. . AI: New models, acquisitions and bugs (2.22.24)

Google, OpenAI, Jasper, Fireworks

Author

Marshall Kirkpatrick
February 22, 2024

Friends, today we cover a number of shorter stories that explore acquisitions, improvements in performance, efficiency, and routing. The acquisition by Jasper is an interesting one as it highlights the rapidly evolving AI landscape and the expansion of capabilities to maintain competitiveness (and relevance).

Here’s today’s news.

Marshall Kirkpatrick, Editor

Fireworks Launches FireFunction V1 Incorporating 3rd Party Functions & Mixtral

First impacted: Software Developers, Data Scientists
Time to impact: Short

Fireworks' FireFunction V1 API offers the ability to access real-time data via 3rd party APIs and produce structured outputs using a Mixtral-based model. They say compared to GPT4 it's as much as four times faster. Dynamic agents that take multiple steps, incorporating data from other services, are highlighted as a use case. [Function calling] Explore more of our coverage of: API Launch, Dynamic Data Access, OpenAI Compatibility. Share this story by email

First impacted: Corporate marketing teams, Business API users
Time to impact: Short

Jasper has acquired Clipdrop, an AI image creation tool, from Stability AI, with the aim of boosting its offerings to corporate marketing teams, according to a blog post by Jasper. Based in Paris, Clipdrop "enables users to edit existing images, create new ones, and develop an expansive number of variations in size, detail, and style." [Jasper Expands by Acquiring Image Platform Clipdrop from Stability AI] Explore more of our coverage of: AI Acquisition, Corporate Marketing, Image Creation Tool. Share this story by email

Stability AI Previews Text-to-Image Model, Stable Diffusion 3

First impacted: AI developers, AI researchers
Time to impact: Short

Stability AI is preparing for the launch of its new text-to-image model, Stable Diffusion 3, which the company says offers "greatly improved performance in multi-subject prompts, image quality, and spelling abilities." The model is currently open for early preview sign-ups, ahead of a detailed technical report from the company. [Stable Diffusion 3 — Stability AI] Explore more of our coverage of: AI Models, Text-to-Image Technology, Stable Diffusion 3. Share this story by email

First impacted: Librarians, Digital Archivists
Time to impact: Short

Simon Willison has posted a blog post showcasing the capabilities of Gemini Pro 1.5 by getting it to accurately identify books on his bookshelf from a brief, fast-moving video. While it had some hiccups, its ability to handle diverse inputs, including variable bitrate videos, is remarkable. [The killer app of Gemini Pro 1.5 is video] Explore more of our coverage of: Google AI, Video Processing, Advanced Modeling. Share this story by email

First impacted: AI developers, AI model users
Time to impact: Short

On February 20, 2024, OpenAI's LLMs encountered a bug leading to incorrect word sequences, according to a release by the company. The issue was traced back to the model incorrectly selecting numbers that map to tokens, resulting in strange word sequences. The problem was especially noticeable when inference kernels were used in certain GPU configurations. OpenAI reports that the bug has since been rectified. [Unexpected responses from ChatGPT] Explore more of our coverage of: Language Processing Bug, GPU Configurations, User Experience Optimization. Share this story by email