. . AI: Open AI Launches Some Fine API's and Meta's Models Get A $100k Open Source Challenger (4.4.24)

Friends, in today's AI news, the spotlight is on OpenAI and their new fine-tuning APIs, which are gaining impressive traction. Additionally, there are updates from Cohere on enterprise-grade models and SDKs, developed in collaboration with Langchain. Google is also making waves with algorithmic optimization, notably by reducing the number of layers in models, and there's an open-source model called JetMoE by MyShell.ai utilizing open data, which is challenging Meta's models with a mere $100k in training costs. Enjoy!

-Marshall

OpenAI Launches Fine-Tuning API

First impacted: AI developers, AI Users
Time to impact:

OpenAI has launched new features for its fine-tuning API, including epoch-based checkpoint creation for training and integration with third-party services (like Weights and Biases), hyperparameter configuration and, options for assisted tuning and custom models. The blog post provides case studies which showcase performance improvements after applying the above and the results are quite impressive as you'd expect. [Introducing improvements to the fine-tuning API and expanding our custom models program] Explore more of our coverage of: OpenAI, Fine-Tuning API, Telecom AI Integration. Share this story by email

Cohere Launches Enterprise-Grade LLM, Command R+ and Python SDK Enhanced AI Interactions with Langchain

First impacted: Cloud Infrastructure Managers, Business Workflow Automation Specialists
Time to impact:

Cohere has launched Command R+, a enterprise grade LLM. The company claims outperforms similar models in scalability as it brings cost effectiveness and performance in the one package. The LLM reportedly offers advanced RAG with citation, supports 10 major languages, and can automate complex business workflows with something they call 'Tool Use'. The press release doesn't provide detailed information about deployment options yet, including whether a private instance on Azure will be offered. In the second story for Cohere today, Langchain announced that they have launched a Python SDK called langchain-cohere, aimed at enhancing human-machine interactions through their natural language processing models. According to LangChainAI, the package includes features that allow developers to build chat bots, generate text, and rank strings based on relevance, following the release of their v5 SDK. Check out more information here. [Introducing Command R+: A Scalable LLM Built for Business] Explore more of our coverage of: Cohere Command R+, Language Learning Models, Business Workflow Automation. Share this story by email

Google's Language Models use Mixture-of-Depths Method to Reduce Compute Requirements

First impacted: AI researchers, AI model developers
Time to impact:

Google researchers have shared a research paper on how to dynamically allocate computing power using 'depth' as a parameter and achieved the same results with fewer FLOPs per forward pass. The models use a technique known as Mixture-of-Depths, which caps the number of tokens involved in calculations at a certain layer, allowing for dynamic allocation of computing power, even with a static computation graph. Authors found that routing fraction of the normal tokens through every second layer worked the best and that the cost of attention for those layers decreases quadratically. This could be an interesting way of making ultra long context length more efficient. [Mixture-of-Depths: Dynamically allocating compute in transformer-based language models] Explore more of our coverage of: Google AI, Transformer Models, Dynamic Compute Allocation. Share this story by email

JetMoE-8B Outperforms Meta AI's Model With $100K in Training

First impacted: AI researchers, Data scientists
Time to impact:

The JetMoE-8B model, developed by MyShell.ai, is said to outperform Meta AI's LLaMA2-7B, despite costing less than $100K to train. The team used only public datasets for training and has open-sourced its code. [JetMoE: Reaching LLaMA2 Performance with 0.1M Dollars] Explore more of our coverage of: AI Models, Open-Source Code, Cost-Efficiency. Share this story by email

AI-Powered Agent AIDE Matches Human Performance in Kaggle Contests

First impacted: Data scientists, AI researchers
Time to impact:

AIDE, an AI data science agent built by pre-seed funded R&D company WeCo.ai, has reportedly matched human-level performance in over 60 Kaggle data science competitions, even outperforming half of the human participants on average, according to a benchmarking process. The AI, which is set to be open-sourced, uses a method called Solution Space Tree Search and is said to have outperformed an industrial-level AutoML system called H2O and ChatGPT with user guidance, all while maintaining cost-efficiency with an LLM inference cost of less than $1 per task. [Introducing Weco AIDE] Explore more of our coverage of: AI Competitions, Data Science, Open-Source AI. Share this story by email

Lambda Secures $500M for AI Cloud Expansion

First impacted: AI developers, Cloud service users
Time to impact:

"Lambda, the GPU cloud company founded by AI engineers and powered by NVIDIA GPUs, today announced that it has secured a special purpose GPU financing vehicle of up to $500 million to fund the expansion of its on-demand cloud offering. The financing was follows Lambda’s $320 million Series C funding round in February 2024. 'This first-of-its-kind financing vehicle unlocks a new way to fund the deployment of tens of thousands of NVIDIA GPUs,' said Mitesh Agrawal, Lambda COO. " [Lambda Announces $500M GPU-Backed Facility to Expand Cloud for AI] Explore more of our coverage of: AI Cloud Services, GPU Technology, Funding News. Share this story by email

Brev Acquires Competitor Agora.io

First impacted: Brev customers, Agora.io customers
Time to impact:

GPU access startup Brev has acquired Agora, which describes itself as an "accessible GPU-powered sandbox environments built for AI engineers." Brev's founder explained on HackerNews a few months ago: "We work with smaller data centers that have single node GPU capacity. So while there's a shortage for clouds that have capacity for clusters, we're focusing on aggregating single nodes behind a unified API." [via @NaderLikeLadder] Explore more of our coverage of: Brev Acquisition, Agora.io, AI Industry Collaboration. Share this story by email