GLOSSARY: Back from the Brink

Can't read or see images? View this email in a browser

OpenAI, which was losing millions of dollars on ChatGPT and was on the verge of bankruptcy, has finally projected its revenue expectation to investors. It is looking to clock a revenue of about $1 billion in the next twelve months. The expected revenue is multi-times more than the projected revenue of $200 million in 2023.

https://media0.giphy.com/media/v1.Y2lkPTc5MGI3NjExZDdudXF5OXRzaGh1dXEwNXUxODFsOHozajBpYzVzajR1ejEzOWIxYSZlcD12MV9pbnRlcm5hbF9naWZfYnlfaWQmY3Q9Zw/VTxmwaCEwSlZm/giphy.gif

OpenAI confidently projected this new revenue after the launch of ChatGPT Enterprise, which claims to offer enterprise-grade security and privacy, unlimited higher-speed GPT-4 access, longer context windows for processing longer inputs, advanced data analysis capabilities, and customisation options.

ChatGPT Enterprise was announced a few weeks after Azure ChatGPT, tailored for businesses that provide a similar user experience to ChatGPT but offered as your private ChatGPT. Microsoft claimed that Azure ChatGPT is built for more control and privacy compared to the public model — sounds exactly like ChatGPT Enterprise.

Interestingly, a day after the announcement, Microsoft silently withdrew the Azure ChatGPT project from GitHub. Two weeks later, the same product was announced, now dubbed ChatGPT Enterprise by OpenAI.

OpenAI is fueling Microsoft's recent AI offerings, encompassing text generation and summarisation functions within Office 365 productivity applications, a chatbot akin to ChatGPT called Bing Chat, and the GitHub Copilot coding assistant. Microsoft anticipates that these AI products will contribute over $10 billion in value.

Under their investment agreement, Microsoft can integrate OpenAI's technology into its products with minimal to no revenue sharing with the startup.

The stark difference between the revenue of Microsoft and OpenAI from AI products paints a clear picture of how the latter is struggling to figure out the monetisation plan. Based on OpenAI products, Microsoft is claiming to build businesses 10 times more than OpenAI’s.

It’s time for OpenAI to re-think its business strategy and consider taking a leaf out of Microsoft’s book to establish a profit-making successful company.

Read the full story here.

Bold and Responsible

Google, trusted by nearly half the world's population with their data, has grappled with data breaches and privacy issues, notably since the 2018 Google+ API breach. In 2023, the company placed its 'bold and responsible' approach at the forefront of all its activities. In just eight months, Google released 49 security blogs covering diverse updates, including AI-driven projects and AndroidOS enhancements.

During Google Cloud's Next '23 conference, the company unveiled GCP's security strategy, which revolves around three pillars: harnessing Mandiant expertise, integrating security into Google Cloud innovations, and offering expertise across various environments, harnessing generative AI to address evolving security challenges.

Read the full story here.

Evaluating Benchmarks

The landscape for LLM evaluation is expanding with the emergence of various benchmarks designed to assess their capabilities across diverse domains. These benchmarks offer insights into LLMs' performance, covering tasks like coding proficiency, natural language understanding, multilingual comprehension, and more.

Here are the top 5 benchmarks for evaluating language model efficiency:

HumanEval: Assesses coding capabilities through a set of 164 programming problems, testing language understanding and algorithmic skills.
MBPP (Mostly Basic Python Programming): Contains 1,000 Python programming problems designed for introductory programmers, evaluating code generation capabilities.
MMLU (Multilingual Multitask Learning for Understanding): Measures LLMs' multilingual natural language understanding skills across 57 tasks, assessing accuracy and fluency.
TriviaQA (1-shot): Evaluates the ability to answer questions with just one training instance, covering a wide range of questions.
BIG (Beyond the Imitation Game) - Bench Hard: Features over 200 tasks across ten categories, challenging LLMs with various language understanding tasks.

Read the full story here.

Alternatives for LangChain Critics

https://media0.giphy.com/media/v1.Y2lkPTc5MGI3NjExbGc5a29iYzVqMGViM2FzM3E5amprNjc3bThhODhqMnMyN3A1aXc4dCZlcD12MV9pbnRlcm5hbF9naWZfYnlfaWQmY3Q9Zw/Y3l2CbTxVd6oC47Bq5/giphy.gif

LangChain, although highly discussed, has faced criticism for complicating the deployment of large language models (LLMs) rather than simplifying it. Alternatives have emerged to address this complexity:

Auto-GPT: Aims to create a self-reliant conversational AI, focusing on executing codes and commands for precise solutions. However, it may encounter logic loops.
LlamaIndex: Offers versatile data management and access, extracting data from various sources and integrating smoothly with applications like LangChain.
BabyAGI: Serves as an AI-driven task manager, prioritising and executing tasks through OpenAI, LangChain, and vector databases.
AgentGPT: Ideal for enterprises, it introduces self-sustaining AI agents through web browsers, engaging in human interactions to fulfil tasks.
MetaGPT: A multi-agent framework transforming software development by generating user stories, competitive analysis, requirements, and more.
AutoChain: Combines LangChain and AutoGPT approaches, offering a nimble framework for agent creation and automated scenario assessment.
PromptChainer: Enables AI-driven flows with traditional programming and models, managing AI-generated insights. Supports multiple models and databases.