When ML meets Product: November ’24 AI Product Updates
Keeping up to date with new AI models, products, use cases, trends, and resources
Welcome to the November edition of “When ML Meets Product — AI Product Updates”. As the Artificial Intelligence field rapidly advances, staying up-to-date with the latest developments and understanding their impact on business and products has never been more essential. This edition brings you November’s top updates:
- GenAI Model Updates: Latest breakthroughs from major foundational model players.
- AI Product Updates: Product trends and cool new products that leverage AI.
- AI Ethics and legislation Updates: Important news and relevant resources to stay updated with AI Ethics and AI legislation topics
- Other Relevant Resources: Stay informed with additional resources.
GenAI Model Updates
Lately, there is a lot of controversy around perceived diminishing returns in generative AI models. This refers to the reduced performance gains achieved in the newest model’s versions despite increasing computational and resource investments. A Reuters report notes that OpenAI and its rivals are exploring new paths to smarter AIs as current approaches hit their limits, while The Batch discusses how new models show only incremental improvements despite growing complexity, training data, and costs.
Still, we continue seeing a lot of news from the main GenAI model providers:
- Anthropic: introduces Model Context Protocol (making it easier to connect securely to data sources, such as Google Drive, Slack, and Github), styles (formal, concise, explanatory), and preferences (similar to ChatGPT’s memory, allows the user to state context and characteristics she wants to preserve over all interactions).
- Mistral: launches their last multimodal Pixtral Large, while adding new features to their platform Le Chat (web search, analyzing documents…).
- Runway: introduces Frames (a new base model for image generation, focused on maintaining stylistic consistency and further creativity) and Act-One (generation of expressive character performances using video and voice as inputs).
- Nvidia: introduces Fugatto, a generative AI sound model that allows users to create high-quality, realistic audio content (music, voices, and sound) for applications like prototyping music productions, marketing, education, and more. This includes creating audios from 0, and editing existing ones (e.g. “isolate the voice from this song”).
AI Product Updates
Extending LLMs to search use cases is yet another trend we are seeing (with the news a while back that OpenAI was launching SearchGPT). This month:
- Meta is also developing an AI search engine
- Perplexity is adding a “shops” feature in their LLM-based searcher (one-click checkout, visual search, helps find and compare for the best product…).
- Google allows asking Maps for inspiration (curated with Gemini), with examples such as “things to do with friends at night”, summarizing reviews, and enabling to ask specific questions about specific places in natural language.
Differentiation in the market won’t come from merely using LLMs, as these are now accessible to everyone, but from the unique experiences, functionalities, and value products can provide through them (if we are all using the same foundational models, what will differentiate us?, carving out your competitive advantage with AI). And what are these unique experiences companies are trying to build through AI?
- Cool Product #1: A fun product (not sure if actually useful against scammers but at least great for marketing and brand purposes), AI granny by O2, an adorable LLM-based grandma who will make scammers lose time and patience.
- Cool Product #2: Marketing and advertising are sectors that have already been greatly disrupted with GenAI. This time, instead of general products for the sector, we see a specific product for a specific application (food market). Ai Palette “generates new concepts for brands based on identified trends, transforming traditional market research in the process”.
- Cool Product #3: JuristAI wants to revolutionize law-related work, by providing dedicated solutions to help augment attorneys.
- Cool Product #4: some products are trying to lower the barrier to use AI to automate and boost specific tasks. Writer “builds generative AI into any business process with Writer’s secure enterprise platform”, while Gumloop “automates any workflow with AI through drag / drop / deploy tools”.
AI Ethics and Legislation Updates
On ideologies and biases
A recent paper proves how Large Language Models reflect the ideology of their creators (which makes sense due to the relationship responses can have with the training data used + the feedback loop during reinforcement steps of the training), and how the same model can reflect differences based on the language it is queried (again probably due to differences in training data available for each language). Having sensibility around this is key to ensure a responsible selection of the model to use in your products, and mitigate risks related to biases and inconsistencies with multi-language use cases!
To tackle this, several frameworks try to evaluate and benchmark LLMs based on ethical principles. For example, Compl-AI evaluates LLMs on the principles stated by the EU AI Act, and LLM Observatory offers metrics and benchmarks on social bias metrics.
Human oversight and supervisions are one of the strategies thought to prevent discrimination due to biased predictive models. However, a recent paper showed how, for AI-aided decision-making in sensitive tasks, human oversight did not prevent discrimination. Thus, the authors recommend a comprehensive approach when designing oversight systems.
On privacy and memorization
On the relationship of LLMs and privacy, an interesting series explores the potential use of personal data in the training of different LLMs. On a fun turn of events, the fact that ChatGPT wasn’t able to answer “who is David Mayer” triggered a lot of discussions on why was that the case (maybe someone from the EU asking the company to be forgotten?). I’ve recently tested the query and seems to be running fine though!
Ensuring people’s privacy also in the training datasets of LLMs is important due to the memorization capability of these models (and the fact that they can just spit out anything if properly prompted by the final user). For more on this topic, check out this great article “Deep learning memorization, and why you should care”.
On the impact to employment
A recent paper explores the potential of AI transforming work processes and flattening hierarchies in the knowledge economy (it was found that the use of AI benefitted especially those with lower abilities). Also, how process moved towards more autonomous work and increased exploration activities.
Other relevant resources
AI Agents are the new hype, and Langchain has just released some great insights about the use of AI agents based on responses from over 1300 professionals. We can see high adoption rates, different types of controls (tracing / observability, guardrails, evaluation…), permissions these agents have (read / write / …), and the main perceived limitations and barriers for these systems.
Another great survey, in this case from Slack, explores the willingness of companies to incorporate AI in their processes, how many workers are looking forward to being augmented with AI (but feel uncomfortable sharing they use AI in their work-related tasks), and how up-skilling in AI is a shared priority for both companies and employees.
Great overview from Ravi Mehta on how we arrived to the current “Learning Era”, where computers can learn from data, how this extended the type of problems that can be solved (“we’re not longer limited by what we can code”), and what is the right mindset to build products in this new era.
And last but not least, an amazing talk by Claire Vo, “Product Management is Dead”: how AI can greatly accelerate certain product management tasks (e.g. documenting a product strategy), but at the same time augment you to do more, more types of tasks, greater impact!
Wrapping it up
That was it from When ML meets Product — November’24 AI Product Updates. Hope you enjoyed the read!
I’ll be happy to hear your thoughts, questions, or suggestions.
And if you want more content, check out my latest blogpost “GenAI is Reshaping Data Science Teams — Challenges, opportunities, and the evolving role of data scientists”. It covers the impact of the GenAI revolution to Data Scientists and Machine Learning teams, and why they remain indispensable to guide organizations toward leveraging AI effectively and responsibly.
