Memory Leak — #29
VC Astasia Myers’ perspectives on machine learning, cloud infrastructure, developer tools, open source, and security. Sign up here.
🚀 Products
System Initiative Code Now Open Source
System Initiative is a collaborative power tool designed to remove the papercuts from DevOps work. The founders created System Initiative because they were driven by frustration that despite the many good tools available for DevOps workflows, stitching all these tools together still doesn’t guarantee success. Using the digital twin concept borrowed from the gaming industry, System Initiative enables DevOps teams to build detailed interactive simulations of their infrastructure to rapidly update their production environments. System Initiative open sourced its entire codebase on GitHub under the Apache License version 2.0. There are no commercial-only features.
Why does this matter? Last week Hashicorp announced it was moving to a Business Source License, a source-available license that allows copying, modification, redistribution, non-commercial use, and commercial use under specific conditions. The CEO of System Initiative is Adam Jacob, the former co-founder and CTO of Chef, which competed with HashiCorp. It is interesting to see System Initiative go the other direction and open source its work, given the historical context.
DialpadGPT is the first LLM designed for business communications. Dialpad trained its model on billions of conversations and transcripts to provide unmatched precision in the context of business interactions. Plus, it’s fully customizable to a specific industry and business need. Unlike open source models, Dialpad’s domain-specific training also minimizes inaccuracies and “hallucinations” across AI-generated outputs. DialpadGPT automatically powers helpful AI features like instant call summaries and in-the-moment coaching for sellers and support agents — right out of the box.
Why does this matter? It is interesting to see a SaaS company offer a fine-tuned LLM to their end customers. Typically, we see LLMs embedded in the incumbent’s product experience. Dialpad takes it one step further. Dialpad finetuned their own language model, which we are starting to see more.
Kyligence Zen + Copilot allows users to communicate intuitively with their data and metrics, even without technical expertise. It leverages the power of a large language model powered by Azure OpenAI, offering users the ability to search for metrics, conduct in-depth analysis, gain valuable insights, and enable the automatic creation of dashboards using intuitive, natural language conversations centered around business metrics. The purpose is to lower the barriers to leveraging data, boosting overall work efficiency.
Why does this matter? There have been a lot of startups focused on self-service data analysis that translates natural language to SQL. The intention is to help people who can’t write SQL to understand their business data. This AI data copilot is a natural extension for Kyligence’s core metrics business.
📰 Content
6 Patterns for Platform Engineering Success
PlatformCon is the main platform engineering community event each summer. This piece highlights some patterns for building a successful team and Internal Development Platform (IDP) including recognizing that each platform will be unique and maintaining one API access point.
Why does this matter? We are seeing the rise of IDPs, which are built by a platform team to build golden paths and enable developer self-service. Platform engineering teams specify what resources start up with what environment or at what request. They also set base-line templates for application configurations and govern permissions. This helps them to automate recurring tasks such as spinning up environments and resources and makes their setup easier to maintain by enforcing standards. Platform engineering is an emerging field to watch.
Llama-2 and the Open Source LLM
Arjun Bansal, the CEO and co-founder of Log10, a developer-centric AI startup, discusses the release of Llama-2, an open source LLM. He highlights that Llama-2 narrows the gap between GPT-3.5 capabilities and open source models. In the piece he walks through how to self-host Llama-2.
Why does this matter? While OpenAI models took the world by storm over the past year, we are starting to see open source models become increasingly popular for builders because the quality of the models is improving. We hear businesses using multiple models as part of their application experience and increasing interest in fine-tuning.
AI Bots Are So Good at Mimicking the Human Brain and Vision That Captchas Are Useless
Computers and Humans Apart (CAPTCHA) are a series of puzzles requiring you to correctly identify traffic lights, buses, or crosswalks to prove that you’re indeed human before you log in. The technology is intended to protect a website from fraud and abuse without creating friction. A study conducted by a group of researchers including three from the University of California, Irvine, one from ETH Zurich, one from the Lawrence Livermore National Laboratory, and one from Microsoft, showed that AI bots are now better than humans at decoding the CAPTCHAs. The researchers recruited 1,400 participants to test websites that used CAPTCHA puzzles, which account for 120 of the world’s 200 most popular websites. “The bots’ accuracy ranges from 85–100%, with the majority above 96%. This substantially exceeds the human accuracy range we observed (50–85%),” the research paper read. “Furthermore the bots’ solving times are significantly lower in all cases, except reCAPTCHA, where human solving time of 18 seconds is nearly similar to the bots’ time of 17.5 seconds.”
Why does this matter? A big trend in AI are agentic systems that can take action given a task using Chain of Thought and Tree of Thought. Since CAPTCHAs are intended to prevent bots on website, mostly for security, there were concerns that CAPTCHAs would block agents from completing their tasks. Now data suggests that AI agents will be able to get around CAPTCHAs. For well-intentioned AI services this is a boon. However, for malicious actors it suggests they will be able to get around website security. The survey’s findings underscore that AI is disrupting the threat landscape such that security controls are less effective.
💼 Jobs
⭐️DragonflyDB — Community Lead
⭐️Chroma — Member of Technical Staff
⭐️Speakeasy — Founding Engineer
