AI Powered Tools

A run down of the latest SaaS products and services which leverage AI to you take back some time.

• ElevenLabs - for media and content delivery, unleash the power of cutting-edge technology to generate realistic, captivating speech in a wide range of languages.

• Scenario - for game designers who want to create high quality game assets quickly. Maintain a consistent character style through different scenarios and model types

• Otter - for professionals. AI meeting assistant that records audio, writes notes, captures action items, and generates summaries.

News

Microsoft's WaveCoder: Innovating Coding Language Models

Microsoft's research introduces WaveCoder, a cutting-edge model for training efficient coding language models with fewer examples. Complemented by CodeOcean, a dataset of 20,000 diverse code examples, this approach aims to balance cost-effectiveness with quality in coding LLMs. WaveCoder outperforms similar-sized models in tasks like code summarization and repair, showcasing the potential of refined and diverse instruction data in enhancing model efficiency. Although not yet released, discussions suggest a possible future availability of WaveCoder and CodeOcean​​.

Read the full article here

Rapid Rise of Nightshade: A Tool Against AI Artwork Misuse

Nightshade, a tool developed by the University of Chicago, aims to protect artists from AI models that use their work without consent. In just five days, it achieved 250,000 downloads globally, surpassing expectations. It works by subtly altering artwork to mislead AI algorithms, a method termed "poisoning." The high demand overwhelmed the university's servers, leading to the provision of mirror links. Its predecessor, Glaze, has been downloaded 2.2 million times since April 2023. The team plans to combine Glaze's defensive approach with Nightshade's offensive strategy, but this requires thorough testing and is expected to take at least a month. Additionally, an open-source version of Nightshade is being considered.

Read the full article on The Verge.

ChatGPT Enhances User Experience with Direct GPT Invocation

OpenAI has introduced a new feature for ChatGPT paid users, allowing them to directly invoke third-party GPTs within chats by simply typing "@" followed by selecting a GPT from a list. This update, aimed at increasing GPTs' discoverability, provides users with the ability to add GPTs to conversations for diverse use cases, each with an understanding of the conversation's full context. Following the recent launch of the GPT Store, this feature encourages the development and utilization of custom GPTs without requiring coding expertise, although challenges such as moderation and traffic growth remain. For more details, visit TechCrunch.

The Dangers of AI-Generated Internet Content

The Mashable article by Cecily Mauran, titled "The era of the AI-generated internet is already here," This is a fantasic writeup shining a light on the growing prevalence of AI-generated content on the internet. It highlights the risks associated with such content, including the spread of misinformation and a decrease in content diversity and quality. Despite these challenges, the article suggests that the internet will evolve, with more content moving behind paywalls and an increase in tools to protect against AI-scraping. What do you think? Are we simply seeing an evolution of the Internet or are we headed for an AI buffet of more and more AI content? Read the full article on Mashable.

Large Language Models

No doubt LLMs are the flavour of the month. We like other model architectures too but this is where there is pace, so we need to help you get to the important stuff.

“Leaked” Open Source AI Model Nears GPT-4 Performance

A new open source AI model, miqu-1-70b, has been “leaked”, drawing significant attention in the AI community. This model, posted on HuggingFace by a user named "Miqu Dev," shares prompt formats with Mistral's Mixtral 8x7b and reportedly approaches OpenAI's GPT-4 in performance. Speculations suggest "Miqu" could be a quantized Mistral model. Mistral's CEO confirmed the leak as an old, quantized model from a customer. This development hints at a major shift in AI, challenging OpenAI's dominance with potentially free, open source alternatives.

Read the full article on VentureBeat.

Meta's Code Llama 70B: Closing the Gap with GPT-4 in AI Programming

Meta introduces its most advanced code generation AI, Code Llama 70B, labeled as their "largest and best-performing model" to date. This model, an upgrade from previous versions, boasts improved query handling and accuracy, outperforming GPT-3.5 in the HumanEval benchmark. Code Llama 70B, built on Llama 2, aids developers in generating and debugging code. Alongside, Meta released specialized tools like Code Llama - Python and Code Llama - Instruct. Despite competition from Amazon's CodeWhisperer and Microsoft's GitHub Copilot, Meta's Code Llama 70B remains free for research and commercial use, and is accessible on Hugging Face with GPU support.

Original article: The Verge

Social Spotlight

Spotlight on some social content which is a must and should be shared. Great YouTube videos, insightful X posts. Whatever it is, you should be getting right to it.

Today we highlight some more great work from Matthew Berman (@MatthewBerman) with his latest video on fully local copilot.