AI/ML news summary: week 35

Here are the articles, guides, and news about AI; Week 35. I read tons of RSS feeds and blogs, so you won't have to scour the internet yourself for the latest AI news of this week.

An explanation in plain English, or my comments, read the quotes.

TL;DR

OpenAI now offers fine-tuning for GPT-4o. This lets developers customize their models for better performance. It has free training tokens available until September 23. Microsoft launched three open-source models in its Phi 3.5 series to improve their multilingual and scientific tasks. OpenAI partnered with Condé Nast to integrate SearchGPT. Whti this act, they want to improve search and content reliability.

AI21 Labs released Jamba 1.5 models. They combined Transformer and State Space Model architectures. Jamba 1.5 Mini is crazy enough, outperforming larger models on benchmarks. Nvidia introduced StormCast. That is an AI model improving weather forecasts by 10%, aiding disaster planning. Anthropic's Claude reached $1 million in app sales in 16 weeks but has a lot of competition as Apple integrates ChatGPT. And Nvidia's Llama-3.1-Minitron 4B, created through pruning and distillation, matches even larger models' performance with greater efficiency.

So in short, this is a crazy week for Short Language Models !

Nous Research published a report on DisTrO, new distributed optimizers that reduce inter-GPU communication significantly. And that boosts multi-location training (I'll explain that below). Amazon's AI tool, Amazon Q, added a code transformation feature, and that is saving 4,500 developer years and $260 million in system upgrades (I wrote about that here: Anticipating AI's next move ? article ③ ? ) . Google DeepMind and Imperial College London developed FermiNet. That is a cool neural network that accurately models molecular energies, which is advancing my favorite topic: quantum chemistry. Jina AI introduced "Late Chunking", and that is a new data retrieval technique that uses contextual embeddings for better search performance.

An explanation in plain English, of all of this mumbo jumbo, down below

Before we start!

If you like this topic and you want to support me:

1. Comment on the article; that will really help spread the word ??
2. Connect with me on Linkedin ??
3. Subscribe to TechTonic Shifts to get your daily dose of tech ??

Latest AI and ML Developments

1. Custom Fine-Tuning for GPT-4o Now Offered by OpenAI

OpenAI has introduced fine-tuning options for GPT-4o. This is only cool for developers because it lets them customize models for specific needs. That means a better performance and less cost. This feature is available to users on a paid plan, with free daily training tokens offered until September 23 (pffff). (Source: Original article on OpenAI's fine-tuning announcement)

2. Microsoft Introduces New Phi 3.5 AI Models

Microsoft launched three new open-source AI models under its Phi 3.5 series: mini-instruct, MoE-instruct, and vision-instruct. These models improve reasoning capabilities for multilingual tasks and scientific research, particularly in handling long documents (like this one). They have challenges with factual accuracy and potential bias. Now Microsoft suggests using these models with retrieval-augmented generation systems for the best results, especially in environments with limited resources. (Source: Original article on Microsoft’s Phi 3.5 models)

In plain English: Microsoft released three AI models in its Phi 3.5 series. These models are good at handling multilingual tasks and long documents but they can struggle with accuracy and bias.

3. OpenAI Partners with Condé Nast for Enhanced Search Features

OpenAI has teamed up with Condé Nast to incorporate SearchGPT into the publisher’s platforms.

This partnership gets OpenAI the much needed dough, and Conde some cool search capabilities and credibility of their content.

The collaboration is a strategic move to help media companies manage the financial impacts of rapid technological changes (read: repay media companies for the stolen content to train their LLM)

(Source: Original article on OpenAI and Condé Nast partnership)

4. AI21 Labs Launches Jamba 1.5 Models for Long-Context AI

AI21 Labs has released a new set of models called Jamba 1.5, which blends Transformer and State Space Model architectures. The series includes two versions: Mini (12B active/52B total) and Large (94B active/398B total) MoE models. Jamba 1.5 Mini is a leader in its class. It has achieved a score of 46.1 on the Arena Hard benchmark. They are moving ahead of larger models like Mixtral 8x22B and Command-R+. The Arena Hard benchmark, measures a model's ability to handle challenging language understanding and reasoning tasks. The benchmark hasn't got a maximum. (Source: Original article on AI21 Labs' Jamba 1.5 models)

In plain English: AI21 Labs released the Jamba 1.5 models, blending Transformer and State Space architectures. The Mini version (12B active) leads its class with a 46.1 score on the Arena Hard benchmark, beating larger models. This benchmark tests how well models handle complex language tasks.

Readers, be honest in the comments.. Do you really care about stuff like the above? Just testing the waters with this little litmus test.

5. Nvidia Unveils StormCast AI for Weather Prediction

Nvidia has introduced StormCast. That is a new AI model on its Earth-2 platform. It wants to improve mesoscale weather forecasting by simulating atmospheric dynamics. This is a lot of blabla, but what it means is that this model improves prediction accuracy by 10% over traditional six-hour forecasts. Which is helping people in more effective disaster planning.

With this release, Nvidia plays the hyprocite, because it wants to build a reputation in AI-powered climate technology (joining the ranks of Google, Microsoft, and IBM). This is because they feel sorry for ruining the climate in the first place with their humongous energy consumption. (Source: Original article on Nvidia's StormCast)

6. Anthropic’s Claude Reaches $1 Million in Mobile App Sales

Anthropic's AI assistant, Claude, has generated over $1 million in revenue from its mobile app on iOS and Android in just 16 weeks. It has seen rapid growth in the U.S., but Claude faces new challenges as Apple plans to integrate ChatGPT directly into iPhones. 16 september, right? (Source: Original article on Anthropic's Claude earnings)

7. Nvidia's Llama-3.1-Minitron 4B Model: Small But Mighty

Nvidia's research team developed Llama-3.1-Minitron 4B by using pruning and distillation to compress the Llama 3 model. This smaller model competes well with larger models and similar-sized small language models and it is far more efficient to train and deploy. (Source: Original article on Nvidia's Llama-3.1-Minitron)

This is a good development, because less training means less energy consumption, because training accounts between 70-80% of all cost of an AI model

8. New Report on DisTrO by Nous Research

Nous Research released a report on DisTrO (Distributed Training Over the Internet). And that is a set of distributed optimizers that are both architecture-agnostic and network-agnostic.

These optimizers reduce inter-GPU communication by 1000x to 10,000x without needing amortized analysis. This breakthrough could be useful for multi-location training in both large tech companies and decentralized, open-source projects. (Source: Original article on Nous Research's DisTrO report)

And now in plain English. Nous Research released a report on their DisTrO. That it has new tools called distributed optimizers. These optimizers significantly reduce communication time between GPUs, making them up to 10,000 times faster. Now this could benefit AI training across multiple locations for both large companies and open-source projects.

9. Amazon Q Improves Software Upgrades with New Code Transformation

Amazon's GenAI tool for software development, Amazon Q, now includes a new feature for code transformation aimed at foundational software hygiene tasks. This update has saved Amazon the equivalent of 4,500 developer years in system upgrades, leading to an estimated $260 million in annual efficiency gains. Over 50% of production Java systems were upgraded to newer versions faster and with less effort. (Source: Original article on Amazon Q's new feature)

In short: Amazon's AI tool, Amazon Q, added a new feature to improve code updates, saving 4,500 developer years and $260 million in system upgrades. It helped upgrade over 50% of Java systems to newer versions more quickly and easily.

10. Google DeepMind Solves Complex Quantum Chemistry Problems

Researchers from Imperial College London and Google DeepMind have proposed an AI-based solution to model molecular states. They developed a neural network called FermiNet to compute atomic and molecular energies with high precision. For the complex carbon dimer molecule, they achieved a mean absolute error (MAE) of 4 meV, five times better than the previous best methods with an MAE of 20 meV. (Source: Original article on Google DeepMind's quantum chemistry research)

In plain English: Researchers created FermiNet, an AI that calculates molecular energies with just a little error, far more accurate than previous methods. This helps quantum chemistry and speeds up drug and material development.

11. Jina AI Develops "Late Chunking" for Improved Data Retrieval

Jina AI has introduced a new technique called "Late Chunking" for embedding data chunks, which enhances retrieval performance by using the contextual information provided by 8192-length embedding models. This method creates chunk embeddings that are conditioned on previous chunks, leading to better context representation. (Source: Original article on Jina AI's Late Chunking)

In plain English: "Late Chunking" improves data retrieval by using context from previous chunks, leading to more accurate and relevant search results for AI models.

Quick learning bytes

1. LLM-Enabled RAG Systems Best Practices: A detailed study on the key components and best practices for Retrieval-Augmented Generation (RAG) systems, a critical use case for large language models (LLMs). (Source: Original study on RAG systems)
2. Training a Workforce on Generative AI: Companies like Synechron and USAA are focusing on training their teams on generative AI to drive innovation and manage risks effectively. (Source: Original article on AI training in companies)
3. AI for Automating Bug Reports: A team built an AI solution to automate bug report creation by extracting messages from Discord, summarizing them with Google Gemini, and adding them to GitLab. (Source: Original article on AI for bug reports)
4. Understanding Linear Regression Coefficients: This post explains how to interpret coefficients in linear regression models, covering scenarios involving both numerical and categorical variables. (Source: Original article on interpreting regression coefficients)
5. Introduction to ggml for Machine Learning: An overview of ggml, a machine learning library focused on transformer inference, written in C and C++. (Source: Original article on ggml library)

And no, GGML does not stand for GarGaMeL (GenZ: Smurfs), although I would very much like it to be

Tools

• Phi-3 CookBook: The official repository for Microsoft’s Phi-3 models, the most cost-effective SLMs.
• Cursor: An AI-powered code editor designed to improve developer productivity. Even an eight year old could code with this thing!
• Haystack: A framework for building LLM-powered applications, supporting transformer models and vector search.
• Helicone: An open-source tool for logging, monitoring, and debugging LLMs.
• N8n: A workflow automation tool that integrates and streamlines various applications.

Noteworthy Scientific Research

1. Benchmarks for Multimodal LLMs: A review of 180 benchmarks used to evaluate Multimodal Large Language Models (MLLMs) and suggestions for future research. (Source: Original paper on MLLM benchmarks)
2. AlphaZero for Circuit Design: ShortCircuit, a new architecture using AlphaZero, improves circuit design efficiency by 14.61% over existing tools. (Source: Original paper on AlphaZero circuit design)
3. RAG Best Practices Study: Research on optimizing Retrieval-Augmented Generation (RAG) approaches, including strategies for improving multimodal question-answering. (Source: Original paper on RAG best practices)
4. Impact of Code on LLM Pre-Training: Analyzes how including code data in LLM pre-training affects performance across tasks, even those unrelated to coding. (Source: Original paper on code in LLM pre-training)
5. Matryoshka-Adaptor for Efficient Embeddings: A new framework that reduces the size of LLM embeddings while maintaining performance, improving computational efficiency. (Source: Original paper on Matryoshka-Adaptor)
6. Plasticity Loss in Deep Learning: Discusses the loss of plasticity in deep learning and a new approach to improve continual learning models. (Source: Original paper on plasticity loss in AI)
7. xGen-MM (BLIP-3) for Open Multimodal Models: Salesforce’s xGen-MM framework for creating Large Multimodal Models (LMMs) with open-source datasets and tools for better in-context learning. (Source: Original paper on xGen-MM)

Quick Updates

1. OpenAI Hires Irina Kofman: Former Meta executive Irina Kofman joins OpenAI to lead strategic initiatives. (Source: Original article on OpenAI's new hire)
2. Google’s AI Studio Gets a Prompt Gallery: Google has launched a free Prompt Gallery in AI Studio to help developers streamline AI model creation. (Source: Original article on Google’s AI Studio update)
3. Anysphere Raises $60 Million for AI Coding Tool: Anysphere secures $60 million in Series A funding for its AI-powered coding assistant, Cursor. (Source: Original article on Anysphere's funding)
4. Together AI’s Rerank API for Search Systems: Together AI releases Rerank API with exclusive access to Salesforce’s LlamaRank model, enhancing enterprise search. (Source: Original article on Together AI's Rerank API)
5. Luma AI’s Dream Machine 1.5 for Video Generation: Luma AI launches an upgraded version of its text-to-video model, offering better realism and motion tracking. (Source: Original article on Luma AI’s Dream Machine)
6. China’s Ambition in Robotics: At the 2024 World Robot Conference, Chinese companies showcased 27 humanoid robots, aiming to lead the robotics industry. (Source: Original article on World Robot Conference 2024)

Signing off - Marco

Well, that's a wrap for today. Tomorrow, I'll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ??

Think a friend would enjoy this too? Share the newsletter and let them join the conversation. LinkedIn appreciates your likes by making my articles available to more readers.

Top-rated articles:

• Fount helped save 13.4M in nine months with AI based work optimization.
• Prepare for the "AI-first" businesses.
• AI based Intelligent Automation: [RPA + BPM] x AI = ROI
• A model for strategy development in the AI age
• We don't want to talk to your chatbot
• Copyright infringements by all major AI players
• Lobotomized AI*
• The AGI hype = Bored Apes
• 150 Digital Transformation templates new ones added
• Chip happens ! This article explains the alphabet soup of chip names (CPU, GPU, LPU, TPU)
• Now why do AI/ML project FAIL!
• The age of hyper-personalization