Weekly AI and NLP News — August 28th 2023
<h1> News From The Web</h1>
<ul>
<li><a href="https://openai.com/blog/gpt-3-5-turbo-fine-tuning-and-api-updates" rel="noopener ugc nofollow" target="_blank">GPT-3.5 Turbo fine-tuning released</a>. OpenAI has released fine-tuning for GPT-3.5 Turbo, offering improved performance on specific tasks. The fine-tuned version can even match or surpass the capabilities of base GPT-4. Early testers have significantly reduced prompt size through fine-tuning. The costs for training and usage input/output are provided at $0.008, $0.012, and $0.016 per 1K tokens, respectively.</li>
<li><a href="https://ai.meta.com/blog/code-llama-large-language-model-coding/" rel="noopener ugc nofollow" target="_blank">Introducing Code Llama, a state-of-the-art large language model for coding</a>. Meta has released Code Llama, an advanced LLM for coding that can generate code and natural language about code. It is available in three models and comes in different sizes to meet various needs.</li>
<li><a href="https://ai.meta.com/blog/seamless-m4t/" rel="noopener ugc nofollow" target="_blank">Introducing a foundational multimodal model for speech translation</a>. Meta has developed a powerful foundational model called SeamlessM4T that is capable of handling various text and speech tasks across 100 languages. It includes automatic speech recognition, speech-to-text translation, speech-to-speech translation, text-to-text translation, and text-to- speech translation, supporting a wide range of input and output languages.</li>
<li><a href="https://www.deeplearning.ai/short-courses/finetuning-large-language-models/" rel="noopener ugc nofollow" target="_blank">New DeepLearning.AI Course on Finetuning Large Language Models</a>. A new free course on “Finetuning Large Language Models” by DeepLearning.AI. The course aims to help AI professionals understand the usage of finetuning on LLMs, including data preparation, training, and evaluation. Finetuning enables training the model on custom data, updating the neural nets’ weights, and improving results by incorporating style, form, and new knowledge.</li>
<li><a href="https://techcrunch.com/2023/08/18/ai2-drops-biggest-open-dataset-yet-for-training-language-models/" rel="noopener ugc nofollow" target="_blank">AI2 drops biggest open dataset yet for training language models</a>. AI2 has released a significant open dataset called Dolma, comprising 3 trillion tokens. Unlike other datasets from OpenAI and Anthropic, Dolma provides transparency in terms of what information was removed, why, and how personal details were handled. The disclosure of these details is uncommon and raises questions about the ethical and legal acquisition of other datasets.</li>
<li><a href="https://huggingface.co/blog/idefics" rel="noopener ugc nofollow" target="_blank">Introducing IDEFICS: An Open Reproduction of State-of-the-art Visual Langage Model</a>. IDEFICS is an impressive open-source visual language model with 9 billion and 80 billion parameters, based on DeepMind’s Flamingo. It offers the ability to describe images, generate stories, and answer image-related questions. It’s trained on diverse open datasets such as Wikipedia, Public Multimodal Dataset, LAION, and OBELICS.</li>
<li><a href="https://www.zdnet.com/article/bings-search-market-share-fails-to-budge-despite-ai-push/" rel="noopener ugc nofollow" target="_blank">Bing’s search market share fails to budge despite big AI push</a>. Despite the introduction of innovative AI features like Bing AI Chat and Bing Image Creator, Bing’s market share remains stagnant at around 3%. Microsoft disputes this data, citing internal growth numbers, but experts believe any missing interactions will not greatly affect the overall situation.</li>
</ul>
<p><a href="https://medium.com/nlplanet/weekly-ai-and-nlp-news-august-28th-2023-9754bb941c51"><strong>Read More</strong></a></p>