It’s August already, so hopefully this edition finds you on a beach 🏖️ somewhere, enjoying some well earned time off!
The biggest announcement of last month was the introduction of the Llama 3.1 family of models, the largest of which is on par with GPT4o, whilst the smaller ones are the best performing models of their category 🔥
In this issue, we also cover a real case study with a retail client of ours, on building an AI shopping assistant, and we share our thoughts around the role generative AI can play in healthcare 🩺
📰 Bites from AI news
✍️ Gradient Labs published a really useful blog post for building agentic workflows. Instead of multiple agents collaborating (which is an approach that is quite common in research papers and some startups) they use a more deterministic flow that combines LLM calls in a modular and auditable, making developing and debugging those workflows easier and less risky 🚀
✍️ Explosion published a practical guide on developing AI models using LLMs to prototype, humans to course correct, and finally training small 🗜️, efficient 🏎️ and more cost effective 💰 and performant 📈 models tailored to your task.
🧪 A new paper entitled “AgentInstruct: Toward Generative Teaching with Agentic Flows” was published on arXiv. The paper features a new recipe for tuning LLMs completely from synthetic data and raw documents by employing a collection of agents or LLMs tasked with formatting the text to increase quality, generate diverse instructions and make the data more complex 💪
🧪 Meta introduced Llama 3.1 🦙, coming in 3 sizes 8B, 70B and 405B. The first two are best open models in their respective categories including distilled versions like Phi and Gemma while the last is on par with GPT4o 🔥
🖱️ Deploying multiple Lora adapters is now even easier thanks to huggingface 🤗 after the introduction of a multi lora adapters workflow in its services and libraries. The economics of multiple Lora adapters are very compelling to the extent that this what Apple used to pack multiple LLM based models into its newest iPhones 📱
✍️ Product recommendations with AI agents
We worked with a big retail company to develop a conversational shop assistant that can recommend products based on what the consumer is looking for, all while ensuring its responses are safe and unbiased. The assistant is built using an agentic workflow consisting of components to retrieve, rank and evaluate among others, with some additional guardrails.
You can read the entire case study here.
🩺 Generative AI for healthcare
We believe generative AI can play a transformative role in healthcare by providing up to date and relevant information from trusted sources to both doctors and patients, accelerating the path to new drugs by assisting in clinical trial designs and suggesting potential compounds, improving education by creating content and simulated environments for doctors to practice. Read more in our latest blog https://medium.com/mantisnlp/generative-ai-for-healthcare-1b12e4d715e4