World’s Smartest 34B LLM is here 😎

Today’s top AI Highlights:

1. Abacus AI releases LLM with Better Math and Reasoning Abilities

2. New GPT-4 Model to Tackle Laziness

3. First-ever Multimodal AI Model to Control Lucid Dreams

4. Code, Prototype, Train and Deploy AI Apps - All in one platform

& so much more!

Read time: 3 mins

Latest Developments 🌍

LLM with Better Math and Reasoning Abilities 🔢

Abacus.AI has released MetaMath-Bagel-DPO-34B which represents a focused effort to elevate the capabilities of LLMs in mathematical and reasoning tasks. Rooted in the foundational technology of the Bagel models by Jon Durbin, which are enhancements of Yi 34B and Mixtral, MetaMath-Bagel-DPO-34B is designed to not only excel in specialized tasks but also to maintain a high level of performance across a range of other benchmarks. The model ranks second in the 34B category on the Open LLM Leaderboard.

Key Highlights:

1. The model development incorporates innovative training methodologies. The process began with supervised fine-tuning (SFT) using specialized datasets, and then employed an interleaving technique with DPO stages. This approach was pivotal in elevating the model's capabilities in mathematical reasoning without compromising its performance in other areas.

2. A critical element in the model's training was the use of the MetaMathFewshot dataset. This new version of the MetaMath dataset introduces a few-shot learning scenario, where a random number of question-answer pairs precede the target question. This design helps the model better understand and adapt to few-shot prompting, a common interaction pattern for end-users of LLMs, thereby directly targeting and improving the model's math and reasoning abilities.

3. In terms of performance, the MetaMath-Bagel-DPO-34B model demonstrates a commendable balance across various benchmarks. While the model achieves a 13% improvement in GSM8K scores, it also maintains a high level of performance across other benchmarks.

OpenAI’s New Models with Reduced Pricing 💸

OpenAI has announced a significant update across its suite of models and developer tools, including new embedding models with reduced prices, a new GPT-4 model in preview to tackle “laziness”, reduced pricing for GPT-3.5 Turbo, and advanced API management features. Notably, this update introduces highly efficient embedding models and offers substantial improvements in terms of performance and cost-effectiveness.

1. Embedding Models: Two new embedding models: text-embedding-3-small and text-embedding-3-large, offer a notable performance upgrade over its predecessor, with a MIRACL score improvement from 31.4% to 44.0% and an MTEB score increase from 61.0% to 62.3%. This model is priced 5x lower than the previous generation, at $0.00002 per 1k tokens. The larger model creates embeddings up to 3072 dimensions, improving the MIRACL score to 54.9% and MTEB score to 64.6%, priced at $0.00013 per 1k tokens.

2. GPT-3.5 Turbo Update: The new GPT-3.5 Turbo model, gpt-3.5-turbo-0125, will feature a 50% reduction in input prices to $0.0005 per 1K tokens and a 25% reduction in output prices to $0.0015 per 1K tokens. This model includes enhancements for higher accuracy in requested formats and fixes a text encoding issue for non-English language function calls.

3. GPT-4 Turbo Preview: The updated GPT-4 Turbo, gpt-4-0125-preview, focuses on improving task completion, specifically addressing issues of "laziness" in the model's responses. It also resolves a bug affecting non-English UTF-8 generations. OpenAI plans to launch GPT-4 Vision in general availability in the coming months.

4. New Developer Platform Features: OpenAI is enhancing its platform with new API management tools, allowing developers to assign permissions to API keys and track usage metrics at an API key level. These features aim to provide developers with more control and visibility over their API usage.

Transformer for Lucid Dreams 🛌

Ever wondered if you could control your dreams? Prophetic's new creation, Morpheus-1, is turning that fantasy into a possibility. Morpheus-1 is the first multi-modal generative ultrasonic transformer, designed specifically to induce and stabilize lucid dreams. This innovative model operates differently from traditional LLMs, using brain states as inputs to generate ultrasonic holograms for neurostimulation. It's a 103 million parameter transformer model, trained over two days on 8 GPUs.

Morpheus-1's approach contrasts with the prevalent methods in ultrasound research, which typically involve collecting fMRI data and using statistical analysis or basic machine learning. This new model aims to capture a broader spectrum of neural activity, likening its function to creating a symphony rather than playing a single note. Morpheus-1 has the potential for practical applications like enhancing creativity, providing therapeutic benefits in mental health, and even aiding in the study of consciousness.

Tools of the Trade ⚒️

1. Lightning AI Studio: A cloud-based AI development platform offering an integrated environment for coding, , prototyping, training, and deploying AI web apps with zero local setup, leveraging cloud GPUs and unlimited storage for scalable, efficient AI project management.

1. style-diff-500 TTS by Neets: New text-to-speech API that is 97% cheaper than ElevenLab’s API, at 0.005/1000 characters. It’s also 3-6x faster: 250ms vs. 1300ms for Eleven “turbo”/2300ms for main.

2. AI Filesystem (AIFS): Designed for local semantic search, offering a straightforward way to search and organize files based on their content and context rather than just filenames.

3. KPI builder GPT: Designed to assist startup founders in identifying the most relevant and impactful Key Performance Indicators (KPIs) for their businesses.

Hot Takes 🔥

1. In 4 years, a child has seen 50 times more data than the biggest LLMs. ~ Yann LeCun

2. AI isn't catching up to "us", because AI makes us smarter. Any externalized cognitive device we invent (calculators, chess engines, LLMs...) becomes a tool for us to use -- a part of our "greater brain". ~ François Chollet

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!