<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Axelle Toni</title>
    <description>The latest articles on DEV Community by Axelle Toni (@anaxelle).</description>
    <link>https://dev.to/anaxelle</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4151930%2Fbe824f35-f159-4e54-accf-424b19f0b1af.png</url>
      <title>DEV Community: Axelle Toni</title>
      <link>https://dev.to/anaxelle</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/anaxelle"/>
    <language>en</language>
    <item>
      <title>Choosing an Open-Source LLM for Our Waste-Sorting Assistant</title>
      <dc:creator>Axelle Toni</dc:creator>
      <pubDate>Tue, 27 Oct 2026 07:04:37 +0000</pubDate>
      <link>https://dev.to/anaxelle/choosing-an-open-source-llm-for-our-waste-sorting-assistant-oni</link>
      <guid>https://dev.to/anaxelle/choosing-an-open-source-llm-for-our-waste-sorting-assistant-oni</guid>
      <description>&lt;ol&gt;
&lt;li&gt;About our project&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Our project is an assistant that tells people which bin an item belongs in, and why. Some items are easy, like a plastic bottle. Others are tricky, like a greasy pizza box or a broken light bulb. We want the assistant to answer in a clear format, for example “Category: Recyclable. Reason: …”.&lt;/p&gt;

&lt;p&gt;In my NLP class I built a next-word predictor using GPT, but I only used the basic version and never reached GPT-2. Researching this blog helped me see that next-word prediction is exactly what every modern LLM does, just at a much bigger scale. An LLM gives a probability to every possible next token and picks one, over and over, until it has a full answer.&lt;/p&gt;

&lt;p&gt;Our team has no budget and uses the free Google Colab T4 GPU, which has 16 GB of memory. So I searched for open models of about 4 billion parameters or fewer.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The four models I looked at&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Qwen3-4B-Instruct-2507 (Alibaba, Qwen team). It has 4 billion parameters and a very large context window of 262,144 tokens. The licence is Apache-2.0. It is built to follow instructions well, and it has no “thinking mode,” so its answers don’t start with long reasoning text.&lt;/p&gt;

&lt;p&gt;Llama-3.2-3B-Instruct (Meta). It has about 3.21 billion parameters and a 128K context window. It was released in September 2024 under the Llama 3.2 Community License. It is good at conversation and summarising. To download it, you need a Hugging Face account and must accept Meta’s terms.&lt;/p&gt;

&lt;p&gt;Phi-4-mini-instruct (Microsoft). It has 3.8 billion parameters and a 128K context window, under the MIT licence. It came out in February 2025 and is especially good at maths and logic for its size.&lt;/p&gt;

&lt;p&gt;SmolLM3-3B (Hugging Face). It has 3 billion parameters. It was trained with a 64K context and can reach 128K using a method called YaRN. The licence is Apache-2.0. It supports six languages and has a reasoning mode you can switch on or off. Its training recipe is fully public.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Comparison table&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fagj0c70wy15i5ncao62e.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fagj0c70wy15i5ncao62e.jpeg" alt=" " width="569" height="356"&gt;&lt;/a&gt;es/71wb73shg6qyxk5gvfbq.jpeg)&lt;/p&gt;

&lt;p&gt;The memory column is my estimate: parameters × 2 bytes. All four fit inside the T4’s 16 GB.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The model I choose: Qwen3-4B-Instruct-2507&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I chose Qwen3-4B-Instruct-2507 for four reasons.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It fits on the free Colab T4. I downloaded and loaded it on a free Colab T4. The model files are 8.04 GB, which matches my estimate of 4 billion × 2 bytes. I loaded it in float16, because the T4 does not support bfloat16. It loaded successfully, so it fits within the 16 GB limit. I did not measure the exact speed. &lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fw03qqli6deied52118uq.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fw03qqli6deied52118uq.png" alt=" " width="766" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;It is free and easy to start. I didn’t need an API key, a paid plan, or to accept any terms. It is Apache-2.0, so I could download and use it right away. Llama needs one extra step, an account and Meta’s agreement, which is more work for a team.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;It follows instructions and gives clean answers. Our assistant must answer in the same format every time. Qwen3-4B-Instruct-2507 was improved for instruction following, and because it has no thinking mode, its answer comes straight, without extra text. That makes it easy to read and easy to use in code.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;It can help our project in three ways.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Sorting accuracy: it sticks to our categories and our format.&lt;br&gt;
Explaining disposal rules: the large context window lets us paste our local recycling rules into the prompt, so it explains using the rules instead of guessing.&lt;br&gt;
Tricky items: we can show it a few examples of hard items and ask for a one-sentence reason.&lt;/p&gt;

&lt;p&gt;Why I did not choose the others. Phi-4-mini is best at maths and logic, which matters less for sorting waste. Llama needs the account and terms step. SmolLM3 is very open and good, but it is built around six languages and a shorter trained context, which helps us less.&lt;/p&gt;

&lt;p&gt;Limits. Qwen3-4B is the biggest of the four, so it is a little slower and uses the most memory. If it ever runs out of memory, our backup is Qwen3-1.7B. Like all LLMs, it can make things up (hallucinate), so for rules that matter, we should put the real rules in the prompt and not trust the model’s memory.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Conclusion&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I started this task knowing only next-word prediction from my class project. Now I understand how bigger models work and how to choose one with real facts: size, licence, memory, and how well it follows instructions. For our project, I choose Qwen3-4B-Instruct-2507.&lt;/p&gt;

&lt;p&gt;References&lt;br&gt;
Qwen team. Qwen3-4B-Instruct-2507 model card. &lt;a href="https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507" rel="noopener noreferrer"&gt;https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507&lt;/a&gt;&lt;br&gt;
Meta. Llama-3.2-3B-Instruct model card. &lt;a href="https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct" rel="noopener noreferrer"&gt;https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct&lt;/a&gt;&lt;br&gt;
Microsoft. Phi-4-mini-instruct model card. &lt;a href="https://huggingface.co/microsoft/Phi-4-mini-instruct" rel="noopener noreferrer"&gt;https://huggingface.co/microsoft/Phi-4-mini-instruct&lt;/a&gt;&lt;br&gt;
Hugging Face. SmolLM3-3B model card. &lt;a href="https://huggingface.co/HuggingFaceTB/SmolLM3-3B" rel="noopener noreferrer"&gt;https://huggingface.co/HuggingFaceTB/SmolLM3-3B&lt;/a&gt;&lt;br&gt;
Qwen Team. Qwen3 Technical Report. arXiv (confirm the number before publishing).&lt;br&gt;
Course lecture notes: LLM_01 Introduction to LLMs.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>machinelearning</category>
      <category>opensource</category>
    </item>
    <item>
      <title>How Computers See: What I Learned in My First Computer Vision Class</title>
      <dc:creator>Axelle Toni</dc:creator>
      <pubDate>Tue, 13 Oct 2026 14:05:33 +0000</pubDate>
      <link>https://dev.to/anaxelle/how-computers-see-what-i-learned-in-my-first-computer-vision-class-157</link>
      <guid>https://dev.to/anaxelle/how-computers-see-what-i-learned-in-my-first-computer-vision-class-157</guid>
      <description>&lt;p&gt;Before my first computer vision class, I thought a computer looked at a photo the way I do. I was wrong. A computer has never seen a photo, not once.&lt;/p&gt;

&lt;p&gt;This week we started computer vision, the part of AI that teaches machines to "see". Here is what I learned, in simple words, for anyone starting from zero like me.&lt;/p&gt;

&lt;h2&gt;
  
  
  A photo is just numbers
&lt;/h2&gt;

&lt;p&gt;When I look at a street photo, I instantly see a bus and some people. A computer gets a grid of numbers instead. Every tiny square (a pixel) has three numbers: how much &lt;strong&gt;red&lt;/strong&gt;, &lt;strong&gt;green&lt;/strong&gt; and &lt;strong&gt;blue&lt;/strong&gt; it has, each from 0 to 255. Why 255? Each value uses 8 bits, and 8 bits can hold 256 different values, from 0 to 255.&lt;/p&gt;

&lt;p&gt;In class we opened a sample photo with Python:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;numpy&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;PIL&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Image&lt;/span&gt;

&lt;span class="n"&gt;img&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;array&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;Image&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;open&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;bus.jpg&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;convert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;RGB&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;img&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;shape&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;   &lt;span class="c1"&gt;# (1080, 810, 3) for our sample photo
&lt;/span&gt;&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;img&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;    &lt;span class="c1"&gt;# 2624400 numbers in total
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That one photo is 1080 rows × 810 columns × 3 colours, which is &lt;strong&gt;2,624,400 numbers&lt;/strong&gt;. The computer never "sees a bus". It only sees those numbers. The big challenge of computer vision is getting from the numbers to the meaning.&lt;/p&gt;

&lt;h2&gt;
  
  
  What can a computer do with a photo?
&lt;/h2&gt;

&lt;p&gt;Almost every vision system answers one of three questions: &lt;em&gt;What is in the image? Where is it? What is happening?&lt;/em&gt; We ran three pretrained models on the same street photo, and each one answered differently:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Classification&lt;/strong&gt; gives one label for the whole image. It said "minibus" with 57.1% confidence and "police van" with 33.7%. It was unsure between two vehicles that look alike, and one label is not very useful for a whole street scene.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Detection&lt;/strong&gt; draws a box around each object. It found 1 bus (94% confident) and 4 people.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Segmentation&lt;/strong&gt; labels every pixel. It marked exactly which pixels belong to each object. In our photo, the objects covered 44% of the image.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;More detailed answers need more detailed training labels, and those cost more to create.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does a computer find an edge?
&lt;/h2&gt;

&lt;p&gt;This was my favourite part, because it is only multiplication and addition. A &lt;strong&gt;filter&lt;/strong&gt; is a small grid of weights that slides over the image. At each spot, you multiply each pixel by its weight and add everything up.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;sobel_x&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;array&lt;/span&gt;&lt;span class="p"&gt;([[&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
                    &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
                    &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]])&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This filter computes "right side minus left side". On a flat patch where every pixel is 100, you get −400 + 400 = &lt;strong&gt;0&lt;/strong&gt;, so there is no edge. On a patch that goes from dark (0) to bright (200), you get 0 + 800 = &lt;strong&gt;800&lt;/strong&gt;, a big number, so there is an edge.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;CNN&lt;/strong&gt; (convolutional neural network) does exactly this, with one difference: nobody chooses the weights. Training learns them from thousands of labelled images. The first layers learn edges and colours, the middle layers learn textures and parts like eyes or wheels, and the last layers recognise whole objects.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why YOLO is fast
&lt;/h2&gt;

&lt;p&gt;For detection we used &lt;strong&gt;YOLO&lt;/strong&gt;, which stands for "You Only Look Once". Older detectors first proposed possible regions and then classified each one, which was accurate but slow. YOLO looks at the whole image one time and predicts every box and label together. That is fast enough for live video and small devices. It was introduced in 2015 by Joseph Redmon and colleagues.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;ultralytics&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;YOLO&lt;/span&gt;

&lt;span class="n"&gt;detector&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;YOLO&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;yolo11n.pt&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;            &lt;span class="c1"&gt;# small pretrained detection model
&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;detector&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;street.jpg&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;conf&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;0.4&lt;/span&gt;&lt;span class="p"&gt;)[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;  &lt;span class="c1"&gt;# only keep boxes it is 40%+ sure about
&lt;/span&gt;&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;box&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;boxes&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;names&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;box&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;cls&lt;/span&gt;&lt;span class="p"&gt;)],&lt;/span&gt; &lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="nf"&gt;float&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;box&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;conf&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  When models fail
&lt;/h2&gt;

&lt;p&gt;This part surprised me most. We made the same photo harder, the way a cheap phone or bad lighting would:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Very dark:&lt;/strong&gt; the model invented a skateboard that was not there.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Upside down:&lt;/strong&gt; it lost the bus.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Low resolution:&lt;/strong&gt; it found nothing.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A model only knows what it was trained on. Many of these models were trained mostly on photos from Europe, North America and Asia. So where might they fail here? On crops, clothing, road scenes, skin tones and lighting. The fix is &lt;strong&gt;local data&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this matters for us
&lt;/h2&gt;

&lt;p&gt;Our course looked at real examples from Africa. Nuru helps farmers diagnose cassava disease from a leaf photo, and it works offline on a basic phone, so it works in fields with no internet. BLAIVA, built by FabLab Rwanda in Kigali, is a wearable camera that describes objects, text and faces through earphones. It is still in beta testing. These systems were built by people like us, which makes this field feel close and possible.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'm taking away
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Images are numbers: height × width × 3.&lt;/li&gt;
&lt;li&gt;Filters find patterns, and CNNs learn those filters on their own.&lt;/li&gt;
&lt;li&gt;Classification, detection and segmentation answer different questions.&lt;/li&gt;
&lt;li&gt;Models fail on what they have not seen, so local data matters.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Next, I'll build my own image classifier on photos from around me, with a separate test set so I can check it honestly.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;kLab TekHer AI Program, "Introduction to Computer Vision" class slides and demo notebook&lt;/li&gt;
&lt;li&gt;Ultralytics YOLO11 pretrained models; ImageNet and COCO datasets&lt;/li&gt;
&lt;li&gt;YOLO: Joseph Redmon et al., 2015&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>python</category>
    </item>
    <item>
      <title>Word Embeddings, Explained Simply (For Developers Who Are New to NLP)</title>
      <dc:creator>Axelle Toni</dc:creator>
      <pubDate>Wed, 30 Sep 2026 14:38:28 +0000</pubDate>
      <link>https://dev.to/anaxelle/word-embeddings-explained-simply-for-developers-who-are-new-to-nlp-4jei</link>
      <guid>https://dev.to/anaxelle/word-embeddings-explained-simply-for-developers-who-are-new-to-nlp-4jei</guid>
      <description>&lt;p&gt;When I first heard the term "word embeddings," it sounded like something only researchers with PhDs would understand. It turns out the core idea is actually simple, and once it clicks, a lot of modern AI (chatbots, search engines, recommendation systems) suddenly makes a lot more sense.&lt;br&gt;
What are word embeddings, really?&lt;/p&gt;

&lt;p&gt;What are word embeddings, really?&lt;/p&gt;

&lt;p&gt;Computers don't understand words the way we do. To a computer, "cat" and "dog" are just strings of letters with no relationship to each other unless we tell it otherwise. Word embeddings are a way of converting words into lists of numbers (called vectors) so that words with similar meanings end up with similar numbers.&lt;/p&gt;

&lt;p&gt;For example, imagine every word gets plotted as a point in space. Words like "cat" and "dog" would land close together, because they're both animals, often pets, and used in similar sentences. A word like "airplane" would land far away, because it's used in a completely different context.&lt;/p&gt;

&lt;p&gt;This matters because it lets computers do things like:&lt;/p&gt;

&lt;p&gt;Understand that "good" and "great" mean similar things&lt;br&gt;
Search for a term and still find related results, not just exact matches&lt;br&gt;
Power translation, chatbots, and recommendation systems&lt;/p&gt;

&lt;p&gt;Without embeddings, a computer would treat every word as completely unrelated to every other word which makes real language understanding almost impossible.&lt;/p&gt;

&lt;p&gt;How Word2Vec works  &lt;/p&gt;

&lt;p&gt;One of the most well-known ways to create embeddings is Word2Vec, developed by researchers at Google. The idea behind it is surprisingly simple: "a word is defined by the company it keeps."&lt;/p&gt;

&lt;p&gt;Word2Vec looks at huge amounts of text and learns, based on which words tend to appear near each other, how to place each word in that numerical space. If "coffee" and "tea" often show up in similar sentences ("I drink coffee every morning," "I drink tea every morning"), the model learns to place them close together — even though it was never explicitly told they're both drinks.&lt;/p&gt;

&lt;p&gt;There are two main approaches inside Word2Vec:&lt;/p&gt;

&lt;p&gt;CBOW (Continuous Bag of Words): predicts a word based on the words around it.&lt;br&gt;
Skip-gram: does the opposite predicts the surrounding words based on one word&lt;/p&gt;

&lt;p&gt;Either way, the model isn't taught definitions. It only learns patterns from context, and meaning emerges from that.&lt;/p&gt;

&lt;p&gt;A concrete example&lt;/p&gt;

&lt;p&gt;Here's a simple example using the gensim library in Python, training a tiny Word2Vec model on a small sample of sentences:&lt;/p&gt;

&lt;p&gt;from gensim.models import Word2Vec&lt;/p&gt;

&lt;p&gt;sentences = [&lt;br&gt;
    ["i", "drink", "coffee", "every", "morning"],&lt;br&gt;
    ["i", "drink", "tea", "every", "morning"],&lt;br&gt;
    ["dogs", "and", "cats", "are", "pets"],&lt;br&gt;
    ["cats", "chase", "mice"],&lt;br&gt;
]&lt;/p&gt;

&lt;p&gt;model = Word2Vec(sentences, vector_size=50, window=2, min_count=1, workers=1)&lt;/p&gt;

&lt;p&gt;print(model.wv.most_similar("coffee"))&lt;/p&gt;

&lt;p&gt;Even with this tiny amount of data, the model starts to notice that "coffee" and "tea" appear in similar contexts, and it will rank "tea" as more similar to "coffee" than something like "dogs."&lt;/p&gt;

&lt;p&gt;A classic (and honestly kind of magical) example from real, large-scale Word2Vec models is:&lt;/p&gt;

&lt;p&gt;King - man + woman = queen&lt;/p&gt;

&lt;p&gt;The model isn't told anything about royalty or gender it just learns this relationship purely from patterns in text.&lt;/p&gt;

&lt;p&gt;What surprised me and what I struggled with&lt;/p&gt;

&lt;p&gt;I haven't worked hands-on with word embeddings yet, but I've always loved anything related to IT, so this topic caught my attention right away. What I enjoyed most was understanding how these models are trained how they "learn" without actually thinking. A model doesn't understand meaning; it only picks up patterns from whatever data it's given, which means how well it performs depends entirely on what you train it with. That part stuck with me: as the person training a model, you have to be careful about the data you feed it, because it will simply repeat whatever patterns exist in that data. What didn't surprise me is that computers don't naturally understand human language, I already knew that from earlier learning.&lt;/p&gt;

&lt;p&gt;What I struggled with was the more practical side. When I looked at example code like the one above, I found it hard to fully understand the output, things like most_similar() results and what the numbers actually mean. I'm also still getting comfortable with writing this kind of code myself, rather than just reading someone else's. Understanding the theory made sense to me quickly, but going from theory to actually building and reading a working model is still something I'm learning to do.&lt;/p&gt;

&lt;p&gt;Final thoughts&lt;/p&gt;

&lt;p&gt;Word embeddings are one of those ideas that feel complicated from the outside but make a lot of sense once you see a small example. If you're starting out in NLP like I am, I would recommend just trying the code above with your own sentences seeing which words the model considers "similar" is a great way to build intuition.&lt;/p&gt;

&lt;p&gt;Sources:&lt;/p&gt;

&lt;p&gt;Sources: Mikolov et al., "Efficient Estimation of Word Representations in Vector Space" (2013); Gensim documentation: &lt;a href="https://radimrehurek.com/gensim/" rel="noopener noreferrer"&gt;https://radimrehurek.com/gensim/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>python</category>
      <category>machinelearning</category>
    </item>
  </channel>
</rss>
