocode</h3>&#xA;Using pseudocode (I know, I should have used Javascript instead), it would look like this.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python">initialize bit_array of size m with all 0s&#xA;choose k hash functions&#xA;&#xA;function add(item):&#xA; for i in 1 to k:&#xA; index = hash_i(item) mod m&#xA; bit_array[index] = 1&#xA;&#xA;function check(item):&#xA; for i in 1 to k:&#xA; index = hash_i(item) mod m&#xA; if bit_array[index] == 0:&#xA; return &#34;not present&#34;&#xA; return &#34;might be present&#34;&#xA;</code></pre></div><h3 id="bloom-filter-flow-chart">Bloom filter flow chart</h3>&#xA;And if you are a fan of flow charts it would look like this&#xA;<div class="mermaid">flowchart TD&#xA; A[Start] --> B[Hash item with k functions]&#xA; B --> C[Get indexes]&#xA; C --> D{Add or Check?}&#xA; D -->|Add| E[Set bits at indexes to 1]&#xA; D -->|Check| F[Are all bits = 1?]&#xA; F -->|Yes| G[Might be present]&#xA; F -->|No| H[Not present]&#xA;</div>&#xA; &#xA; &#xA; <h2 id="why-not-use-hashmaps-instead">Why not use hashmaps instead?</h2>&#xA;Well you could use some hashmaps approaches like buckets or &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/go/golang-maps-or-dictionaries/">Swiss tables</a>&#xA; but consider that you have to deal with collisions, also you can use many less intensive hash functions in a bloom filter whereas in hashmaps you need a strong hash function to avoid collisions which can consume some resources.&#xA;Also in a hashmap you store all the information as key-value pairs, whereas in a bloom filter you only store the results of the hash function, which can save a lot of space in memory.&#xA;<h2 id="when-you-shouldnt-use-them">When you shouldn&rsquo;t use them?</h2>&#xA;<ul>&#xA;<li>If you need to delete items, standard bloom filters only support inserting new data.</li>&#xA;<li>If data is small, there is no sense in using a probabilistic approach with false positives, use a hashmap instead.</li>&#xA;<li>As I told you before, the more data you add, the chances of getting false positives increases, so don&rsquo;t use it for data that grows a lot.</li>&#xA;</ul>&#xA;And that&rsquo;s all, I wrote this entry because I read about this data structure in the &#xA;&#xA;<a class="markdown-link" href="https://amzn.to/41rodp3" rel="noopener nofollow" target="_blank">System Design Interview Book</a>&#xA;, found interesting the fact that there is a probabilistic data structure out there, when must of the times you don&rsquo;t deal with probability when you deal with data structures, quite the contrary, you look for determinism.&#xA;<h2 id="some-bloom-filter-libraries">Some bloom filter libraries</h2>&#xA;Chances are there are already community-maintained libraries, so don&rsquo;t worry about reinventing the wheel.&#xA;<ul>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://github.com/bits-and-blooms/bloom" rel="noopener nofollow" target="_blank">bits and blooms (go)</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://github.com/KenanHanke/rbloom" rel="noopener nofollow" target="_blank">rbloom(python)</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://www.npmjs.com/package/bloom-filters" rel="noopener nofollow" target="_blank">Bloom filters(javascript)</a>&#xA;</li>&#xA;</ul>&#xA;

Suppose you want to check whether a piece of data belongs to a larger set. Let&rsquo;s say you&rsquo;re Google and you want to check if certain url has been marked as spam, the dumb approach would be to iterate over every url marked as spam to see if you can find it. Sure, maybe you think, &ldquo;I&rsquo;ll save the spammy sites in a hashmap&rdquo;, but then, a hashmap of the gazillion sites that exists in the internet? There must be a way that uses less space.

I Built a Bloom Filter Data Structure Simulator

<h2 id="why-would-we-want-to-create-a-mcp-server">Why Would We Want to Create a MCP Server?</h2>&#xA;Creating a MCP server allows us to connect an LLM or AI to real-time data, personal data, or other data sources.&#xA;I previously wrote a post where I explain &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ill-help-you-understand-the-model-context-protocol-or-mcp/">how the Model Context Protocol (MCP) works internally</a>&#xA;; you can check it out if you&rsquo;re not satisfied with just a recipe and want to learn more.&#xA;In this post, I’ll detail how to create a MCP server. We’re going to make one that solves &#xA;&#xA;<a class="markdown-link" href="https://community.openai.com/t/incorrect-count-of-r-characters-in-the-word-strawberry/829618" rel="noopener nofollow" target="_blank">one of the most embarrassing LLM errors: not being able to count the number of r&rsquo;s</a>&#xA; in variations of the word strawberry. For example: strawberrrry or strawberrrrrrry.&#xA;If you’re not a bot, you already know that counting letters is a pretty trivial task for a human.&#xA;On the other hand, for an LLM it&rsquo;s almost impossible due to how it works—based on tokens. &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-generated-art-and-ai-generated-code-are-treated-differently/">AI can create art and code</a>&#xA;, even though it &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/chat-gpt-searles-chinese-room-and-consciousness/">isn’t a consciou entity</a>&#xA;, but it can’t count letters. Contradictory, isn’t it?&#xA;For now, we&rsquo;ll turn this into an educational exercise on how to deploy a MCP server.&#xA;<h2 id="requirements-to-create-a-mcp-server-in-javascript">Requirements to Create a MCP Server in JavaScript</h2>&#xA;We’ll start with a Node installation—you know, the result of running: npm init -y, and we’ll create the index file using &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/linux/linux-basic-commands-grep-ls-cd-cat-cp-rm-scp/">the Linux touch command</a>&#xA;&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-bash" data-lang="bash">.&#xA;├── index.js&#xA;├── package.json&#xA;└── package-lock.json&#xA;</code></pre></div><h3 id="set-packagejson-as-a-module">Set package.json as a Module</h3>&#xA;Inside the package.json file, change or create the type property and assign it the value module, so npm treats it as a module.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">&#34;type&#34;: &#34;module&#34;&#xA;</code></pre></div><h3 id="install-the-library">Install the Library</h3>&#xA;A MCP server requires &#xA;&#xA;<a class="markdown-link" href="https://github.com/modelcontextprotocol/typescript-sdk" rel="noopener nofollow" target="_blank">the official MCP server SDK</a>&#xA;, just install it via npm.&#xA;We also need zod, which, <del>if you’re unlucky enough to use JavaScript</del>, is used to validate data.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-typescript" data-lang="typescript">npm install @modelcontextprotocol/sdk@1.16.0 zod@3.25.76&#xA;</code></pre></div><h2 id="create-a-mcp-server">Create a MCP Server</h2>&#xA;Afterward let&rsquo;s create a MCP server just initiliazing a new McPServer object&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">import {McpServer} from &#34;@modelcontextprotocol/sdk/server/mcp.js&#34;&#xA;&#xA;const server = new McpServer({&#xA; name: &#39;Strawberry Count&#39;,&#xA; version: &#39;1.0.0&#39;&#xA;})&#xA;</code></pre></div><h2 id="create-mcp-tools-in-a-mcp-server">Create MCP Tools in a MCP Server</h2>&#xA;Next, we&rsquo;ll define the tools, or mcp tools, that allow us to receive a parameter from the LLM so we can do whatever we want with it—in this case, count the r&rsquo;s. I already talked about mcp tools in my introduction to the Model Context Protocol.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">// server.tool(&#34;&lt;Name&gt;&#34;, &#34;&lt;Description&gt;&#34;, {&#xA;server.tool(&#34;Strawberry Count&#34;, &#34;Count the number of r&#39;s present in an strawberry variant&#34;, {&#xA; param: z.string().describe(&#34;The strawberry variant, it can have a variable number of r&#39;s&#34;)&#xA;},&#xA; // callback with the param&#xA; // AI will detect this param&#xA; async ({ param }) =&gt; {&#xA; return {&#xA; content: [&#xA; {&#xA; type: &#39;text&#39;,&#xA; text: `The word has: ${param.toLowerCase().split(&#34;&#34;).filter(letter =&gt; letter == &#34;r&#34;).length} r&#39;s`&#xA; }&#xA; ]&#xA; }&#xA; }&#xA;)&#xA;</code></pre></div>One thing to notice here, we’re counting the r’s manually—no tricks whatsoever. We process the string we receive and return the number of r’s.&#xA;Also, I want to highlight how awesome is the dark magic that the LLM uses to get our custom param. Our LLM detects the param that we want to receive from the user&rsquo;s input using as context the description information that we passed to the describe method.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753317103/are-you-a-wizard_hlrgoi.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753317103/are-you-a-wizard_hlrgoi.webp"&#xA; loading="lazy"&#xA; alt="MCP meme">&#xA;</figure>&#xA;</a>&#xA;Obviously, your mcp tool won&rsquo;t do something as useless as counting r’s—you could fetch data from an API, process it, get info from the file system, or even from another LLM—whatever you want.&#xA;<h2 id="where-does-the-llm-get-its-context-from">Where Does the LLM Get Its Context From?</h2>&#xA;For this example, we’ll use STDIO (Standard Input Output). But in the MCP introduction I talked you a little bit about its alternatives: SSE and HTTP Stream.&#xA;In this case, it’s basically the user’s text input.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">const transport = new StdioServerTransport()&#xA;</code></pre></div><h2 id="run-the-mcp-server-to-read-from-stdio">Run the MCP Server to Read from STDIO</h2>&#xA;Now we’ll connect this transport to the MCP server so it can read directly from STDIO.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">await server.connect(transport)&#xA;</code></pre></div><h3 id="connecting-the-mcp-to-other-applications-like-claude-desktop">Connecting the MCP to Other Applications, like Claude Desktop</h3>&#xA;We can connect this MCP server to any program that supports MCP so we can use it.&#xA;For example, for Claude Desktop we would fill out the claude_desktop_config.json file.&#xA;Notice how we declare that the MCP server should run with the <code>npx</code> command followed by the list of arguments in order.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">{&#xA; &#34;mcpservers&#34;: {&#xA; &#34;name&#34;: {&#xA; &#34;commando&#34;: &#34;npx&#34;,&#xA; &#34;args&#34;: [&#xA; &#34;-y&#34;,&#xA; &#34;node&#34;,&#xA; &#34;&lt;full_path_to_js_file&gt;&#34;&#xA; ]&#xA; }&#xA; }&#xA;}&#xA;</code></pre></div>If you wanted to use TypeScript instead of JavaScript, you could do something like this:&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">{&#xA; &#34;mcpservers&#34;: {&#xA; &#34;name&#34;: {&#xA; &#34;commando&#34;: &#34;npx&#34;,&#xA; &#34;args&#34;: [&#xA; &#34;-y&#34;,&#xA; &#34;tsx&#34;,&#xA; &#34;&lt;full_path_to_ts_file&gt;&#34;&#xA; ]&#xA; }&#xA; }&#xA;}&#xA;</code></pre></div><h2 id="analyzing-the-mcp-server-with-the-inspector">Analyzing the MCP Server with the Inspector</h2>&#xA;How we test that our MCP works? Claude’s inspector lets you test and debug MCP servers. You can view the tools you&rsquo;ve created and simulate their function—however, these don’t use an LLM. It’s a plain server, functioning as if we were the LLM.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-gui.jpg_oyzri4.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-gui.jpg_oyzri4.webp"&#xA; loading="lazy"&#xA; alt="MCP Inspector GUI">&#xA;</figure>&#xA;</a>&#xA;To run it, you can use npx and pass the server command:&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">npx -y @modelcontextprotocol/inspector &lt;mcp_server_command&gt;&#xA;</code></pre></div>I wasn’t able to run the command using substitution. I tried using the relative and absolute directory, but I couldn’t connect to the server—I’m not sure why.&#xA;However it did work by manually filling in the data in the Inspector’s graphical interface.&#xA;After running the inspector, we’ll have a server running on port 6274 (by default, though this can be changed).&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-javascript" data-lang="javascript">npx -y @modelcontextprotocol/inspector npm index.js&#xA;npx -y @modelcontextprotocol/inspector npm /home/dir/mcp-server/index.js&#xA;</code></pre></div>Next, fill in the transport type and commands in case you had the same issue I did.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-panel.jpg_k2td1x.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-panel.jpg_k2td1x.webp"&#xA; loading="lazy"&#xA; alt="MCP inspector configuration panel">&#xA;</figure>&#xA;</a>&#xA;If we click on “list tools,” it’ll detect the tool we just created.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-tools-panel.jpg_fhvq4v.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-inspector-tools-panel.jpg_fhvq4v.webp"&#xA; loading="lazy"&#xA; alt="MCP inspector tools panel">&#xA;</figure>&#xA;</a>&#xA;By clicking on List tools, and then on the name of the tool, will activate and we’ll be able to pass it our parameter: a strawberry variant with as many r’s as we want.&#xA;Just remember, this parameter is what we’d receive directly from our LLM—it would be the one responsible for parsing the user’s input and returning the parameter accordingly, so don&rsquo;t mess with the input.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-strawberry-count-result.jpg_pnon99.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753313069/coffee-bytes/mcp-strawberry-count-result.jpg_pnon99.webp"&#xA; loading="lazy"&#xA; alt="MCP inspector input box">&#xA;</figure>&#xA;</a>&#xA;As you can see, the code works. The MCP server returns our message. In a real-world setting, this would be received by the LLM and would tell the user the correct number of r’s in their strawberry variant—without making a fool of itself, as it usually would normally do.&#xA;

<h2 id="why-would-we-want-to-create-a-mcp-server">Why Would We Want to Create a MCP Server?</h2>&#xA;Creating a MCP server allows us to connect an LLM or AI to real-time data, personal data, or other data sources.&#xA;I previously wrote a post where I explain &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ill-help-you-understand-the-model-context-protocol-or-mcp/">how the Model Context Protocol (MCP) works internally</a>&#xA;; you can check it out if you&rsquo;re not satisfied with just a recipe and want to learn more.&#xA;In this post, I’ll detail how to create a MCP server. We’re going to make one that solves &#xA;&#xA;<a class="markdown-link" href="https://community.openai.com/t/incorrect-count-of-r-characters-in-the-word-strawberry/829618" rel="noopener nofollow" target="_blank">one of the most embarrassing LLM errors: not being able to count the number of r&rsquo;s</a>&#xA; in variations of the word strawberry. For example: strawberrrry or strawberrrrrrry.

How to create a MCP server and MCP tools from scratch?

Picture this: It&rsquo;s 2 AM, you&rsquo;re deploying a &ldquo;simple&rdquo; column type change to production, and suddenly your entire application is down because the table lock is taking forever. Your phone starts buzzing with angry Slack notifications, and you&rsquo;re frantically trying to explain to your team why the &ldquo;5-minute migration&rdquo; has been running for 30 minutes.&#xA;<h2 id="whats-the-shadow-table-strategy">What&rsquo;s the Shadow Table Strategy?</h2>&#xA;The shadow table strategy is like having a stunt double for your database table. Instead of modifying your original table directly (and potentially bringing your application to its knees), you create a <del>shadow clone</del> new table with the desired structure, gradually copy data over, and then perform a lightning-fast switcheroo.

Zero Downtime Migrations: Shadow Table Strategy Explained

All the tech influencers are spreading FOMO and talking about tools like Bolt, Vercel&rsquo;s V0, and Lovable. But I get it, &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">AI is overhyped right and we could be in a bubble</a>&#xA;.&#xA;Despite their non-realistic valuation, I&rsquo;m astonished about what these tools can do without needing &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/software-architecture/basic-characteristics-of-an-api-rest-api/">REST APIs</a>&#xA;, nor &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ill-help-you-understand-the-model-context-protocol-or-mcp/">Model Context Protocol</a>&#xA; integrations from the user.&#xA;<h2 id="bolt-vs-lovable-vs-v0-pricing">Bolt vs Lovable vs V0 pricing</h2>&#xA;Let&rsquo;s compare the pricing plans and features of each one, from V0, Lovable, and Bolt, focusing on aspects that may be relevant to those interested in creating <del>vulnerable</del> sites using AI.

My Bolt vs Lovable vs V0 Vercel Comparison

Today, while doom scrolling on Zuckerberg&rsquo;s social network, the algorithm recommended me an image of Sakura Card Captors in the style of the Spanish painter Remedios Varo.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1748206520/coffee-bytes/sakura-card-captors-remedios-varo_mwazan.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1748206520/coffee-bytes/sakura-card-captors-remedios-varo_mwazan.webp"&#xA; loading="lazy"&#xA; alt="Sakura Card Captors image in the style of Remedios Varo created with AI"><figcaption>&#xA; Sakura Card Captors image in the style of Remedios Varo created with AI&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;When I checked the comments—I don’t even know what I gain from doing that—I noticed that the most liked ones expressed strong disdain for artificial intelligence. A normal behavior if you&rsquo;re in the &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">middle of AI overhype</a>&#xA;&#xA;<h2 id="artists-hate-ai-devs-are-forced-to-accept-it">Artists Hate AI, Devs Are Forced to Accept It</h2>&#xA;There’s a strong consensus among artists regarding AI.&#xA;Movements like #NOAI or &#xA;&#xA;<a class="markdown-link" href="https://www.forbes.com/sites/lesliekatz/2024/07/17/human-intelligence-art-movement-takes-defiant-stand-against-ai/" rel="noopener nofollow" target="_blank">Made with human intelligence</a>&#xA; have made &#xA;&#xA;<a class="markdown-link" href="https://notbyai.fyi/es/" rel="noopener nofollow" target="_blank">their rejection of this technology very clear</a>&#xA;. The contempt seems to stem from how it was trained—using artists&rsquo; work without their permission, accused of plagiarizing something as personal as a style—and the fear of being replaced by this technology, which affects them economically.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753062961/coffee-bytes/no-ai-movement_g772l5.jpg">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753062961/coffee-bytes/no-ai-movement_g772l5.jpg"&#xA; loading="lazy"&#xA; alt="NO AI movement"><figcaption>&#xA; NO AI movement&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;On the other hand, there’s been a strong trend (whether organic or orchestrated, I don’t know) in the software world, to adopt AI as just another tool to increase the quality and quantity of code produced. Even when there’s resistance, it’s more focused on technical aspects and potential flaws rather than the ethical concerns of its training process.&#xA;This contrast fascinates me.&#xA;<h2 id="my-thoughts-on-why-ai-generated-code-and-ai-generated-art-are-treated-differently">My Thoughts on Why AI-Generated Code and AI-Generated Art Are Treated Differently</h2>&#xA;It’s not about one being right or wrong—it just really catches my attention.&#xA;Perhaps it’s because art is a direct form of visual and emotional communication. The fact that it’s generated by a machine might be perceived as an inferior form of expression or a threat to something considered purely human.&#xA;In contrast, code, by its functional and abstract nature, lacks that emotional weight, and the fact that a computer program generates it goes completely unnoticed.&#xA;Humans evolved using sight as a survival tool, so anything perceived visually will have a more significant impact. Code, as an abstraction of logic, is very recent and wasn’t part of our species&rsquo; evolutionary pressures or material reality. Plus, given its nature, it’s far less visual. Remember, code isn’t the final product—it’s an abstraction.&#xA;<h3 id="code-is-pragmatic-art-isnt">Code Is Pragmatic, Art Isn’t</h3>&#xA;Code was created to abstract, in a language closer to human speech, a series of instructions for a computer, with a well-defined and deterministic purpose.&#xA;But art tells a different story.&#xA;Oscar Wilde would say:&#xA;<blockquote>&#xA;&ldquo;All art is quite useless.&rdquo;</blockquote>&#xA;His statement doesn’t mean art is useless but that it exists for beauty and expression, not survival or moral instruction.&#xA;If there’s any truth to that, art is far more flexible than code and logical-mathematical systems.&#xA;I believe it’s precisely this flexibility that allows AI, combined with randomness, to function so differently for art and code.&#xA;<h2 id="the-role-of-randomness-in-generative-models-and-llms">The Role of Randomness in Generative Models and LLMs</h2>&#xA;AI can abstract patterns from its training dataset. Once it identifies those patterns or archetypes, it uses randomness to generate results that &ldquo;fit&rdquo; within them.&#xA;<h3 id="ai-generated-art">AI-Generated Art</h3>&#xA;An artist gathers stimuli from other artists, personal experiences, books, music… and &ldquo;mixes&rdquo; them to create something new. The result is something original that retains certain references or characteristics of the works that inspired it, no matter how subtle.&#xA;All of this happens consciously, whereas &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/chat-gpt-searles-chinese-room-and-consciousness/">AI does it unconsciously—or at least that’s what Searle would suggest</a>&#xA;.&#xA;<h3 id="ai-generated-code">AI-Generated Code</h3>&#xA;In the case of AI-generated code, the process is practically the same. The AI detects patterns in numerous code samples and then generates new code, different but based on the patterns the LLMs have &ldquo;learned.&rdquo;&#xA;In both cases, randomness allows for something new—not an exact copy of the originals it was trained on (as long as it doesn’t overfit).&#xA;<h2 id="randomness-impacts-the-differences-between-ai-generated-art-and-code">Randomness Impacts the Differences Between AI-Generated Art and Code</h2>&#xA;As I mentioned earlier, art is far more subjective and variable than coding. There are millions (or more) ways to present a concept or image, which aligns perfectly with how LLMs or generative models work.&#xA;When it comes to writing code, however, the number of ways to achieve the same thing is much more limited. Writing a loop that counts from 0 to 100 in the same programming language might have a dozen solutions.&#xA;This number starkly contrasts with the potential ways to paint a concept, which could be practically limitless.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1753063920/coffee-bytes/artificial-intelligence-art-randomness_pozdsb.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1753063920/coffee-bytes/artificial-intelligence-art-randomness_pozdsb.webp"&#xA; loading="lazy"&#xA; alt="An emotional photograph in the style of a Fujifilm disposable camera (like &#39;Utsurundesu&#39;) showing a medium shot of a Japanese woman in her 20s, crying at a wedding ceremony. She stands among guests, wearing a soft pastel-colored dress. Her makeup is slightly smudged by tears. The scene is imbued with deep nostalgia and melancholy. Soft, dim lighting casts a gentle glow on her dark hair. The focus is on her expressive face, and in the reflection of her eyes, a bride and groom can be seen standing together. The background shows a beautiful wedding venue with warm lighting. The image features pronounced film grain, slightly muted tones, and cinematic blur, amplifying the emotional intensity of the moment from a wider perspective. The composition evokes longing and fleeting memories, characteristic of heartfelt disposable camera photography. The background is softly blurred.">&#xA;</figure>&#xA;</a>&#xA;With code, it’s slightly different. Obviously, as the codebase grows, the number of possible solutions increases too. But here’s a crucial difference: unlike art, not all solutions are equally valid.&#xA;<h2 id="the-differences-between-ai-generated-art-and-code">The Differences Between AI-Generated Art and Code</h2>&#xA;Let me expand on this. The subjective nature of art means all representations of the same concept are valid—a Picasso-style apple is just as valid as one painted by Velázquez or Carrington.&#xA;This isn’t the case with software. A variation of any complex program (a web crawler, a database, etc.) will behave differently in different scenarios, and there will be a superior or inferior version based on the code’s requirements.&#xA;These differences can be measured quantitatively: execution time, memory usage, scalability, &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/python/how-to-measure-requests-per-second-with-locust-in-python/">requests per second</a>&#xA; etc., leaving us with some solutions better than others.&#xA;Given this, AI will &ldquo;get it right&rdquo; less often and require more iterations when generating code. However, when creating art, the user will likely consider the result valid after just a few iterations.&#xA;This makes me think that, from my perspective, the non-deterministic nature of AI works better for creating art than for creating code.&#xA;Now, there’s another aspect I want to touch on regarding the differences between AI-generated code and AI-generated art.&#xA;<h2 id="the-contrast-in-perception-between-ai-generated-code-and-ai-generated-images">The Contrast in Perception Between AI-Generated Code and AI-Generated Images</h2>&#xA;The final product of AI-generated art is the same product the user perceives. In contrast, the code AI creates doesn’t interact with the end user.&#xA;There’s no public backlash or protests over AI-generated code because people don’t see it. Code, in itself, doesn’t evoke passion or anger—unlike &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/devin-ai-will-this-ai-replace-programmers/">when Devin AI promised to replace programmers</a>&#xA;. Maybe it’s because the visual stimulus of art is far more intense than a wall of text expressing relationships between abstract entities.&#xA;For proof, just look at the response to tools like &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/my-bolt-vs-lovable-vs-v0-vercel-comparison/">Bolt, Lovable, or V0</a>&#xA;, which were trained using developers&rsquo; code—also without their permission.&#xA;<h3 id="ai-generated-websites">AI-Generated Websites</h3>&#xA;But what about websites? I’d argue that even if we observe the crystallization of code into pixels on a webpage, the reaction is entirely different from what we’d experience with art—despite both being perceived visually.&#xA;An AI-generated website (and perhaps the code itself) is completely indistinguishable from one made by a human. Plus, a website isn’t usually perceived as the result of human expression (especially in today’s internet).&#xA;But with art, it’s much easier to tell—hands with too many fingers, logical inconsistencies, even that default, generic style that’s become so characteristic. Or the now-popular Ghibli style that’s turned into a commodity.&#xA;AI-generated art stands out whereas AI-generated code is invisible.&#xA;

Today, while doom scrolling on Zuckerberg&rsquo;s social network, the algorithm recommended me an image of Sakura Card Captors in the style of the Spanish painter Remedios Varo.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1748206520/coffee-bytes/sakura-card-captors-remedios-varo_mwazan.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1748206520/coffee-bytes/sakura-card-captors-remedios-varo_mwazan.webp"&#xA; loading="lazy"&#xA; alt="Sakura Card Captors image in the style of Remedios Varo created with AI"><figcaption>&#xA; Sakura Card Captors image in the style of Remedios Varo created with AI&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;When I checked the comments—I don’t even know what I gain from doing that—I noticed that the most liked ones expressed strong disdain for artificial intelligence. A normal behavior if you&rsquo;re in the &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">middle of AI overhype</a>&#xA;

AI-generated Art and AI-generated code are treated differently

<h2 id="the-best-books-to-learn-python">The Best books to learn Python</h2>&#xA;I used these curated resources to learn Python, most of them are books, since I read them, I can heartily recommend them to you. If this little guide can help you to not get lost in the sea of options out there I&rsquo;ll be thrilled to share my experience learning Python with you.&#xA;<style>&#xA; .affiliate-link-button{&#xA; background-color: #FED400;&#xA; color: #020100;&#xA; padding: 10px 24px;&#xA; border-radius: 4px;&#xA; &#xA; }&#xA; .hide-on-sm {&#xA; display: inline;&#xA; }&#xA; .hide-on-lg {&#xA; display: none;&#xA; }&#xA; @media (max-width: 720px) {&#xA; .hide-on-sm {&#xA; display: none;&#xA; }&#xA; .hide-on-lg {&#xA; display: inline;&#xA; }&#xA; }&#xA; &#xA; </style>&#xA;<h3 id="dive-into-python-3">Dive into Python 3</h3>&#xA;This book is my favorite of all, it starts from scratch and begins to teach you the syntax of the language by creating small simple programs that even a fool <del>,like me,</del> can understand. The programs are basic but the author doesn&rsquo;t stop there, throughout the book he delves into topics such as:&#xA;<ul>&#xA;<li>Explanation of how Python transforms bytes into strings</li>&#xA;<li>Efficiency of algorithms at a very basic level</li>&#xA;<li>Generators</li>&#xA;<li>Iterators</li>&#xA;<li>Closures</li>&#xA;<li>Regular expressions</li>&#xA;<li>XML handling</li>&#xA;<li>HTTP requests</li>&#xA;</ul>&#xA;In my humble opinion this book is my favorite source for learning Python, I also &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/python/dive-into-python-the-best-book-to-learn-python/">wrote a review of dive into Python</a>&#xA; in the past.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742360/coffee-bytes/dive-into-python-e-reader-cover_ohoah5.jpg">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742360/coffee-bytes/dive-into-python-e-reader-cover_ohoah5.jpg"&#xA; loading="lazy"&#xA; alt="Dive into Python book cover"><figcaption>&#xA; Dive into Python Book cover&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;&#xA;&#xA;&#xA;&#xA;&#xA;<a href="https://amzn.to/42HBn0l" class="hide-on-sm affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Dive into Python on Amazon&#xA;</a>&#xA;<a href="https://amzn.to/42HBn0l" class="hide-on-lg affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Check price on Amazon&#xA;</a>&#xA;&#xA;&#xA;<h3 id="beggining-python-from-novice-to-professional">Beggining Python from novice to Professional</h3>&#xA;Beggining Python from novice to Professional is an introduction to the language from scratch, basically starts with the installation of the Python binary and takes you step by step from syntax, through each of the most common uses of the language, like how to create native graphical interfaces, web servers from scratch using sockets and even some basic frameworks, like &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/django/why-should-you-use-django-framework/">Django web framework</a>&#xA;.&#xA;Highly recommended if you want to get an overview of Python capabilities.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742217/coffee-bytes/Beginning_Python_from_novice_to_pro_tlfayl.jpg">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742217/coffee-bytes/Beginning_Python_from_novice_to_pro_tlfayl.jpg"&#xA; loading="lazy"&#xA; alt="Beggining Python from Novice to Pro Book cover"><figcaption>&#xA; Beggining Python from Novice to Pro Book cover&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;&#xA;&#xA;&#xA;&#xA;&#xA;<a href="https://amzn.to/44tPuZZ" class="hide-on-sm affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Beginning Python from Novice to Pro on Amazon&#xA;</a>&#xA;<a href="https://amzn.to/44tPuZZ" class="hide-on-lg affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Check price on Amazon&#xA;</a>&#xA;&#xA;&#xA;<h3 id="cracking-codes-with-python">Cracking Codes with Python</h3>&#xA;A bit unusual book, it explains the basics of Python while teaching you the history of cryptography. This book also assumes that you know absolutely nothing about Python. The theme of the book is to teach you an encryption method each chapter.&#xA;You will learn how to encrypt and decrypt using Caesar cipher, Vigenere, single-use notebooks and, to wrap the last chapters, the author dwells into the basics of RSA cipher.&#xA;If you are looking for something short, sweet and enjoyable to master the most basic parts of the language this is your book.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742522/coffee-bytes/cracking_codes_with_python_mnovzk.jpg">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1750742522/coffee-bytes/cracking_codes_with_python_mnovzk.jpg"&#xA; loading="lazy"&#xA; alt="Cracking with Python book cover"><figcaption>&#xA; Cracking with Python book cover&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;&#xA;&#xA;&#xA;&#xA;&#xA;<a href="https://amzn.to/4ixeQcE" class="hide-on-sm affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Cracking Codes with Python on Amazon&#xA;</a>&#xA;<a href="https://amzn.to/4ixeQcE" class="hide-on-lg affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Check price on Amazon&#xA;</a>&#xA;&#xA;&#xA;<h3 id="python-pocket-reference">Python Pocket Reference</h3>&#xA;This little book of about 260 pages is the whole Python programming language in a nutshell, but it is not didactic and is more like a cheatsheet or summary of all the features and subtleties of this programming language.&#xA;I can recommend it, but only if you already have notions of the language, in which case it is an excellent way to reinforce concepts and learn something new.&#xA;The strength of this book is its portability, given its pocket-size you can read it at any leisure time or when you are on the go.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1745616632/coffee-bytes/python-pocket-reference_vmfikn.jpg">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1745616632/coffee-bytes/python-pocket-reference_vmfikn.jpg"&#xA; loading="lazy"&#xA; alt="Python Pocket Reference Book cover"><figcaption>&#xA; Python pocket reference book cover&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;&#xA;&#xA;&#xA;&#xA;&#xA;<a href="https://amzn.to/4jNfdRh" class="hide-on-sm affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Python Pocket Reference on Amazon&#xA;</a>&#xA;<a href="https://amzn.to/4jNfdRh" class="hide-on-lg affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Check price on Amazon&#xA;</a>&#xA;&#xA;&#xA;<h3 id="python-tricks">Python tricks</h3>&#xA;If you already have some notions of the language, this book will take your skills even further. Python tricks is a compilation of some little known, or overlooked, features of this programming language. What do you mean? Well, things like:&#xA;<ul>&#xA;<li>The use of zip</li>&#xA;<li>Context managers</li>&#xA;<li>Handling iterators</li>&#xA;<li>Magic methods or dunders</li>&#xA;<li>Closures</li>&#xA;<li>Stacks and Queues</li>&#xA;<li>Named tuples</li>&#xA;</ul>&#xA;And too many things to cover in a short post. But if you have a chance and a little more experience in the language, I highly recommend you to acquire it.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1745688775/coffee-bytes/python-tricks_zjtir8.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1745688775/coffee-bytes/python-tricks_zjtir8.webp"&#xA; loading="lazy"&#xA; alt="Python tricks book cover">&#xA;</figure>&#xA;</a>&#xA;&#xA;&#xA;&#xA;&#xA;&#xA;<a href="https://amzn.to/45QvNfz" class="hide-on-sm affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Python Tricks on Amazon&#xA;</a>&#xA;<a href="https://amzn.to/45QvNfz" class="hide-on-lg affiliate-link-button" target="_blank" rel="nofollow sponsored">&#xA; Check price on Amazon&#xA;</a>&#xA;&#xA;&#xA;

<h2 id="the-best-books-to-learn-python">The Best books to learn Python</h2>&#xA;I used these curated resources to learn Python, most of them are books, since I read them, I can heartily recommend them to you. If this little guide can help you to not get lost in the sea of options out there I&rsquo;ll be thrilled to share my experience learning Python with you.&#xA;<style>&#xA; .affiliate-link-button{&#xA; background-color: #FED400;&#xA; color: #020100;&#xA; padding: 10px 24px;&#xA; border-radius: 4px;&#xA; &#xA; }&#xA; .hide-on-sm {&#xA; display: inline;&#xA; }&#xA; .hide-on-lg {&#xA; display: none;&#xA; }&#xA; @media (max-width: 720px) {&#xA; .hide-on-sm {&#xA; display: none;&#xA; }&#xA; .hide-on-lg {&#xA; display: inline;&#xA; }&#xA; }&#xA; &#xA; </style>&#xA;<h3 id="dive-into-python-3">Dive into Python 3</h3>&#xA;This book is my favorite of all, it starts from scratch and begins to teach you the syntax of the language by creating small simple programs that even a fool <del>,like me,</del> can understand. The programs are basic but the author doesn&rsquo;t stop there, throughout the book he delves into topics such as:

My Curated Best Resources To Learn Python Programming Language

Previously I told you how &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/seo/my-technical-seo-mistakes-when-i-migrated-my-site-from-wordpress/">I made many mistakes in SEO</a>&#xA; when I migrated my website from Wordpress to Hugo, after that I started watching a lot of videos about SEO, especially Romuald&rsquo;s videos (He is one of the biggest SEO influencers in Spanish speaking countries), I&rsquo;ve also read &#xA;&#xA;<a class="markdown-link" href="https://amzn.to/42uNZrv" rel="noopener nofollow" target="_blank">The art of SEO</a>&#xA; and I tried to summarize everything I learned from him and the book in a short post, just for you, before &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">AI overhype</a>&#xA; explodes, taking SEO with it.

My Technical SEO Basics Checklist Only For Web Developers

If you&rsquo;re struggling to understand the Model Context Protocol, I feel you, I watched a lot of videos on it, but I just couldn&rsquo;t understand it, so I took my notes about it and turned it into a post so you can understand it effortless, also I wrote a post where I talk about &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/how-to-create-a-mcp-server-and-mcp-tools-from-scratch/">how to create a MCP server and MCP tools from scratch</a>&#xA;.&#xA;<h2 id="why-was-the-model-context-protocol-created">Why was the Model Context Protocol created?</h2>&#xA;One of the most convenient uses of LLMs is to use them to analyse or process our own data, for which there are several options, such as &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/fine-tuning-a-llm-small-practical-guide-with-resources/">LLM fine-tuning</a>&#xA; or RAG, or manually passing our data as context as part of the prompt.&#xA;Anthropic members noticed this and proposed to standardise this process by creating a protocol for providing context and the ability to interact with systems to LLMs. Where does that context come from? Well, pretty much anywhere, Github, a database, your file system, an API (&#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/software-architecture/fast-and-performant-apis-using-go-lang-grpc-and-protobuffers/">gPRC type</a>&#xA;, REST or others), basically any source that can return information.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1745694716/coffee-bytes/modex-context-protocol-basic-summary_jy2nct.png">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1745694716/coffee-bytes/modex-context-protocol-basic-summary_jy2nct.png"&#xA; loading="lazy"&#xA; alt="Example of MCP usage, user asks for the content of its own files, LLM reads them and replies">&#xA;</figure>&#xA;</a>&#xA;This protocol, called the Model Context Protocol, is being sold as the next AI revolution, will it be &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">just AI overhype and bubble</a>&#xA; or a completely groundbreaking paradigm?&#xA;MCP exists to act as a bridge between an LLM and contextual information, in a standardised and agnostic way. The key word here is standardisation.&#xA;<h2 id="how-does-the-model-context-protocol-or-mcp-work">How does the Model Context Protocol or MCP work?</h2>&#xA;The MCP conforms to a client-server architecture.&#xA;<h3 id="client-role-in-mcp">Client role in MCP</h3>&#xA;A client that implements the Model Context Protocol can connect to an LLM and a set of MCP services or servers that provide it with the appropriate context it needs.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1743559566/coffee-bytes/screenshot-claude-ai-ui_xdtpmk.png">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1743559566/coffee-bytes/screenshot-claude-ai-ui_xdtpmk.png"&#xA; loading="lazy"&#xA; alt="A potential MCP Client"><figcaption>&#xA; A potential MCP Client&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;A MCP client represents the user interface to which we would normally pass our prompts; this interface may be Claude Desktop or some other equivalent.&#xA;<div class="mermaid">architecture-beta&#xA; group api(logos:anthropic-icon)[MCP]&#xA;&#xA; service llm(logos:serverless)[LLM] in api&#xA; service whatsapp(logos:whatsapp-icon)[MCP server Whatsapp] in api&#xA; service mcp(logos:anthropic-icon)[MCP Client] in api&#xA; service telegram(logos:telegram)[MCP server Telegram] in api&#xA; service github(logos:github-icon)[MCP server Github] in api&#xA; junction junctionCenter&#xA;&#xA; llm:L <--> R:mcp&#xA; mcp:L <-- R:junctionCenter&#xA; whatsapp:B <-- T:junctionCenter&#xA; telegram:T <-- B:junctionCenter&#xA; github:R <-- L:junctionCenter&#xA;</div>&#xA; &#xA; &#xA; <h3 id="server-role-in-mcp">Server role in MCP</h3>&#xA;A server is responsible for reading information, whether from a file, a database, a web service (via an API), or any other accessible source. To do this, it will use certain predefined resources or tools that restrict what the client can or cannot read or do.&#xA;<div class="mermaid">architecture-beta&#xA; group api(logos:anthropic-icon)[MCP]&#xA;&#xA; service server(server)[Whatsapp Server] in api&#xA; service whatsapp(logos:whatsapp-icon)[MCP Whatsapp] in api&#xA;&#xA;&#xA; server:L <--> R:whatsapp&#xA;</div>&#xA; &#xA; &#xA; <h4 id="how-does-the-mcp-know-what-it-can-read-or-do-from-a-server">How does the MCP know what it can read or do from a server?</h4>&#xA;To know what resources a service or server can access or modify, they implement a similar functionality &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/go/go-with-reflect-discover-how-reflect-can-boost-your-programs-flexibility/">Reflect in Go and GraphQL</a>&#xA;, with which they expose information about themselves.&#xA;For example, in the &#xA;&#xA;<a class="markdown-link" href="https://github.com/modelcontextprotocol/servers/tree/main/src/github" rel="noopener" target="_blank">CCM Github repository</a>&#xA; we see the list of actions that can be performed, as well as the required inputs for each of them.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1743551599/coffee-bytes/github-mcp-screenshot-list_ad1rel.png">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1743551599/coffee-bytes/github-mcp-screenshot-list_ad1rel.png"&#xA; loading="lazy"&#xA; alt="Screenshot of the available actions of a Github MCP server&#34;"><figcaption>&#xA; Screenshot of the available actions of a Github MCP server&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;Internally there is no magic at all, the services interact with the servers or information sources through their respective APIs. Take a look at the code that executes the listCommits action shown in the image above.&#xA;Notice how there is an initial validation using the familiar Zod library, but at the end of the snippet there is a function call that makes a request to a Github URL.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;display:grid;"><code class="language-javascript" data-lang="javascript"> 1import { z } from &#34;zod&#34;;&#xA; 2import { githubRequest, buildUrl } from &#34;../common/utils.js&#34;;&#xA; 3&#xA; 4export const ListCommitsSchema = z.object({&#xA; 5 owner: z.string(),&#xA; 6 repo: z.string(),&#xA; 7 sha: z.string().optional(),&#xA; 8 page: z.number().optional(),&#xA; 9 perPage: z.number().optional()&#xA;10});&#xA;11&#xA;12export async function listCommits(&#xA;13 owner: string,&#xA;14 repo: string,&#xA;15 page?: number,&#xA;16 perPage?: number,&#xA;17 sha?: string&#xA;18) {&#xA;19 return githubRequest(&#xA;20 buildUrl(`https://api.github.com/repos/${owner}/${repo}/commits`, {&#xA;21 page: page?.toString(),&#xA;22 per_page: perPage?.toString(),&#xA;23 sha&#xA;24 })&#xA;25 );&#xA;26}&#xA;</code></pre></div><h4 id="how-does-the-llm-knows-what-action-should-peform">How does the LLM knows what action should peform?</h4>&#xA;This is the magic part, the LLM can &ldquo;read&rdquo; the available actions from the MCP and, based on the prompt, decide which action should be executed, then the MCP will perform the action an return the answer to the LLM, afterwards the answer will be read by the LLM, which will generate a final response for the user.&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1751225840/coffee-bytes/MCP-flow-diagram_jjziao_sfst5j.webp">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1751225840/coffee-bytes/MCP-flow-diagram_jjziao_sfst5j.webp"&#xA; loading="lazy"&#xA; alt="MCP Flow Diagram">&#xA;</figure>&#xA;</a>&#xA;So far I have been talking about resources and actions, but in the MCP these two actions which are in charge of giving context to an LLM have names. Of course they were not going to miss the opportunity to give some branding to their protocol.&#xA;&#xA;<h2 id="context-types-resources-and-tools-in-the-model-context-protocol">Context types: Resources and Tools in the Model Context Protocol</h2>&#xA;The MCP context types are divided into:&#xA;<ul>&#xA;<li>Resources.</li>&#xA;<li>Tools.</li>&#xA;<li>Sampling, performs queries to other models, less common.</li>&#xA;<li>Prompts, prompts templates, less common.</li>&#xA;</ul>&#xA;<h3 id="what-is-the-resources-context-type">What is the Resources context type?</h3>&#xA;You can think of Resources as &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/software-architecture/basic-characteristics-of-an-api-rest-api/">a GET request to a REST API</a>&#xA;, intended to get information, without modifying anything.&#xA;<h3 id="what-is-the-type-of-context-tools">What is the type of context Tools?</h3>&#xA;On the other hand, a tool would be the equivalent of a POST, UPDATE, DELETE or PATCH request.&#xA;Think of tools as a mechanism to allow a client to make modifications to your servers or services.&#xA;If you are more familiar with &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/django/how-to-create-a-graphql-api-in-django-rapidly-using-graphene/">GraphQL-like APIs</a>&#xA;, you can think of resources and tools as queries and mutations, respectively.&#xA;<h2 id="how-does-the-mcp-communicate-with-servers">How does the MCP communicate with servers?</h2>&#xA;The MCP has two communication paradigms, one is via STDIO or Standard Input Output, <del>right out of your C programming nightmares #include &lt;stdio.h&gt;</del> ideal for local communication, for example with a local database such as Postgres or SQLite.&#xA;The other type is SSE or Server Sent Events, which does POST request streaming (similar &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/software-architecture/fast-and-performant-apis-using-go-lang-grpc-and-protobuffers/">to the gRPC streaming requests I&rsquo;ve already told you about</a>&#xA;), ideal of course for communications that are not in the same environment.&#xA;<h2 id="the-mcp-protocol-is-stateful">The MCP protocol is stateful</h2>&#xA;Another aspect worth noting is that, to date, &#xA;&#xA;<a class="markdown-link" href="https://github.com/modelcontextprotocol/specification/discussions/102" rel="noopener nofollow" target="_blank">the protocol requires a persistent connection between client and server</a>&#xA;, so this can be a drawback for auto-scalable architectures that adapt to demand, as well as moving in the opposite direction in a world that is striving to become stateless.&#xA;Just as a reminder:&#xA;<table>&#xA; <thead>&#xA; <tr>&#xA; <th>Aspect</th>&#xA; <th>Stateless Application</th>&#xA; <th>Stateful Application</th>&#xA; </tr>&#xA; </thead>&#xA; <tbody>&#xA; <tr>&#xA; <td>Server State</td>&#xA; <td>No stored client state between requests</td>&#xA; <td>Maintains client state across requests</td>&#xA; </tr>&#xA; <tr>&#xA; <td>Session Data</td>&#xA; <td>Client-side storage (e.g., cookies, tokens)</td>&#xA; <td>Server-side storage (e.g., session DB, memory)</td>&#xA; </tr>&#xA; <tr>&#xA; <td>Scalability</td>&#xA; <td>Horizontally scalable (no affinity needed)</td>&#xA; <td>Requires sticky sessions or state replication</td>&#xA; </tr>&#xA; <tr>&#xA; <td>Complexity</td>&#xA; <td>Simpler to implement and scale</td>&#xA; <td>More complex due to state management</td>&#xA; </tr>&#xA; <tr>&#xA; <td>Fault Tolerance</td>&#xA; <td>Resilient (failed requests can go to any server)</td>&#xA; <td>Vulnerable to server failures</td>&#xA; </tr>&#xA; <tr>&#xA; <td>Examples</td>&#xA; <td>&#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/software-architecture/rest-api-best-practices-and-design/">REST APIs (when properly designed), HTTP/HTTPS</a>&#xA;</td>&#xA; <td>Traditional monoliths, WebSocket apps</td>&#xA; </tr>&#xA; </tbody>&#xA;</table>&#xA;At the moment it is a matter of debate whether it will stay that way, there is no certainty whether stateless versions will be implemented or if there will be changes to it. After all, this protocol is new and who knows what the future holds for the AI world, giants can rise and fall in months, &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/devin-ai-will-this-ai-replace-programmers/">as Devin AI</a>&#xA; and Rabbit R1 did.&#xA;If you want to know more read the &#xA;&#xA;<a class="markdown-link" href="https://github.com/modelcontextprotocol" rel="noopener" target="_blank">official Model Context Protocol documentation</a>&#xA; and the introduction to the &#xA;&#xA;<a class="markdown-link" href="https://www.anthropic.com/news/model-context-protocol" rel="noopener" target="_blank">MCP</a>&#xA;&#xA;

If you&rsquo;re struggling to understand the Model Context Protocol, I feel you, I watched a lot of videos on it, but I just couldn&rsquo;t understand it, so I took my notes about it and turned it into a post so you can understand it effortless, also I wrote a post where I talk about &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/how-to-create-a-mcp-server-and-mcp-tools-from-scratch/">how to create a MCP server and MCP tools from scratch</a>&#xA;.&#xA;<h2 id="why-was-the-model-context-protocol-created">Why was the Model Context Protocol created?</h2>&#xA;One of the most convenient uses of LLMs is to use them to analyse or process our own data, for which there are several options, such as &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/fine-tuning-a-llm-small-practical-guide-with-resources/">LLM fine-tuning</a>&#xA; or RAG, or manually passing our data as context as part of the prompt.

I'll Help You Understand the Model Context Protocol or MCP

How to use Localhost using https instead of http, fast and without having to issue certificates manually nor dealing with netstat?&#xA;<h2 id="why-to-use-https-on-localhost-instead-of-http">Why to use https on localhost instead of http?</h2>&#xA;It is quite common to perform Oauth testing while performing integrations or some kind of integration with a third party application, the problem is that some third party providers are quite restrictive in this regard, they do not accept integrations or callbacks to URLs that do not use https.&#xA;This is the case for apps on X (formerly Twitter) or Facebook developer platforms. <del>Zuckerberg!</del>&#xA;<a hx-boost="false" href="https://res.cloudinary.com/dwrscezd2/image/upload/v1739482326/facebook-developer-app-https-callback_hmhesu.png">&#xA;<figure class="md-local-image"><img src="https://res.cloudinary.com/dwrscezd2/image/upload/v1739482326/facebook-developer-app-https-callback_hmhesu.png"&#xA; loading="lazy"&#xA; alt="Facebook developer portal callback needs https"><figcaption>&#xA; Facebook developer portal callback needs https&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;There are several ways of dealing with this, one of them is to sign our own certificates manually, which can be cumbersome in my opinion, but let&rsquo;s skip that and do it the quick and easy way.&#xA;<h2 id="running-https-on-localhost-with-caddy">Running https on localhost with Caddy</h2>&#xA;&#xA;&#xA;<a class="markdown-link" href="https://github.com/caddyserver/caddy" rel="noopener nofollow" target="_blank">Caddy</a>&#xA; is a server written in &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/go/go-programming-language-introduction-to-variables-and-data-types/">Go programming language</a>&#xA;, known to be easy peasy to configure (Unlike &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/linux/nginx-keepalive-gzip-http2-better-performance-on-your-website/">configuring Nginx</a>&#xA;), and it also includes https by default.&#xA;It is so simple to set up that it only requires a simple file called Caddyfile, with no extension. You can think of the latter as the equivalent of a &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/docker/how-to-write-a-docker-file-from-scratch/">Dockerfile in Docker</a>&#xA;.&#xA;Install it following the instructions according to your OS and create the Caddyfile.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-bash" data-lang="bash">your-sub-domain.localhost {&#xA; reverse_proxy http://localhost:5000&#xA;}&#xA;</code></pre></div>&#xA;The sub-domain part is just to be able to isolate your app from the rest of the applications serving on localhost.&#xA;The above configuration will create a reverse proxy that will redirect traffic on your-sub-domain.localhost to port 5000 of your localhost. Remember to replace the port with the port of your choice.&#xA;<div class="mermaid">graph TD;&#xA; your-sub-domain.localhost-->http:localhost:5000;&#xA;</div>&#xA; &#xA; &#xA; Finally just run caddy run, or caddy start if you want a process running in terminal or detached run, respectively, in the directory where the Caddyfile is located and caddy will create a reverse proxy to your localhost on the port you specified and we&rsquo;ll be running https on localhost as intended.&#xA;<h3 id="solution-1-to-error-caddy-listen-tcp-port_number-bind-permission-denied">Solution 1 to error: Caddy “listen tcp :&lt;port_number&gt;: bind: permission denied”.</h3>&#xA;If you get a permission denied error while trying to run caddy, it is because Linux prevents non-root processes from listening on major ports such as 443 or 80.&#xA;To allow caddy to listen on those ports and allow &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/linux/understand-permissions-in-gnu-linux-and-the-chmod-command/">linux to assign the required permissions</a>&#xA;, run the following command.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-bash" data-lang="bash">sudo setcap CAP_NET_BIND_SERVICE=+eip $(which caddy)&#xA;</code></pre></div><h3 id="solution-2-to-error-caddy-listen-tcp-port_number-bind-permission-denied">Solution 2 to error: Caddy “listen tcp :&lt;port_number&gt;: bind: permission denied”.</h3>&#xA;Another posibility is that caddy was already running, if you install it could run as a service, to which you have to stop it and execute it again.&#xA;<div class="highlight"><pre tabindex="0" style="color:#e2e4e5;background-color:#282a36;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-bash" data-lang="bash">caddy stop&#xA;caddy start&#xA;</code></pre></div>Afterwards you should be able to access your-sub-domain.localhost or your-sub-domain.localhost:443 via https in your browser, &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/linux/linux-basic-commands-lsof-top-ps-kill-systemctl-chown-chroot/">the curl command</a>&#xA; or whatever tool you prefer to use.&#xA;

How to use Localhost using https instead of http, fast and without having to issue certificates manually nor dealing with netstat?&#xA;<h2 id="why-to-use-https-on-localhost-instead-of-http">Why to use https on localhost instead of http?</h2>&#xA;It is quite common to perform Oauth testing while performing integrations or some kind of integration with a third party application, the problem is that some third party providers are quite restrictive in this regard, they do not accept integrations or callbacks to URLs that do not use https.

The Easiest Way To Use Https In Localhost (my opinion)

I tried to fine-tune a DeepSeek model but I failed, then I tried to do it with Llama and I couldn&rsquo;t make it work, finally I succeeded with a Mistral model, specifically the 7B parameter model.&#xA;I felt completely lost, lots of tutorials but none seemed to work, hence I decided to share with my readers my notes and some resources that I found useful during the process.&#xA;Still, don&rsquo;t get confused by this, I still believe that &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">AI is overhyped, but that will change soon</a>&#xA;.&#xA;<h2 id="where-to-learn-the-basics-of-ai-and-neural-networks">Where to learn the basics of AI and Neural networks?</h2>&#xA;If you have no idea about artificial intelligence or what a neural network is, the most complete and didactic videos belong to 3blue1brown.&#xA;However, it is important to clarify that these videos do not start from scratch, it is necessary that you have a background in linear algebra, linear regression and differential and integral calculus. I know it can be a little discouraging for those who come from non-technical backgrounds but it is not as difficult as it seems.&#xA;<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">&#xA; <iframe allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen" loading="eager" referrerpolicy="strict-origin-when-cross-origin" src="https://www.youtube.com/embed/aircAruvnKk?autoplay=0&amp;controls=1&amp;end=0&amp;loop=0&amp;mute=0&amp;start=0" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" title="YouTube video"></iframe>&#xA; </div>&#xA;&#xA;If you only speak english you probably don&rsquo;t know DotCSV, but what you may not know is that DotCSV has an introduction to neural networks and their videos are excellent, the bad part of course is that his videos are in Spanish. Time to turn on the subtitles.&#xA;<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">&#xA; <iframe allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen" loading="eager" referrerpolicy="strict-origin-when-cross-origin" src="https://www.youtube.com/embed/MRIv2IwFTPg?autoplay=0&amp;controls=1&amp;end=0&amp;loop=0&amp;mute=0&amp;start=0" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" title="YouTube video"></iframe>&#xA; </div>&#xA;&#xA;But if you already know everything about neural networks, then, let&rsquo;s start with the fine-tuning concept.&#xA;<h2 id="what-is-fine-tuning">What is fine-tuning?</h2>&#xA;Fine-tuning is the procedure of selecting an already trained model and continuing its training on a particular data set from the field. Like for example a LLM that has been trained specifically in scientific articles about diabetes, which you can do through &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/understand-the-model-context-protocol-or-mcp-once-and-for-all/">the Model Context Protocol</a>&#xA;.&#xA;Because, why use prompt engineering when you can fine-tune a model and get the same results?&#xA;<a hx-boost="false" href="https://i.imgflip.com/9kjbo2.jpg">&#xA;<figure class="md-local-image"><img src="https://i.imgflip.com/9kjbo2.jpg"&#xA; loading="lazy"&#xA; alt="Drake hotline bling meme up: prompt engineering down: fine-tuning"><figcaption>&#xA; I&rsquo;m just kidding, always try prompt engineering first&#xA; </figcaption>&#xA;</figure>&#xA;</a>&#xA;Most current LLM models exhibit quite positive overall performance, but fail to solve particular problems focused on specific tasks. The fine-tuning procedure provides significant benefits, such as decreased computational costs and the opportunity to use advanced models without the need to build one from scratch. &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/devin-ai-will-this-ai-replace-programmers/">It may not replace programmers</a>&#xA; but fine-tuning can turn a model into the ultimate prediction tool in an area of knowledge.&#xA;<h3 id="steps-needed-to-fine-tune-an-llm">Steps needed to fine-tune an LLM</h3>&#xA;The steps may vary according to the model but generally the process looks something like this:&#xA;<ol>&#xA;<li>Load a dataset: Gather your dataset, either existing or original.</li>&#xA;<li>Data preprocessing: Tokenize your dataset using model tokenizer and split it into training, validation and test sets.</li>&#xA;<li>Model selection: Choose a pre-trained LLM.</li>&#xA;<li>Parameter configuration: Configure the hyperparameters.</li>&#xA;<li>Training or Fine-Tuning: Train the model on the customized dataset.</li>&#xA;<li>Evaluation: Test it with the test set.</li>&#xA;<li>Inference: Deploy the fine-tuned model for inference on new data, making sure it generalizes well to unseen inputs.</li>&#xA;</ol>&#xA;Here are some resources about fine-tuning:&#xA;I borrowed this &#xA;&#xA;<a class="markdown-link" href="https://github.com/brevdev/notebooks/blob/main/mistral-finetune-own-data.ipynb" rel="noopener" target="_blank">notebook because it was the only one that worked and modified accodringly</a>&#xA;, you can thank brevdev for it.&#xA;I loved this resource for understanding the &#xA;&#xA;<a class="markdown-link" href="https://www.datacamp.com/tutorial/fine-tuning-large-language-models" rel="noopener" target="_blank">general steps to fine-tuning an LLM</a>&#xA;.&#xA;The &#xA;&#xA;<a class="markdown-link" href="https://www.kaggle.com/code/eduardomzepeda/fine-tuning-mistral-7b-with-linkedin-job-posting" rel="noopener nofollow" target="_blank">notebook that I used to fine-tune</a>&#xA; the model is here and you can check it.&#xA;<h2 id="where-to-get-datasets-for-fine-tuning-an-llm">Where to get datasets for fine-tuning an LLM?</h2>&#xA;In the same way you need data to train an AI model, fine-tuning requires that you provide it with the best data can you find, the best source is your own data.&#xA;I used datasets from &#xA;&#xA;<a class="markdown-link" href="https://huggingface.co/datasets/" rel="noopener nofollow" target="_blank">HuggingFace</a>&#xA;, however it is important to consider that pre-trained models that are freely available were probably trained using the same data, so if you decide to use them, don&rsquo;t expect to have a substantial improvement over normal model performance, it would be best to use your own data or original content.&#xA;Now we should have the data, what LLM should we use?&#xA;<h3 id="fine-tuning-data-sources">Fine-tuning data sources</h3>&#xA;<ul>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://github.com/awesomedata/awesome-public-datasets" rel="noopener nofollow" target="_blank">Awesome public datasets</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://kaggle.com" rel="noopener nofollow" target="_blank">Kaggle</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://opendatainception.io/" rel="noopener nofollow" target="_blank">opendatainception</a>&#xA;</li>&#xA;</ul>&#xA;Make sure your dataset is in a proper format. Eg, I had to convert one using Python to JSONL, because each line was a separate file, which made the import incredibly slow.&#xA;<a hx-boost="false" href="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/o49qurb343h802weo8cv.png">&#xA;<figure class="md-local-image"><img src="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/o49qurb343h802weo8cv.png"&#xA; loading="lazy"&#xA; alt="HuggingFace dataset screenshot">&#xA;</figure>&#xA;</a>&#xA;Now we should have the data, what LLM should we use now?&#xA;<h2 id="which-models-are-best-for-fine-tuning-an-llm">Which models are best for fine-tuning an LLM?</h2>&#xA;To do fine-tuning, a good starting point is using a pre-trained LLM, the most parameters the better results but its performance will be worse.&#xA;As a rule of thumb, it&rsquo;s a good idea to browse HuggingFace, and look for large models (the ones with more parameters), among the most popular ones I could find are Llama, DeepSeek, Mistral, Falcon, among others.&#xA;I tested the following models, not because they were the best, but because they were light and did not require as much computing power (I&rsquo;m working with limited resources, leave me alone), but you can use their larger versions with more parameters:&#xA;<ul>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://huggingface.co/meta-llama/Llama-3.1-8B" rel="noopener nofollow" target="_blank">Llama-3.1B</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" rel="noopener nofollow" target="_blank">DeepSeek-R1-Distill-Qwen-1.5B</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://huggingface.co/mistralai/Mistral-7B-v0.3" rel="noopener nofollow" target="_blank">Mistral-7B-v0.x</a>&#xA;</li>&#xA;</ul>&#xA;I got better results with the Mistral LLMs, probably because the DeepSeek ones were trained using Chinese sources.&#xA;<a hx-boost="false" href="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/xuywp28gqcln00ydopo2.jpg">&#xA;<figure class="md-local-image"><img src="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/xuywp28gqcln00ydopo2.jpg"&#xA; loading="lazy"&#xA; alt="Is this Searle&#39;s Chinese room meme">&#xA;</figure>&#xA;</a>&#xA;I was actually able to train and deploy it but the model returned Chinese characters, (&#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/chat-gpt-searles-chinese-room-and-consciousness/">just like Searle&rsquo;s room</a>&#xA;), from time to time as part of the response, although it was incredibly fast.&#xA;Now that we have the data and the model, where can we train our LLM? The cheaper the better.&#xA;<h2 id="where-to-fine-tune-an-llm-for-free">Where to fine-tune an LLM for free?</h2>&#xA;It would be best to have your own GPUs, but in case you can&rsquo;t, Google, through &#xA;&#xA;<a class="markdown-link" href="https://colab.research.google.com/" rel="noopener nofollow" target="_blank">Google Colab</a>&#xA; has GPUs available for free for a limited amount of hours per day, &#xA;&#xA;<a class="markdown-link" href="https://www.kaggle.com/" rel="noopener nofollow" target="_blank">Kaggle</a>&#xA; also offers you 30 hours of GPU usage per week to run your experiments. <del>And create your furry images using the forbidden models.</del>&#xA;<a hx-boost="false" href="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/4o5rmh1dv8il7pdml2gr.png">&#xA;<figure class="md-local-image"><img src="https://dev-to-uploads.s3.amazonaws.com/uploads/articles/4o5rmh1dv8il7pdml2gr.png"&#xA; loading="lazy"&#xA; alt="Kaggle free GPUs">&#xA;</figure>&#xA;</a>&#xA;In my fine-tuning adventure I also found a company called &#xA;&#xA;<a class="markdown-link" href="https://salad.com/" rel="noopener nofollow" target="_blank">salad</a>&#xA;, which serves as an intermediary between users and gamers who want to rent their GPUs, the prices are incredibly competitive. &#xA;&#xA;<a class="markdown-link" href="https://vast.ai/" rel="noopener nofollow" target="_blank">VastAI</a>&#xA; is also another option. &#xA;&#xA;<a class="markdown-link" href="https://m.do.co/c/a22240ebb8e7" rel="noopener nofollow" target="_blank">Digital Ocean</a>&#xA; also has very competitive prices and provices a wide variety of GPUs.&#xA;We have everything we need now, let&rsquo;s review the overall process.&#xA;<h2 id="llm-fine-tuning-tutorials">LLM Fine-tuning Tutorials</h2>&#xA;For Fine-tuning I found these videos quite enjoyable, short and straight to the point, however I could not produce a working model, it is possible that the notebooks were outdated and probably my inexperience make me fail.&#xA;This is a good introduction but does not share the jupyter notebook, although I was able to find it on Google Colab. I put the link in the next section.&#xA;<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">&#xA; <iframe allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen" loading="eager" referrerpolicy="strict-origin-when-cross-origin" src="https://www.youtube.com/embed/pxhkDaKzBaY?autoplay=0&amp;controls=1&amp;end=0&amp;loop=0&amp;mute=0&amp;start=0" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" title="YouTube video"></iframe>&#xA; </div>&#xA;&#xA;He is a bit more specific and concise.&#xA;<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">&#xA; <iframe allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen" loading="eager" referrerpolicy="strict-origin-when-cross-origin" src="https://www.youtube.com/embed/Q9zv369Ggfk?autoplay=0&amp;controls=1&amp;end=0&amp;loop=0&amp;mute=0&amp;start=0" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" title="YouTube video"></iframe>&#xA; </div>&#xA;&#xA;&#xA;Now we know the overall process, where can we find some template for fine-tuning?&#xA;<h2 id="notebooks-for-llm-fine-tuning">Notebooks for LLM fine-tuning.</h2>&#xA;This was &#xA;&#xA;<a class="markdown-link" href="https://github.com/brevdev/notebooks/blob/main/mistral-finetune-own-data.ipynb" rel="noopener nofollow" target="_blank">the notebook that allowed me to fine-tuning the Mistral model</a>&#xA;, with a couple of modifications, of course.&#xA;Most of the notebooks just need you to change the model and adapt the dataset to the model input, so you can use them as a starting point for your particular case.&#xA;Others that I found particularly well documented but did not work for me or produced unwanted results are these:&#xA;<ul>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://www.kaggle.com/code/kingabzpro/fine-tuning-deepseek-r1-reasoning-model" rel="noopener nofollow" target="_blank">Fine-tuning DeepSeek R1</a>&#xA;</li>&#xA;<li>&#xA;&#xA;<a class="markdown-link" href="https://colab.research.google.com/drive/135ced7oHytdxu3N2DNe1Z0kqjyYIkDXp" rel="noopener nofollow" target="_blank">Alpaca + Flame fine-tuning</a>&#xA;</li>&#xA;</ul>&#xA;<h2 id="where-to-deploy-an-llm-model">Where to deploy an LLM model?</h2>&#xA;After creating your model you may want to make it available to others.&#xA;Generally there are tools like Gradio that abstract the process of generating a chat server like OpenAI, you can set them up on your own or use the ones available from HuggingFace, although obviously for a monthly cost, the free plans only have CPU and RAM, so it is usually not enough to run your LLM, especially the big ones.&#xA;Since my project was for educational purposes and I don&rsquo;t want to pay for a premium plan, I decided to put the code necessary to run my model in a &#xA;&#xA;<a class="markdown-link" href="https://colab.research.google.com/drive/1Fe348rmXbDyvjoDPGEKrBtPurpfwnFgG" rel="noopener nofollow" target="_blank">GoogleColab notebook with the necessary code to run it.</a>&#xA;&#xA;

I tried to fine-tune a DeepSeek model but I failed, then I tried to do it with Llama and I couldn&rsquo;t make it work, finally I succeeded with a Mistral model, specifically the 7B parameter model.&#xA;I felt completely lost, lots of tutorials but none seemed to work, hence I decided to share with my readers my notes and some resources that I found useful during the process.&#xA;Still, don&rsquo;t get confused by this, I still believe that &#xA;&#xA;<a class="markdown-link" href="https://coffeebytes.dev/en/artificial-intelligence/ai-is-overhyped-when-will-the-bubble-burst/">AI is overhyped, but that will change soon</a>&#xA;.

Eduardo Zepeda's latests posts