Memory Leak — #14
VC Astasia Myers’ perspectives on machine learning, cloud infrastructure, developer tools, open source, and security. Sign up here.
LangChain is an open source Python library aimed at assisting in the development of LLMs by coordinating tasks in an LLM workflow. LangChain provides many modules that can be used to build language model applications. Modules can be combined to create more complex applications, or be used individually for simple applications. There are six types of modules: 1) LLMs: get predictions from a language model, 2) Prompt templates: manage prompts for LLMs, 3) Chains: combine LLMs and prompts in multi-step workflows, 4) Agents: dynamically call chains based on user intent, 5) Memory: add state to chains and agents, 6) Utilities: actions. LangChain enables users to build many use cases with LLMs including agents, chatbots, data augmented generation, question answering, summarization, generate similar examples, compare models, and evaluation. It is used by software developers and data scientists.
Why does this matter? Large Language Models (LLMs) are emerging as a transformative technology exemplified by OpenAI GPT-3, Cohere, RoBERTa, among others. An LLM often needs to be tied to other processes for pre- and post-processing as well as database infrastructure or SaaS products for context to provide value. LangChain helps with this. We believe the Machine Learning (ML) world is divided into two camps. Camp one are ML practitioners who train and fine-tune ML models. Camp two are those that leverage prompt engineering to get results from a hosted foundational model via API. LangChain supports ML practitioners in camp two. We can imagine a world where LangChain extends beyond LLMs to foundational models of other modalities like Stable Diffusion.
GPT Index is a project consisting of a set of data structures designed to make it easier to use large external knowledge bases with LLMs. GPT Index helps to provide the following advantages:
- Remove concerns over prompt size limitations.
- Abstract common usage patterns to reduce boilerplate code in your LLM app.
- Provide data connectors to your common data sources (Google Docs, Slack, etc.).
- Provide cost transparency + tools that reduce cost while increasing performance.
Why does this matter? As mentioned above, tying LLMs to external data sources is helpful for getting value out of the service. We’ve heard GPT-3’s 4K token limit and lack of connection to internal data management solutions made it hard to use.
MusicLM, a model generating high-fidelity music from text descriptions such as “a calming violin melody backed by a distorted guitar riff”. MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. There are some fun examples on the website.
Why does this matter? Advancements in ML have primarily focused on natural language, images, and video. Generative AI applied to audio is still emerging. We imagine the same discussion around visual generative AI copyright violations will start in the music industry.
A Framework for Prioritizing Tech Debt
Max Countryman, Senior Director of Engineering at Lob, provides a step-by-step process for prioritizing technical debt. He suggests answering the following questions:
- If we choose to do nothing, will this issue become worse, remain the same, or improve?
- If it’ll become worse, how quickly will it degrade?
- If it remains the same, how much disruption is it causing today?
- If it’ll improve, at what point will it improve to the degree it’s no longer an obstruction?
Why does this matter? Technical debt is an enduring challenge for businesses that holds them back from development velocity. We can imagine a world where technical debt is minimized continuously with intelligent solutions like Grit.io.
From Batch to Streams: Building Value from Data In-Motion
Ricardo Ferreira, a Senior Developer Advocate at AWS, underscores that the rise of streaming-based systems. He states that streaming data architectures are very good to provide tools for companies to start reacting to opportunities as they present themselves. Ferreira discusses the streaming data journey and traps to avoid.
Why does this matter? We believe competitive business pressures and customers expectations will continue to push companies to become more near real-time or real-time in their data and ML infrastructure.
After Inking Its OpenAI Deal, Shutterstock Rolls Out a Generative AI Toolkit to Create Images Based on Text Prompts
Customers of Shutterstock’s Creative Flow online design platform will now be able to create images based on text prompts, powered by OpenAI and Dall-E 2. Key to the feature is that Shutterstock says the images are “ready for licensing” right after they’re made.
Why does this matter? We believe that foundational models will benefit incumbents and lead to a new generation of foundational model-native SaaS/prosumer products. Similar to the transition from on-premises software to the cloud, there will be a group of existing vendors that innovate like Shutterstock and a group that will cease to be relevant.
⭐️Claypot — Founding Engineer (Infra)
⭐️ Grit — Founding ML Engineer
⭐️ Omni — Senior Data Visualization Engineer