• AI Time to Impact
  • Posts
  • . . AI: New Model Architecture, Training, and a New Code Interpretation Tool (4.30.24)

. . AI: New Model Architecture, Training, and a New Code Interpretation Tool (4.30.24)

LangChain, Cohere, StarCoder

Friends, in today's AI news, the spotlight is on model training and architecture as we see updates from Cohere and BigCode's Starcoder as they innovate in this space. We also see a cool new feature from LangChain and their code interpreter feature along with some headlines.

Editorial leadership change: Today will be my last day as Editor of AI Time to Impact. I will be passing the torch to Sasha Krecinic, who speaks to hundreds of AI companies each year as part of the team at VC firm Headline.com and has been helping me edit this newsletter for the past few months. Sasha is going to take it from here. I'd love to connect with you on LinkedIn and I hope you'll continue to make great use of this newsletter!

Thanks for all the support and engagement,

Marshall Kirkpatrick

First impacted: AI researchers, AI model evaluators
Time to impact: Short

Researchers from Cohere propose using a Panel of LLM Evaluators (PoLL), composed of multiple smaller models, as a more cost-effective and less biased method for assessing LLMs. According to their research, a PoLL is over seven times less costly and exhibits less intra-model bias compared to using a single large model like GPT-4. [Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models] Explore more of our coverage of: AI Research, LLM Evaluation, Cost-Effective Models. Share this story by email

First impacted: AI researchers, Software developers
Time to impact: Short

StarCoder2-15B-Instruct-v0.1, a large language model (LLM), has achieved a score of 72.6 on the HumanEval test, slightly surpassing the 72.0 score of CodeLlama-70B-Instruct, according to a blog post. This is also interesting as it represents the first entirely self-aligned code LLM, trained using a unique three-step data generation pipeline that includes extracting high-quality seed functions from a large corpus of licensed source code, creating diverse code instructions, and generating a high-quality response through execution-guided self-validation. [StarCoder2-Instruct: Fully Transparent and Permissive Self-Alignment for Code Generation] Explore more of our coverage of: Large Language Models, AI Benchmarking, Open-Source Training. Share this story by email

First impacted: Software Developers, Cloud Computing Specialists
Time to impact: Short

LangChainAI has launched a feature that utilizes e2b_dev's Code Interpreter SDK to comprehend and execute code in a secure cloud sandbox, powered by Firecracker. [e2b-cookbook/examples/langchain-python at main · e2b-dev/e2b-cookbook] Explore more of our coverage of: LangChainAI, Code Interpreter SDK, Cloud Sandbox. Share this story by email