Accelerating into the Future
Hardware, models, algorithms and learning
In this edition:
Quick Poll on Model Usage
My thoughts
Notable headlines
Learning, Tools, and Experiments
AI MISTAKES
Quick Poll
Ethan Mollick (Wharton School at the University of Pennsylvania) recently shared the following on social media:
In every group I speak to, from business executives to scientists, including a group of very accomplished people in Silicon Valley last night, much less than 20% of the crowd has even tried a GPT-4 class model.
Less than 5% has spent the required 10 hours to know how they tick.
I’d like to know how many of my readers have experienced the incredible power of today's most advanced Generative AI Models. Please don’t feel like you need to pretend one way or the other. We’re all friends here.
If you answered ‘no,’ I’d love to know if I can help change this.
I’d also like to know if you would be interested in helping turn your boss, friends, or co-workers from a ‘no’ to a ‘yes’.
Don’t be shy—you can reply to this email. I love reading your replies to this newsletter!
My Thoughts
Over the last several years, I've thought deeply about what it means to accelerate into the future. My most memorable encounter with Accelerate comes from my work in DevOps. It's an excellent book.
ACCELERATE The Science of Lean Software and DevOps: Building and Scaling High-Performing Technology Organizations by Nicole Forsgren, Ph.D., Jez Humble, and Jean Kim. If you're an IT practitioner, I would wholeheartedly recommend this book. It contains some really great information about best practices and organizational transformation.
I know what you’re thinking: “What does this have to do with AI?”
I'm a firm believer that books like this, thoughts contributed by the DevOps community, have ultimately led to the incredible pace of change that we're experiencing in generative AI right now.
The evolution of the automation process and systems needed to build these cutting-edge AI capabilities are possible because of the practices that have been rigorously improved through building cross-functional teams that enable flow, faster feedback, and continuous learning
AI improvements and breakthroughs happen every day. Whatever you do, don't go on vacation; you'll get left behind because the pace of change is so rapid.
In all seriousness, though, I’ve settled on a monthly cadence because with all the announcements and developments coming at us, we must constantly take the time to zoom out and see what really matters. So that’s what I’ll continue to do here. I’ll share the news items that I think are relevant and share what I think is important.
Without a doubt, we are accelerating into the future, and the future of AI is very bright. My excitement grows with every new use case and application release because it means that more people will benefit from the use of these cutting-edge tools. This will provide an abundance of opportunities for all of us to learn, make mistakes, and accelerate the success of ourselves, our organization, and our customers.
Notable Headlines
Anthropic’s release of Claude 3 has dramatically impacted the scene by outperforming GPT-4 in some cases. Many consider it to be at least as good as, if not better than, GPT -4. I’m using it as my default, but I have yet to cancel anything with OpenAI.
NVIDIA launched the Blackwell Platform, adding new and improved processing capabilities, and with some compromises, enabling 25x better efficiency than prior solutions. The Blackwell architecture includes several new elements designed to accelerate various tasks, and I was particularly impressed with the new NVLINK capabilities that enable the network to not only relay information across GPUs but actually to contribute to the mix. They also have a drop-in version so existing Hopper cabinets can be upgraded to Blackwell.
Stability AI CEO Resigned, among other Key Talents. Emad Mostaque, the CEO and founder of Stability AI, has stepped down from his role for wanting to pursue decentralized AI. Peter Diamandis interviewed him on YouTube a few days ago if you’re interested in what he’s thinking.
Sakana AI’s New Foundation Models by Merging Open Models. Sakana AI has introduced Evolutionary Model Merge, which uses evolutionary techniques to efficiently discover the best ways to combine different open-source models with diverse capabilities to create new foundation models.
Google's system, SAFE, evaluates factual accuracy in text, especially from large language models.
It uses Google Search to verify claims and shows promise for reliable information. Check out the paper on Arxiv here. I suspect this will get much of my attention in the coming weeks.
X.ai released Grok-1 as open-source last month, but this month, it was announced that Grok 1.5 is on the way and will be available on X ‘soon’. It excels in math and coding tasks, showcasing improved performance.
OpenAI and Microsoft are reportedly collaborating on an ambitious project to build a supercomputer named "Stargate," with an estimated cost of up to $100 billion. This initiative is part of a broader five-phase plan, with Stargate representing the fifth phase, expected to launch as early as 2028
Did I make a MISTAKE by missing a news story you found pivotal? I’d love to hear from you!
Learning, Tools, and Experiments
Last week’s LIVE broadcast covered the evolving role of prompt engineering. In the video, I covered the DSPy framework for algorithmically optimizing LM prompts and weights. While many of my readers aren’t likely to need this level of prompt optimization, automation, or rigorous process separation, it is a project that bears watching.
What I didn’t know a few days when I did the live broadcast was another tool called gpt-prompt-engineer that enables prompt optimization/automation with much less effort than DSPy. The creator, Matt Shumer, made it easy to access and use with a one-click run in Google Colab.
Andreeson Horowitz rates another inclusion in my newsletter. This time, they’ve shared an excellent report 16 Changes to the Way Enterprises Are Building and Buying Generative AI. This report is valuable to me because I have spent many years inside a large Fortune 50 company, and seeing that perspective through this report is helping me to continue improving my holistic view of the Generative AI industry. This report also calls out that if large companies are looking to invest, small independents and startups can reap some of the benefits at their size and scale by being choosy about testing and experimenting with this technology at a size and risk envelope they can tolerate. If there’s a message I’d like for you to take away from this newsletter, it is that you don’t have to be an enterprise to reap incredible value from Generative AI. It’s time to get started experimenting if you haven’t already.
AI MISTAKES
The excitement around Generative AI is reaching new heights. Some are now calling it a bubble. I don’t think we’re in a bubble at this moment, but it’s always important to temper enthusiasm with a dose of reality, so here are some ways that AI MISTAKES have been made.
As always, deploying and using this technology is essential, but we must do so with an eye toward creating value and mitigating unnecessary risks that lead to sloppy mistakes.
First up this month is a sobering story about the use of AI to identify mushrooms. In 2022, after a significant rise in mushroom poisoning, Australian poison researchers tested out a set of applications that identify mushrooms using photos. The results were not impressive. The worst outcome was that some of the apps identified toxic mushrooms as edible. It’s essential to recognize that these tools are getting better all the time, but it may be best to choose a use case that doesn’t include accidental death by poisoning.
The second story I came across this month is a use case intended to help people better understand legal compliance issues in New York City, which may be providing information contrary to its stated intention. NYC announced a program in partnership with Microsoft to help businesses understand and comply with the law regarding buildings, tenants, and other relevant business items, such as employment. So far, it’s not going perfectly. The article calls out some examples of where things go wrong. Because the chatbots are new, this could also be a user interface and training issue. It’s possible to apply these generative AI systems to help people find reliable information quickly without sacrificing quality. We can do better.