. . AI: Cheaper, Faster, Smaller and Smarter Models (2.13.24)

A universal forecasting model

Friends, there was so much news in AI today that we didn't cover the late breaking news of Andrej Karpathy leaving OpenAI (we may tomorrow). But there's lots more big news and new resources covered below.

It really feels like we are talking about the release of a new model every day lately. Today is no different with 3 new models and 2 significant improvements to existing ones. From being able to remember things better to being able to predict the future, today's edition has you covered!

Here’s today’s news.

-Marshall Kirkpatrick, Editor

First impacted: Enterprise users, OpenAI customers
Time to impact: Short

OpenAI says it is testing memory enhancements for ChatGPT to improve its ability to recall past interactions and tailor future conversations according to user preferences. The feature, which is being rolled out to a small portion of ChatGPT free and Plus users, enables ChatGPT to remember user-specific details, with a broader launch planned for the future. According to the company, GPTs will also have their own distinct memory and developers will have the option to enable memory for their GPTs. [Memory and new controls for ChatGPT] Explore more of our coverage of: OpenAIChatGPTAI Memory Enhancements. Share this story by email

First impacted: Artificial Intelligence Developers, Graphic Designers
Time to impact: Short

StabilityAI has launched Stable Cascade, a new text-to-image model, which the company says significantly reduces costs by 16 times when compared to similar-sized Stable Diffusion models. The model was built on the Würstchen architecture, which was released in a research paper on Arxiv in September 2023. It reportedly excels in prompt alignment and aesthetic quality based on several tests against models such as Playground v2 and SDXL. [Introducing Stable Cascade — Stability AI] Explore more of our coverage of: StabilityAIText-to-Image ModelCost Efficiency. Share this story by email

First impacted: Content Creators, Cybersecurity Professionals
Time to impact: Medium

Hugging Face has compiled and shared a collection of tools aimed at identifying and preventing the proliferation of AI-generated "deepfake" content. Among these are watermarking tools and techniques known as "Photoguard" for protecting images from generative model manipulation, LLM Watermarking, and image "poisoning," which is a strategy intended to disrupt the development of facial recognition models. [via @mmitchell_ai] Explore more of our coverage of: Deepfake PreventionAI WatermarkingFacial Recognition Disruption. Share this story by email

First impacted: Salesforce AI users, Data Scientists
Time to impact: Medium

Salesforce AI Research has announced the launch of Moirai, an AI model trained to do universal forecasting. They say it was trained on Large-scale Open Time Series Archive (LOTSA), a massive open time series dataset with 27 billion observations across nine domains, and has outshone other models in forecasting tests. Fields experimented in include forecasting in energy, transport, climate, sales, economics, and healthcare. [Unified Training of Universal Time Series Forecasting Transformers] Explore more of our coverage of: Salesforce AIMoirai ModelTime Series Analysis. Share this story by email

First impacted: AI developers, Edge Compute
Time to impact: Medium

Reka AI has launched a new language model, Reka Flash, which they say rivals larger models such as Gemini Pro and GPT-3.5 in a number of benchmarks, including tasks involving multiple languages and visual inputs. It does this despite the smaller 21B parameters model size which is many times smaller than GPT-3.5 and Gemini Pro. Reka AI has further expanded its offerings with Reka Edge, a compact 7B model suitable for local deployments and latency-sensitive applications, which they say outperforms similar scaled models like Llama-2 7B and Mistral 7B on standard language benchmarks. [Reka Flash: An Efficient and Capable Multimodal Language Model - Reka AI] Explore more of our coverage of: Reka AILanguage ModelsMultilingual AI. Share this story by email

That’s it! More AI news tomorrow.