Overview
Microsoft (MSFT) is a household name for providing Microsoft Office Suite, the Windows Operation System, GitHub, LinkedIn, Xbox, Bing search. Microsoft has cloud service call Azure that power other businesses.
Recent News
2/26/2024, Microsoft announce partnership with Mistral AI, a 9 month old French start up company cofounded by Google Deepmind alums. https://azure.microsoft.com/en-us/blog/microsoft-and-mistral-ai-announce-new-partnership-to-accelerate-ai-innovation-and-introduce-mistral-large-first-on-azure/ Azure AI infrastructure will support Mistral Large LLM model for training and inference. According to Mistral AI ( https://mistral.ai/technology/#models), their Mistral Large model scored 81.2% on MMLU, placing it second to GPT-4 at 86.4% but ahead of Anthropic which scored 78.5% and Google Gemini Pro which scored 71.8%. GPT-3.5 scored 70.0% and LLaMA 2 70B is last in this comparison at 69.9%. Measuring Massive Language Understanding (MMLU) is benchmark of 57 tasks covering math, US history, computer science, law and more. It stresses a model’s knowledge and problem solving ability. Mistral AI’s focus on performance and efficiency. They believe it is possible to do more with less compute which makes their solution more cost effective. Mistral has free open models call Mistral 8x7B.
AI Opportunity with Generative AI
In 2023, ChatGPT brought AI to the masses and everyone is now aware of its capabilities. ChatGPT uses a large language model (LLM) and knows more than the average person. It can assist with school work, creative tasks like writing and compute programming. Stanford professor, former Tesla Autopilot executive and AI researcher Andrej Kaparthy gave a high level introduction lecture on LLM on his YouTube channel. The talk is titled: Intro to Large Language Models. OpenAI’s model is closed private and not open source. Facebook/Meta’s Llama-2 is open source. OpenAI ChatGPT has a lot in common with Llama. Through Llama-2 we can see how ChatGPT may have been built.
An LLM consists of the model, the trained parameters of the model, the dataset used to built the model, and the thin C++ or Python client that runs interference by running input through the model using the parameters. The Llama 2 is trained on 2 trillion tokens and 1 million human preferences, and 100k supervised fine-tuning. The transformer neural network model is initially trained to predict the next word given an initial set of words and trained on 10 TB of internet text and books. This is the pertaining step and due to the volume of data that need to be processed, it requires upwards of 6,000 GPUs and about $2 millions of cloud computing cost running around the clock for 2 weeks. The output is a base model. This model can predict next word but isn’t very helpful. To train the assistant, the model continues the training on a new set of data consisting of human entered prompts and answers conversations. This dataset has about 100k high quality entries and the stage is call fine-tuning and the focus of this stage is alignment. The model that is produced is the assistant model and due to the training, when given a new prompt asking for help with something, the model will output helpful answers levering the knowledge it has obtained in the first pretraining step. After that there is 3rd stage call Reinforcement Learning Human Feedback (RLHF) to compare generated outputs and a human evaluates which is better. Increasingly as LLMs get better, some of the inputs to the model are LLM generated along with humans to speedup the process.
Andrej highlighted in his lecture the LLM Scaling Laws. It states that LLM accuracy has not reached its limits yet and as number of parameters and amount of text increases, LLM intelligence increases. So without a new breakthrough in algorithm or model, LLMs can get better. As it gets better, it will find more use cases and can perform more tasks. The race to generate better LLMs is what’s fueling the demand for more compute.
In the second half of Andrej’s video, he demonstrated how ChatGPT can extends its capabilities by using tools such as the web and Python. It can dynamically fetch new pages and extract the information from those web pages to answer the original question. ChatGPT is also able to generate Python code and execute what it generated to produce a nice looking graph.
In Daniel Kaheman’s book Thinking Fast and Slow introduces the idea of two level of thinking in humans. System 1 is quick and no effort, and System 2 is logical and takes effort. Current LLM behave like System 1 level thinking where each next word output at the same rate. There’s research to expand LLM to have System 2 where we can tell the model to take more time to come up with a higher quality answer.
Generative AI Ranking leaderboard page on Hugging Face currently shows Open AI’s GPT-4-Turbo, and GPT-04-0314 leading Claude-1 as of December 17, 2023.
https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard
It is computed using this notebook: https://colab.research.google.com/drive/1KdwokPjirkTmpO_P1WByFNFiqxWQquwH#scrollTo=o_CpbkGEbhrK
How Does Microsoft Benefit from Generative AI
Through its 49% ownership of OpenAI the creator of ChatGPT, Microsoft is integrating AI capability across all its products. It will add a feature call CoPilot to Office Suite and Github. With GitHub Copilot, a programmer can chat with AI and ask it to generate code for a certain programming language. The AI had been trained with all the publicly available GitHub repositories and generates workable code and enhances programmer productivity. As of 11/27/2024, the cost of Copilot Business is $19 per user / month. Copilot Enterprise is $39 per user /month to be available Feb 2024. Copilot is also available for Microsoft 365. It integrates AI to Teams, Word, Outlook, Powerpoint, Excel. This capability enhances productivity and costs $30 user/month with 1 year commitment. https://www.microsoft.com/en-us/microsoft-365/enterprise/copilot-for-microsoft-365#Pricing
Training AI requires large amount of compute and data. Microsoft Azure can provide the compute and data storage for customers who want to leverage AI by creating AI models specific for their business requirements using their own secret source or private data.
Microsoft Azure provides access to Meta’s Llama 2 LLM models as pay as you go or hosted on one’s Azure subscription using Azure VMs. https://learn.microsoft.com/en-us/azure/ai-studio/how-to/deploy-models-llama?tabs=azure-studio. AI Azure Studio also provides access to OpenAI models for businesses to build custom CoPilots or other AI products.
Key Health Stats
| F24_Q2 | F24 Q1 | |
| Revenue (billions) GAAP | $62.0 | $56.5 |
| YoY Revenue Growth | 18% | 13% |
| Income | $21.9 | $22.3 |
| YoY Income Growth | 33% | 27% |
| Productivity and Business Process Revenue | $19.2 | $18.6 |
| YoY Productivity and Business Process Revenue Growth | 13% | 13% |
| Intelligent Cloud Azure Revenue | $25.9 | $24.3 |
| YoY Intelligent Cloud Azure Revenue Growth | 20% | 19% |
| More Personal Computing Revenue | $16.9 | $13.7 |
| YoY Personal Computing Revenue Growth | 19% | 3% |
| Free Cash Flow | $9.1 | $20.6 |
| Diluted Earnings per share | $2.93 | $3.00 |
| Earnings per Share TTM | $11.07 ($2.93+$3.00+$2.69+$2.45) | $10.33 |
| Recent Stock Price | $397.58 (1/31/2024) | $378.61 (11/27/2024) |
| Price / Earnings PE Ratio | 35.92 | 36.66 |