Scaling Down LLMs for a Safer Future ⏬

Today’s top AI Highlights:

1. Microsoft’s Textbooks-only Approach to Build a Small yet Mighty LLM

2. Expanded AI Features in Slack

3. Any-to-Any Multimodal LLM

4. New Feature in Runway and Pika Labs for Image-to-Video

& so much more!

Read time: 3 mins

Latest Developments 🌍

How Small can a Capable LLM be? 🤏

As we witness the growth of LLMs primarily stemming from scale with the most powerful models nearing trillions of parameters and tokens, this poses problems in cost of training, deploying, maintaining and environment impact, as well as hallucinations, toxic content generation, and bias. Amidst these issues, Microsoft has released a technical report introducing phi-1.5, a 1.3 billion parameter language model.

Key Highlights:

• The model follows the "Textbooks Are All You Need" approach, achieving remarkable results in common sense reasoning, surpassing models 5x its size on complex tasks like mathematics and coding.

• The absence of web data in phi-1.5's training dataset shows promise in reducing issues related to toxic and biased content generation.

• phi-1.5 is open-sourced, empowering the research community to explore in-context learning, interpretability, and strategies to address AI model challenges.

Slack Supercharged with AI 🚀

Salesforce has announced powerful AI capabilities to Slack aimed at enhancing productivity and transforming teamwork.

Key Highlights:

• Slack AI: Delivers instant channel recaps, thread summaries, and search answers, streamlining work and saving time.

• Workflow Automation: With improved Workflow Builder and custom app development, users can automate tasks effortlessly, connecting various tools and centralizing automation resources.

• Slack Lists: Empowers users to manage and track work within the flow of communication, making project tracking and task prioritization more efficient.

Any-to-Any Multimodal LLM 🔃

Researchers at NExT++ Lab and National University of Singapore introduce NExT-GPT, an any-to-any multimodal LLM, enhancing AI's ability to understand and communicate across diverse modalities.

Key Highlights:

• NExT-GPT seamlessly combines an LLM (Vicuna) with multimodal adaptors and diverse diffusion decoders to process inputs and generate outputs in text, images, videos, and audio, mirroring human-like multimodal communication.

• It efficiently leverages pre-trained encoders and decoders, and requires minimal parameter tuning (just 1% of certain projection layers), making it cost-effective and adaptable for potential expansion into additional modalities.

• Through modality-switching instruction tuning, NExT-GPT enhances its capabilities and controllability.

Jumping the Obstacle 🧗

Researchers have introduced a cutting-edge system that enables low-cost robots to autonomously navigate complex environments through vision-based parkour learning.

Key Highlights:

• Robots can now learn diverse parkour skills through an end-to-end vision-based policy, which is transferred to a quadrupedal robot equipped with an egocentric depth camera, eliminating the need for reference motion data.

• The system utilizes a novel reinforcement learning approach inspired by direct collocation to generate versatile parkour abilities, from climbing high obstacles to squeezing through narrow slits.

Google’s Digital Futures Project 💡

Google is launching Digital Futures Project, a global initiative aimed at addressing AI's opportunities and challenges. Google.org will establish a $20 million fund, providing grants to leading think tanks and academic institutions worldwide to promote responsible AI development and research.

Tools of the Trade ⚒️

• Runway and Pika Labs: Director Mode in GEN-2 and Camera Movement Parameter in Pika Labs now offer more control on the direction and intensity of movement while converting an image to video.

• NoiseGPT: A decentralized AI platform that provides unbiased and uncensored generative models for hyper-realistic text-to-speech, dialogue bots, and voice cloning capabilities.

• QuillO: Unlock the potential of your data by transforming it into dynamic knowledge graphs and create context-aware content.

• M1-Project: Use AI to create detailed Ideal Customer Profiles, including company details and buyer personas, and save time to research your ideal customer.

• ClearCypherAI: Real-time voice intelligence platform offering audio-to-text, text-to-audio, AI audio-to-audio, finetuned GPT models, voiceprint and synthesis, and more.

Hot Takes 🔥

• We’ll build AGI with Python. Think about that for a second. ~ Santiago

• March 2023: GPT-4 blows your mind. March 2024: GPT-4 is the bare minimum. ~ Pete

• If your goal in life is to maximize money, do LLMs over quant salary for all levels for LLMs is above quant Major talent shortage rn for AI ~ Owen

Meme of the Day 🤡

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!