Briefing on AI Developments: Mistral AI, Record-breaking ChatGPT, and Thinking Language Models
Mistral AI Unveils New Language Models
French AI innovator Mistral AI announced the launch of two new language models, Ministral 3B and 8B. These models are engineered for deployment on devices and in edge-computing scenarios, boasting the ability to handle context lengths up to 128,000 tokens. They excel in tasks like on-device translation, offline assistants, local data analytics, and autonomous robotics. The models reportedly outperform comparable ones in benchmarks, and the Mistral 3B even outshines its predecessor, the Mistral 7B. While APIs give access to these models, the weights for research purposes are available for Ministral 8B Instruct.
ChatGPT Hits New Milestones
OpenAI's ChatGPT has achieved a new record, exceeding 3.1 billion visits in September 2024, which marks a 112% increase from the previous year. This milestone puts ChatGPT ahead of its competitors and makes it more popular than Bing in the U.S. The mobile app is also seeing a surge in monthly active users, partly due to OpenAI's expansion of free offerings to include top models with limited image generation and access to its new logic model \textit{o1}. Moving forward, OpenAI aims to position ChatGPT as a competitive search engine, though it faces the challenge of converting free users into paying customers.
Usernames Impact ChatGPT's Responses
Research by OpenAI indicates that usernames can systematically influence ChatGPT's responses, a phenomenon termed "First-Person Bias." For instance, in story-writing tasks, female-associated usernames tended to lead to narratives with female protagonists and more emotion, while male-associated usernames resulted in darker plotlines. Though slight biases are observed with ethnically associated names, through reinforcement learning, ChatGPT's newer versions have reduced such biases to as low as 0.2%.
Aleph Alpha and AI in Public Services
The Federal Employment Agency of Germany is set to invest up to €19 million in AI products and services from the startup Aleph Alpha. Over four years, this contract aims to enhance processes, increase staff productivity, and mitigate workforce reduction due to retirements. AI will primarily aid in automating administrative processes, document preparation, application assessment, and benefitting knowledge management and customer consulting. Notably, human staff will continue making final decisions. Meanwhile, Aleph Alpha revealed that its AI assistant F13, now available nationwide, offers features such as research, fact-checking, translations, and text generation for public authorities.
New Methods in AI "Thinking"
A collaborative team from Meta and academic institutions has created a novel approach called "Thought Preference Optimization" (TPO) for improving large language models. This method fosters "thinking" before answering, enhancing model performance in diverse tasks, not limited to just arithmetic or logic. Unlike traditional methods requiring extensive data, TPO leverages human-like thought processes to iteratively refine model outputs, resulting in an elevated performance in areas such as general knowledge and marketing, albeit with a slight drop in problem-solving related to math.
Combating Spam with AI
German email providers GMX and Web.de are now filtering a massive 1.9 billion emails weekly as spam, predominantly phishing attempts, marking an increase from 1.4 billion in the previous quarter. These fraudulent emails often masquerade as messages from parcel services, demanding fees for supposedly held packages. In response, these providers deploy AI to quickly assess the volume of emails being sent from a certain server to determine potential spam, effectively relegating 99.9% of malicious emails to spam folders.
This news roundup is originally from heise online.