- The Deep View
- Posts
- ⚙️ A new LLM for Microsoft
⚙️ A new LLM for Microsoft
Ian Krietzberg
May 07, 2024
Good morning. My name is Ian Krietzberg. As of today, I am coming to you as The Deep View’s Editor-in-Chief. I previously worked as a tech reporter at TheStreet, where I built up a beat in technology and artificial intelligence, which led me here, to you. I’ve covered just about every aspect of AI over the past year, including ethical impacts and considerations, regulatory efforts, positive use cases and that rose-tinted intersection of investing and AI.
You can check out some of my prior coverage right here. I’m thrilled to be diving into the newsletter arena with The Deep View team.
Connect with me on Twitter (calling it ‘X’ still feels wrong).
In today’s newsletter:
Meet MAI-1, Microsoft’s up-and-coming heavyweight AI model
Palantir reports another revenue beat
Stack Overflow partners up with OpenAI
Nvidia and another AI copyright lawsuit
Palantir reports another revenue beat
AI and software giant Palantir reported a 21% year-over-year revenue growth – to $634 million – for the first quarter of 2024 after the bell Monday. This quarter, according to the company, marks its sixth consecutive quarter of GAAP profitability.
Palantir reported a net income of $106 million for the quarter and earnings of eight cents per share, in line with analyst expectations.
Though governments remain Palantir’s biggest source of revenue, it noted steadily increasing growth in new areas. U.S. commercial revenue, for instance, spiked 40% year-over-year to $150 million. The company’s U.S. customer headcount spiked nearly 70% to 262.
But the stock, which closed Monday up 8%, fell around 8% in after-hours trading (in the hour after the report’s publication) on weaker-than-anticipated second-quarter and full-year guidance.
In his letter to shareholders, CEO Alex Karp said that the company’s net income represents the largest quarterly profit in Palantir’s history.
“Our business is now growing at a scale and pace that even our most ardent advocates would have been hard pressed to say was likely, or even possible,” Karp said. “The profits that we promised would come, and that skeptics repeatedly and confidently said were beyond us, have now arrived.”
Stack Overflow teams up with OpenAI
Stack Overflow, the Q&A forum for techies and computer programmers, on Monday announced a new API partnership with OpenAI.
The collaboration, according to OpenAI, will – over time – enhance its models’ ability to answer highly technical, programming-related questions. OpenAI will, as part of its collaboration, “surface validated technical knowledge from Stack Overflow directly into ChatGPT.”
Stack Overflow, meanwhile, will leverage OpenAI’s models as part of its ongoing integration of generative AI within its platform.
The companies did not disclose the financial terms of the arrangement.
The partnership comes just a few months after Stack Overflow announced a similar API partnership with Google.
Stack Overflow CEO Prashanth Chandrasekar said last year that the “platforms that fuel LLMs should be compensated for their contributions so that companies like us can reinvest back into our communities to continue to make them thrive.”
The partnership, coming right in the midst of ongoing copyright concerns, comes about a week after OpenAI announced yet another media partnership (this time, with the FT).
Speaking of Copyright… Another lawsuit for Nvidia
The question of how copyright law relates to generative AI has become central to the business of AI over the past year. Two camps seem to be under construction on the media side of things: one group (like the FT, Business Insider and Stack Overflow) that are willing to take a check in exchange for allowing OpenAI and its peers to continue consuming their content, and another, that has shirked licensing deals and filed lawsuits against those companies instead.
Last week, four (more) authors filed a class action lawsuit against Nvidia and Databricks. Their claims are pretty similar to existing lawsuits in the space – that the companies violated copyright law by training AI models (without the knowledge, permission or compensation of the author) on authors’ content.
The answer to that legal question of whether content scraping for training purposes is fair use has not yet been provided by the U.S. Copyright Office.
“We respect the rights of all content creators and believe we created our models in full compliance with copyright law,” Nvidia said in a statement.
Read the lawsuit here.
TOGETHER WITH CAMUNDA
Unlock the potential of AI to orchestrate your business processes.
Where does AI fit within your hyper-automation tech stack?
Predictive, generative, and augmented artificial intelligence can power innovation within your team and make everybody’s lives easier.
However, these types of automation can also come with some unique challenges. Want to learn how to avoid them?
❓️ Misc:
Sam Altman talks about the future of AI at Stanford (link)
Synopsys sells software unit to Clearlake Capital in $2.1 billion deal (link)
📊 What we’re reading:
The 'rat race of AI development (link)
The DOJ’s case against Google (link)
💰Jobs in AI:
Generative AI Consultant: Correlation One · United States · Remote · Contract (Apply here)
Generative AI Scientist: Mission Cloud · United States · Remote · Full-time · Mid-Senior level (Apply here)
NLP Annotation Specialists: Rylem Staffing · United States · Remote · Contract · Mid-Senior level (Apply here)
TOGETHER WITH OMNIPILOT
Imagine an AI... that can type in every app on macOS with full context on what's on your screen.
Omnipilot is a Mac app that brings the power of AI to every application on your computer.
It uses the current context of the app you're working in to provide intelligent suggestions and assistance, just a shortcut away.
Whether you're writing documents, composing emails, or looking for quick answers to questions, Omnipilot is the fastest way to use AI in your work.
Microsoft is training a new LLM
Since investing more than $10 billion in OpenAI, Microsoft has had its hands in more than a few AI-related projects and applications. The company’s work with Copilot, according to Wedbush tech analyst Dan Ives, for example, has been “transformative” in proving the financial value of AI.
And while Microsoft has lately seemed to be a fan of partnering up with startups to leverage existing tech, The Information reported Monday that Microsoft is currently training a new in-house large language model internally referenced as MAI-1.
The model will have about 500 billion parameters, according to the report, which cited two Microsoft employees with knowledge of the project. OpenAI’s GPT-4 model, in comparison, has more than one trillion parameters.
Microsoft declined a request for comment, though Kevin Scott, the company’s CTO, responded to the report on LinkedIn, saying that he’s “not sure why this is news.”
Scott said that for years now, Microsoft has built supercomputers which OpenAI has then used to train models. He added that, for “years and years and years,” Microsoft has built its own in-house AI models.
“AI models are used in almost every one of our products, services and operating processes at Microsoft, and the teams making and operating things on occasion need to do their own custom work, whether that's training a model from scratch, or fine tuning a model that someone else has built,” Scott said. “There will be more of this in the future too.”
The construction of this new model, according to The Information, is being overseen by Mustafa Suleyman, the CEO of Microsoft AI, the company’s newly bundled consumer-facing AI organization. Microsoft pulled Suleyman – the founder of DeepMind and co-founder of Inflection AI – over to its side in March as part of a ($650 million) deal with Inflection that brought the startup’s models (and most of its staff) to Microsoft.
Microsoft, according to the report, has been setting aside clusters of Nvidia-powered servers to train the model, and has been compiling a dataset for training that includes public data from the internet.
This push into large models comes alongside an ongoing effort from Microsoft to invest in small language models; the company in April announced its Phi-3 family of “tiny but mighty” models. Phi-3 mini has only 3.8 billion parameters.
Shares of Microsoft closed Monday up 1.69%, at $413.54.
My take
Two things come to mind when I’m thinking about Microsoft’s simultaneous investments in both large and small models. The first is training data, and the second is electricity usage.
Companies have been running out of high quality data to train their LLMs on for months now, a feat that apparently happens pretty quickly when companies train their models on the bulk of the available internet. Largely because the amount of usable data out there is finite (and because LLMs are quite closely linked to their training data) most new LLMs out there are relatively similar to one another.
With the purpose behind MAI-1 still unclear, I’m wondering what makes this project a worthwhile expense, especially considering the fact that OpenAI already has large models which Microsoft has availed itself of.
In terms of electricity, multiple researchers have told me over the past year that there is a lot of promise in small language models, which can be tuned to handle specific tasks (well) without using an exorbitant amount of electricity (and therefore emitting a lot of carbon dioxide). The bigger the model, the more power it needs to run, and the higher its emissions are.
Purpose-driven ventures in this arena feel important – it remains unclear what, exactly, Microsoft is hoping to accomplish with MAI-1.
Image 1
Which image is real? |
Image 2
Have cool resources or tools to share? Submit a tool or reach us by replying to this email (or DM us on Twitter).
*Indicates a sponsored link
SPONSOR THIS NEWSLETTER
The Deep View is currently one of the world’s fastest-growing newsletters, adding thousands of AI enthusiasts a week to our incredible family of over 200,000! Our readers work at top companies like Apple, Meta, OpenAI, Google, Microsoft, and many more.
One last thing👇
Trying to be useful lately!
— Tesla Optimus (@Tesla_Optimus)
7:53 AM • May 5, 2024
That's a wrap for now! We hope you enjoyed today’s newsletter :)
What did you think of today's email? |
We appreciate your continued support! We'll catch you in the next edition 👋
-Ian Krietzberg, Editor-in-Chief, The Deep View