July 2023 Summaries

Five startups are shaping the landscape of Language AI in Africa by addressing the lack of data and resources for African languages. VulaVula is a language technology service that uses text and speech technologies to increase computational understanding of underrepresented languages, especially South African languages. Gotbot provides conversational AI chatbots for businesses, automating tasks and providing individualized customer service. Botlhale AI offers voice biometrics, voice enabled chatbots, and APIs for African languages, eliminating language barriers for users. Mymanu enables real-time translation through earbuds in 47 different languages. Lesan AI is an automated translation service for Amharic and Tigrinya, with plans to expand to other Ethiopian languages. These startups are working to ensure the representation of African languages in NLP and encourage research for these languages.

The article discusses 11 key papers that have shaped the development of AI's language journey from its inception to the present day. These papers include Alan Turing's "Computing Machinery and Intelligence" (1950), which introduced the concept of a machine passing the Turing test, Rumelhart et al.'s "Learning Representations by Back-propagating Errors" (1986), which described backpropagation for training neural networks, Bengio et al.'s "A Neural Probabilistic Language Model" (2003), which introduced the idea of using neural networks to learn distributed word representations, and many more. The article also highlights major milestones in AI language development such as the introduction of sequence-to-sequence models, Transformer models, BERT, GPT-3, and ChatGPT. It emphasizes that we are still at the beginning of this journey with much more innovation to come.

In 2023, various AI tools are available for transcribing meetings, improving productivity and saving time. These tools offer features such as speaker diarization, action item extraction, automatic meeting summaries, and support for multiple languages. Some of the top AI transcription tools include Otter.AI, Rev.com, Sonix, Fireflies, AssemblyAI, Amazon Transcribe, Speechmatics, and Deepgram. These tools can help businesses improve efficiency and profitability by providing accurate and fast transcriptions of meetings.

Training large language models (LLMs) requires high-quality data ingestion to ensure robust generative outputs. Data ingestion is a complex process involving collection, curation, preprocessing, and tokenization of natural language data. The quality and relevance of the training data directly impact the LLM's performance. Proper data preparation is crucial for foundation models and fine-tuning existing models for domain-specific tasks. Tools like Unstructured API help streamline data ingestion by connecting complex data hierarchies into clean JSON outputs, making it easier for organizations to leverage the power of LLMs in their operations.

Deepgram has released a domain-specific language model (DSLM) for speech summarization, specifically tailored for call center and sales enablement interactions. The DSLM is fine-tuned using over 200K domain-specific conversations and offers benefits such as personalization, superior accuracy on specialized topics, low inference costs, speed, and no token length or audio duration limits. This model automates the process of summarizing customer interactions and extracting pertinent information at scale, improving efficiency and enhancing the overall customer experience. Deepgram's DSLM-powered Summarization Model delivers superior results compared to alternative summarization methods and is now publicly available for pre-recorded, English audio.

The field of Machine Learning is constantly evolving with new developments emerging from various sources. To effectively follow ML content, one must first identify their areas of interest within the broad domain of ML and determine their technical predilection for the type of content they consume. A key principle to efficiently keep up with academic literature in this dynamic field is "Quantity over Quality," which encourages maximizing exposure to a broad spectrum of topics before honing in on those most aligned with your interests and objectives. There are two types of places on the internet that provide the latest advancements in machine learning: community-based forums and professional publications. Community-based forums like Reddit, Hacker News, and Research Hub allow anyone to post their findings or opinions, while professional publications such as arXiv and IEEE offer more technically-inclined resources. To make the most out of an academic paper or a longer article, one can leverage AI tools like Talk to PDF and ChatGPT for assistance. Additionally, reading actively by asking questions about relevance, researcher's approach, reproducibility, and novel contributions can help evaluate the significance and potential impact of the content on your learning or research.

Meta has unveiled its latest open source large language model family, Llama 2, which comes in various sizes with pre-trained versions available for 7B, 13B, and 70B parameters. The Llama 2-chat variant is fine-tuned for chat-style interactions through supervised fine-tuning and reinforcement learning with human feedback (RLHF). Meta has trained Llama 2 on 2 trillion tokens of data from publicly available sources, increasing its performance on common LLM benchmarks. The model's license allows for research and commercial use, setting a new standard for open source large language models and potentially impacting the AI industry as much as ChatGPT's launch.

This blog post provides a solution to delete multiple entries from a FaunaDB database that match a given regex pattern. The author shares their experience with the lack of clear documentation and offers a script that can be used for this purpose. The script requires the user's secret, table name, and regex pattern as input. It initializes faunadb, retrieves matching documents using pagination, and then iterates through the results to delete matching entries. The author recommends printing the results before deleting anything to ensure accuracy.

Deepgram introduces a self-serve on-premise solution, allowing users to manage their API keys and distribution credentials independently without going through a representative. This feature enhances project management by streamlining the process and saving valuable time. The self-serve solution is available through both Console UI and the Deepgram API. A hands-on tutorial guide provides step-by-step instructions on how to create on-prem API keys and distribution credentials within the Console. Users can provide feedback or discuss further details with Deepgram's product experts via GitHub discussions or direct contact.

The text discusses the top five YouTubers who teach LangChain, a machine learning framework designed for connecting large language models with other tools at high levels of abstraction. These YouTubers are Sam Witteveen, James Briggs, Greg Kamradt (Data Indy), Abdul Majed Raja (1littlecoder), and Mayo Oshin (Chat with Data). Each channel offers different approaches to teaching LangChain, catering to various experience levels and coding languages. The text also encourages readers to explore other learning options as more educators emerge in the field of LangChain.

The rise of Generative AI has led to an explosion of creative, often bizarre projects on the internet. These projects range from using hallucinations of language models like ChatGPT to create cursed images and videos, to jailbreaking chatbots on Reddit forums. One notable project is Stochastic Parrot, a bot that uses empty prompts to generate media through GPT-4, Dall-E, and ElevenLabs API. Another interesting creation is the podcast Myself, I Am and That, where AI clones of a sound engineer have conversations about life. Additionally, websites like This Person Does Not Exist use GANs to create hyper-realistic portraits of nonexistent people. While many of these projects are self-funded or run by one person, they offer a glimpse into the potential and strange possibilities of Generative AI.

This blog post discusses the use of various AI coding assistants such as StarCoder, GitHub Copilot, ChatGPT, and Gorilla to help software engineers code faster and more efficiently. The author tested these models using two Python coding questions related to string manipulation and API usage. They found that while none of the models produced perfect code every time, they all provided useful starting points for developers, especially those new to certain APIs or programming languages. The author recommends using GitHub Copilot with ChatGPT on the side for debugging purposes. StarCoder can also be used as an alternative to Copilot but may have higher latency. Gorilla is a newer model that specializes in making API calls, but it is currently limited to Hugging Face APIs only. In conclusion, AI coding assistants are becoming increasingly useful for experienced programmers who want to accelerate their development process while maintaining code readability and efficiency. However, these models should not be relied upon as the sole source of learning how to code, as they are still in their early stages of development.

Deepgram has partnered with Genesys to integrate its industry-leading speech-to-text (STT) technology into the Genesys Cloud platform, enabling users to test and utilize Deepgram's services in their contact center environment. The integration allows Genesys customers to power next-generation conversational AI solutions, improving customer experiences, enhancing operational efficiency, and achieving substantial cost savings. Through this strategic partnership, businesses can leverage the advanced speech recognition technology offered by Deepgram, revolutionizing their contact center operations and providing exceptional customer experiences.

This article provides an overview of deep learning, its history, and key terminology. Deep learning is a subset of neural networks that involves multiple layers of interconnected neurons. It is particularly good at tasks involving complex, unstructured data such as images, audio, and text. The training process for neural nets typically includes both a learning mode (training) and a running mode (inference). Deep learning models are often viewed as black boxes, making it difficult to explain why outputs are arrived at. The article also covers the history of deep learning, tracing its roots back to the 1940s and 50s with the proposal of the first artificial neural net (McCulloch-Pitts neuron). It highlights key advancements in the field, such as the introduction of error backpropagation by Paul Werbos in 1974 and the coining of the term "deep learning" by Geoffrey Hinton in 2006. The article delves into various types of deep learning models, including Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Transformers. It also discusses the role of GPUs in enabling faster training of neural networks and the importance of large amounts of data for deep learning research. Finally, the article mentions popular deep learning code frameworks like PyTorch and TensorFlow, as well as Deepgram's use of Rust for production purposes to bypass Python's Global Interpreter Lock (GIL) optimization that negatively impacts multithreaded computation.

The text discusses six foundational courses for learning about Large Language Models (LLMs), which are revolutionizing the tech landscape with models like ChatGPT, Bard, LLaMA, and Claude. LLMs have potential implications beyond entertainment, such as replacing human jobs and introducing new professions dedicated to training and optimizing them. The text emphasizes the importance of understanding machine learning, deep learning, and natural language processing (NLP) fundamentals for comprehending LLMs. The six courses are divided into three phases: 1. Theory and Foundations: This phase includes 3Blue1Brown's series on linear algebra, calculus, and neural networks, which provide an intuitive understanding of the mathematical foundations required for machine learning. 2. NLP Basics: Sequence Models from Coursera by Andrew Ng focuses on sequence models used in processing text, while NLP – Natural Language Processing with Python by Jose Portilla covers text data preparation and transformation techniques. 3. Deep Dive into Modern Language Models: DS-GA 1008 - Deep Learning, NYU delves into the theoretical aspects of deep learning, CS324 - Large Language Models, Stanford University explores specific applications and LLM architectures, and COS 597G - Understanding Large Language Models, Princeton University investigates prompting, reasoning, in-context learning, and addressing issues like bias and toxicity. These courses provide a comprehensive understanding of the foundational theory, implementation, and recent developments in LLMs, catering to learners at various levels of expertise.

The article discusses seven voice synthesis AI startups that have gained attention within the community. These companies include ElevenLabs, Lyrebird (now a part of Descript), WellSaid Labs, Papercup, MURF.AI, deepdub, and Typecast AI. Each startup offers unique features and services in the field of text-to-speech conversion, voice cloning, and multilingual communication. The article highlights how these companies are shaping the future of audio content creation and manipulation, as well as breaking language barriers to enable effective communication across linguistic boundaries.