(Source)

Given my current constraints, I can’t publish GenAI Short & Sweet daily. And so I recommend subscribing to the Rundown AI as a valuable supplement. Unlike our weekly blend of prime news, in-depth analyses, and occasional demos, the Rundown AI offers frequent, concise updates.

Stay up-to-date with AI.

The Rundown is the world’s fastest-growing AI newsletter, with over 500,000+ readers staying up-to-date with the latest AI news, tools, and tutorials.

Our research team spends all day learning what’s new in AI, then distills the most important developments into one free email every morning.

Sign up today.

(Source)

Mistral's new open-source AI model, "Miqu-1-70B," has been leaked; confirmed by Arthur Mensch, CEO of Mistral, it is attributed to an overzealous employee of an early access customer. The model appeared on platforms like HuggingFace and 4chan.

Tech Background

Benchmarks suggest Miqu-1-70B surpasses existing open-source models in capability, setting a new standard for open-source AI. The model was quantized (reducing precision from 32-bit to 16-bit or even 8-bit parameters), indicating Mistral's focus on optimizing AI efficiency without significant performance loss. The model's retraining from Llama 2, utilizing Mistral's entire cluster, shows a commitment to continuous enhancement.

We will continue to monitor this space, especially Mistral, for further developments. 🥸 

(Source)

Unimpressed by the Mistral-Miqu-Leak discussion, a subtle development has occurred in the shadows. Abacus AI's Smaug-70B-v0.1 has stealthily jumped the ranks, with an average score of 80.48 across key benchmarks like ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, and GSM8K, positioning it significantly ahead of its nearest competitor, MoMo-72B-lora-1.8.7-DPO. The latter, enhanced via Direct Preference Optimization (DPO) from its precursor and various hyperparameter optimizations, scores 78.55.

In comparison, Miqu registers at 76.59…. 76.59! What is all that buzz about? 🤷 Well, the Mistral community believes strongly in the model’s performance trajectory.

I am optimistic about the future of open-source LLMs due to the increasing competitiveness in the field. See the graph below, which demonstrates a significant uptrend in performance, breaking away from previous plateaus.

Think different? Join the discussion on LinkedIn or X.

(Source)

OpenAI is developing an early warning system to assess LLMs' potential to facilitate biological threat creation (e.g., viruses, toxins, agricultural threats, etc.).

Initial evaluations indicate that current models, including GPT-4, assist with such misuse. Specifically, on a 10-point scale measuring the accuracy of responses, they observed a mean score increase of 0.88 for experts and 0.25 for students compared to the internet-only baseline.

The main concern is the increased access to information on biothreat creation. This groundbreaking evaluation, the first of its kind, is a foundation for further research and monitoring in this critical area.

(Source)

Midjourney just dropped a new image generation model called Niji v6; the results are stunning.

It is a significant addition to Midjourney's suite of tools, designed to create rich, detailed, faithful anime characters, backgrounds, and scenes.

This is Anime Martin (uploaded a foto) - for some reason, with a skirt:

Turn it on with the following parameter

(Source)

Hugging Face launches HuggingChat, allowing users to build personalized chatbots in two steps, similar to OpenAI’s GPTs.

Customize it with a name, avatar image, and description, choose from LLMs like Llama2 or Mistral, set behavior with system messages, and select message starters. Also, very similar to OpenAI’s GPTs. Just click here to get started.

Haha, my favorite is Italian Grandma. So accurate! 🧺 

The Apple Vision Pro, a mixed-reality headset, was officially launched in the United States on February 2, 2024. This might be a defining moment not only for VR but also for potentially all applications.

As it opens up a whole new world for generative AI applications, we would like to shed some light on already existing Vision Pro applications.

What we have found:

Learning will never be the same.

Watching sports (here F1) will be so much better.

Much more immersive selling, here a Cybertruck experience.

Not there yet, but cooking will eventually be an effortless job for everyone.

Meditation/ exploration: be anywhere in the world.

New dimensions for artistic expression, i.e. DJing.

As I am writing it, it is day 2 of the Apple Vision Pro.

I am on a positive high, thinking about the potential this has. What apps will come, and what will GenAI unlock?