- Starter AI
- Posts
- Mistral’s Large 2, Bing gets an AI boost, Are AI models close to collapsing?
Mistral’s Large 2, Bing gets an AI boost, Are AI models close to collapsing?
More open models are coming our way!
Hello, Starters!
It is incredible to think about the times we're living in. Since AI became more prominent in the industry, there hasn't been a day when we took a break from innovation, and we're just at the very beginning!
Here’s what you’ll find today:
Mistral presents Large 2
Bing steps up its AI game
AI models are prone to collapsing
OpenAI’s Rule-Based Rewards for model safety
Stability AI unveils Stable Video 4D
And more.
💥 Mistral presents Large 2 (3 min)
With open source having its moment in the industry, another big player has decided to make a move too. Mistral has unveiled Large 2, a 123B parameter model with a 128k context window and multilingual support that also includes more than 80 coding languages like Python, Java, and C. However, its "openness" is restricted to non-commercial use.
The benchmark results shared by Mistral showcase the model nearly matching Meta's Llama 3.1 405B. Large 2 has enhanced performance in tasks like reasoning, code generation, and maths.
🔍 Bing steps up its AI game (1 min)
Microsoft has given Bing, its popular search engine, a new AI revamp. They're introducing a twist to the search experience by mixing Gen-AI and LLMs. The feature is being launched on a small batch of user queries and includes an AI-generated response, a link to the source, and regular search results.
After Google's attempt at AI-powered search didn't go well, it's not surprising that fellow tech companies are trying the approach with a bit more caution.
🤖 AI models are prone to collapsing (17 min)
Model collapse has been a recurrent topic as AI technologies keep on developing. Scientists are constantly sharing insights into what would happen if the data on which models are trained became redundant to the point where it starts affecting them. In other words, as AI-generated text becomes more prominent online, and companies keep getting their data from websites, models are prone to affecting themselves with their own content.
This is where the importance of valuable data resides, and why we're also witnessing plenty of attempts by most tech giants to obtain it from different sources. The effects of "model collapse" are something that must be on the lookout for as technology keeps advancing.
📄As part of its commitment to transparency and responsible development, OpenAI has shared research on Rule-Based Rewards, a method that they've been employing in the training of GPT-4 to GPT-4o Mini to ensure that they're aligned with their safety demands. RBRs are based on step-by-step rules to monitor the model's outputs.
🎥Stability AI revolutionises video generation with the launch of Stable Video 4D, standing apart from other models with its ability to process video input and then create new videos from different angles based on the original one. This tool could be a great aid in fields like gaming or AR/VR.
⚡️Quick links
What did you think of today's newsletter? |