Introducing Devin, the first AI software engineer

Meet Devin, the groundbreaking AI software engineer from Cognition, designed to work alongside humans or independently on various software engineering tasks. Devin can plan, execute, and learn from complex engineering challenges and is equipped with developer tools within a sandboxed environment. It enhances productivity by handling mundane tasks, allowing engineers to focus on more complex issues.

Devin's impressive capabilities include learning new technologies, building and deploying applications, finding and fixing bugs, and even training AI models. Notably, Devin demonstrated its prowess on the SWE-bench benchmark, solving 13.86% of real-world GitHub issues, a significant leap from the previous best of 1.96%.

Cognition, the AI lab behind Devin, is dedicated to advancing AI reasoning to create AI teammates that surpass current AI tools. Supported by notable industry leaders and substantial Series A funding, Cognition aims to revolutionize various fields by enhancing human capabilities with AI.

Related: Gold-Medalist Coders Build an AI That Can Do Their Job for Them

Reach your work goals with Wave’s AI coach

Today, top-level talent use coaching to deal with the challenges they face in the workplace around

• Leadership

• Time Management

• Problem-solving skills

Current apps and methodologies for professional growth are outdated

Wave has developed an innovative way to improve your skills by building daily routines.

It is measurable and easy 🔥

Leaders from Amazon, Stripe, Google, or Strapi are already using it.

Get started now and get priority access 🙌

Meta reveals details of two new 24k GPU AI clusters

Meta has unveiled two new data center scale clusters, each equipped with 24,576 Nvidia Tensor Core H100 GPUs, a significant upgrade from the previous 16,000 Nvidia A100 GPUs. These clusters, developed for AI research in fields like natural language processing and image generation, are part of Meta's AI Research SuperCluster initiative. The company plans to expand its infrastructure to include 350,000 Nvidia H100s by the end of 2024, enhancing its capability to support more complex AI models.

The two clusters differ in their network infrastructure, with one using an RDMA over Ethernet solution and the other featuring Nvidia Quantum2 InfiniBand fabric. Both utilize Meta's Grand Teton, an open GPU hardware platform designed for large AI workloads, offering increased bandwidth and power compared to its predecessor. The clusters also incorporate Meta's Open Rack architecture, enabling flexible power shelf placement and optimal server balance for power efficiency and throughput.

Meta continues to advance its AI software framework, PyTorch, to accommodate large-scale GPU training and has introduced the AI Alliance to foster an open AI development ecosystem. This initiative aims to ensure transparency, safety, and responsibility in AI evolution, reflecting Meta's commitment to continual improvement and innovation in its infrastructure to meet the demands of future AI research and development.

Blend LLMs to Make Best Performing AI Model

In this video, Maya discusses a new experimental technique called model blending, which allows individuals, even non-experts in machine learning, to combine different models to enhance their performance on various tasks. The narrator, an amateur in machine learning, successfully blended around 20 models in two weeks, improving their performance on the open LLM leaderboard. They share insights on model blending, including the importance of selecting compatible models and the impact of blending on model performance.

The video further delves into the technical process of model blending, explaining different methods like task arithmetic, slurp, ties, dare, and pass-through. It provides a step-by-step guide on how to use MergeKit, a Python toolkit, for blending models. The narrator emphasizes the necessity of choosing models with the same architecture and highlights the use of YAML files to define blending parameters. The goal is to enable viewers to efficiently blend their models, optimizing their utility across various tasks.

Lastly, the video addresses the issue of data contamination in model blending, which can skew the performance of models on benchmarks, misleadingly portraying them as more effective than they are. The narrator encourages ensuring the purity of data used in training to maintain the integrity and utility of blended models. The video aims to empower viewers to create their own high-performing LLMs through model blending while avoiding common pitfalls.

Other stuff

• In case you were wondering just how cracked the cognition labs team is...

• AI models on CPUs: accurate audio transcriptions without breaking the bank

• Anti-AI sentiment gets big applause at SXSW 2024 as moviemaker dubs AI cheerleading as ‘terrifying bullsh**’

• How AI is disrupting the demand for software engineers

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.