<?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 Yepher on Medium]]></title>
        <description><![CDATA[Stories by Yepher on Medium]]></description>
        <link>https://medium.com/@yepher?source=rss-6e80bd29e24------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*T0eIzergSPYtDm7ZtWlaMA.jpeg</url>
            <title>Stories by Yepher on Medium</title>
            <link>https://medium.com/@yepher?source=rss-6e80bd29e24------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 14:14:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yepher/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[Revolutionizing Speech End-Point Detection: A Data Science Perspective]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">In a world where voice assistants have become integral to everyday life&#x200A;&#x2014;&#x200A;powering smart homes, wearables, and augmented reality devices &#x2014;&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/revolutionizing-speech-end-point-detection-a-data-science-perspective-e3ddf8f6c483?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/revolutionizing-speech-end-point-detection-a-data-science-perspective-e3ddf8f6c483?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3ddf8f6c483</guid>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Thu, 09 Jan 2025 03:07:26 GMT</pubDate>
            <atom:updated>2025-01-09T03:07:26.755Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Limitations of The Linelings]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@yepher/the-limitations-of-the-linelings-12b61db6e903?source=rss-6e80bd29e24------2"><img src="https://cdn-images-1.medium.com/max/1024/1*OChY18ALAzT2R9zwgK71Mg.png" width="1024"></a></p><p class="medium-feed-snippet">When I hear folks say that an AI will never be able to reason. I think about how I struggle to think beyond the three dimensions of my&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/the-limitations-of-the-linelings-12b61db6e903?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/the-limitations-of-the-linelings-12b61db6e903?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/12b61db6e903</guid>
            <category><![CDATA[dimensional-intelligence]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 19:49:47 GMT</pubDate>
            <atom:updated>2025-01-07T19:49:47.338Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Dark Side Of Data]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@yepher/the-dark-side-of-data-e003e631ad3d?source=rss-6e80bd29e24------2"><img src="https://cdn-images-1.medium.com/max/1100/1*wSZifgtBGwCUt6gjRwNIig.png" width="1100"></a></p><p class="medium-feed-snippet">Update 2025&#x2013;01&#x2013;14: Heard a firware udpate has been pushed to update this issue.</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/the-dark-side-of-data-e003e631ad3d?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/the-dark-side-of-data-e003e631ad3d?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/e003e631ad3d</guid>
            <category><![CDATA[automaticlicenseplate]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 05:32:37 GMT</pubDate>
            <atom:updated>2025-01-14T18:29:46.269Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Pinpoint Where a Photo Was Taken Using Only the Time it was Taken]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@yepher/pinpoint-where-a-photo-was-taken-using-only-the-time-and-shadows-3c563ec5a730?source=rss-6e80bd29e24------2"><img src="https://cdn-images-1.medium.com/max/944/1*odjWyz2N3xdxZBSW7X5nMA.png" width="944"></a></p><p class="medium-feed-snippet">Have you ever looked at a photo and wondered where it was taken? Could something as simple as a shadow provide clues to pinpoint a&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/pinpoint-where-a-photo-was-taken-using-only-the-time-and-shadows-3c563ec5a730?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/pinpoint-where-a-photo-was-taken-using-only-the-time-and-shadows-3c563ec5a730?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c563ec5a730</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[geolocation]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Sun, 05 Jan 2025 18:54:47 GMT</pubDate>
            <atom:updated>2025-01-05T20:14:31.414Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Reimagining Operating Systems with AI: A Vision for the AIOS-Agent Ecosystem]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@yepher/reimagining-operating-systems-with-ai-a-vision-for-the-aios-agent-ecosystem-0cd41f6d6b57?source=rss-6e80bd29e24------2"><img src="https://cdn-images-1.medium.com/max/1538/1*iIPQn4vQwG1H6vEw9tXJYw.png" width="1538"></a></p><p class="medium-feed-snippet">In a groundbreaking paper, researchers from Rutgers University present a visionary concept: transforming Large Language Models (LLMs) into&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/reimagining-operating-systems-with-ai-a-vision-for-the-aios-agent-ecosystem-0cd41f6d6b57?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/reimagining-operating-systems-with-ai-a-vision-for-the-aios-agent-ecosystem-0cd41f6d6b57?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/0cd41f6d6b57</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[oss-llm]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Wed, 05 Jun 2024 14:24:06 GMT</pubDate>
            <atom:updated>2024-06-05T14:24:06.751Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Current State of AI: A Comprehensive Overview — February 2024]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@yepher/current-state-of-ai-a-comprehensive-overview-february-2024-a55118345108?source=rss-6e80bd29e24------2"><img src="https://cdn-images-1.medium.com/max/600/1*SMWtzJ6mD0Dl_0rxmkqeQA.gif" width="600"></a></p><p class="medium-feed-snippet">As we navigate through February 2024, the artificial intelligence landscape continues to evolve at an unprecedented pace, marked by&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@yepher/current-state-of-ai-a-comprehensive-overview-february-2024-a55118345108?source=rss-6e80bd29e24------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@yepher/current-state-of-ai-a-comprehensive-overview-february-2024-a55118345108?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/a55118345108</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[sora]]></category>
            <category><![CDATA[groq]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Thu, 22 Feb 2024 22:13:48 GMT</pubDate>
            <atom:updated>2024-02-22T22:23:46.930Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[A Quick Overview of Large Language Models (LLM)]]></title>
            <link>https://medium.com/@yepher/a-quick-overview-of-large-language-models-llm-9118bf256c7a?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/9118bf256c7a</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[llama-2]]></category>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Wed, 29 Nov 2023 00:06:40 GMT</pubDate>
            <atom:updated>2023-11-29T00:06:40.944Z</atom:updated>
            <content:encoded><![CDATA[<p>Our previous article was a quick look at some <a href="https://medium.com/@towardsdatascience_19372/ive-been-playing-with-many-ai-building-blocks-for-several-months-b840e187efff">AI Building Blocks</a>; I mostly brushed over the core of those tools, the Large Language Models (LLM).</p><p>Large language models (LLMs) like <a href="https://chat.openai.com/">ChatGPT</a> and <a href="https://claude.ai/chat">Claude</a> are becoming increasingly popular. But what exactly are they, and how do they work? In this post, I’ll provide a quick overview of LLMs. I’ve added links to more detailed resources in case you want to dive deeper and see the inner workings of an LLM.</p><p>We will cover:</p><ul><li>What is an LLM</li><li>How parameter files are used and where they come from</li><li>How LLMs Actually Work</li><li>The Future of LLMs</li></ul><p><em>If you like this sort of high-level technology overview hold the clap button down at the end to let me know to write more like this</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RL5gldECFvrjNoqR" /></figure><p><strong>What is an LLM?</strong></p><p>At its core, an LLM is just two files on your computer:</p><ol><li>A <a href="https://ollama.ai/library">parameters file</a> containing the “weights” or parameters of a neural network. This file stores all the knowledge learned during training and can be over 100GB.</li><li>A run file with <a href="https://github.com/karpathy/llama2.c/blob/master/run.c">code</a> to run the neural network using those parameters. This can be as simple as 500 lines of C code, or really written in any programming language.</li></ol><p>That’s it! With just those two files, you have a fully self-contained LLM that can generate text. For example, the <a href="https://ai.meta.com/llama/">Llama2</a> LLM could take a prompt like “Write a poem about climate change” and start generating relevant text.</p><p>— — — — — — — — -</p><p><em>You can try it on your own machine using </em><a href="https://ollama.ai/"><em>Ollama</em></a><em>. No ChatGPT, OpenAI or Athrop\c required. Once installed, you do not even need an internet connection.</em></p><p>— — — — — — — — -</p><p><strong>Where Do Those Parameters Come From?</strong></p><p>Producing a parameters file requires extensive training of the neural network on massive amounts of text data. For example, training the 70 billion parameter Llama2 model took:</p><ul><li>10 terabytes of internet text data</li><li>6,000 GPUs for 12 days</li><li>~$2 million in compute costs</li></ul><p>You can think of this process as “<a href="https://en.wikipedia.org/wiki/Data_compression">compressing</a>” a chunk of internet data into a smaller parameters file. The result is a lossy compression, not an exact copy. The parameters capture the overall patterns and knowledge, but not every specific detail.</p><p>Once the expensive training is done and you have the parameters file, running the model just requires a laptop or phone. But training them takes data centers worth of GPUs.</p><p><strong>How Do LLMs Actually Work?</strong></p><p>LLMs are focused on a single task: predicting the next word in a sequence.</p><p>You feed text into the neural network, word by word, and the model predicts what word should come next. To do this well, the model needs to build up knowledge about language, the world, and more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*igHQkqvSRJ12vAtqg1LmLQ.png" /><figcaption>The next word is predicted based on the context of previous words</figcaption></figure><p>The parameters are tuned over training to make better and better predictions. But we don’t really know how they collaborate inside the neural network for this task. So LLMs remain quite mysterious and inscrutable even as they get better at generation.</p><p><strong>From Internet Scrapers to Helpful Assistants</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sJjpL1_bW_YmPdAx1lIzZg.png" /></figure><p>LLMs trained on internet data act more like document generators, spewing out streams of text. To create a helpful “assistant” model, they get fine-tuned further on curated question-and-answer data.</p><p>So for example, a company might hire human labelers to generate ideal responses for questions like “Can you recommend some good pasta recipes?” and “What causes inflation?”</p><p>Training the LLM model further on this Q&amp;A dataset steers the model away from generating documents and towards being a conversational assistant that provides helpful, relevant answers to questions.</p><p><strong>The Future of LLMs</strong></p><p>Large language models still have major limitations in areas like logical reasoning and building on earlier statements in a conversation. But they are improving rapidly as models scale up in size and training data.</p><p>Some key frontiers researchers are working on:</p><ul><li>Giving LLMs more of a “<a href="https://en.wikipedia.org/wiki/Thinking%2C_Fast_and_Slow">Slow Think</a>” or “System 2” for complex reasoning, not just instinctual “<a href="https://en.wikipedia.org/wiki/Thinking%2C_Fast_and_Slow">Fast Think</a>” or “System 1” text generation</li><li>Self-improvement through reinforcement learning, similar to how <a href="https://deepmind.google/technologies/alphago/">AlphaGo</a> surpassed human Go players</li><li>Increased <a href="https://openai.com/blog/introducing-gpts">customization</a> for different domains</li><li>LLM as the <a href="https://memgpt.ai/">operating system</a></li></ul><p>On the security side, attackers are also exploring ways to trigger undesirable behavior in LLMs through techniques like backdoors in training data and carefully crafted input text or images. Defense <a href="https://arxiv.org/pdf/2302.12173.pdf">research</a> is ongoing as well to combat these types of vulnerabilities.</p><p>So large language models present amazing new opportunities along with risks and challenges that require careful navigation in this emerging field. But the pace of progress makes it a space full of potential.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9118bf256c7a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Navigating the Tides of Change: From StackOverflow to AI-Powered Solutions]]></title>
            <link>https://medium.com/@yepher/navigating-the-tides-of-change-from-stackoverflow-to-ai-powered-solutions-cb7e70dc3d5a?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb7e70dc3d5a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[stackoverflow]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[gpt]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Fri, 03 Nov 2023 21:56:14 GMT</pubDate>
            <atom:updated>2023-11-03T21:56:14.966Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*esehHoHImGrG0R_ZpDz79w.png" /><figcaption>The rise of LLM as code assistants</figcaption></figure><p>The software development landscape is experiencing seismic shifts, with AI emerging as a pivotal force. Recent layoffs of 28% at StackOverflow underscore a broader trend: developers, including myself, are increasingly turning to AI for help. As a Staff Data Engineer who has seen the ebb and flow of coding resources, I’ve felt the undercurrents of this change personally. The story is twofold; it’s about the transformation in the resources we use and the very nature of our challenges and their resolutions.</p><h3>The Fossilization of Online Resources</h3><p>StackOverflow has been an invaluable resource for years, a testament to the collective intelligence of the developer community. However, as reported in an article discussing the recent layoffs at the company, the robustness of this resource is being questioned as the software world rapidly evolves. “Layoffs hit coding community StackOverflow as software engineers increasingly shift to AI for help” — this headline not only signals a moment of change but also resonates with my journey away from traditional problem-solving avenues.</p><p>The thread-based knowledge that once felt like a living library shows signs of age. The rate at which technology advances has rendered some solutions obsolete, more relics of a bygone era than tools for today’s developers.</p><h3>Embracing the AI Revolution</h3><p>As the article suggests, many of us in the coding community seek new horizons, and GPT and GitHub’s CoPilot solutions are at the forefront of this exploration. For me, GPT has been a game-changer. It is not about replacing one resource with another but rather about adapting to a tool that better suits the dynamic and fast-paced nature of today’s software development environment.</p><p>It is not uncommon to see continuous build tools call out to an LLM to offer inline solutions to the root cause of the problem right in the build log.</p><h3>Why GPT Has the Edge</h3><p>The immediacy of GPT’s responses and the ability to engage in a back-and-forth dialogue is where its actual value lies. It feels akin to pair programming with a knowledgeable peer who is always on call. This interaction is not just about getting answers but about navigating through a problem space with a guide who can provide contextually relevant advice.</p><h3>Google’s Waning Magic</h3><p>Not long ago, a quick Google search could unravel the mysteries behind an error message. But as the web becomes increasingly cluttered, the effectiveness of this approach has diminished. The precise targeting of past years has given way to a morass of less relevant content, making troubleshooting more tedious than it once was.</p><h3>The Overlap in AI and Human Knowledge Gaps</h3><p>The article on StackOverflow’s layoffs indirectly points to a fascinating aspect of AI: it reflects our own knowledge. When GPT falls short, StackOverflow often does as well, suggesting that these tools are mirrors of our collective expertise — and our collective blind spots.</p><h3>Looking to the Future</h3><p>What does this mean for the future of platforms like StackOverflow? It’s not about extinction but evolution. StackOverflow will continue to be a historical and cultural touchstone within the developer community, but its role is changing. Tools like GPT, on the other hand, represent the leading edge of a new paradigm in information exchange and problem-solving in software development.</p><h3>Conclusion</h3><p>The switch from StackOverflow to GPT is not merely a personal choice but a reflection of broader industry trends. While the article on StackOverflow’s situation is a stark reminder of the rapidly changing tech landscape, it also reinforces the adaptability of the developer community. As we face new challenges, the tools we use will continue to evolve. For now, GPT solutions stand out as a beacon, guiding us through the complexities of modern software development with an AI-driven approach that’s hard to beat.</p><p>The evolution of developer tools is a testament to the innovative spirit of the field. While we embrace new technologies, it’s crucial to recognize the foundational role platforms like StackOverflow have played. As the tides turn, it’s our adaptability and willingness to embrace AI like GPT that will keep us afloat in the ever-changing sea of technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb7e70dc3d5a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I’ve been playing with many AI building blocks for several months.]]></title>
            <link>https://medium.com/@yepher/ive-been-playing-with-many-ai-building-blocks-for-several-months-b840e187efff?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/b840e187efff</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[memgpt]]></category>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[autogen]]></category>
            <category><![CDATA[langchain]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Sat, 28 Oct 2023 18:25:32 GMT</pubDate>
            <atom:updated>2023-10-30T15:50:29.076Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TmUrf6EAVdeO2hyXuNQrew.png" /></figure><p>I’ve been playing with many AI building blocks for several months. It is incredible how fast things are moving, and new ones are showing up every day! Here are a couple of gems I have nothing to do with except that they are beneficial. There are a lot of different LLM models like <a href="https://openai.com/product">OpenAI GPT</a>, <a href="https://ai.meta.com/llama/">LLAMA2</a>, <a href="https://gpt4all.io/index.html">GPT4All</a>, etc. But this is not about those. This is about tools that use those models to build something beneficial quickly.</p><p>There are countless more interesting AI tools out there. But these are my favorites that I’ve had a chance to spend time on.</p><p>I am sharing this list for selfish reasons because there are a lot of really great tools that I have not heard of yet or had a chance to try out. Please share your favorites with me, how you have applied these, or the limitations you find challenging to overcome with LLMs.</p><h3>LangChain</h3><p><a href="https://docs.langchain.com/docs/">LangChain</a> is the OG of AI building blocks, and I assume everyone reading this already knows about this tool. It is a framework for developing applications powered by language models. If you don’t know about LangChain, this is the place to start before looking at anything else I mention.</p><h3>LangFlow</h3><p><a href="https://docs.langflow.org/">LangFlow</a> is an effortless way to experiment and prototype <a href="https://docs.langchain.com/docs/">LangChain</a> configurations. With a simple pip install langflow command and OpenAI API key, you can drag and drop LLM models together in seconds. You can quickly run local LLMs or connect to OpenAI GPT APIs. It has some common pre-build examples for starting points. With its first commit back in May, it already has a fantastic level of functionality. A few features I can&#39;t wait to see enabled are AutoGen and support for the <a href="https://www.getzep.com/">Zep</a> vector database.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r6laXDDWRpXa8Z2x.png" /><figcaption>Screen shot of LangFlow</figcaption></figure><h3>AutoGen</h3><p><a href="https://arxiv.org/pdf/2308.08155.pdf">AutoGPT</a> is probably the most mind-opening research I’ve read. It enables next-gen LLM applications via multi-agent Conversation. Since I’ve read this, I feel like Neo after taking the red pill. When I describe the contents of this paper to people who are not into LLM, they tend to give me a face of fear, which demonstrates the power of this sort of multi-agent. <a href="https://github.com/yepher/autogen_demo">Here</a> is a tip-of-the-iceberg demo that shows a little bit of the power of multi-agent conversations.</p><p>Early testing of AutoGen reduces LLM hallucinations substantially, increases the quality of output, and gets past roadblocks that a single agent LLM can do by itself. I also found information that is found on <a href="https://arxiv.org/pdf/2308.08155.pdf">page 14,15</a> of the AutoGPT paper that gives an excellent summary of related work and the shortcomings of AutoGPT, ChatGPT+ (with code interpreter or plugin), LangChain Agents, Transformers Agent, BabyAGI, CAMEL, Multi-Agent Debate, and MetaGPT. AutGen allows you to use a plurality of these solutions to each solution&#39;s strengths.</p><p>Here is a tangential video about <a href="https://www.youtube.com/watch?v=Zlgkzjndpak&amp;ab_channel=TwoMinutePapers">OpenAI’s ChatGPT Makes A Game For $1!</a> from Two Minute Papers. I find the multi-agent chat more interesting in that video than they used OpenAI’s ChatGPT. <a href="https://arxiv.org/pdf/2307.07924.pdf">see Communicative Agents for Software Development Paper</a></p><p>My fork of AutoGen is based on this Medium article, <a href="https://levelup.gitconnected.com/integrate-autogen-into-your-chatbot-code-interpreter-clone-2ba71e264ad3">Integrate AutoGen into your Chatbot: Code Interpreter Clone</a>, which did not work as-is for me.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zMq3RK_RsTlym5H3.png" /><figcaption>Example AutoGen usage</figcaption></figure><h3>MemGPT</h3><p><a href="https://github.com/cpacker/MemGPT">MemGPT</a> This is the newest paper on this list, so it is also the least fully baked building block. It is a system that intelligently manages different memory tiers in LLMs to effectively provide an extended context within the LLM’s limited context window. Something I like about this paper is the idea of treating AI as an OS, and many of the limitations of the models are solved with the concept of device drivers.</p><p>This is another idea that can help us overcome some significant shortcomings that are experienced with LLM solutions. If you have a large body of content that the LLM was not directly trained on, but you want to have a conversation with that content, it can be challenging. LLM has a fixed context that it can accept before <a href="https://en.wikipedia.org/wiki/Catastrophic_interference">catastrophic forgetting</a>, which is a well-known issue for AI.</p><p>Find out more <a href="https://www.youtube.com/watch?v=4aOLxPdx1Dg&amp;ab_channel=RunLLM">here</a>, where the professor interviews the paper’s lead author.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kuu-cateF6-UaviU.png" /><figcaption>MemGPT Diagram</figcaption></figure><h3>ChatGPT 4 — Native Function Calling</h3><p><a href="https://openai.com/blog/function-calling-and-other-api-updates">ChatGPT 4 Function calling</a> makes many of the previously mentioned tools possible. If you’ve not read about it, it is foundational for many solutions you will likely make going forward.</p><h3>The rising wave of copyright lawsuits targets ChatGPT developer OpenAI</h3><p>This is not a tool or framework but is a <a href="https://www.iam-media.com/article/rising-wave-of-copyright-lawsuits-targets-chatgpt-developer-openai">paper</a> by Aaron Gin and Mike Borella. For many years, I have relied on how Mike Borrella can make complex intellectual legal concepts understandable. It is a good discussion of many of the copyright issues LLM models are facing.</p><h3>LLMFarm</h3><p><a href="https://llmfarm.site/">LLMFarm</a> lets you do LLM locally as a native MacOS or iOS app. It lets you load many of the latest LLMS and does not require a network connection. I’ve not had a lot of luck running directly on my iPhone, but it works great on my M2 Mac.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/438/0*bqRXVqhwl1CV8lkj.png" /><figcaption>LLMFarm running on iOS</figcaption></figure><h3>Zep Vector</h3><p><a href="https://www.getzep.com/">Zep</a> is a vector DB that I’ve seen a lot of information from Daniel Chalef about. I am looking forward to spending more time on it. Unlike other vector databases, Users, Sessions, Memories, and Documents are all first-class citizens in Zep.</p><p>Medium Article: <a href="https://medium.com/llamaindex-blog/zep-and-llamaindex-a-vector-store-walkthrough-564edb8c22dc">Zep and LlamaIndex: A Vector Store Walkthrough</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b840e187efff" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Dreaded: Inbox Full Message]]></title>
            <link>https://medium.com/@yepher/the-dreaded-inbox-full-message-50c7d87fd626?source=rss-6e80bd29e24------2</link>
            <guid isPermaLink="false">https://medium.com/p/50c7d87fd626</guid>
            <category><![CDATA[python-programming]]></category>
            <category><![CDATA[email-productivity]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Yepher]]></dc:creator>
            <pubDate>Thu, 09 Mar 2023 05:12:43 GMT</pubDate>
            <atom:updated>2023-03-09T05:12:43.699Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3Cml7MTlkUiAvc4k" /><figcaption>Mailbox Overflow</figcaption></figure><p>Through the years, I’ve studied how people use communication tools, especially Email. I’ve found three basic types of Email users in the World: <em>Organizers</em>, <em>Deleter</em>, and <em>Collectors</em>. I’ve also found that folks from one camp, like <em>Deleter</em>, can&#39;t understand how folks can stand to manage email like those from another camp, like <em>Collectors</em>. Either Deleters or Collectors can be organizers too.</p><p>I am very much in the unorganized “Collector” camp regarding email. The only time my email ends up in another folder besides the Inbox is through filters and automation.</p><blockquote>A deleter of the email will find little to no value in this post except to use it as proof that their way of email management I superior. It would be interesting to see comments below on which camp you feel you belong to. (I hear you laughing Jef and Harry)</blockquote><p>For all of us Collectors, there comes a day when you have to decide how dedicated you are to your collection of emails. You get the dread message “You’ve used 98% of your storage,” and you have to decide if you want to pony up $€₩$ to get more storage space. Or, delete those emails from <em>MySpace.com</em> that you received well over a decade ago and did not even realize were in your mailbox still (true story). For me, the decision is easy. My affinity for collecting emails is only surpassed by my affinity for not spending money to get more drive space to sock more useless emails away. The first few times I’ve done the exercise, I found plenty of low-hanging fruit with searches like “<em>larger_than:10m</em>” which in GMail will show all emails larger than 10 megabytes. Or maybe find all emails more than fourteen years old with a search like this one “<em>Before:2009/01/01</em>”.</p><p>The first time I did this exercise with Gmail, I deleted so many emails at once GMail thought I was doing a denial of service attack and refused to delete any more emails. I opened a ticket and described my use case, and a few months later, they fixed it, so it worked. I’ve not had that problem since.</p><p>The low-hanging fruit of easy-to-find email to delete is gone. Now I need to be more sophisticated in my tactics to delete enough emails to have a real impact so I can ignore this problem for another ten years. Once the large emails are gone, you must delete <strong><em>A LOT</em></strong> of emails to regain a gigabytes of space.</p><blockquote>According to a report by the email security company Mimecast, the average email size is about 75KB, with about 60% of emails being less than 25KB.</blockquote><p>This is what my mailbox looks like as I write this:</p><p>Minimum email size: 1.2 KB</p><p>Maximum email size: 10.7 MB</p><p>Average email size: 63.1 KB</p><p>I started a search for a tool to help me identify what sorts of emails I could delete that would give me a good amount of storage back with as little work as possible. I’ve found plenty of tools to analyze a hard drive and to help find more storage space. I’ve found almost no tools for doing something similar for your mailbox. I expect a tool for this is out there, but my Google-foo is not up to snuff to find it.</p><p>My solution is to write a simple IMAP mail client that dumps some details about my email into a CSV file. Then create tools to process that and report on those emails to help me find where all my space has gone. I thought about making a commercial application for this, but it is too close to the area I work in professionally, so I was afraid it might be a conflict of interest to do so.</p><p><strong>The Search for what to delete</strong></p><p>When I look at all the “from addresses” in my mailbox and count them, this is what it looks like sorted by most emails to least: <em>(partial data)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/704/1*5zUKtxsV5X2PLm1D-6fang.png" /><figcaption>Noisy Senders</figcaption></figure><p>If I parse the SPF header and extract out the base domain, I get a similar result: <em>(partial data)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/838/1*g5TNN1pvwBuu5FAkZatViA.png" /><figcaption>SPF TLD</figcaption></figure><p>Here I look at the size of all the emails from each domain: <em>(partial data)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/528/1*AEQsj2FlE6bF0Y41Zdt1Sg.png" /><figcaption>Size of email stored from domain</figcaption></figure><p>This is the aggregate size based on the email subject: <em>(partial data)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hcInO2MkUnWHNWY-IsKrZQ.png" /></figure><p>Running the script to size up all the emails for those domains</p><pre>python estimateSizeForDomains.py github.com barrington220.org quora.com ebay.com</pre><p>The size of emails from specified domains is approximately 500 MB</p><p>This is only a dent in my mailbox, but now I can use my tool to slice and dice it to trim it down to size.</p><p>In my next post, I will be more technical and share a GitHub project (or a Colab Notebook) that you can point at your IMAP mailbox to understand and extract the details of your emails into a CSV file. It will also have the tools to convert that data into data similar to what is shown above.</p><p>Join me on my journey to delete 50 GB of email from my mailbox.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50c7d87fd626" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>