<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dr. Tânia Frazão, M.C.S., D.V.M. on Medium]]></title>
        <description><![CDATA[Stories by Dr. Tânia Frazão, M.C.S., D.V.M. on Medium]]></description>
        <link>https://medium.com/@missgorgeoustech?source=rss-a8170c8c5dae------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*eNKmX6HtppeyIfPrralU0g.jpeg</url>
            <title>Stories by Dr. Tânia Frazão, M.C.S., D.V.M. on Medium</title>
            <link>https://medium.com/@missgorgeoustech?source=rss-a8170c8c5dae------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:10:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@missgorgeoustech/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Taming Python Warnings: Essential Techniques for Clean ML Development]]></title>
            <link>https://medium.com/the-abcs-of-ai/taming-python-warnings-essential-techniques-for-clean-ml-development-4865a9e4dd3b?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/4865a9e4dd3b</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 21:10:51 GMT</pubDate>
            <atom:updated>2026-10-29T21:10:51.443Z</atom:updated>
            <content:encoded><![CDATA[<h4>How to keep your notebooks readable without hiding the warnings that actually matter (updated for 2026).</h4><figure><img alt="An open laptop showing code in a dark-themed editor, on a white wooden windowsill next to two small potted green plants." src="https://cdn-images-1.medium.com/max/1024/0*Al9Rl5J52priYGIt" /><figcaption>Photo by <a href="https://unsplash.com/@safarslife?utm_source=medium&amp;utm_medium=referral">Safar Safarov</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>TL;DR</h3><ul><li><strong>Warnings are not errors,</strong> but they’re hints about future breakage, bad numerics, or deprecated APIs.</li><li><strong>Never blanket-ignore everything.</strong> Filter by category, module, or message instead.</li><li><strong>Use context managers</strong> to silence noise only where you need to.</li><li><strong>Log, don’t lose.</strong> Route warnings to a log file with logging.captureWarnings(True) so you can review them later.</li><li><strong>In tests, turn warnings into errors</strong> so deprecations never surprise you in production.</li></ul><p>If you’ve worked with data science libraries or deep learning frameworks, you’ve seen it: a cascade of warning messages flooding your Jupyter notebook. They’re usually not fatal, but they make it hard to focus on the actual output.</p><p>Here’s the thing: the goal isn’t to make warnings disappear. The goal is to <strong>control</strong> them, so the noise goes away and the important messages still reach you. Let me show you how.</p><h3>Understanding Python’s warning system</h3><p>A warning is Python’s way of saying: “This isn’t an error, but you should know about it.”</p><pre>import warnings<br> <br>warnings.warn(&quot;This feature will be removed in version 2.0&quot;, DeprecationWarning)</pre><p>The most common categories you’ll meet:</p><ul><li><strong>DeprecationWarning:</strong> a feature will be removed in a future version (shown by default only in code run directly, like scripts and notebooks).</li><li><strong>FutureWarning:</strong> behavior will change in a future version. Libraries like pandas use it a lot.</li><li><strong>UserWarning:</strong> the generic category libraries use for “heads up” messages.</li><li><strong>RuntimeWarning:</strong> suspicious runtime behavior, like dividing by zero in NumPy or overflow.</li></ul><h3>How Python decides what to do with a warning</h3><p>Every time a warning is raised, Python walks through a list of <strong>filters</strong>, top to bottom. The first filter that matches decides the action. That’s the whole mental model, and it explains why the order of your filters matters.</p><figure><img alt="Flow diagram: a call to warnings.warn goes to the filter list, which Python checks from top to bottom; the first matching filter decides. Example filters: 1, ignore FutureWarning from pandas; 2, default UserWarning from torch.nn; 3, ignore all UserWarning; 4, Python’s default filters. Three outcomes: ignore drops the warning silently; error raises it as an exception; default, once, module or always show it on screen, or in a log file with logging.captureWarnings(True). Note: filterwarnings adds" src="https://cdn-images-1.medium.com/max/1024/1*qwex8YA8tkygmXgRTG_AQg.png" /><figcaption>Diagram by the author.</figcaption></figure><h3>The sledgehammer (and why to avoid it)</h3><pre>import warnings<br>warnings.filterwarnings(&quot;ignore&quot;)  # silences everything from here on</pre><p>It works, but it also hides the warning that tells you your model is silently producing NaNs. Use it only for a quick throwaway experiment, never in shared or production code.</p><h3>Targeted filtering</h3><p>Be as specific as you can:</p><pre># Ignore only deprecation warnings<br>warnings.filterwarnings(&quot;ignore&quot;, category=DeprecationWarning)<br> <br># Ignore warnings coming from one package<br>warnings.filterwarnings(&quot;ignore&quot;, module=&quot;some_noisy_package&quot;)<br> <br># Ignore a specific message (it&#39;s a regular expression)<br>warnings.filterwarnings(&quot;ignore&quot;, message=&quot;.*is experimental.*&quot;)</pre><h3>Silence noise only where you need to</h3><p>The catch_warnings context manager saves the filter state and restores it when the block ends. Since Python 3.11 you can pass the action directly:</p><pre>import warnings<br> <br>with warnings.catch_warnings(action=&quot;ignore&quot;, category=FutureWarning):<br>    import some_noisy_module   # quiet here<br> <br># outside the block, warnings behave normally again</pre><h3>Log warnings instead of hiding them</h3><p>This is my favorite pattern for notebooks and pipelines. One line sends every warning to the standard logging system (under the logger name py.warnings), so you can write them to a file and keep your output clean:</p><pre>import logging<br>import warnings<br> <br>logging.basicConfig(filename=&quot;ml_warnings.log&quot;, level=logging.WARNING)<br>logging.captureWarnings(True)   # warnings now go to the log file<br> <br>warnings.warn(&quot;Learning rate scheduler is deprecated&quot;, DeprecationWarning)</pre><p>Nothing disappears: you just moved the messages somewhere you can read them when you want to.</p><h3>Collect and summarize warnings during training</h3><p>Fine-tuning runs can produce hundreds of repeated warnings. Instead of scrolling through them, record them and print a short summary at the end:</p><pre>import warnings<br>from collections import Counter<br> <br>with warnings.catch_warnings(record=True) as caught:<br>    warnings.simplefilter(&quot;always&quot;)   # record every occurrence<br>    trainer.train()                    # your training loop or Trainer<br> <br>summary = Counter((w.category.__name__, str(w.message)[:80]) for w in caught)<br>for (category, message), count in summary.most_common(10):<br>    print(f&quot;{count:&gt;5}x  {category}: {message}&quot;)</pre><p>You get a clean training log, plus a top-10 list of what actually went wrong, with counts.</p><h3>Taming the usual suspects in ML</h3><h4>Hugging Face Transformers</h4><pre>import os<br>os.environ[&quot;TOKENIZERS_PARALLELISM&quot;] = &quot;false&quot;   # avoids the fork warning<br> <br>from transformers import AutoModelForCausalLM, AutoTokenizer<br>from transformers import logging as hf_logging<br> <br>hf_logging.set_verbosity_error()   # only show errors from transformers<br> <br>tokenizer = AutoTokenizer.from_pretrained(&quot;gpt2&quot;)<br>model = AutoModelForCausalLM.from_pretrained(&quot;gpt2&quot;)</pre><p>You can do the same without touching code by setting the environment variable TRANSFORMERS_VERBOSITY=error.</p><h4>PyTorch</h4><pre>import warnings<br> <br># Hide generic UserWarnings, but keep the ones from torch.nn visible<br>warnings.filterwarnings(&quot;ignore&quot;, category=UserWarning)<br>warnings.filterwarnings(&quot;default&quot;, category=UserWarning, module=&quot;torch.nn&quot;)</pre><p>Notice the order: filterwarnings inserts each new filter at the <strong>front</strong> of the list, so the more specific rule (added last) is checked first.</p><h4>TensorFlow and Keras</h4><p>Much of TensorFlow’s noise comes from its C++ backend, which doesn’t use Python warnings at all:</p><pre>import os<br>os.environ[&quot;TF_CPP_MIN_LOG_LEVEL&quot;] = &quot;2&quot;   # 0=all, 1=no INFO, 2=no WARNING, 3=errors only<br> <br>import tensorflow as tf<br>tf.get_logger().setLevel(&quot;ERROR&quot;)</pre><p>Set the environment variable <strong>before</strong> importing TensorFlow, or it won’t take effect.</p><h3>Without touching your code</h3><p>Sometimes you can’t (or shouldn’t) edit the code. Python reads filters from the command line and the environment:</p><pre># Run a script with deprecation warnings turned into errors<br>python -W error::DeprecationWarning train.py<br> <br># Or for every Python process in this shell<br>export PYTHONWARNINGS=&quot;ignore::FutureWarning&quot;</pre><h3>Make warnings fail your tests</h3><p>This is the best way to stay ahead of breaking changes. In pytest, add this to your pyproject.toml:</p><pre>[tool.pytest.ini_options]<br>filterwarnings = [<br>    &quot;error&quot;,<br>    &quot;ignore::DeprecationWarning:some_third_party_lib.*&quot;,<br>]</pre><p>Your own code can’t ship with new deprecations, while known third-party noise stays tolerated, on purpose and documented.</p><h3>Bonus: mark your own code as deprecated</h3><p>Since Python 3.13, the standard library has a decorator for this (PEP 702). Type checkers and IDEs understand it too:</p><pre>from warnings import deprecated<br> <br>@deprecated(&quot;Use load_dataset_v2() instead&quot;)<br>def load_dataset(path):<br>    ...</pre><p>On older Python versions, the same decorator is available from the typing_extensions package.</p><h3>Best practices</h3><ol><li>Never globally ignore all warnings in shared or production code.</li><li>Filter as specifically as possible: category, module, message.</li><li>Prefer context managers over global filters.</li><li>Log warnings instead of deleting them.</li><li>Turn warnings into errors in your test suite.</li><li>Leave a comment explaining <em>why</em> each filter exists.</li></ol><h3>Bottom line</h3><p>Warnings are hints, not enemies. They tell you about deprecated APIs, numerical problems, and changes that will break your code next year. The trick is not to hide them, but to manage them: quiet where they’re noise, loud where they matter, and always recorded somewhere.</p><p>Your notebooks get cleaner, your training logs get readable, and your future self gets fewer surprises.</p><h3>Subscribe</h3><p>Follow <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> for more practical, beginner-friendly articles about AI and Python.</p><p>Watch the videos on the <a href="https://www.youtube.com/@ABCs-of-GenAI">ABCs of GenAI YouTube channel</a> and listen to the podcast <a href="https://www.youtube.com/@ABCs-of-GenAI/podcasts">AI Agents and Coffee</a>.</p><p>Find me on X: <a href="https://x.com/GorgeousTech">x.com/GorgeousTech</a></p><p>#Python #MachineLearning #DeepLearning #DataScience #Programming</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4865a9e4dd3b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/taming-python-warnings-essential-techniques-for-clean-ml-development-4865a9e4dd3b">Taming Python Warnings: Essential Techniques for Clean ML Development</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Truth About Dog Grooming: A Vet’s Guide to Coat, Nails, Ears and Teeth]]></title>
            <link>https://medium.com/better-pets/the-truth-about-dog-grooming-a-vets-guide-to-coat-nails-ears-and-teeth-48db1e1062fb?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/48db1e1062fb</guid>
            <category><![CDATA[pet-care]]></category>
            <category><![CDATA[dogs]]></category>
            <category><![CDATA[veterinary]]></category>
            <category><![CDATA[pet-health]]></category>
            <category><![CDATA[dog-grooming]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 21:03:39 GMT</pubDate>
            <atom:updated>2026-10-29T21:03:39.871Z</atom:updated>
            <content:encoded><![CDATA[<h4>Grooming isn’t just about looking good. It’s one of the easiest ways to keep your dog healthy, if you do it the right way.</h4><figure><img alt="A small fluffy white dog wrapped in a blue plaid blanket while a person gently brushes its fur with a pink slicker brush." src="https://cdn-images-1.medium.com/max/1024/0*BeLg8MtHy0WuTxcd" /><figcaption>Photo by <a href="https://unsplash.com/@abject?utm_source=medium&amp;utm_medium=referral">benjamin lehman</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>TL;DR</h3><ul><li><strong>Grooming is health care.</strong> It’s your weekly chance to find lumps, parasites, skin problems, and sore spots early.</li><li><strong>Match the routine to the coat.</strong> A Labrador, a Husky, and a Poodle need very different care.</li><li><strong>Don’t forget nails, ears, and teeth.</strong> Dental disease is one of the most common problems I see in adult dogs.</li><li><strong>Watch your dog’s body language.</strong> A growl is communication, not disobedience. Never punish it.</li></ul><h3>Here’s the thing</h3><p>When people hear “grooming”, they often think of a fancy haircut. As a vet, I think of something else: a regular, hands-on health check. Every time you brush your dog, you can notice a new lump, a tick, a hot spot, a painful ear, or a broken nail long before it becomes a big problem.</p><p>And there’s a bonus. Calm, regular grooming sessions build trust between you and your dog, and make vet visits much easier.</p><h3>Brushing: match the routine to the coat</h3><ul><li><strong>Short, smooth coats</strong> (Labradors, Beagles, Boxers): brush about once a week with a rubber curry mitt or bristle brush to remove loose hair and spread natural oils.</li><li><strong>Double coats</strong> (Huskies, German Shepherds, Golden Retrievers): brush a few times a week, and almost daily during shedding season, with an undercoat rake. Avoid shaving double-coated dogs unless your vet recommends it for a medical reason, because the coat helps protect them from heat, cold, and sun, and may not grow back the same way.</li><li><strong>Long, silky coats</strong> (Yorkshire Terriers, Shih Tzus, Cavaliers): brush daily with a slicker brush and finish with a metal comb. Pay extra attention behind the ears, under the legs, and around the collar, where mats form first.</li><li><strong>Curly and wire coats</strong> (Poodles, Bichons, many Terriers): brush several times a week and plan regular professional trims, usually every 4 to 8 weeks.</li></ul><h4>Why mats matter</h4><p>Mats are not just a cosmetic issue. They pull on the skin every time your dog moves, trap moisture and dirt, and can hide skin infections and parasites. Please don’t cut mats out with scissors close to the skin: it’s very easy to cut your dog by accident, and I see these wounds in the clinic regularly. Small mats can be teased apart with a detangling spray and comb; severe matting is a job for a professional groomer with clippers.</p><h3>Bathing: less often than you think</h3><ul><li>Most dogs need a bath every 4 to 8 weeks, or when they’re dirty or smelly. Too many baths can dry out the skin.</li><li><strong>Use a dog shampoo.</strong> Dog skin has a different pH from ours, and human shampoos can irritate it.</li><li>Rinse thoroughly: shampoo residue is a common cause of itching.</li><li>Dry well, especially in skin folds, between the toes, and in thick coats, to prevent skin infections.</li><li>If your vet prescribed a medicated shampoo, follow the instructions about contact time; it often needs several minutes on the skin to work.</li></ul><h3>Nails: the most skipped step</h3><p>A simple rule: if you can hear your dog’s nails clicking on the floor, they’re too long. Long nails change the way a dog stands and walks, which can strain joints over time, and they’re more likely to split or break.</p><ul><li>Trim every 3 to 4 weeks, taking small bits at a time.</li><li>Avoid the <strong>quick</strong>, the living part of the nail with blood vessels and nerves. In light nails you can see it as a pink area; in dark nails, trim tiny slices and stop when you see a darker dot in the center.</li><li>Keep styptic powder or cornstarch nearby in case of bleeding.</li><li>A nail grinder is a good option for dogs who dislike clippers.</li></ul><h3>Ears: check weekly, clean only when needed</h3><p>Look at and gently smell your dog’s ears once a week. Healthy ears are pale pink, with little or no wax and no strong smell. Clean them only when they’re dirty, using an ear cleaner recommended by your vet, and never push cotton swabs deep into the ear canal.</p><p>Head shaking, scratching, redness, discharge, or a bad smell usually mean an ear infection, and that needs a vet. Ear infections are painful, and home remedies can make them worse.</p><h3>Teeth: the hidden problem</h3><p>Dental disease is extremely common in dogs: most adult dogs already show some signs of it. It causes bad breath, pain, and tooth loss, and the bacteria can affect other organs too.</p><ul><li><strong>Daily brushing is the gold standard.</strong> Use a soft dog toothbrush or finger brush and a <strong>dog toothpaste</strong>. Never use human toothpaste: it can contain fluoride and xylitol, which are dangerous for dogs.</li><li>Start slowly: let your dog lick the toothpaste, then touch the teeth, then brush a few teeth, building up over days.</li><li>Dental chews and diets can help, ideally ones accepted by the Veterinary Oral Health Council (VOHC).</li><li>Regular dental checks at the vet are important, and some dogs will need a professional cleaning under anesthesia.</li></ul><h3>Reading your dog during grooming</h3><p>Grooming involves touching sensitive areas, so it’s important to watch how your dog feels about it. Signs of stress include lip licking, yawning, turning the head away, showing the whites of the eyes, a tucked tail, freezing, or trying to leave.</p><p>A <strong>growl is a warning</strong>, and it’s valuable information. Your dog is telling you “I’m uncomfortable”. If you punish growling, you don’t remove the fear; you only remove the warning, and the next step can be a bite. Instead:</p><ul><li>Stop, give your dog space, and make the next session shorter and easier.</li><li>Pair grooming with high-value treats, and break it into tiny steps: one paw today, two paws tomorrow.</li><li>Teach a calm “chin rest” or a mat to settle on, so your dog can choose to participate (this is often called cooperative care).</li><li>Remember that sudden sensitivity can be a sign of pain, such as arthritis, an ear infection, or a skin problem. If your dog suddenly hates being touched somewhere, book a vet check.</li></ul><p>And no, your dog is not trying to be “the alpha”. The old dominance theory has been widely abandoned by behavior experts. Fear, pain, and stress explain most grooming struggles, and they respond to patience and positive training, not force. If your dog shows aggression during grooming or guards objects or food, talk to your vet and a qualified behaviorist.</p><h3>Choosing a professional groomer</h3><ul><li>Ask how they handle nervous dogs. Look for low-stress or fear-free handling.</li><li>Ask how they dry dogs. Unattended heated cage dryers can cause heatstroke, especially in flat-faced breeds like Pugs and French Bulldogs.</li><li>Sedation should only be given by a vet, never by a groomer.</li><li>Tell them about any health issues, lumps, or sore spots before they start.</li></ul><h3>Your basic grooming kit</h3><ul><li>A brush and comb suited to your dog’s coat</li><li>Dog shampoo and a towel</li><li>Nail clippers or a grinder, and styptic powder</li><li>A vet-recommended ear cleaner and cotton pads</li><li>A dog toothbrush and dog toothpaste</li><li>Plenty of small, tasty treats</li></ul><h3>Bottom line</h3><p>Grooming is one of the simplest, cheapest ways to protect your dog’s health. A few minutes a week of brushing, a monthly nail trim, weekly ear checks, and daily tooth brushing go a long way, especially when every session ends with a treat and a happy dog.</p><p><em>This article is general information and doesn’t replace a consultation with your own veterinarian.</em></p><p>What’s the hardest part of grooming your dog? Tell me in the comments and I’ll try to help.</p><p>Follow <a href="https://medium.com/better-pets">Better Pets</a> for more vet-approved tips for happy, healthy pets.</p><p>#DogGrooming #DogCare #PetHealth #Veterinary #Dogs</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48db1e1062fb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/better-pets/the-truth-about-dog-grooming-a-vets-guide-to-coat-nails-ears-and-teeth-48db1e1062fb">The Truth About Dog Grooming: A Vet’s Guide to Coat, Nails, Ears and Teeth</a> was originally published in <a href="https://medium.com/better-pets">Better Pets</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Saying Goodbye: What Happens During Pet Euthanasia, Explained by a Vet]]></title>
            <link>https://medium.com/better-pets/saying-goodbye-what-happens-during-pet-euthanasia-explained-by-a-vet-e2c6f1f3f588?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/e2c6f1f3f588</guid>
            <category><![CDATA[pet-loss]]></category>
            <category><![CDATA[pets]]></category>
            <category><![CDATA[grief]]></category>
            <category><![CDATA[veterinary]]></category>
            <category><![CDATA[dogs]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 20:58:13 GMT</pubDate>
            <atom:updated>2026-10-29T21:14:43.913Z</atom:updated>
            <content:encoded><![CDATA[<h4>Knowing what to expect can turn the hardest day into a calm and loving goodbye.</h4><figure><img alt="An elderly yellow Labrador with a white muzzle lying calmly on a soft rug at home, looking to the side." src="https://cdn-images-1.medium.com/max/1024/0*3zEpRIXs-vjcVvgM" /><figcaption>Photo by <a href="https://unsplash.com/@rebaspike?utm_source=medium&amp;utm_medium=referral">Reba Spike</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I’m a veterinary doctor, and one thing I consider a gift is being deeply empathic to animal suffering. Over the years, I’ve been with many families on the day they say goodbye to their pet. And I’ve noticed something: much of the fear and pain of that day comes from not knowing what will happen.</p><p>So I decided to write this article. It explains the procedure step by step, including the reactions that are normal and expected, so you can prepare and make this moment what it should be: an act of love and peace, with minimal stress for your pet and for you.</p><h3>TL;DR</h3><ul><li><strong>Euthanasia means “a good death”.</strong> Done well, it is peaceful and painless.</li><li><strong>It usually happens in two steps:</strong> a sedative so your pet is calm and relaxed, then a final anesthetic injection.</li><li><strong>Some things that look worrying are normal:</strong> open eyes, small muscle twitches, a reflex breath, or releasing the bladder. They do not mean your pet is suffering.</li><li><strong>You can plan ahead:</strong> where it happens, who is present, and what happens afterwards.</li></ul><h3>What euthanasia really means</h3><p>The word comes from Greek and means “good death”. In veterinary medicine, it means ending life in a way that is calm, painless, and respectful, when an animal’s suffering can no longer be relieved.</p><p>Medically, euthanasia is an overdose of anesthetic. The same family of drugs we use to put animals under anesthesia for surgery is given at a much higher dose, so your pet falls deeply asleep and does not wake up.</p><h3>How do I know it’s time?</h3><p>This is the question I hear most, and there is rarely a perfect moment. Some questions that help:</p><ul><li>Is my pet in pain that medication no longer controls?</li><li>Are they still eating and drinking?</li><li>Can they move, get up, and keep themselves clean?</li><li>Do they still enjoy the things that made them happy?</li><li>Are there now more bad days than good days?</li></ul><p>Quality-of-life scales, such as the one created by veterinary oncologist Dr. Alice Villalobos, can help you look at this more objectively. Above all, talk openly with your vet. We can tell you what we see, and we can support whatever decision you make.</p><blockquote><em>Many owners tell me they wish they hadn’t waited until the very worst day. Choosing a peaceful goodbye a little early is often kinder than one day too late.</em></blockquote><h3>Before the appointment: decisions you can make in advance</h3><p>If the situation allows, planning ahead makes the day much calmer.</p><ul><li><strong>Where:</strong> at the clinic, or at home. Many vets offer home euthanasia, which can be less stressful for pets who are anxious in the car or at the clinic.</li><li><strong>When:</strong> ask for a quiet time of day, so nobody is rushed.</li><li><strong>Who will be present:</strong> you can stay for the whole procedure, for part of it, or say goodbye beforehand. There is no wrong choice. If you can, I gently encourage staying: pets often look for their people.</li><li><strong>Aftercare:</strong> burial (check your local rules) or cremation, either communal or individual with the ashes returned to you.</li><li><strong>Paperwork and payment:</strong> ask to handle these beforehand, so you don’t have to think about them afterwards.</li><li><strong>Comfort:</strong> bring a favorite blanket or toy. Ask your vet if a special treat is allowed, because it depends on the sedation plan.</li></ul><h3>Step 1: Sedation</h3><p>The first step is to sedate your pet. This has several goals:</p><ul><li>Your pet becomes calm and relaxed, so the vet can handle them gently.</li><li>It provides pain relief, which matters if your pet is already in pain.</li><li>It prevents bites or scratches caused by fear.</li><li>For you, it looks like your pet is falling asleep in your arms.</li></ul><p>While the sedative takes effect, some reactions can happen, and I always warn owners about them. You may see the head trembling slightly. If there is food in the stomach, your pet may vomit. When moved, some animals vocalize or growl, and a few may have brief muscle spasms. Just like people, animals react differently to drugs and doses. The good news is that modern sedatives have fewer and fewer side effects, so for most pets this phase is quiet and peaceful.</p><p>Your vet may also place a small catheter in a vein, usually in a leg. This makes the next step smoother.</p><h3>What a sedated pet looks like</h3><p>Once sedated, your pet is still alive, in a state similar to deep rest. The body is completely relaxed. The eyes are often open. You will see the chest moving, because they are still breathing.</p><p>This is the moment for your last goodbyes. Talk to them, stroke them, hold their paw. Take the time you need; your vet will wait for you to say you’re ready, within the time the sedation lasts.</p><h3>Step 2: The final injection</h3><p>When you’re ready, the vet gives a concentrated anesthetic into the vein. Your pet loses consciousness within seconds, then breathing stops, and shortly after, the heart stops. The vet will listen with a stethoscope and gently tell you when your pet has passed.</p><p>The whole step usually takes only a minute or two.</p><h3>What’s normal afterwards (and why it’s not suffering)</h3><p>Some things can be surprising if nobody told you about them. All of these are normal reflexes of the body and do not mean your pet is in pain or aware:</p><ul><li><strong>The eyes stay open.</strong> Closing the eyelids needs muscle activity, which has stopped.</li><li><strong>Small muscle twitches</strong> can happen for a few minutes.</li><li><strong>One or two deep breaths</strong> (called agonal breaths) can occur even after the heart has stopped. This is a reflex of the nervous system, not a sign of life or distress.</li><li><strong>The bladder or bowels may release,</strong> because the muscles fully relax. Placing a towel or pad underneath helps.</li><li><strong>The tongue may come slightly out of the mouth.</strong></li></ul><p>Knowing these things in advance helps you stay calm and present in those last minutes.</p><h3>Children and other pets</h3><p>With children, use simple and honest words: “Our dog was very sick, and the vet helped him die peacefully, without pain.” With young children, try to avoid saying “put to sleep”, because it can make them afraid of going to sleep themselves.</p><h3>After the goodbye</h3><p>Grief for a pet is real grief. You lost a family member, a daily routine, and an unconditional friend. Give yourself time. Many vets, shelters, and organizations offer pet loss support, and talking about your pet with people who understand really helps.</p><h3>Bottom line</h3><p>Euthanasia is one of the most difficult decisions we make as pet owners, and one of the last gifts we can give: a calm end, without pain, surrounded by love. Knowing what will happen lets you focus on what matters most in that moment: being there.</p><p><em>This article is general information and doesn’t replace a conversation with your own veterinarian, who knows your pet and can guide you through every step.</em></p><p>If you’ve been through this, I’m sorry for your loss. You’re welcome to share your pet’s name in the comments.</p><p>Follow <a href="https://medium.com/better-pets">Better Pets</a> for more vet-approved guidance for pet owners.</p><p>#PetLoss #PetEuthanasia #Veterinary #PetCare #Grief</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2c6f1f3f588" width="1" height="1" alt=""><hr><p><a href="https://medium.com/better-pets/saying-goodbye-what-happens-during-pet-euthanasia-explained-by-a-vet-e2c6f1f3f588">Saying Goodbye: What Happens During Pet Euthanasia, Explained by a Vet</a> was originally published in <a href="https://medium.com/better-pets">Better Pets</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Got a New Puppy? A Vet’s Guide to the First Weeks at Home]]></title>
            <link>https://medium.com/better-pets/got-a-new-puppy-a-vets-guide-to-the-first-weeks-at-home-3f568f7a8c7d?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f568f7a8c7d</guid>
            <category><![CDATA[veterinary]]></category>
            <category><![CDATA[pet-health]]></category>
            <category><![CDATA[pet-care]]></category>
            <category><![CDATA[dogs]]></category>
            <category><![CDATA[puppies]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 20:55:43 GMT</pubDate>
            <atom:updated>2026-10-29T20:55:43.082Z</atom:updated>
            <content:encoded><![CDATA[<h4>Vet visits, vaccines, food, puppy-proofing, first aid and training: what really matters (and what to skip).</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kVISLSE4dS-xh7TH" /><figcaption>Photo by <a href="https://unsplash.com/@mtsjrdl?utm_source=medium&amp;utm_medium=referral">mtsjrdl</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>TL;DR</h3><ul><li><strong>Book a vet check in the first week.</strong> Bring any vaccine and deworming records you received.</li><li><strong>Vaccines come as a series,</strong> usually every 2 to 4 weeks until your puppy is at least 16 weeks old. Your vet will adapt the plan to your country and lifestyle.</li><li><strong>Keep the same food at first,</strong> then change gradually over about a week.</li><li><strong>Puppy-proof your home</strong> and save your vet’s and the emergency clinic’s phone numbers today.</li><li><strong>Never give human medicines</strong> or try to make your puppy vomit unless a vet tells you to.</li></ul><p>As a veterinary doctor, I’ve welcomed many new puppies (and their slightly overwhelmed humans) into the clinic. The questions are almost always the same: <em>What should I do first? What’s normal? When should I worry?</em> This guide answers those questions in the order you’ll actually need them.</p><h3>Day 1: Set up a calm, safe space</h3><p>Moving to a new home is a big change for a young puppy. Everything smells different, the family is new, and mum and littermates are gone. Your first job is to make the world feel safe.</p><ul><li><strong>Create a quiet den.</strong> A crate with the door open, or a small playpen, with a soft bed, water, and a chew toy. Place it where the family spends time, not isolated in a garage.</li><li><strong>Keep the first days low-key.</strong> No big welcome party. Short play sessions, lots of naps. Puppies sleep 18 to 20 hours a day, and that’s normal.</li><li><strong>Bring something that smells like home.</strong> If the breeder or shelter agrees, a blanket from the litter helps a lot during the first nights.</li></ul><h3>Week 1: Your first vet visit</h3><p>Book a check-up within the first few days, even if your puppy looks perfectly healthy. At this visit, your vet will usually:</p><ul><li>Do a full physical exam (heart, eyes, ears, teeth, skin, hernias, weight).</li><li>Review the vaccine and deworming records you received and plan the next doses.</li><li>Check or place a microchip and register it in your name. In many countries, including across the EU and the UK, microchipping dogs is a legal requirement.</li><li>Talk about parasite prevention (worms, fleas, ticks), food, and when to plan neutering.</li></ul><p>Write your questions down before you go. There are no silly questions at a puppy visit, I promise.</p><h3>Vaccines and deworming: the basics</h3><p>Puppies get some protection from their mother’s milk, but it fades over the first months. Because we can’t know exactly when, vaccines are given as a <strong>series</strong>: typically starting at 6 to 8 weeks of age and repeated every 2 to 4 weeks until your puppy is <strong>at least 16 weeks old</strong>, as recommended by international guidelines such as the WSAVA vaccination guidelines. A booster follows later, between 6 and 12 months.</p><p>Core vaccines protect against canine distemper, parvovirus, and adenovirus (infectious hepatitis). Rabies is required in many countries, and vaccines like leptospirosis or kennel cough are recommended depending on where you live and your dog’s lifestyle. Your vet will tailor this.</p><p><strong>Deworming</strong> is just as important. Puppies are commonly dewormed every 2 weeks until about 12 weeks of age, then monthly until 6 months. Follow your vet’s product and dose: puppies are small, and doses matter.</p><blockquote><em>Until the vaccine series is complete, avoid places where many unknown dogs go, like dog parks. Parvovirus survives in the environment for a long time.</em></blockquote><h3>Food and water</h3><ul><li><strong>Don’t change the food on day one.</strong> Ask what your puppy was eating and keep it for the first days. Then switch gradually over about 7 days, mixing more of the new food each day.</li><li><strong>Choose a complete and balanced puppy food</strong> that matches your puppy’s expected adult size. Large and giant breeds need specific large-breed puppy diets to control growth.</li><li><strong>Feed 3 to 4 small meals a day</strong> until around 4 to 6 months, then move to 2 to 3 meals.</li><li><strong>Fresh water, always.</strong></li><li><strong>No cooked bones,</strong> and keep human snacks out of reach (more on that below).</li></ul><h3>Puppy-proof your home</h3><p>Puppies explore with their mouths. Before your puppy arrives, get down to puppy height and look around:</p><ul><li>Hide electrical cables and chargers.</li><li>Store medicines, cleaning products, and rodent and slug poisons in closed cupboards.</li><li>Keep these foods out of reach: chocolate, xylitol (in sugar-free gum and some peanut butters), grapes and raisins, onions and garlic, macadamia nuts, and alcohol.</li><li>Remove small objects that can be swallowed: socks, hair ties, children’s toys, batteries.</li><li>Check that toys are the right size: <strong>too big to swallow</strong>, sturdy, and without small parts that can break off.</li></ul><h3>A simple first aid kit</h3><p>You don’t need a full ambulance at home. You need the right numbers and a few basics:</p><ul><li>Your vet’s phone number, the nearest 24-hour emergency clinic, and an animal poison helpline for your country.</li><li>Sterile gauze, a non-stick bandage, and adhesive tape.</li><li>Sterile saline to rinse eyes or wounds.</li><li>A digital thermometer and a tick remover.</li><li>Blunt-tipped scissors and a clean towel.</li></ul><p>Two important rules:</p><ol><li><strong>Never give human medicines.</strong> Common painkillers like ibuprofen and paracetamol can be very dangerous for dogs.</li><li><strong>Never try to make your puppy vomit at home</strong> unless a vet or poison helpline specifically tells you to, and how. Some substances cause more damage coming back up, and home remedies like salt or bleach can be deadly. If your puppy swallowed something it shouldn’t, call your vet right away.</li></ol><h3>Small breeds: watch for low blood sugar</h3><p>Very small and toy-breed puppies can develop <strong>hypoglycemia</strong> (low blood sugar), especially if they skip meals, get cold, or are stressed. Signs include weakness, trembling, wobbly walking, a glassy look, and in severe cases seizures.</p><p>If you notice these signs and your puppy is conscious and able to swallow, you can rub a small amount of honey or glucose syrup on the gums. Don’t pour liquid into the mouth of a puppy that is weak or can’t swallow, because it may go into the lungs. Then go to the vet immediately, even if your puppy seems to improve. Prevention is simple: regular small meals and a warm place to rest.</p><h3>Training: start gently, start now</h3><p>The period between roughly 3 and 14 weeks of age is a key socialization window. What your puppy experiences calmly now shapes how confident they’ll be as an adult.</p><ul><li><strong>Use positive reinforcement.</strong> Reward what you want with treats, play, and praise. Punishment creates fear and damages trust.</li><li><strong>House training is about timing.</strong> Take your puppy outside after waking up, after eating, after playing, and every couple of hours. Reward right after they go. Accidents happen; just clean them with an enzymatic cleaner and move on.</li><li><strong>Keep sessions short.</strong> Two to five minutes, a few times a day, is perfect for a puppy brain.</li><li><strong>Practice handling.</strong> Gently touch paws, ears, and mouth, and reward. Future vet visits, nail trims, and grooming will be much easier.</li><li><strong>Join a puppy class</strong> run by a qualified trainer that uses reward-based methods. Ask your vet when it’s safe to start based on your puppy’s vaccines.</li></ul><h3>When to call the vet</h3><p>Call your vet the same day if your puppy:</p><ul><li>Has vomiting or diarrhea, especially with blood, or more than once or twice.</li><li>Is unusually tired, won’t eat, or won’t drink.</li><li>Has difficulty breathing, a swollen belly, or pale gums.</li><li>Ate something toxic or something that could block the intestines.</li></ul><p>With young puppies, things can change quickly. It’s always better to call and hear “it’s fine” than to wait.</p><h3>Bottom line</h3><p>The first weeks with a puppy are a mix of joy, chewed shoes, and very little sleep. Focus on the essentials: an early vet visit, a complete vaccine and deworming plan, a safe home, good food, and kind, consistent training. The rest you’ll learn together.</p><p><em>This article is general information and doesn’t replace a consultation with your own veterinarian, who knows your puppy and the rules in your country.</em></p><p>Do you have a new puppy at home? Tell me their name in the comments!</p><p>Follow <a href="https://medium.com/better-pets">Better Pets</a> for more vet-approved tips for happy, healthy pets.</p><p>#Puppy #DogCare #PetHealth #Veterinary #NewPuppy</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f568f7a8c7d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/better-pets/got-a-new-puppy-a-vets-guide-to-the-first-weeks-at-home-3f568f7a8c7d">Got a New Puppy? A Vet’s Guide to the First Weeks at Home</a> was originally published in <a href="https://medium.com/better-pets">Better Pets</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why You Should Read Computer Science Books More Often]]></title>
            <link>https://medium.com/the-abcs-of-ai/why-you-should-read-computer-science-books-more-often-400f024615a2?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/400f024615a2</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[books]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 20:51:31 GMT</pubDate>
            <atom:updated>2026-10-29T20:51:31.829Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI can write the code now. Books teach you whether that code is any good.</h4><figure><img alt="A desk with a tall stack of programming books, including Clean Code, in front of two monitors showing lines of code, with a laptop and a potted plant nearby." src="https://cdn-images-1.medium.com/max/1024/0*dsMcDxm3m0Hr0WU2" /><figcaption>Photo by <a href="https://unsplash.com/@stoman?utm_source=medium&amp;utm_medium=referral">Nangialai Stoman</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>TL;DR</h3><ul><li><strong>AI writes a lot of our code in 2026.</strong> That makes fundamentals more valuable, not less, because someone still has to judge the output.</li><li><strong>Books give you mental models.</strong> Tutorials and chatbot answers give you fragments. A good book gives you the whole map.</li><li><strong>You don’t need to read 50 books.</strong> A handful of well-chosen ones, read slowly and with code open, will change how you think.</li></ul><p>Keep reading for the five reasons I think every developer (and every curious beginner) should read computer science books, plus a short list of books worth your time and a reading routine that fits a busy week.</p><h3>Here’s the thing</h3><p>Today you can ask an AI assistant to write a function, a test suite, or even a small app, and it will happily do it in seconds. So it’s fair to ask: why spend weeks reading a 400-page book about algorithms or operating systems?</p><p>Because you can’t review what you don’t understand.</p><p>AI tools are fantastic accelerators. But they are confident, fast, and sometimes wrong. The person who knows <em>why</em> a hash map is fast, <em>why</em> a database needs indexes, or <em>why</em> a recursive function blows the stack is the person who catches the bug before it reaches production. That knowledge mostly lives in books.</p><h3>Why books, and not just docs, videos, or AI answers?</h3><p>Docs tell you <em>how</em> to use a tool. Videos are great for a quick demo. AI answers are perfect for a specific question. But all three give you pieces.</p><p>A well-written book is different. An author spent years deciding what to teach, in which order, and what to leave out. You get a structured path from the basics to the hard parts, with the connections between ideas made explicit. That structure is exactly what builds intuition.</p><blockquote><em>Tutorials teach you to follow steps. Books teach you to think.</em></blockquote><h3>5 reasons to read CS books more often</h3><h4>1. You become a better reviewer of AI-generated code</h4><p>When an assistant hands you 200 lines of code, your job shifts from writing to reviewing. Reviewing well means spotting the O(n²) loop hiding inside an innocent-looking list comprehension, the race condition in async code, or the SQL query that will crawl once the table grows. Books on algorithms, concurrency, and databases train exactly that eye.</p><h4>2. Fundamentals age slowly</h4><p>Frameworks change every year. Data structures, networking, memory, and complexity barely change at all. Time spent on fundamentals keeps paying you back for decades, while yesterday’s framework tutorial is often outdated before you finish it.</p><h4>3. Debugging needs a mental model</h4><p>Debugging is detective work. You form a hypothesis about what the machine is doing and test it. If you don’t have a model of how memory, processes, or the network actually behave, you’re left guessing (or pasting the error into a chatbot and hoping). Books give you that model.</p><h4>4. Better vocabulary means better prompts</h4><p>This one surprises people. The more precise your technical vocabulary, the better your results with AI. Asking for “a function that’s fast” gets you something. Asking for “an O(log n) lookup using a balanced tree, with the edge cases for empty input handled” gets you something much closer to what you need.</p><h4>5. Books teach trade-offs, not just answers</h4><p>Real engineering is a series of trade-offs: speed versus simplicity, consistency versus availability, flexibility versus safety. Good books walk you through <em>why</em> each choice exists. That’s the kind of judgment AI can’t make for you, because it doesn’t know your constraints, your users, or your budget.</p><h3>A short list of books worth your time</h3><p>You don’t need a huge library. Pick the one that matches your current goal.</p><ul><li><strong>Understand how computers really work:</strong> <em>Code: The Hidden Language of Computer Hardware and Software</em> by Charles Petzold. Friendly, visual, and surprisingly fun.</li><li><strong>Think like a professional programmer:</strong> <em>The Pragmatic Programmer</em> by David Thomas and Andrew Hunt. Short chapters full of habits that still hold up.</li><li><strong>Algorithms without fear:</strong> <em>Grokking Algorithms</em> by Aditya Bhargava, with illustrations on almost every page. When you want the deep version, <em>Introduction to Algorithms</em> by Cormen, Leiserson, Rivest, and Stein is the classic reference.</li><li><strong>Write better Python:</strong> <em>Fluent Python</em> by Luciano Ramalho. It explains how the language actually works under the hood.</li><li><strong>Systems and data:</strong> <em>Designing Data-Intensive Applications</em> by Martin Kleppmann. If you build anything with a database, read it.</li><li><strong>AI and machine learning in practice:</strong> <em>Designing Machine Learning Systems</em> and <em>AI Engineering</em> by Chip Huyen, and <em>Build a Large Language Model (From Scratch)</em> by Sebastian Raschka.</li><li><strong>The timeless classic:</strong> <em>Structure and Interpretation of Computer Programs</em> by Abelson and Sussman, which you can read for free online.</li></ul><h3>How to actually read them (when you’re busy)</h3><p>Let’s be honest: most CS books die on the nightstand after chapter two. Here’s a routine that works:</p><ol><li><strong>20 minutes a day beats 4 hours on Sunday.</strong> Consistency builds understanding; marathons build fatigue.</li><li><strong>Read with your editor open.</strong> Type the examples. Break them. Fix them. Reading code is not the same as running it.</li><li><strong>Use AI as a study buddy, not a shortcut.</strong> After a chapter, ask your assistant to quiz you, or ask it to explain a concept in a different way. Then check its answer against the book.</li><li><strong>Skip without guilt.</strong> You don’t need to read every chapter in order. Read what solves today’s problem, then come back.</li><li><strong>Teach what you learned.</strong> Write a short post, explain it to a colleague, or record a two-minute video. If you can explain it simply, you understand it.</li></ol><h3>Bottom line</h3><p>AI is changing how we write software, and that’s exciting. But it raises the bar on understanding, it doesn’t lower it. The developers who thrive will be the ones who can direct AI tools <em>and</em> judge their output, and that judgment comes from fundamentals.</p><p>So pick one book from the list. Read 20 minutes today. Your future self (and your code reviewers) will thank you.</p><p>Which CS book changed the way you think? Tell me in the comments, I’m always looking for my next read.</p><h3>Subscribe</h3><p>Follow <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> for more practical, beginner-friendly articles about AI.</p><p>Watch the videos on the <a href="https://www.youtube.com/@ABCs-of-GenAI">ABCs of GenAI YouTube channel</a> and listen to the podcast <a href="https://www.youtube.com/@ABCs-of-GenAI/podcasts">AI Agents and Coffee</a>.</p><p>Find me on X: <a href="https://x.com/GorgeousTech">x.com/GorgeousTech</a></p><p>#ComputerScience #Programming #SoftwareEngineering #LearnToCode #AI</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=400f024615a2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/why-you-should-read-computer-science-books-more-often-400f024615a2">Why You Should Read Computer Science Books More Often</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI Enters the Browser Wars with ChatGPT Atlas]]></title>
            <link>https://medium.com/the-abcs-of-ai/openai-enters-the-browser-wars-with-chatgpt-atlas-70afe3c0cb68?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/70afe3c0cb68</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Mon, 03 Nov 2025 22:09:54 GMT</pubDate>
            <atom:updated>2025-11-06T18:06:18.256Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why OpenAI’s New Browser Should Terrify Google</h3><h4>OpenAI’s new AI-powered web browser integrates ChatGPT directly into your browsing experience — here’s everything you need to know about features, privacy, and what it means for the future of web browsing</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WtyldaPDI23LbJQ8smrhoA.png" /><figcaption>From OpenAI offical website accessed 03/1172025.</figcaption></figure><h3>TL;DR</h3><p>OpenAI just launched ChatGPT Atlas, a new web browser for macOS that puts AI directly into your browsing experience. Built on Chromium, it lets you chat with ChatGPT directly from any webpage, summarize content with a single click, and even search the web using AI. The catch? It can remember your browsing habits to become smarter (which you can turn off), and like all AI, it sometimes makes mistakes. Think of it as Microsoft Edge’s Copilot, but from OpenAI.</p><p>So here’s the thing: OpenAI just jumped into the browser game. And they’re not just making another Chrome clone — they’re trying to reimagine what browsing the web actually means in an AI-powered world.</p><h3>What Actually Is ChatGPT Atlas?</h3><p>Let’s start with the basics. ChatGPT Atlas is a web browser. You know, like Chrome or Firefox, or Safari. It’s got tabs, bookmarks, password filling, and incognito mode. All the usual stuff.</p><p>But here’s where it gets interesting.</p><p>OpenAI built Atlas on top of Chromium — that’s the same engine Chrome uses — but they’ve woven ChatGPT directly into the browsing experience. Not as an extension you download. Not as a separate app. It’s <em>built in</em>.</p><p>Think about how you normally use ChatGPT. You’re reading an article, you find something interesting, you copy it, switch to ChatGPT, paste it, ask your question, get an answer, then switch back. It’s clunky, right?</p><p>Atlas fixes that.</p><h3>How Does It Actually Work?</h3><p>OpenAI showed off a few clever tricks during their demo:</p><p><strong>Selecting text anywhere on a webpage.</strong> You can highlight text — maybe it’s an email you’re writing, maybe it’s a paragraph you’re reading — and a little ChatGPT button appears. Click it. Ask the AI to improve your writing, explain the concept, translate it, whatever. No copying and pasting required.</p><p><strong>The sidebar.</strong> There’s an “Ask ChatGPT” button in the corner. Click it, and ChatGPT opens in a sidebar. The clever part? It can <em>see</em> the webpage you’re on. So you can ask “summarize this article” or “what does this code do?” and it knows what you’re talking about.</p><p><strong>Search, reimagined.</strong> When you search from the URL bar, Atlas can use ChatGPT’s search engine instead of Google. The results come organized into categories: text, images, videos, and news. It’s different. Whether it’s better? That’s for you to decide.</p><blockquote>But here’s something important: during their own demo, OpenAI showed that Atlas sometimes gets things wrong. The AI hallucinated. Made mistakes. This isn’t a bug — it’s just how these systems work. You still need to think critically about what it tells you.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FEj6hnsQgV_c&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DEj6hnsQgV_c&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2FEj6hnsQgV_c%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/df288d043af9ac886d2a13b60a52e325/href">https://medium.com/media/df288d043af9ac886d2a13b60a52e325/href</a></iframe><h3>The Memory Thing</h3><p>Now we get to the part that’s both really cool and potentially concerning.</p><p>Atlas has something called “browser memory.” When you turn it on, ChatGPT remembers things from your browsing. Not just from your chats — from the actual websites you visit.</p><p>Imagine this: You’ve been browsing Christmas gifts. Atlas notices. Later, when you open the browser, it might suggest “Want to continue looking at gifts?” or even “Here’s a comparison of those three products you looked at yesterday.”</p><p>Or maybe you’re working on a project. Atlas sees you visiting certain sites, reading certain docs. It could automatically generate a to-do list based on what you’ve been researching.</p><p>Pretty smart, right?</p><p>But also… kind of personal. OpenAI knows this raises questions. So here’s what they did:</p><ul><li>It’s <strong>off by default</strong>. You have to turn it on.</li><li>It doesn’t remember passwords or payment info.</li><li>It’s separate from cookies and site storage.</li><li>You can delete these memories anytime.</li></ul><blockquote>Oh, and one more thing: OpenAI says this browsing data <em>might</em> be used to improve their AI models. So if privacy is important to you — and it should be — you’ll want to think carefully about enabling this feature.</blockquote><h3>Why Does This Matter?</h3><p>Here’s the bigger picture: OpenAI isn’t just making a browser for fun. They’re making a move for control.</p><p>Right now, most of us use Chrome. We search with Google. We browse websites that Google’s algorithms decided we should see. Google controls a huge chunk of how we access information online.</p><p>Microsoft saw this and integrated Copilot into Edge. Now OpenAI is doing the same thing with Atlas.</p><p>The browser wars aren’t about speed anymore. They’re not about which one uses less RAM (though, let’s be honest, they all use too much). They’re about AI. Whoever controls the browser controls how AI mediates your relationship with information.</p><p>That’s what’s really at stake here.</p><h3>Should You Try It?</h3><p>If you’re on macOS and you’re curious about AI? Sure, why not.</p><p>If you’re someone who’s always copying things to ChatGPT anyway? This might actually save you time.</p><p>If you’re privacy-conscious? Maybe wait and see. Or at least keep that browser memory turned off.</p><blockquote>The truth is, we don’t know yet if Atlas will become a serious Chrome competitor or if it’ll be a niche tool for AI enthusiasts. What we do know is that it represents a vision of the future — one where browsing the web and talking to AI are the same thing.</blockquote><p>Whether that future is exciting or concerning probably depends on who you ask.</p><p>But one thing’s for sure: the way we browse the internet is changing. And ChatGPT Atlas is OpenAI’s bet on what comes next.</p><h3>Want to Learn More About Generative AI?</h3><p>If you found this article helpful, I invite you to explore more content on my YouTube channel <a href="https://www.youtube.com/@ABCs-of-GenAI"><strong>ABCs of GenAI</strong></a>. The channel offers informative videos about the latest developments in generative artificial intelligence, making complex AI concepts accessible to everyone — from beginners to experts.</p><p>Whether you’re curious about new AI tools, want to understand how models like ChatGPT work, or stay updated on the rapidly evolving world of GenAI, you’ll find simple explanations and practical insights designed to help you navigate this exciting technology.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FISzH135yYdc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DISzH135yYdc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FISzH135yYdc%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8297fb4ae2a07d2460ac90c99d64f538/href">https://medium.com/media/8297fb4ae2a07d2460ac90c99d64f538/href</a></iframe><p>Subscribe and click on the notifications bell. It will help me grow and reach more people.</p><p>Thank you for reading and watching my content.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=70afe3c0cb68" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/openai-enters-the-browser-wars-with-chatgpt-atlas-70afe3c0cb68">OpenAI Enters the Browser Wars with ChatGPT Atlas</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Workflows vs Agents: Which AI System Should You Actually Build?]]></title>
            <link>https://medium.com/the-abcs-of-ai/workflows-vs-agents-which-ai-system-should-you-actually-build-9f433ab11c83?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f433ab11c83</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Tue, 14 Oct 2025 14:37:39 GMT</pubDate>
            <atom:updated>2025-10-14T14:37:39.937Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stop over-engineering your AI systems. Here’s how to choose between single LLMs, workflows, and autonomous agents.</h4><figure><img alt="workflows schema on the left and a pink neon brain with the word agents below" src="https://cdn-images-1.medium.com/max/1024/1*4aYICdNDoXBSgUhSkxqCLw.png" /></figure><h3>TL;DR</h3><ul><li><strong>Most developers are choosing the wrong AI architecture</strong> (and wasting time and money because of it).</li><li>You have three options: <strong>Single LLMs</strong>, <strong>Structured Workflows</strong>, and <strong>Autonomous Agents</strong>. Each solves different problems, and picking the wrong one is expensive.</li><li><strong>The mistake teams make:</strong> Building an agent when they needed a workflow. Or worse — building a workflow when a single LLM call would’ve done the job.</li></ul><p><strong>Want to know which one you actually need?</strong> Keep reading. I’ll show you the three questions that make this decision obvious.</p><p>Here’s the thing: when you’re building with AI, the architecture you choose matters way more than most people realize. Pick the wrong one, and you’re either over-engineering something simple or under-engineering something complex. Neither is good.</p><p>So which AI system should you actually use?</p><p>It depends on three big questions:</p><ol><li><strong>How complex is your task?</strong> Is it something simple like “summarize this article” or something more complex like “plan an entire vacation”?</li><li><strong>Does your system need to adapt?</strong> Does it need to remember things or change based on what’s happening in real time?</li><li><strong>What are your operational requirements?</strong> Do you need predictability, or can you accept some unpredictability in exchange for flexibility?</li></ol><p>Answer those three questions, and you’ll know exactly which system to build.</p><p>Let’s dive into the three main approaches: <strong>single LLM features</strong>, <strong>structured workflows</strong>, and <strong>autonomous agents</strong>.</p><h3>Single LLM Features: The Simplest AI Design</h3><p><strong>What does this actually mean?</strong></p><p>You give it some input, and it gives you some output. That’s it. No memory of what happened before, no complicated multi-step reasoning. Just input goes in, output comes out.</p><p><strong>When to use it:</strong></p><p>This is best for simple, well-defined, one-shot tasks like:</p><ul><li>Text summarization</li><li>Sentiment classification</li><li>Translation</li><li>Information extraction</li></ul><p><strong>How it works:</strong></p><p>Imagine you have a news article about AI in healthcare. You send it to Claude and say, “Summarize this in one sentence.” You get back: “AI is revolutionizing healthcare through improved diagnostics and personalized treatment.” Done. No memory between calls, no multi-step reasoning — just a direct input-output flow.</p><p><strong>The pros:</strong></p><ul><li><strong>Super fast</strong> to build and run</li><li><strong>Deterministic</strong>: same input = same output every time</li><li><strong>Cheap</strong>: minimal compute and orchestration overhead</li></ul><p><strong>The cons:</strong></p><ul><li><strong>Can’t adapt</strong>: no context or dynamic decision-making</li><li><strong>No memory</strong>: each request is brand new, as if you’re talking to it for the first time</li></ul><p><strong>Bottom line:</strong> If you’ve got a basic task like “translate this” or “summarize that,” just use a single LLM. Don’t build some fancy system when a simple input-output will do the job.</p><h3>Structured Workflows: For Complex but Predictable Tasks</h3><p>Now let’s level up. <strong>Structured workflows</strong> are when you chain things together: Step 1 happens, then Step 2, then Step 3. You write the code that says exactly what happens and in what order.</p><p>It’s <strong>deterministic</strong>, meaning you know exactly what’s going to happen every single time. No surprises.</p><p><strong>When to use it:</strong></p><p>When you’ve got repetitive tasks that need to happen the same way every time. Think about:</p><ul><li><strong>Processing insurance claims</strong>: Scan the documents → extract the data → validate it → store it. Same steps, same order, every time.</li><li><strong>Batch report generation</strong>: generating 100 reports overnight</li><li><strong>Financial and healthcare transaction processing</strong>: where you need consistency, traceability, and proof for compliance</li></ul><p><strong>The pros:</strong></p><ul><li><strong>Rock solid</strong>: you know exactly what’s going to happen</li><li><strong>Consistent</strong>: works the same way every time</li><li><strong>Cost-efficient</strong>: no unnecessary exploration</li><li><strong>Compliance-ready</strong>: easy to monitor, debug, audit, and version</li></ul><p><strong>The cons:</strong></p><ul><li><strong>Rigid</strong>: If something new comes up that you didn’t plan for, you’re stuck. You have to go back, write new code, and handle that new edge case.</li><li><strong>Development overhead</strong>: you need to code each exception or variant</li></ul><p><strong>Bottom line:</strong> Use workflows when you absolutely need things to happen the same way every time. When compliance matters, when you need an audit trail, when efficiency is everything — that’s workflow territory.</p><h3>Autonomous Agents: The Big Guns</h3><p>Here’s where things get really interesting.</p><p><strong>Autonomous agents</strong> are built to tackle the messy, complicated stuff that doesn’t have a clear playbook. They adapt on the fly and solve problems you didn’t even know you had.</p><p><strong>What does this actually mean?</strong></p><p>Think of an autonomous agent like a really smart assistant that can figure things out as it goes:</p><ol><li>It makes a plan</li><li>Tries something</li><li>Sees what happens</li><li>Adjusts</li></ol><p>It’s got a whole toolbox at its disposal, and it picks the right tool for the job. If the first approach doesn’t work? No problem. It pivots, tries something else, learns from what just happened, and keeps going.</p><p><strong>When to use it:</strong></p><p>When you don’t know the answer up front. Complex stuff where there’s no clear path:</p><ul><li><strong>Research agents</strong> synthesizing information across dozens of sources</li><li><strong>Adaptive customer support</strong> where every conversation is different</li><li><strong>AI-driven virtual assistants</strong> that help plan vacations by gathering preferences, researching destinations, suggesting accommodations, and adapting based on feedback</li><li><strong>Automation that iteratively refines results</strong> based on real-time feedback</li></ul><p><strong>The pros:</strong></p><ul><li><strong>Highly adaptable</strong>: handles unforeseen situations</li><li><strong>Dynamic decision-making</strong>: iterates and improves over time</li><li><strong>Reduces human intervention</strong>: manages complexity autonomously</li></ul><p><strong>The cons:</strong></p><ul><li><strong>Unpredictable outcomes</strong>: you never quite know what they’ll do</li><li><strong>Complex to build and debug</strong>: harder to guarantee compliance</li><li><strong>More expensive to run</strong></li></ul><blockquote><strong>Bottom line:</strong> Agents are NOT your default choice. I know they’re trendy and there’s a lot of hype, but you should only deploy them when you actually need that adaptability — when the problem is genuinely messy and complex, with no clear solution path.</blockquote><h3>Three Simple Rules to Choose Wisely</h3><p>Let me give you three rules that will save you a ton of time and headaches:</p><h3>Rule #1: Start Simple</h3><p>Seriously, don’t overthink it. If you’ve got a basic task like “translate this,” “summarize that,” or “edit my report,” just use a single LLM. Don’t build some fancy system when a single input-output will do the job.</p><h3>Rule #2: Use Workflows for Predictability</h3><p>When you absolutely need things to happen the same way every time — when compliance matters, when you need an audit trail, when efficiency is everything — that’s workflow territory.</p><h3>Rule #3: Agents Are Not Your Default</h3><p>Yes, they’re cool. Yes, they’re powerful. But deploy them only when you truly need that adaptability. When the problem is genuinely complex and messy, when you need real reasoning — then and only then, bring out the agents.</p><h3>The Real World: Hybrid Architectures</h3><p>Here’s the secret: you don’t actually have to choose just one design.</p><p>In the real world, smart teams are mixing and matching. They use workflows for the boring, predictable stuff, then put an agent on top when needed. The ecosystem is exploding with integration standards like:</p><ul><li><strong>Model Context Protocol (MCP)</strong> from Anthropic</li><li><strong>Agent Communication Protocol (ACP)</strong> from IBM</li></ul><p>Frameworks like <strong>LangChain</strong> also do a lot of the heavy lifting for you.</p><p>The point is: you don’t have to reinvent the wheel. Just pick your foundation, add what you need, and use the standards that are already out there. The tools are getting better every day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YuCtbYuhLEEpMThfJu7mXw.png" /></figure><h3>The Bottom Line</h3><p>Here’s what it all comes down to: <strong>the most complex solution is almost never the best solution.</strong></p><p>Pick the simplest thing that actually solves your problem. Your future self will thank you.</p><p>Start with single LLM features for atomic problems. Use workflows when you need rock-solid predictability. Deploy agents only when you truly need that adaptability and the problem is genuinely complex and messy.</p><p>Focus on solving your actual problem. That’s my best advice.</p><p>📺 <em>Watch the video version of this article (published October 8, 2025) on my YouTube channel</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FISzH135yYdc%3Fstart%3D4%26feature%3Doembed%26start%3D4&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DISzH135yYdc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FISzH135yYdc%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/411463224db916d182000281a8025850/href">https://medium.com/media/411463224db916d182000281a8025850/href</a></iframe><h3>Related Reading from ABCs of Gen AI</h3><ul><li><a href="https://medium.com/the-abcs-of-ai/chatgpt-agent-automate-full-workflows-without-losing-control-46d680bce10e">ChatGPT Agent: Automate Full Workflows Without Losing Control</a></li><li><a href="https://medium.com/the-abcs-of-ai/agentic-ai-vs-ai-agents-understanding-the-critical-differences-to-avoid-confusion-and-scams-d272b5f34f66">Agentic AI vs AI Agents: Understanding the Critical Differences to Avoid Confusion and Scams</a></li></ul><h3>Subscribe</h3><ul><li><a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a></li><li><a href="https://www.youtube.com/@ABCs-of-GenAI">ABCs of GenAI</a></li></ul><p><a href="https://x.com/GorgeousTech">x.com/GorgeousTech</a></p><p>NEW! <a href="https://www.youtube.com/@ABCs-of-GenAI/podcasts">Podcast — AI Agents and Coffee</a></p><p>#ArtificialIntelligence #AgenticAI #AIAgents #AITerminology #TechExplained #workflows</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f433ab11c83" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/workflows-vs-agents-which-ai-system-should-you-actually-build-9f433ab11c83">Workflows vs Agents: Which AI System Should You Actually Build?</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Python Global Variables: Complete Guide for Functions, Classes, and Modules in AI Development]]></title>
            <link>https://medium.com/the-abcs-of-ai/python-global-variables-complete-guide-for-functions-classes-and-modules-in-ai-development-970addb6d03c?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/970addb6d03c</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[guides-and-tutorials]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Tue, 02 Sep 2025 10:47:59 GMT</pubDate>
            <atom:updated>2025-09-02T10:47:59.029Z</atom:updated>
            <content:encoded><![CDATA[<h4>Master the Read/Modify Distinction and Avoid Common Pitfalls</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*C6st6olti0OYW0Vc" /><figcaption>Photo by <a href="https://unsplash.com/@kylejglenn?utm_source=medium&amp;utm_medium=referral">Kyle Glenn</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction</h3><p>Understanding Python global variables is crucial for AI development work. When I first started working with Python projects, global variables were a frequent source of confusion and bugs. Let’s explore how Python treats global variables differently when you’re reading versus modifying them, how to use them across functions, classes, and modules, and see how this applies to generative AI code.</p><h3>TL;DR: Python Global Variables Quick Reference</h3><ul><li><strong>Global variables</strong> are defined at the top level of a module or explicitly declared with global keyword</li><li><strong>Reading globals</strong>: No global keyword needed - Python automatically checks outer scopes</li><li><strong>Modifying globals</strong>: Must use global keyword before any assignment or modification</li><li><strong>Common mistake</strong>: Forgetting to declare global before modifying, causing UnboundLocalError</li><li><strong>In AI development</strong>: Useful for tracking metrics, sharing hyperparameters, and maintaining state</li><li><strong>Best practice</strong>: Limit globals to constants, configuration, or use classes/parameters for complex state</li></ul><h3>What Are Global Variables in Python?</h3><p>Global variables in Python are variables defined at the top level of a module or explicitly declared global within a function. They can be accessed from any part of the code, including inside functions, classes, and even across modules when properly imported.</p><h3>Example of Python Global Variables</h3><pre>balance = 0  # Our global account balance starts at zero<br><br>def main():<br>    # Notice here we&#39;re just READING the global balance - no &#39;global&#39; keyword needed!<br>    print(&quot;Balance:&quot;, balance)  <br>    deposit(100)  # Let&#39;s add some money<br>    withdraw(50)  # And take some out<br>    print(&quot;Balance:&quot;, balance)  # Reading again - still no &#39;global&#39; needed<br><br>def deposit(n):<br>    # Here&#39;s where it gets interesting - we need to declare we&#39;re using the global variable<br>    global balance  # This tells Python &quot;I want to modify the global balance, not create a local one&quot;<br>    balance += n    # Now we can safely modify it - this changes the global value<br><br>def withdraw(n):<br>    global balance  <br>    balance -= n    # This changes the same global balance that main() reads from<br><br>if __name__ == &quot;__main__&quot;:<br>    main()  </pre><p>(example from CS50 Python course)</p><h3>Python Global Variables Across Functions</h3><h4>Reading vs. Modifying</h4><h4><strong>1. Reading global variables:</strong></h4><ul><li>No global keyword required</li><li>Python automatically checks outer scopes when a variable isn’t found locally</li><li>Think of this as “Python will look everywhere to find a variable you’re reading”</li></ul><p><strong>2. Modifying global variables:</strong></p><ul><li>Must use global keyword</li><li>Without it, Python assumes you’re creating a local variable</li><li>This is Python asking you to be explicit about your intentions to modify global state</li></ul><h4>Common Error When Using Global Variables in Functions</h4><p>If you tried to write the deposit function without the globalkeyword:</p><pre>def deposit(n):<br>    # Without global, Python thinks we want a local variable named balance<br>    balance += n  # ERROR: UnboundLocalError - Python can&#39;t find a local balance to modify!</pre><p>Running this code would give you: UnboundLocalError: local variable &#39;balance&#39; referenced before assignment</p><p>This happens because when Python sees balance += n, it thinks: &quot;They&#39;re assigning to balance, so it must be a local variable.&quot; But += needs to read the value first, and since Python now thinks it&#39;s local (but hasn&#39;t been initialized), it raises an error.</p><h4>Correct Syntax for Global Variables in Functions</h4><pre>def deposit(n):<br>    global balance    # First, tell Python we want the global balance<br>    balance += n      # Now we can safely modify it</pre><h4>Invalid Syntax for Global Variables</h4><p>You might be tempted to try shortcuts like:</p><pre># Don&#39;t try these - they won&#39;t work!<br>global balance += n           # ❌ Python expects variable names only after global<br>global (balance += n)         # ❌ Still not valid syntax<br>global balance = balance + n  # ❌ Can&#39;t combine declaration and assignment</pre><p>The global statement must stand alone, declaring which names should be treated as global variables.</p><h3>Python Global Variables in Classes</h3><p>When working with classes, global variables interact with instance and class variables in specific ways:</p><pre>counter = 0  # Global variable<br><br>class MyCounter:<br>    class_count = 0  # Class variable<br>    <br>    def __init__(self, initial=0):<br>        self.count = initial  # Instance variable<br>        <br>    def increment(self):<br>        self.count += 1  # Modifying instance variable<br>        MyCounter.class_count += 1  # Modifying class variable<br>        <br>        # To modify the global variable<br>        global counter<br>        counter += 1<br>        <br>    def display_all(self):<br>        # No global needed to read<br>        print(f&quot;Instance count: {self.count}&quot;)<br>        print(f&quot;Class count: {MyCounter.class_count}&quot;)<br>        print(f&quot;Global count: {counter}&quot;)</pre><h3>Python Global Variables Across Modules</h3><p>To use global variables across different Python modules:</p><p><strong>File: config.py</strong></p><pre># Global variables defined in a module<br>API_KEY = &quot;your_api_key_here&quot;<br>MODEL_PARAMETERS = {<br>    &quot;temperature&quot;: 0.7,<br>    &quot;max_tokens&quot;: 100<br>}<br><br># Counter that can be modified by other modules<br>request_counter = 0</pre><p><strong>File: api_client.py</strong></p><pre># Import the entire module<br>import config<br><br># Or import specific variables<br>from config import API_KEY, MODEL_PARAMETERS, request_counter<br><br>def make_api_call(prompt):<br>    # Reading imported globals (no &#39;global&#39; needed)<br>    response = call_api(prompt, config.API_KEY, config.MODEL_PARAMETERS)<br>    <br>    # To modify the imported counter<br>    global request_counter  # This refers to config.request_counter when imported with &#39;from&#39;<br>    request_counter += 1<br>    <br>    # Or directly modify through the module<br>    config.request_counter += 1  # Alternative approach<br>    <br>    return response</pre><h3>Applications in Generative AI Development</h3><h4>1. Training Loop Metrics with Global Variables</h4><p>When training generative models, we often need to track metrics across iterations:</p><pre>iteration_count = 0          # We&#39;ll track how many training iterations we&#39;ve done<br>best_fid_score = float(&#39;inf&#39;)  # And the best score we&#39;ve achieved (lower is better for FID), this is to keep track and update the best  fid score.<br><br>def train_generation_model(dataset, epochs):<br>    # We need to modify these counters, so declare them as global<br>    global iteration_count, best_fid_score<br>    <br>    for epoch in range(epochs):  # For each complete pass through the dataset<br>        for batch in dataset:    # For each batch of training examples<br>            # Training code here - this would update our model weights<br>            <br>            # Every 500 iterations, let&#39;s evaluate our model&#39;s performance<br>            if iteration_count % 500 == 0:<br>                current_fid = evaluate_fid_score(generate_samples())<br>                <br>                # If we found a better model, let&#39;s save it!<br>                if current_fid &lt; best_fid_score:<br>                    # Update our global best score with the new best score<br>                    best_fid_score = current_fid  # This is why we needed &#39;global&#39;<br>                    save_checkpoint(&quot;best_model.pt&quot;)  # Save the improved model<br>                    <br>            # Don&#39;t forget to increment our counter!<br>            iteration_count += 1  # This is also why we needed &#39;global&#39;<br>            <br>def print_training_stats():<br>    # Here we&#39;re just READING the globals, not modifying them<br>    # So no &#39;global&#39; keyword is needed - Python will find them automatically<br>    print(f&quot;Trained for {iteration_count} iterations&quot;)<br>    print(f&quot;Best FID score: {best_fid_score:.4f}&quot;)</pre><p>You will encounter this pattern frequently when experimenting with GANs and diffusion models. Tracking the best model across training runs is essential, and using global variables makes this convenient (though not always ideal for larger projects).</p><h4>2. User Settings for Image Generation with Global Variables</h4><p>In interactive generation apps, we often have global settings:</p><pre># Global settings for our image generator<br>generation_settings = {<br>    &quot;style&quot;: &quot;photorealistic&quot;,  # What style of image to generate<br>    &quot;steps&quot;: 50,                # How many denoising steps to use<br>    &quot;guidance_scale&quot;: 7.5,      # How closely to follow the prompt<br>    &quot;seed&quot;: 42                  # Random seed for reproducibility<br>}<br><br>def update_style(new_style):<br>    # We want to modify the global dictionary, so we need to declare it<br>    global generation_settings<br>    generation_settings[&quot;style&quot;] = new_style  # Update just the style key<br>    <br>def update_steps(step_count):<br>    global generation_settings  # Same idea - modifying the global dictionary<br>    generation_settings[&quot;steps&quot;] = step_count  # Update just the steps key<br>    <br>def generate_image(prompt):<br>    # Here&#39;s the interesting part - we&#39;re only READING from the dictionary<br>    # So we don&#39;t need the &#39;global&#39; keyword! Python will find it for us<br>    return diffusion_model.generate(<br>        prompt=prompt,<br>        num_inference_steps=generation_settings[&quot;steps&quot;],<br>        guidance_scale=generation_settings[&quot;guidance_scale&quot;],<br>        seed=generation_settings[&quot;seed&quot;]<br>    )</pre><p>This pattern is common in web apps and interfaces for Stable Diffusion or similar models. The global settings make it easy to maintain user preferences across multiple function calls.</p><h4>3. Conversation Memory in LLM Applications</h4><p>When building chatbots with LLMs, maintaining conversation history is crucial:</p><pre># Our global conversation state<br>message_history = []  # This will store the back-and-forth messages<br>conversation_context = &quot;You are a helpful Python assistant.&quot;  # System prompt<br><br>def add_user_message(message):<br>    # We need to modify the history list, so declare it global<br>    global message_history<br>    # Append adds to the list, which is a modification<br>    message_history.append({&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: message})<br>    <br>def add_assistant_message(message):<br>    global message_history  # Same idea - modifying the global list<br>    message_history.append({&quot;role&quot;: &quot;assistant&quot;, &quot;content&quot;: message})<br>    <br>def get_llm_response():<br>    # First we&#39;re just READING the globals to build the context<br>    # No &#39;global&#39; needed for reading!<br>    context = [{&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: conversation_context}]<br>    context.extend(message_history)<br>    <br>    # Get the AI&#39;s response<br>    response = llm.generate(context)<br>    <br>    # But now we want to UPDATE the history, so we need &#39;global&#39;<br>    global message_history<br>    message_history.append({&quot;role&quot;: &quot;assistant&quot;, &quot;content&quot;: response})<br>    <br>    return response</pre><p>You can use this pattern when prototyping conversational agents. The global history makes it easy to maintain context across turns, but in production systems, I’d recommend more structured approaches.</p><h3>Debugging Global Variables in Python</h3><h4>Common Global Variable Bugs and How to Fix Them</h4><ol><li><strong>Missing </strong><strong>global declaration</strong></li></ol><ul><li><strong>Symptom</strong>: UnboundLocalError: local variable &#39;x&#39; referenced before assignment</li><li><strong>Diagnosis</strong>: You’re trying to modify a global variable without declaring it</li><li><strong>Fix</strong>: Add global x at the beginning of your function</li></ul><pre># Bug<br>def update_counter():<br>    counter += 1  # UnboundLocalError<br><br># Fix<br>def update_counter():<br>    global counter<br>    counter += 1</pre><p><strong>2. Global variables not updating</strong></p><ul><li><strong>Symptom</strong>: Your function runs without errors, but the global value doesn’t change</li><li><strong>Diagnosis</strong>: You might be creating a local variable with the same name</li><li><strong>Fix</strong>: Ensure you’ve declared the variable as global</li></ul><pre># Bug (creates local variable instead of modifying global)<br>def reset_counter():<br>    counter = 0  # Creates local counter, doesn&#39;t affect global<br><br># Fix<br>def reset_counter():<br>    global counter<br>    counter = 0  # Now modifies the global</pre><p><strong>3. Name shadowing between modules</strong></p><ul><li><strong>Symptom</strong>: Updates to a variable aren’t reflected in other modules</li><li><strong>Diagnosis</strong>: You may have variables with the same name in different modules</li><li><strong>Fix</strong>: Use module-qualified names or import the module, not just its variables</li></ul><pre># config.py<br>counter = 0<br><br># bug.py<br>from config import counter<br>counter = 5  # Creates local copy, doesn&#39;t modify config.counter<br><br># fix.py<br>import config<br>config.counter = 5  # Properly updates the variable in config module</pre><p><strong>4. Global changes affecting unexpected parts of code</strong></p><ul><li><strong>Symptom</strong>: Making a change in one place breaks seemingly unrelated functionality</li><li><strong>Diagnosis</strong>: Multiple functions depend on the same global variable</li><li><strong>Fix</strong>: Refactor to use more localized state (classes, parameters)</li></ul><p><strong>5. Issues with multithreading</strong></p><ul><li><strong>Symptom</strong>: Race conditions, inconsistent values</li><li><strong>Diagnosis</strong>: Multiple threads accessing/modifying the same global</li><li><strong>Fix</strong>: Use thread-safe approaches like threading.Lock or avoid globals</li></ul><pre>import threading<br><br>counter = 0<br>counter_lock = threading.Lock()<br><br>def thread_safe_increment():<br>    global counter<br>    with counter_lock:<br>        counter += 1</pre><h3>Debugging Tips for AI Development</h3><p>In AI applications, global variables often store model parameters, metrics, or application state. Here are specific debugging strategies:</p><ol><li><strong>Use print statements strategically</strong></li></ol><pre>def train_step():<br>    global learning_rate, best_loss<br>    print(f&quot;Before update: lr={learning_rate}, best={best_loss}&quot;)<br>    # ... training code ...<br>    print(f&quot;After update: lr={learning_rate}, best={best_loss}&quot;)</pre><p>2. <strong>Add assertions to catch invalid states</strong></p><pre>def generate_image():<br>    # Verify our global settings are in valid ranges<br>    assert 0 &lt; generation_settings[&quot;guidance_scale&quot;] &lt; 20, &quot;Invalid guidance scale&quot;<br>    assert 10 &lt;= generation_settings[&quot;steps&quot;] &lt;= 150, &quot;Invalid step count&quot;<br>    # ... generation code ...</pre><p>3. <strong>Implement logging for global state changes</strong></p><pre>import logging<br>logging.basicConfig(level=logging.INFO)<br><br>def update_model_parameters(new_params):<br>    global model_parameters<br>    logging.info(f&quot;Updating parameters: {new_params}&quot;)<br>    model_parameters.update(new_params)<br>    logging.info(f&quot;New parameters state: {model_parameters}&quot;)</pre><p><strong>4. Create debug helper functions</strong></p><pre>def debug_globals():<br>    &quot;&quot;&quot;Print the state of all important global variables&quot;&quot;&quot;<br>    print(&quot;=== GLOBAL STATE ===&quot;)<br>    print(f&quot;Training iterations: {iteration_count}&quot;)<br>    print(f&quot;Best score: {best_score}&quot;)<br>    print(f&quot;Learning rate: {learning_rate}&quot;)<br>    print(f&quot;Model parameters: {model_parameters}&quot;)<br>    print(&quot;===================&quot;)</pre><p><strong>5. Convert problematic globals to a debug-friendly class</strong></p><pre>class AIState:<br>    def __init__(self):<br>        self.iteration_count = 0<br>        self.best_score = float(&#39;inf&#39;)<br>        <br>    def update_score(self, new_score):<br>        old_score = self.best_score<br>        if new_score &lt; self.best_score:<br>            self.best_score = new_score<br>            print(f&quot;Score improved: {old_score:.4f} → {new_score:.4f}&quot;)<br>        <br>    def __str__(self):<br>        return f&quot;AIState(iterations={self.iteration_count}, best_score={self.best_score:.4f})&quot;<br><br># Replace individual globals with an instance<br>state = AIState()</pre><h3>Global Variables Best Practices in Python</h3><h3>When to Use Global Variables</h3><p><strong>Appropriate uses:</strong></p><ul><li>Configuration constants</li><li>Application-wide settings</li><li>Read-only reference data</li></ul><p><strong>When to avoid:</strong></p><ul><li>Multi-threaded code (can lead to race conditions)</li><li>Complex state management</li><li>Code that will be imported by other modules</li></ul><p>I’ve learned the hard way that overusing globals can make code difficult to debug and test, especially in larger AI systems.</p><h3>Better Alternatives to Global Variables</h3><h4>1. Class-Based Approach</h4><pre>class ImageGenerator:<br>    def __init__(self):<br>        # Here we&#39;re using instance variables instead of globals<br>        # &#39;self&#39; refers to the specific instance of this class<br>        self.settings = {<br>            &quot;style&quot;: &quot;photorealistic&quot;,  # Default style<br>            &quot;steps&quot;: 50                 # Default steps<br>        }<br>        <br>    def update_settings(self, **kwargs):<br>        # No &#39;global&#39; needed! We&#39;re modifying the instance variable<br>        # &#39;**kwargs&#39; lets us pass in any number of keyword arguments<br>        self.settings.update(kwargs)  # Updates the dictionary with new values<br>        <br>    def generate(self, prompt):<br>        # Again, no &#39;global&#39; needed - we access our settings via &#39;self&#39;<br>        return self._call_model(prompt, self.settings)</pre><p>Classes provide encapsulation, making it clearer which functions depend on which data. This becomes increasingly important as your AI projects grow.</p><h4>2. Function Parameters Instead of Globals</h4><pre>def train_model(dataset, epochs, initial_fid=float(&#39;inf&#39;)):<br>    # Instead of using a global, we accept parameters and track locally<br>    best_fid = initial_fid  # Start with the provided initial value<br>    <br>    for epoch in range(epochs):<br>        # Training logic would go here<br>        current_fid = evaluate()  # Calculate current performance<br>        best_fid = min(best_fid, current_fid)  # Keep the better score<br>        <br>    # Return the result instead of relying on a global<br>    return best_fid  # This makes the function more self-contained</pre><p>Explicit parameters make functions more predictable and testable — important qualities when debugging complex AI systems.</p><h3>Python Global Variables Between Files</h3><p>When you need to share variables between multiple Python files in your project, there are several approaches:</p><ol><li><strong>Create a dedicated configuration module:</strong></li></ol><pre># config.py<br>MODEL_PATH = &quot;models/diffusion_v2.pt&quot;<br>BATCH_SIZE = 16<br>LEARNING_RATE = 0.0001</pre><p><strong>2. Import the config in other files:</strong></p><pre># train.py<br>from config import MODEL_PATH, BATCH_SIZE, LEARNING_RATE<br><br>def train():<br>    model = load_model(MODEL_PATH)<br>    # Use the imported constants<br>    optimizer = Adam(model.parameters(), lr=LEARNING_RATE)<br>    # ...</pre><p><strong>3. For modifiable globals between files:</strong></p><pre># state.py<br>training_progress = 0<br>finished_batches = 0<br><br># trainer.py<br>import state<br><br>def process_batch(batch):<br>    # Process the batch...<br>    <br>    # Update the shared state<br>    state.finished_batches += 1<br>    state.training_progress = state.finished_batches / total_batches</pre><h3>Summary</h3><ul><li>Reading global variables requires no special syntax in Python</li><li>Modifying global variables requires the global keyword</li><li>The global declaration must come before any assignments</li><li>Global variables can be accessed across functions, classes and modules</li><li>In AI development, globals are useful for tracking metrics and settings</li><li>For larger projects, consider using classes or function parameters instead of globals</li></ul><p>Remember that while globals can be convenient for quick prototypes and scripts, as your generative AI projects grow more complex, more structured approaches to state management will save you debugging time in the long run.</p><h3>Follow the ABCs of GenAI publication for more focused posts. Thank you</h3><p>You may also like:</p><p><a href="https://medium.com/the-abcs-of-ai/handling-multi-line-strings-in-python-a-primer-for-deep-learning-enthusiasts-22be0aba7973">Handling Multi-Line Strings in Python: A Primer for Deep Learning Enthusiasts</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=970addb6d03c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/python-global-variables-complete-guide-for-functions-classes-and-modules-in-ai-development-970addb6d03c">Python Global Variables: Complete Guide for Functions, Classes, and Modules in AI Development</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ChatGPT Agent: Automate Full Workflows Without Losing Control]]></title>
            <link>https://medium.com/the-abcs-of-ai/chatgpt-agent-automate-full-workflows-without-losing-control-46d680bce10e?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/46d680bce10e</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[chatgpt-agent]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Wed, 23 Jul 2025 17:23:01 GMT</pubDate>
            <atom:updated>2025-07-23T17:23:01.687Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uEJj-7Agz2S332FuwN4jlQ.png" /></figure><p>OpenAI just plugged a virtual laptop into ChatGPT. <a href="https://openai.com/index/introducing-chatgpt-agent/">ChatGPT Agent</a> clicks through websites, runs terminal commands, and spits out editable slide decks — tackling prompts like “analyze three competitors and create slides” end‑to‑end. Crucially, it stops for your approval before any risky move, so you stay in the driver’s seat.</p><h3><strong>TL;DR — ChatGPT Agent in 90 seconds</strong></h3><ul><li><strong>What launched?</strong> A unified Operator + deep research agent that uses a virtual computer to finish multi‑step tasks end‑to‑end.</li><li><strong>Toolbox inside one chat:</strong> visual &amp; text browsers, Unix‑style terminal, API calls, optional Gmail / GitHub connectors.</li><li><strong>Human‑first control:</strong> asks before risky actions; you can pause, take over, or get progress pings any time.</li><li><strong>State‑of‑the‑art results:</strong> tops Humanity’s Last Exam (41.6 → 44.4 w/ parallel), DSBench, SpreadsheetBench.</li><li><strong>Who gets it now?</strong> Pro (400 msgs/mo today), Plus &amp; Team (40 msgs) rolling out; EEA (European Economic Area) “coming soon.”</li><li><strong>Why care?</strong> Automate drudgery, but beware of prompt injection and beta quirks — keep a review loop in place before shipping outputs.</li></ul><figure><img alt="Infographic: three ChatGPT Agent benefits — End‑to‑End Automation, Human‑in‑Loop Control, Benchmark‑Level Accuracy" src="https://cdn-images-1.medium.com/max/1024/0*3tU85tJG7e70vqMK" /><figcaption>Three super powers infographic made by author.</figcaption></figure><h3>Benefits</h3><h4>End-to-End Automation</h4><p>ChatGPT Agent is more than a chatbot upgrade; it’s a full-stack assistant that owns the entire workflow for you. Ask it to “analyze three competitors and create a slide deck.” It will hop between a visual browser, text browser, terminal, and direct API calls — scraping sites, running Python for data wrangling, then exporting editable PowerPoint slides. Because each tool lives inside a persistent “virtual computer,” the context of earlier clicks and code carries forward seamlessly, letting the model fluid-switch between reasoning and action without you writing any glue code. For developers, that means routine research-plus-build chores collapse from an afternoon of tab-juggling into a single prompt — and the same toolbox is ready for cron-like tasks via connectors to Gmail, GitHub, and more.</p><h4>You Stay the Pilot</h4><p>Power without guardrails breeds anxiety, so the agent ships with human-in-the-loop controls by default. Before it makes a high-stakes move — checking out an online cart, sending an email, or pushing code — it pauses to ask for explicit approval. You can intervene at any moment: take over the live browser session, request a progress snapshot, pause a runaway task, or abort mid-stream and still retrieve partial results. For critical steps like “send this to my mailing list,” Watch Mode forces you to click “Send” yourself, blocking silent misfires.</p><p>These design choices mean you gain speed while retaining ultimate authority — crucial for developers who must safeguard credentials, client data, and production repos.</p><h4>Benchmark Beast</h4><p>The flashy toolbox is backed by hard numbers: the model powering ChatGPT Agent sets a new state-of-the-art 41.6 pass@1 on Humanity’s Last Exam, climbing to 44.4 with simple parallel runs. It reaches 27.4 % on FrontierMath — problems that stump many postgraduate mathematicians — and beats previous models on real-world simulations like preparing finance models or competitive analyses. On DSBench, which grades data-science pipelines end-to-end, the agent surpasses top human baselines; on SpreadsheetBench, it nearly doubles Copilot-in-Excel (45.5 % vs 20 %). These wins translate into fewer hallucinations and higher-quality deliverables, giving developers the confidence that delegated tasks won’t come back as cleanup fire drills.</p><h3>Real‑world use‑cases</h3><p><strong>1 | Hands‑free slide &amp; spreadsheet prep</strong><br>Picture the weekly grind of turning product analytics dashboards into board‑ready decks. In Agent Mode, you paste a screenshot URL, say “convert this into an editable slide deck and roll last‑week’s CSV into the revenue model,” and the agent scrapes the image, pulls the latest numbers, runs code in its terminal, and drops a polished .pptx plus an updated spreadsheet — no copy‑pasting, no formula surgery.</p><p><strong>2 | Connector‑powered stand‑up digests</strong><br>Hook up Gmail and GitHub, then ask: “Summarize yesterday’s PR comments and flag blockers for the 9 AM stand‑up.” The agent fetches your repo activity, cross‑checks calendar invites, and returns a five‑bullet status you can paste straight into Slack — while still pausing for approval before any email goes out.</p><p><strong>3 | Personal concierge tasks</strong><br>Outside work, the same toolbox plans and books a full Japanese breakfast for four, picks restaurants that deliver the right ingredients, and schedules delivery windows that don’t collide with meetings. Or it strings together flights, hotels, and museum tickets into a shareable itinerary — saving you the multi‑tab hassle.</p><p><strong>4 | Set‑and‑forget automation</strong><br>Because the agent remembers tool state, you can tell it: “Every Monday at 6 AM, run last week’s metrics query, refresh the spreadsheet, and email me the deck.” Flip the schedule toggle once, and it re‑runs the whole chain weekly, pinging your phone when the new report is ready.</p><h3>Risk and Limitations</h3><figure><img alt="Infographic: three ChatGPT Agent cautions — Prompt‑Injection Traps, Beta Rough Edges, Access &amp; Cost Limits" src="https://cdn-images-1.medium.com/max/1024/0*SI4yVDTX7VH4pUk6" /><figcaption>Three caution flags made by the author.</figcaption></figure><p><strong>Prompt‑Injection Traps</strong><br>Because the agent can browse, click, and even log in to sites on your behalf, malicious pages gain a new attack surface: hidden meta‑tags or invisible elements can smuggle instructions that hijack the workflow, exfiltrate data from your Gmail connector, or trick the agent into hazardous actions. OpenAI hardened the model against these exploits — training it to spot suspicious content, monitoring for tell‑tale patterns, and always pausing for your explicit OK before any consequential step — but it still urges users to</p><blockquote>disable connectors they don’t need and keep a watchful eye on unfamiliar domains.</blockquote><p><strong>Beta Rough Edges</strong><br> Slideshows are powerful but still rough‑cut. Formatting can feel “rudimentary,” and exported PowerPoint files sometimes diverge from what you preview in the browser. Spreadsheet uploads work, but you can’t yet feed an existing deck back for polishing. OpenAI is already training the next iteration, but for now,</p><blockquote>they recommend a human QA pass before you ship decks to clients or executives.</blockquote><p><strong>Access &amp; Cost Limits</strong><br> The agent rolls out first to paying subscribers — Pro gets 400 messages per month; Plus and Team plans get 40, with flexible credits for overages. Enterprise and Education tiers arrive “in the coming weeks,” while the European Economic Area and Switzerland must wait for regulatory clearance. Factor those caps and regional gaps into any rollout plan.</p><h3>How to try it today</h3><ol><li><strong>Toggle Agent Mode</strong><br> In any ChatGPT conversation, open the <strong>Tools</strong> dropdown and select <strong>“agent mode.”</strong> A heads‑up narration bar will appear once the agent starts working.</li><li><strong>Give it a concrete task &amp; watch the run</strong><br> Describe what you want — “conduct deep research on X,” “create a slideshow,” or “submit my expenses.” You’ll see each click and command, and you can grab control of the browser at any moment.</li><li><strong>(Optional) Connect Gmail, GitHub, etc.</strong><br>Head to <strong>Settings → Connectors</strong> and authenticate the apps you’d like the agent to query. Once linked, it can do things like summarise your inbox or scan PR comments — but it still pauses for a manual login before acting on those sites.</li><li><strong>Schedule a recurring run</strong><br>After a job finishes, type something like “repeat this every Monday at 6 AM.” The agent will re‑run the entire chain automatically — ideal for weekly metrics decks or inbox digests.</li><li><strong>Keep “Watch Mode” on for sensitive actions</strong><br> Tasks that could have a real‑world impact (sending an email, making a purchase) require an explicit <strong>Send / Approve</strong> click from you. The agent also asks for confirmation before any irreversible move.</li><li><strong>Pause, resume, or request a progress summary</strong><br>If the run stalls — or you just need a status check — use the pause button or type “progress?” to get a quick snapshot, then let it continue or stop and harvest partial results.</li><li><strong>Tidy up your browsing data</strong><br>One click in <strong>Settings → Privacy</strong> wipes all agent cookies and logs you out of active site sessions; handy when switching projects or sharing a machine.</li><li><strong>Know your plan limits</strong><br>Roll‑out is live for paid tiers only: <strong>Pro = 400 msgs/mo; Plus &amp; Team = 40.</strong> Enterprise/Education arrive “in the coming weeks,” and the EEA + Switzerland are still pending approval.</li></ol><h3>Competitive landscape &amp; what’s next</h3><p>Where does ChatGPT Agent sit in the fast‑forming “agentic stack,” and how might things evolve over the next few quarters?</p><h4>- Microsoft Copilot Studio — multi‑agent orchestration inside M365</h4><p>At Build 2025, Microsoft unveiled <strong>Copilot Studio’s “multi‑agent orchestration” preview</strong>: makers can wire several Copilot agents together (including ones built on Azure AI Agents Service) so each hands sub‑tasks to the next — e.g., sales data → Word proposal → Outlook follow‑ups — all in a single canvas. Early testers can even let these agents control desktop apps through a new “computer use” mode. <a href="https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/multi-agent-orchestration-maker-controls-and-more-microsoft-copilot-studio-announcements-at-microsoft-build-2025/">Microsoft</a></p><p><strong>Take‑away:</strong> Studio emphasises enterprise plumbing — tying agents to the Microsoft 365 stack and Azure Foundry models — whereas ChatGPT Agent focuses on an all‑purpose virtual computer that any end user can drive.</p><h4>- Anthropic Claude — context‑rich connectors via MCP</h4><p>Just days ago, Anthropic rolled out a <strong>tool directory that plugs Claude directly into Notion, Canva, Stripe, Figma, and more</strong>. Thanks to its Model Context Protocol (MCP), Claude can pull live project data or payment records without the copy‑paste ritual that usually derails chatbots. Paid users simply authorise each app once, then issue natural‑language commands. <a href="https://www.techradar.com/computing/artificial-intelligence/you-dont-have-to-explain-everything-to-claude-anymore-its-finally-in-your-apps">TechRadar</a></p><p><strong>Take‑away:</strong> Claude is betting on secure, fine‑grained app permissions; ChatGPT Agent instead gives you a full browser/terminal/API sandbox but requires you to approve risky steps manually.</p><h4>Google Cloud — Agent2Agent protocol + AI Agent Marketplace</h4><p>At Google Cloud Next, the company staked its claim on <strong>multi‑vendor, cross‑framework agents</strong>. A new <strong>Agent2Agent protocol</strong> lets AI agents from Atlassian, Box, SAP, ServiceNow, and others talk to one another, while an <strong>AI Agent Marketplace</strong> inside Google Cloud Marketplace showcases partner‑built agents ready to buy and deploy. <a href="https://www.channelfutures.com/cloud/google-cloud-next-spotlights-agentic-ai">Channel Futures</a></p><p><strong>Take‑away:</strong> Google is pursuing ecosystem scale — standardising how third‑party agents interoperate — while OpenAI is perfecting a unified agent that can already research, click, and code on its own.</p><h3>What’s next for ChatGPT Agent</h3><ul><li><strong>Operator sunset &amp; feature parity:</strong> the old Operator preview site disappears “in a few weeks,” with deep‑research mode now a drop‑down option inside ChatGPT.</li><li><strong>Enterprise/Education rollout:</strong> after Pro/Plus/Team, larger org tiers arrive “in the coming weeks,” alongside regional expansion to the EEA and Switzerland.</li><li><strong>Richer slide decks:</strong> OpenAI is already training the next iteration to close formatting gaps and accept existing decks as templates.</li><li><strong>Reduced oversight over time:</strong> expect a gradual dial‑back of mandatory confirmations as safety systems mature, making agent runs feel more seamless.</li></ul><p>ChatGPT Agent currently leads on single‑chat breadth — research, code, file export — while rivals race to specialise in enterprise orchestration (Microsoft), data‑native workflows (Anthropic), or open multi‑agent ecosystems (Google).</p><p>For developers deciding where to invest, the near‑term calculus is tooling fit: if you live in VS Code and Gmail, ChatGPT Agent’s virtual laptop may already feel like home; if your org standardises on M365 or Google Cloud, keep an eye on those multi‑agent previews.</p><h3><strong>Conclusion</strong></h3><p>ChatGPT Agent turns one prompt into a full, end‑to‑end workflow — yet still keeps you firmly in control. Spin it up on a safe task, watch how it handles research, code, and slides, then decide where it belongs in your stack.</p><h3><strong>Follow the ABCsofGenAI for more focused articles. Thank you.</strong></h3><p>#ChatGPT‑Agent #AI‑Workflows #DeveloperTool #Productivity #OpenAI</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46d680bce10e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/chatgpt-agent-automate-full-workflows-without-losing-control-46d680bce10e">ChatGPT Agent: Automate Full Workflows Without Losing Control</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Google I/O 2025: A Technical Deep Dive Into the AI Platform That Changes Everything]]></title>
            <link>https://medium.com/the-abcs-of-ai/google-i-o-2025-a-technical-deep-dive-into-the-ai-platform-that-changes-everything-d1356a2c6dd7?source=rss-a8170c8c5dae------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1356a2c6dd7</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[free]]></category>
            <category><![CDATA[flow]]></category>
            <dc:creator><![CDATA[Dr. Tânia Frazão, M.C.S., D.V.M.]]></dc:creator>
            <pubDate>Sat, 24 May 2025 11:49:34 GMT</pubDate>
            <atom:updated>2025-05-24T11:49:34.707Z</atom:updated>
            <content:encoded><![CDATA[<h4>Breaking down the revolutionary announcements from Google’s developer conference — with insights from Josh Woodward and Tulsee Doshi and how to experiment all these new features</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YwE7WLG2hKY90W2cMDlNOw.png" /></figure><p><em>Based on: </em><a href="https://www.youtube.com/watch?v=DPQF_6o4G_E">Release Notes: Google I/O 2025 Recap with Josh Woodward and Tulsee Doshi</a></p><h3>TLDR</h3><ul><li><strong>VEO 3</strong>: Video generation now includes synchronized audio and dialogue</li><li><strong>Native Audio</strong>: Gemini speaks naturally in 70+ languages with emotion</li><li><strong>Jules</strong>: Autonomous coding agent handles bugs, tests, and features</li><li><strong>Flow</strong>: Professional filmmaking platform built with Hollywood input</li><li><strong>New Models</strong>: Deep Think for reasoning, Diffusion for speed, updated Flash/Pro</li><li><strong>General Availability in June 2025</strong>: Stable models with 1-year support commitment</li><li><strong>AI Ultra Plan</strong>: New tier with all premium features (50% off for 3 months)</li></ul><h3>What’s Inside this article</h3><ul><li>The Big Picture: Gemini Becomes the Platform</li><li>VEO 3: When Video Generation Learned to Speak</li><li>Flow: Professional Tools Built by Professionals</li><li>Native Audio: The End of Robotic AI Voices</li><li>The Gemini Model Portfolio</li><li>Jules: Your Autonomous Development Partner</li><li>Democratizing Development: Stitch and AI Studio</li><li>Implementation Strategy</li><li>The 2030 Vision</li><li>Why This Matters: The Inflection Point</li></ul><p>Google I/O 2025 wasn’t just another developer conference. It was the moment Google transformed from an AI experimenter into the architect of a comprehensive AI platform that touches every aspect of computing.</p><p>In an exclusive Release Notes session, Josh Woodward and Tulsee Doshi revealed the technical details behind announcements that had the crowd cheering and developers scrambling to update their roadmaps. Here’s what you need to know.</p><h3>The Big Picture: Gemini Becomes the Platform</h3><p>Let’s start with the most fundamental shift:</p><blockquote><strong>“Gemini is no longer just an AI model. It’s now the connective tissue of Google’s entire ecosystem.”</strong></blockquote><p>Think about what this means:</p><ul><li>Gemini in Android Studio for code generation</li><li>Gemini in Chrome for web interactions</li><li>Gemini in development tools across the stack</li><li>Gemini in enterprise solutions</li></ul><p>As Doshi explained:</p><blockquote>“We’re not just telling a story about Gemini as a model being awesome. We’re telling a story about Gemini in our products and actually landing with real users in real ways.”</blockquote><p>This isn’t incremental improvement. This is architectural transformation.</p><h3>VEO 3: When Video Generation Learned to Speak</h3><h4>The Technical Breakthrough</h4><p>VEO 3 doesn’t just generate video — it creates complete audiovisual experiences. Here’s what makes it revolutionary:</p><p><strong>Input</strong>: Natural language description<br><strong>Output</strong>:</p><p>Video with</p><ul><li>Synchronized dialogue</li><li>Contextual sound effects</li><li>Appropriate background audio</li><li>Coherent narrative flow</li></ul><p>Josh Woodward’s reaction perfectly captured it:</p><blockquote>“You can describe something in natural language, and it just comes to life.”</blockquote><h4>How It Works</h4><p>Building on VEO 2’s foundation from December 2024, VEO 3 integrates multimodal generation at the model level. This means audio and video aren’t generated separately and stitched together — they’re created as a unified output.</p><h4>How to Access</h4><p>Access Veo3 through Vertex AI on Google Cloud or through the new AI Ultra Google plan, see below for more details.</p><h3>Flow: Professional Tools Built by Professionals</h3><p>Google didn’t just release a model and hope for the best. In 100 days, they built <strong>Flow</strong> — a complete video creation platform designed with actual filmmakers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mweI_s98HED4avEf6-qrVA.gif" /><figcaption>Flow demonstration video by Google.</figcaption></figure><h4>What Makes Flow Different</h4><p><strong>Co-creation Process</strong>:</p><ul><li>Hollywood professionals provided input</li><li>Emerging AI filmmakers tested workflows</li><li>Every feature was validated by real users</li></ul><p><strong>Key Capabilities</strong>:</p><ul><li>Intelligent clip trimming that understands narrative</li><li>Scene extension that maintains coherence</li><li>Professional-grade output quality</li></ul><h4>How to Access Flow</h4><h4>- Google AI Ultra</h4><p>Flow is available via Google’s new premium tier called Ultra<strong>.</strong></p><p>Right now it’s 50% off for three months, and you also get:</p><ul><li>Access to VEO 3</li><li>The highest rate limits (which means you can use it more)</li><li>YouTube Premium</li><li>30TB of storage (that’s a LOT)</li><li>All the other premium AI stuff, you can check on <a href="https://blog.google/products/google-one/google-ai-ultra/">https://blog.google/products/google-one/google-ai-ultra/</a></li></ul><p><em>Note: Demand is so high that Google is actively managing compute capacity.</em></p><h4><strong>- Google AI Pro</strong></h4><p>However, if the Ultra price is inconvenient you can also access it through <strong>Google AI Pro</strong> subscription also having significant new features.</p><ul><li>Flow Access: Full AI filmmaking capabilities powered by the Veo 2 model</li><li>Gemini in Chrome: Early access to browser-integrated AI features</li><li>Same Price: All new benefits included at no additional cost</li></ul><p><strong>Availability Timeline</strong></p><ul><li><strong>Phase 1</strong>: United States (available now)</li><li><strong>Phase 2</strong>: International rollout (coming soon)</li></ul><p>This positions Google AI Pro as the middle tier between the basic free tier and the new Ultra plan, offering substantial creative tools and early feature access for existing subscribers.</p><p><em>Note: If you are outside US and you want to experiment, you can do it with a VPN. I advise you to create a new gmail account while using the VPN and then subscribe the free trial for Google AI Pro.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1G93eHP6fTVcu8cx9r853g.png" /><figcaption>Differences between different plans Pro and Ultra.</figcaption></figure><h3>Native Audio: The End of Robotic AI Voices</h3><p>Here’s where things get technically fascinating. <strong>Native Audio represents a fundamental shift in how AI generates speech.</strong></p><h4>Traditional Approach vs. Native Audio</h4><p><strong>Old Way</strong>: Generate text → Convert to speech<br><strong>Native Audio</strong>: Generate speech directly</p><p>This isn’t just a performance optimization. It enables:</p><p>✓ <strong>Natural emotion and inflection</strong><br>✓ <strong>70+ languages at launch</strong><br>✓ <strong>Seamless language switching</strong><br>✓ <strong>Contextual awareness of conversations</strong></p><p>As Doshi noted:</p><blockquote>“The model can actually tell when we’re speaking to each other versus when we’re speaking to the model. It knows when to respond, when to interrupt, when to hang back.”</blockquote><h4>How to access</h4><p>With some limitations regarding the number of generations, you can check Google AI Studio. Although it is a playground for developers to test Google AI services, you can use it for individual tasks. And the price is free.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I1_I2f3o9u4ro4fCri_p7w.png" /><figcaption>Google AI Studio shows where to find the native speech generation. You can also find the live audio on the side.</figcaption></figure><h3>The Gemini Model Portfolio: Right Tool for Every Job</h3><p>Google announced three distinct models, each optimized for specific use cases:</p><h4>1. Gemini 2.5 Pro Deep Think</h4><p><strong>Purpose</strong>: Complex reasoning and analysis</p><p><strong>How it works</strong>: Explores multiple solution paths before synthesizing an answer</p><p><strong>Key features</strong>:</p><ul><li>State-of-the-art performance on coding and math</li><li>“Thinking budgets” for cost control</li><li>Multimodal reasoning capabilities</li></ul><p><strong>Best for</strong>: Complex problem-solving, code architecture, mathematical proofs</p><h4>2. Gemini Diffusion</h4><p><strong>Purpose</strong>: Ultra-fast responses</p><p><strong>How it works</strong>: Non-sequential generation with iterative refinement</p><p><strong>Performance</strong>: “So fast that if you blinked, you missed it”</p><p><strong>Best for</strong>: Real-time applications, interactive experiences</p><h4>3. Gemini 2.5 Flash &amp; Pro Updates</h4><p><strong>Purpose</strong>: Production workloads</p><p><strong>What’s new</strong>:</p><ul><li>Significantly improved coding performance</li><li>Better multimodal handling</li><li>1-year stability commitment</li></ul><h4><strong>How to access these models</strong></h4><p>General release in <strong>June 2025.</strong></p><p><em>Note: After release, check the AI Google Studio for updates. At the moment of the writing of this article, Pro has been updated to include code improvements, but I couldn’t find the thinking mode for it. Flash on Google AI Studio has a thinking mode available.</em></p><p>Check all the models available: <a href="https://ai.google.dev/gemini-api/docs/models">https://ai.google.dev/gemini-api/docs/models</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PPT3f3dio6VsBqC_Q3Qh9Q.png" /><figcaption>Screenshot of the web page models documentation.</figcaption></figure><h3>Jules: Your Autonomous Development Partner</h3><p>Jules represents a paradigm shift in development tools. Starting as a Halloween 2024 experiment, it’s now a fully autonomous coding agent.</p><h4>Core Capabilities</h4><p><strong>What Jules Does</strong>:</p><ol><li><strong>Bug Fixing</strong>: Identifies and resolves issues independently</li><li><strong>Test Generation</strong>: Creates comprehensive test suites</li><li><strong>Documentation</strong>: Writes context-aware docs</li><li><strong>Feature Development</strong>: Implements new functionality</li><li><strong>Code Cast</strong>: Generates audio summaries of changes</li></ol><h4>How to Use Jules</h4><pre>1. Connect your GitHub repository<br>2. Assign tasks in natural language<br>3. Review generated pull requests<br>4. Approve, modify, or reject changes</pre><p><strong>Current Access</strong>: Open beta at jules.google.com (5 tasks/day limit).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UD98rxZD5WkP8ArhmCfzyg.png" /><figcaption>Jules UI screenshot.</figcaption></figure><h3>Democratizing Development: Stitch and AI Studio</h3><h3>Stitch: Design-First Development</h3><p><strong>The Concept</strong>: Describe an interface in plain English, get production-ready design files.</p><p><strong>What’s Revolutionary</strong>: Output includes proper markup, not just mockups.</p><p><strong>Target Users</strong>: Product managers, designers, and anyone with ideas but limited technical skills.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fq39uczIW_MY%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dq39uczIW_MY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fq39uczIW_MY%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/62e248ee7c5ab1fbb22041c1b52b84d2/href">https://medium.com/media/62e248ee7c5ab1fbb22041c1b52b84d2/href</a></iframe><h4>How to access</h4><p>Through Google Labs.</p><p><a href="https://stitch.withgoogle.com/">https://stitch.withgoogle.com/</a></p><h3>AI Studio: Now with Native Code Editing</h3><p>The platform now supports in-line code generation and editing. In a meta-moment, users discovered they could screenshot AI Studio and use it to generate new features for AI Studio itself.</p><h3>Implementation Strategy: What You Should Do Now</h3><h4>For Individual Developers</h4><ol><li><strong>Start experimenting</strong> with the current models</li><li><strong>Plan for June 2025</strong> GA (General Availability) releases</li><li><strong>Learn to use thinking budgets</strong> for cost optimization</li><li><strong>Integrate Jules</strong> into your workflow for routine tasks</li></ol><h4>For Organizations</h4><ol><li><strong>Pilot the AI Ultra plan</strong> to evaluate capabilities</li><li><strong>Identify use cases</strong> for each model variant</li><li><strong>Prepare infrastructure</strong> for increased AI workloads</li><li><strong>Train teams</strong> on AI-assisted development</li></ol><h4>For Content Creators</h4><ol><li><strong>Explore Flow</strong> for video projects</li><li><strong>Experiment with VEO 3</strong> for content generation</li><li><strong>Focus on creative direction</strong> over technical execution</li></ol><h3>The 2030 Vision: Where This Is Heading</h3><p>The session participants painted a compelling picture of the future:</p><ul><li><strong>Scientific breakthroughs</strong> enabled by non-specialists using AI</li><li><strong>Ubiquitous AI assistance</strong> integrated into daily life</li><li><strong>New content paradigms</strong> we can’t yet imagine</li><li><strong>Democratized access</strong> to capabilities currently reserved for experts</li></ul><blockquote>“By 2030, so much of what Demis presented on this vision of the universal virtual assistant is going to be actually realized materially,” Doshi predicted.</blockquote><h3>Why This Matters: The Inflection Point</h3><p>Several factors indicate we’re at a critical juncture:</p><ol><li><strong>Cursor chose Gemini</strong> for their sophisticated developer base</li><li><strong>Model improvements are accelerating,</strong> not plateauing</li><li><strong>Google is meeting developers where they are</strong>, not forcing proprietary tools</li><li><strong>The ecosystem has reached critical mass</strong></li></ol><h3>The Bottom Line</h3><p>Google I/O 2025 represents a fundamental shift in how we’ll build and interact with technology. These aren’t experimental features anymore — they’re production-ready tools that will reshape entire industries.</p><p>The question isn’t whether to adopt these technologies, but how quickly you can integrate them into your workflow. The companies and developers who move first will have significant advantages as these capabilities become table stakes.</p><h3>Final Thought</h3><p>As the session wrapped up, Logan presented commemorative TPU v4 chips to the speakers — a humorous acknowledgment of the massive compute demands these services create. The joke about needing to “put them back in circulation” underscores an important reality: we’re pushing the boundaries of what’s computationally possible, and Google is committed to making it accessible.</p><p>The future isn’t coming. It’s here. And it’s running on Gemini.</p><p><em>What aspects of these announcements will most impact your work? Share your thoughts in the comments or connect with me to discuss.</em></p><h4>Don’t forget to subscribe to ABCsofGenAI!</h4><h4>Thank you.</h4><p><em>Source: </em><a href="https://www.youtube.com/watch?v=DPQF_6o4G_E">Release Notes: Google I/O 2025 Recap with Josh Woodward and Tulsee Doshi</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1356a2c6dd7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-abcs-of-ai/google-i-o-2025-a-technical-deep-dive-into-the-ai-platform-that-changes-everything-d1356a2c6dd7">Google I/O 2025: A Technical Deep Dive Into the AI Platform That Changes Everything</a> was originally published in <a href="https://medium.com/the-abcs-of-ai">The ABCs of GenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>