Weights & Biases Announces W&B Weave - the Lightweight Toolkit for Developers to Deploy Generative AI Applications with Confidence
18 Abril 2024 - 12:00PM
Business Wire
Fully Connected – Weights & Biases, the AI developer
platform, today announced W&B Weave at their annual conference
Fully Connected. W&B Weave is a lightweight toolkit for
software developers who want to deploy generative AI applications
with confidence, providing software developers with a system of
record for the experimental large language model (LLM) application
development process.
W&B Weave extends the Weights & Biases AI developer
platform beyond machine learning practitioners who want to build
and train large scale models to serve software developers who want
to build applications with LLMs. Over the last 6 years, Weights
& Biases has enabled the most innovative foundation model
builders to create the generative AI industry. Today, Weights &
Biases is used by over 30 foundation model builders and 1,000
companies to productionize machine learning at scale.
Since the emergence of LLMs such as OpenAI’s GPT 3.5,
Anthropic’s Claude 2, Meta’s Llama 2, and Mistral’s Mixtral 8x7B,
every organization has been challenged to define and execute their
generative AI strategy. While it’s easy to create a generative AI
demo, it’s hard to confidently deploy a generative AI application
into production given the risks of abuse, misalignment, and
hallucinations.
Building software with LLMs is fundamentally different from
traditional software development as LLMs are non-deterministic by
nature. The solution to this challenge is to treat the model as a
closed system where only the inputs and outputs are visible and
follow a scientific workflow. This experimental workflow is similar
to the workflow machine learning practitioners use to build these
LLMs in the first place.
W&B Weave was built to enable this experimental
workflow:
- Log everything: Capture every interaction with LLMs,
from development to production. This data is expensive to produce,
so log it. Use it to improve applications, and build up
evaluations.
- Experiment: Try many different configurations and
parameters to explore what improves the application.
- Evaluate: Build up suites of evaluations to measure
progress. Use evaluations to determine what improves the
application. Evaluations are to LLM development what Unit Tests are
to traditional software development.
“We built Weave for developers to support the scientific
workflow of building AI applications,” said Shawn Lewis, CTO at
Weights & Biases. “We made it lightweight and practical with
minimal abstractions. We built the tools that helped bring LLMs
into the world. Now we're building tools to make them work in
production applications.”
W&B Weave has two components:
- Traces. Add one line of code to log the behavior of the
LLM application to pinpoint exactly what went wrong. Build RAG
applications with full observability into retrieved documents,
functions used, and chat messages served to the LLM.
- Evaluations. Score the LLM application in a lightweight,
customizable way. Build rigor with a systematic and organized
evaluation framework. Spot trends, identify regressions, and make
informed decisions about future iterations.
“I like Weave both conceptually and in practice,” said Jonathan
Whitaker, AI Researcher at Answer.AI. “I want a lightweight tool
that fits into my workflow, rather than a tool that imposes a new
workflow on me. Weave achieves this by making it easy to decorate a
few functions and then persistently capture the LLM inputs and
outputs.”
To learn more about W&B Weave, visit http://wandb.me/weave
and get started today.
About Weights & Biases
Weights & Biases is the leading AI developer platform
supporting end-to-end MLOps and LLMOps workflows. Used by over 30
foundation model builders and 1,000 companies to productionize
machine learning at scale including teams at OpenAI, Toyota, and
Microsoft. Weights & Biases is part of the new standard of best
practices for machine learning.
View source
version on businesswire.com: https://www.businesswire.com/news/home/20240418287123/en/
pr@wandb.com