Fireside | From Supercomputing to Software feat. Kevin from AMD

Sep 28, 2024

Watch on YouTube

Takeaways

  • Building data center is similar to planning a small town, requiring careful planning and consideration of various factors, including power consumption, space requirements, and cooling solutions.
  • Understanding the AI landscape and identifying the positioning is crucial for success in AI software development.
  • How to allocate the resources between industry and academia? The lack of computing resources in academia make people struggle to continue their foundational research.
  • Utilize available AI tools like Morphic, Scispace and NotebookLM to accelerate her research and learning process.

Insights

Can you describe your personal experience of how AI has changed from 2016 to now?

Back in 2017, when I visited Google Brain, I was surprised to see that the labeling process was still heavily manual. That same year, I also had the opportunity to ride in a pre-release Tesla Autopilot vehicle and experienced firsthand the advancements they have made in self-driving technology since then. Now, you can see Waymo in San Francisco.

With your experience in AI, do you believe that the visions you had then are being realized now? Are there any visions that have been achieved and others that haven't yet?

I believe some of the visions I had are being realized in commercial applications. My team, for instance, used AI for marketing and advertising to generate images and improve conversion rates. However, I believe that wider adoption of AI across industries will take time as technology and businesses need to integrate. Advancements in both software and hardware are needed to achieve more ambitious visions, such as the metaverse, which require significant computational resources and technological breakthroughs.

Could you describe the AI landscape and how it has changed, especially for new individuals entering the field?

The AI landscape, particularly the application layer, is in a constant state of flux. I would recommend that individuals new to the field focus on developing specific or hyperlocal AI models rather than attempting to build general or foundational models like those created by large corporations. It's also essential to understand the overall AI landscape and identify a niche to focus on. One of the biggest challenges in AI development, especially for those in academia, is the limited access to resources, particularly expensive hardware like GPUs.

What inference platforms do you use and recommend for working with AI models, especially in terms of speed and efficiency?

I recommend 'Together AI' and 'groq,' particularly the latter, for its speed and configurability. When working with large language models like 'llama3.1 405b,' selecting the right inference platform is critical as it directly affects inference time. Choosing an inference platform should be based on your project's requirements, specifically speed, resource needs, and model compatibility.