<?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 Tomas Vykruta on Medium]]></title>
        <description><![CDATA[Stories by Tomas Vykruta on Medium]]></description>
        <link>https://medium.com/@tom_21755?source=rss-c2339a6e30bf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hrw8t81XZGeEznXtPEu7JA.jpeg</url>
            <title>Stories by Tomas Vykruta on Medium</title>
            <link>https://medium.com/@tom_21755?source=rss-c2339a6e30bf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 10:20:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tom_21755/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[Understanding Causal LLM’s, Masked LLM’s, and Seq2Seq: A Guide to Language Model Training…]]></title>
            <link>https://medium.com/@tom_21755/understanding-causal-llms-masked-llm-s-and-seq2seq-a-guide-to-language-model-training-d4457bbd07fa?source=rss-c2339a6e30bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4457bbd07fa</guid>
            <dc:creator><![CDATA[Tomas Vykruta]]></dc:creator>
            <pubDate>Sun, 30 Apr 2023 15:39:33 GMT</pubDate>
            <atom:updated>2023-04-30T15:39:33.184Z</atom:updated>
            <content:encoded><![CDATA[<h3>Understanding Causal LLM’s, Masked LLM’s, and Seq2Seq: A Guide to Language Model Training Approaches</h3><p>In the world of natural language processing (NLP), choosing the right training approach is crucial for the success of your language model. In this article, we’ll delve into the differences between Causal Language Modeling (CLM), Masked Language Modeling (MLM), and Sequence-to-Sequence (Seq2Seq) approaches, discussing their importance, and exploring best practices for using them.</p><h4>Why it’s important to understand CLM, MLM, and Seq2Seq</h4><p>Understanding these training methods allows you to select the most appropriate approach for your specific NLP task, ultimately enhancing your model’s performance. Each method has its unique strengths and weaknesses and is suited to different types of problems. By understanding the fundamentals of each approach, you can optimize your model’s training and fine-tuning, leading to better outcomes.</p><h4>Causal Language Modeling (CLM)</h4><p>CLM is an autoregressive method where the model is trained to predict the next token in a sequence given the previous tokens. CLM is used in models like GPT-2 and GPT-3 and is well-suited for tasks such as text generation and summarization. However, CLM models have unidirectional context, meaning they only consider the past and not the future context when generating predictions.</p><h4>Masked Language Modeling (MLM)</h4><p>MLM is a training method used in models like BERT, where some tokens in the input sequence are masked, and the model learns to predict the masked tokens based on the surrounding context. MLM has the advantage of bidirectional context, allowing the model to consider both past and future tokens when making predictions. This approach is especially useful for tasks like text classification, sentiment analysis, and named entity recognition.</p><h4>Sequence-to-Sequence (Seq2Seq)</h4><p>Seq2Seq models consist of an encoder-decoder architecture, where the encoder processes the input sequence and the decoder generates the output sequence. This approach is commonly used in tasks like machine translation, summarization, and question-answering. Seq2Seq models can handle more complex tasks that involve input-output transformations, making them versatile for a wide range of NLP tasks.</p><h4>Key Differences in CLM, MLM, Seq2Seq</h4><p>key differences in the implementation, architecture, and output models for causal language modeling (CLM), masked language modeling (MLM), and sequence-to-sequence (seq2seq) modeling.</p><h4>Causal Language Modeling (CLM):</h4><ul><li>Implementation: In CLM, the model is trained to predict the next token in the sequence, given the previous tokens. During training, the input tokens are fed into the model, and the model predicts the probability distribution of the next token. The loss is calculated based on the model’s predictions and the actual target tokens, which are just the input tokens shifted by one position.</li><li>Architecture: CLM is typically used with autoregressive models like GPT. These models use a unidirectional (left-to-right) Transformer architecture, where each token can only attend to the tokens that come before it. This prevents the model from “cheating” by attending to the target tokens during training.</li><li>Output Model: A fine-tuned CLM model can generate coherent text by predicting one token at a time, making it suitable for text generation tasks. However, it may not be as effective at capturing bidirectional context compared to MLM models.</li></ul><h4>Masked Language Modeling (MLM):</h4><ul><li>Implementation: In MLM, the model is trained to predict masked tokens within the input sequence. During preprocessing, a certain percentage of tokens are randomly masked, and the model is trained to predict the original tokens at those masked positions. The loss is calculated based on the model’s predictions and the actual target tokens (the original tokens that were masked).</li><li>Architecture: MLM is used with models like BERT, which use a bidirectional Transformer architecture. Unlike CLM models, MLM models can attend to all tokens in the input sequence during training, allowing them to capture context from both left and right.</li><li>Output Model: A fine-tuned MLM model is better at understanding context and relationships between words in a sequence, making it suitable for tasks like text classification, sentiment analysis, named entity recognition, or question answering.</li></ul><h4>Sequence-to-Sequence (seq2seq) Modeling:</h4><ul><li>Implementation: In seq2seq modeling, the model is trained to generate output sequences based on input sequences. The model consists of two parts: an encoder that encodes the input sequence into a latent representation, and a decoder that generates the output sequence based on this latent representation. The loss is calculated based on the model’s predictions and the actual target output tokens.</li><li>Architecture: Seq2seq models typically use an encoder-decoder architecture, where both the encoder and decoder can be based on the Transformer architecture (e.g., T5, BART) or other architectures like LSTMs (e.g., the original seq2seq model). The encoder processes the input sequence and generates a context representation, while the decoder generates the output sequence based on the encoder’s output and its own hidden state.</li><li>Output Model: A fine-tuned seq2seq model is better at tasks where the model needs to generate coherent output text based on input text, such as summarization, translation, or question answering.</li></ul><p>In summary, the main differences between these approaches lie in their training objectives, model architectures, and the tasks they are suited for. CLM models focus on predicting the next token and are good for text generation tasks, MLM models focus on predicting masked tokens and are good for tasks requiring bidirectional context understanding, and seq2seq models focus on generating output sequences based on input sequences and are good for tasks requiring the generation of coherent output text.</p><h4>MLM parameter in Huggingface selects MLM or CLM</h4><p>The mlm parameter in the DataCollatorForLanguageModeling class is a boolean flag that indicates whether you want to train a masked language model (MLM) or a causal language model (CLM).</p><ul><li>mlm=True: The data collator will prepare the training data for masked language modeling. In this mode, the input sequences will have a certain percentage of tokens masked, and the model will be trained to predict the original tokens at those masked positions. This is the training objective used by models like BERT.</li><li>mlm=False: The data collator will prepare the training data for causal language modeling, also known as autoregressive language modeling. In this mode, the model will be trained to predict the next token in the sequence given its previous tokens. This is the training objective used by models like GPT.</li></ul><h4>what are the drawbacks of MLM vs CLM for a LLM chatbot?</h4><p>Each language modeling technique, Masked Language Modeling (MLM) and Causal Language Modeling (CLM), has its own advantages and drawbacks when used for building a chatbot. Here are some key points to consider:</p><h4>Masked Language Modeling (MLM):</h4><ol><li>Pros:</li></ol><ul><li>MLM can potentially capture bidirectional context, as the model learns to predict masked tokens based on both the preceding and following tokens. This can help the model understand the context more effectively, which might be useful in some chatbot scenarios.</li></ul><ol><li>Cons:</li></ol><ul><li>MLM models like BERT are not designed to generate text autoregressively. While they can be fine-tuned for various NLP tasks, they are not inherently built for text generation like chatbot responses. Adapting MLM models for text generation typically requires additional architectural changes, such as adding a decoder or using a seq2seq model.</li><li>During inference, MLM models cannot predict tokens incrementally, as the training involves predicting masked tokens in parallel. This may result in a less coherent chatbot response, as the model has not been trained to generate text sequentially.</li></ul><h4>Causal Language Modeling (CLM):</h4><ol><li>Pros:</li></ol><ul><li>CLM models like GPT are designed for autoregressive text generation, making them more suitable for chatbot applications. They predict the next token in the sequence given the previous tokens, which aligns with how chatbot responses are generated.</li><li>CLM models can generate coherent and contextually relevant responses because they are trained to predict the next token in a sequence based on the preceding tokens, taking into account the context provided by the input.</li></ul><p>Cons:</p><ul><li>CLM models do not explicitly capture bidirectional context, as they only generate tokens based on the preceding tokens. This may lead to a slightly less nuanced understanding of context compared to MLM models, which consider both preceding and following tokens during training.</li><li>Due to their autoregressive nature, CLM models may have slower inference times than MLM models, especially when generating long sequences, as they have to predict each token one at a time.</li></ul><h4>Which popular LLM’s were trained with DataCollatorForSeq2Seq collator?</h4><p>DataCollatorForSeq2Seq is typically used for sequence-to-sequence (seq2seq) models, where the model is designed to generate output sequences based on input sequences. Some popular seq2seq models in the Hugging Face Transformers library include:</p><ol><li>BART (Bidirectional and Auto-Regressive Transformers): BART is a denoising autoencoder with a seq2seq architecture, which has been shown to perform well on a variety of natural language understanding and generation tasks. It is pre-trained using a denoising objective, where the model learns to reconstruct the original text from corrupted versions of it.</li><li>T5 (Text-to-Text Transfer Transformer): T5 is a seq2seq model that frames all NLP tasks as text-to-text problems. It is pre-trained using a denoising objective similar to BART. T5 has been shown to perform well on various NLP tasks, including translation, summarization, and question-answering.</li><li>MarianMT: MarianMT is a seq2seq model specifically designed for neural machine translation. It is part of the Marian NMT framework and has been trained on translation tasks for various language pairs.</li><li>Pegasus (Pre-training with Extracted Gap-sentences for Abstractive Summarization): Pegasus is a seq2seq model designed for abstractive summarization. It is pre-trained using a gap-sentence generation task, where important sentences are removed from the input, and the model learns to generate these missing sentences.</li><li>ProphetNet: ProphetNet is a seq2seq model with a novel self-supervised objective called future n-gram prediction. It has been shown to perform well on tasks like abstractive summarization and question generation.</li></ol><h4>Conclusion</h4><p>Understanding the differences between CLM, MLM, and Seq2Seq is crucial for selecting the most appropriate training approach for your language model. By following best practices and leveraging the strengths of each method, you can optimize your model’s performance and achieve better results in your NLP tasks.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4457bbd07fa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Optimization: Layer-wise Optimal Rank Adaptation (LORA)]]></title>
            <link>https://medium.com/@tom_21755/llm-optimization-layer-wise-optimal-rank-adaptation-lora-1444dfbc8e6a?source=rss-c2339a6e30bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/1444dfbc8e6a</guid>
            <dc:creator><![CDATA[Tomas Vykruta]]></dc:creator>
            <pubDate>Sun, 30 Apr 2023 15:27:41 GMT</pubDate>
            <atom:updated>2023-04-30T15:27:41.541Z</atom:updated>
            <content:encoded><![CDATA[<p>Layer-wise Optimal Rank Adaptation (LORA) is a technique that aims to reduce the computational cost and memory footprint of large-scale Transformer models while preserving their performance. The intuition behind LORA is based on the observation that many matrices in deep neural networks, especially the attention mechanism of Transformers, can be effectively approximated by low-rank matrices. A low-rank matrix is a matrix that can be decomposed into the product of two smaller matrices. By leveraging this property, LORA reduces the number of parameters and computational complexity, making it more efficient and practical to use large models in various applications.</p><p>Here is a high-level explanation of why LORA works:</p><ol><li>Low-rank approximation: In deep neural networks, many weight matrices have a low-rank structure. This means that they can be decomposed into the product of two smaller matrices, which significantly reduces the number of parameters required to represent the original matrix. LORA focuses on approximating these weight matrices by their low-rank counterparts, thus reducing the model’s size and computational cost.</li><li>Layer-wise adaptation: LORA adapts the rank of the low-rank approximation on a per-layer basis. This is based on the observation that different layers in a Transformer may have different optimal ranks. By choosing the rank for each layer individually, LORA allows for a more efficient trade-off between model size, computational cost, and performance.</li><li>Fine-tuning: After the low-rank approximation is applied to the target layers, LORA fine-tunes the model to compensate for potential loss of performance. This fine-tuning step helps the model to adapt to the low-rank approximation and maintain its original performance.</li></ol><p>LORA is particularly useful for large-scale Transformer models, as these models can have billions of parameters and require significant computational resources. By using LORA, it is possible to create more efficient models that are easier to deploy in real-world applications, such as on-device machine learning or low-latency inference tasks.</p><p>In summary, LORA leverages the low-rank structure of the weight matrices in Transformer models to reduce the number of parameters and computational cost. The layer-wise adaptation and fine-tuning steps ensure that the model’s performance is maintained while achieving these efficiency gains.</p><h4>how does the layer-wise adaptation work?</h4><p>The layer-wise adaptation in LORA refers to the process of selecting an optimal rank for the low-rank approximation on a per-layer basis. The intuition behind this is that different layers in a Transformer model may have different optimal ranks for the low-rank approximation, which allows for a more efficient trade-off between model size, computational cost, and performance. Here’s a step-by-step explanation of how layer-wise adaptation works:</p><ol><li>Identify target layers: First, the target layers for applying LORA are identified. These are usually the layers that contribute the most to the model’s size and computational cost, such as the query and value projection matrices in the multi-head self-attention mechanism.</li><li>Compute low-rank approximation for each layer: For each target layer, a low-rank approximation is computed. This can be done using various methods, such as singular value decomposition (SVD) or other matrix factorization techniques. The objective is to find two smaller matrices whose product closely approximates the original weight matrix in the target layer.</li><li>Select optimal rank: The optimal rank for each layer is chosen by considering the trade-off between the model’s performance and the reduction in parameters and computational cost. This can be done by evaluating the performance of the model on a validation set using different ranks and selecting the rank that provides the best balance between performance and efficiency. In some cases, this process can be automated using techniques like cross-validation or Bayesian optimization.</li><li>Replace target layers: Once the optimal rank for each layer is determined, the original weight matrices in the target layers are replaced with their low-rank approximations. This reduces the number of parameters in the model and the computational cost associated with the target layers.</li><li>Fine-tuning: After the low-rank approximation is applied to the target layers, the model is fine-tuned to adapt to the changes and recover any potential loss of performance. This can be done using standard training techniques, such as gradient descent, with a smaller learning rate to avoid large updates that could harm the performance.</li></ol><p>By performing layer-wise adaptation, LORA can effectively balance the trade-off between model size, computational cost, and performance. It allows for more efficient models that are tailored to the specific requirements of each layer in the Transformer architecture, resulting in more efficient and practical models for real-world applications.</p><h4>why was r=16 chosen for the OPT6.7B model? what does this number mean and why not use a smaller or larger number?</h4><p>The r=16 parameter in the LORA configuration for the OPT6.7B model refers to the rank of the low-rank approximation. The rank is the dimensionality of the smaller matrices resulting from the low-rank decomposition. By using a lower rank, the number of parameters in the model is reduced, which leads to a decrease in computational cost and memory requirements.</p><p>Choosing the optimal rank is a trade-off between model efficiency and performance. A smaller rank results in a more compact model with fewer parameters, leading to faster inference and lower memory usage. However, if the rank is too small, the low-rank approximation may not capture the essential information in the original weight matrices, which can lead to a significant degradation in the model’s performance.</p><p>On the other hand, a larger rank provides a better approximation of the original weight matrices, which helps preserve the model’s performance. However, a larger rank also means more parameters and higher computational cost, which can negate the benefits of using LORA in the first place.</p><p>The choice of r=16 for the OPT6.7B model likely comes from experimentation and validation on a held-out dataset. The developers would have tested different values of r to find the optimal balance between model efficiency and performance. The r=16 value was probably chosen because it provided a good trade-off between maintaining the model&#39;s performance and reducing its size and computational cost.</p><p>It is essential to understand that the optimal value of r may vary depending on the specific model, dataset, and task. If you are working with a different model, you might need to perform your own experiments to find the best value of r for your use case.</p><h4>Intuitively, what does r=16 means, and how it would differ from r=2?</h4><p>When using LORA, the parameter r represents the rank of the low-rank approximation applied to the target layers in the Transformer model. It determines the dimensionality of the smaller matrices resulting from the low-rank decomposition. The choice of r directly affects the trade-off between model efficiency (fewer parameters and lower computational cost) and performance (accuracy and generalization).</p><p>Here’s a step-by-step explanation of what r=16 means and how it would differ from r=2:</p><ol><li>Identify target layers: First, identify the target layers to which LORA will be applied. These are usually the layers that contribute the most to the model’s size and computational cost, such as the query and value projection matrices in the multi-head self-attention mechanism.</li><li>Compute low-rank approximation: For each target layer, compute a low-rank approximation using the specified rank (r). This is typically achieved using methods like singular value decomposition (SVD) or other matrix factorization techniques. The objective is to find two smaller matrices whose product closely approximates the original weight matrix in the target layer.</li></ol><ul><li>For r=16, the low-rank approximation will result in two matrices with a dimensionality of 16. The product of these matrices will approximate the original weight matrix.</li><li>For r=2, the low-rank approximation will result in two matrices with a dimensionality of 2. The product of these matrices will approximate the original weight matrix.</li></ul><ol><li>Replace target layers: Replace the original weight matrices in the target layers with their low-rank approximations. This reduces the number of parameters in the model and the computational cost associated with the target layers.</li></ol><ul><li>With r=16, the model will have more parameters than with r=2, but the approximation will be closer to the original weight matrices, which can help maintain the model&#39;s performance.</li><li>With r=2, the model will have fewer parameters and lower computational cost, but the approximation may be less accurate, potentially leading to a significant degradation in the model&#39;s performance.</li></ul><ol><li>Fine-tuning: After the low-rank approximation is applied to the target layers, the model is fine-tuned to adapt to the changes and recover any potential loss of performance. This can be done using standard training techniques, such as gradient descent, with a smaller learning rate to avoid large updates that could harm the performance.</li></ol><p>To summarize, when using r=16, the low-rank approximation will have a higher rank, resulting in more parameters and higher computational cost compared to r=2. However, the approximation will likely be more accurate, preserving the model&#39;s performance better. In contrast, when using r=2, the model will be more compact, with fewer parameters and lower computational cost, but the performance may degrade due to the less accurate approximation. The optimal value of r depends on the specific model, dataset, and task, and it should be chosen based on a balance between maintaining performance and achieving efficiency gains.</p><h4>What is a typical rank without Lora?</h4><p>The original rank of the weight matrices in a Transformer model without LORA is equal to the dimensionality of those matrices. For the query and value projection matrices in the multi-head self-attention mechanism, their dimensions depend on the model architecture and the number of attention heads.</p><p>In a typical Transformer architecture, the input embeddings and the hidden states have a dimensionality of d_model, and the model has num_heads attention heads. Each attention head has a separate query and value projection matrix. The dimensions of the query and value projection matrices are d_model x (d_model // num_heads).</p><p>For example, if you have a Transformer model with d_model = 768 and num_heads = 12, the dimensions of the query and value projection matrices would be 768 x 64. In this case, the original rank of the weight matrices is 64 (equal to d_model // num_heads).</p><p>When applying LORA, the original rank is replaced with a lower rank (r) to achieve the low-rank approximation, which reduces the number of parameters and the computational cost. The choice of r depends on the desired trade-off between model efficiency and performance preservation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1444dfbc8e6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is Generative AI Really Living Up to the Hype? Debunking the Myths and Setting the Record Straight]]></title>
            <link>https://medium.com/@tom_21755/is-generative-ai-really-living-up-to-the-hype-debunking-the-myths-and-setting-the-record-straight-ff445552a21?source=rss-c2339a6e30bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff445552a21</guid>
            <dc:creator><![CDATA[Tomas Vykruta]]></dc:creator>
            <pubDate>Thu, 02 Feb 2023 16:47:58 GMT</pubDate>
            <atom:updated>2023-02-02T16:58:38.278Z</atom:updated>
            <content:encoded><![CDATA[<h4>Generative AI has recently been facing backlash on Twitter, with negative sentiment on the rise. In this article, I aim to address some of the common concerns raised and shed light on the issue. To illustrate this, I will reference a report by Gordon Ritter and Jake Saper, General Partners at Emergence Capital. Note that their views are not unique, and similar opinions are widely held among others. In my opinion, this opposition is a typical response from the late adopter community.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*32azyk1Miemcxt15.png" /></figure><p>Lately, Twitter has been buzzing with negative sentiment towards Generative AI. But is it really all it’s cracked up to be? In this explosive article, we’re diving into the common concerns raised and shining a light on the truth behind the technology.</p><p>Don’t believe the naysayers — the criticism Generative AI is facing is just a natural reaction from the late adopter community. The more disruptive a technology is, the more pushback it will receive from those who are resistant to change. But here’s the thing — they’re not always right.</p><p>To bust the myth and set the record straight, we’ll be referencing a report by Gordon Ritter and Jake Saper of Emergence Capital. Their views, while widely held among others, are not representative of the whole truth. In this article, we’ll be addressing their specific points, point by point:</p><ul><li><strong>Not accurate: </strong>False. Medical GPT models have <a href="https://interestingengineering.com/innovation/google-deepmind-medpalm-language-model">already passed medical school exams</a> and are proving to be in the middle of the pack, outperforming some practicing doctors. With the ability to answer both multiple choice and written questions, these models are demonstrating their accuracy and knowledge.</li><li><strong>Not outcome-oriented:</strong> False. With the rise of <a href="https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html">IFT (Instruction Fine Tuning) and CoT (Chain of Thought),</a> Generative AI is showing impressive results in solving complex logic problems through multi-step reasoning. And this is just the beginning — imagine what’s to come in the next few years.</li><li><strong>Broad, not deep</strong>: False. FLAN (Fine-Tuned Language Models) with few-shot prompt tuning are achieving remarkable results in specific domains, even with significantly reduced parameter sizes. <a href="https://ai.googleblog.com/2023/02/the-flan-collection-advancing-open.html?m=1">A recent paper by Google </a>provides further proof of this technology’s potential.</li><li><strong>Not ready for enterprise</strong>: False. With Microsoft’s 10B investment and their announcement that OpenAI is a major strategy for their office and Azure suite, it’s clear that Generative AI is ready for the big leagues. OpenAI has already launched its first GPT3-powered office product and has enterprise contracts starting at 100k, with a dedicated cloud environment for enterprise customers.</li></ul><p>But here’s the thing — these critics are already behind the times. The field of Generative AI is advancing at a breakneck pace, and limitations that exist today may be solved in the coming weeks or months. These laggard investors are missing out on the second derivative — the rate of change — and their rearview mirror approach could cost them big time.</p><p>The rise of Generative AI has also brought with it a new risk — political bias. If AI becomes perceived as a tool of the two-party political system, it could slow innovation, spark endless debates about censorship, and even lead to regulatory intervention. This is a concern that deserves attention and careful consideration.</p><p>So, is Generative AI really all it’s cracked up to be? The truth is, it’s already changing the game and pushing the limits of what’s possible. Don’t believe the myths — discover the truth for yourself in this must-read article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff445552a21" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[This Company is Calling All Engineers to Help Take Enterprise Software to the Next Level]]></title>
            <link>https://medium.com/@tom_21755/this-company-is-calling-all-engineers-to-help-take-enterprise-software-to-the-next-level-57f5865589f1?source=rss-c2339a6e30bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/57f5865589f1</guid>
            <dc:creator><![CDATA[Tomas Vykruta]]></dc:creator>
            <pubDate>Mon, 24 Jan 2022 21:00:32 GMT</pubDate>
            <atom:updated>2022-01-24T21:00:32.134Z</atom:updated>
            <content:encoded><![CDATA[<p>“With the insurance industry often overwhelmed by ever-increasing volumes of claims, the company’s artificial intelligence (AI) software offers insurance carriers a new approach to claims management.”</p><p>Interview with <a href="https://www.linkedin.com/in/ACoAAAjlMz4BWjTu3bs__BkIQebXWh8f25DD7AM">Karan Uppal</a> and me about:</p><p>-tech culture during the Great Resignation<br>-the future of AI in vertical SaaS<br>-why now is an amazing time to join <a href="https://www.linkedin.com/company/evolution-iq/">EvolutionIQ</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=57f5865589f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>