Shaping the Future of Artificial Intelligence

When my dad was in high school, they taught him how to use a typewriter.

?

When I was in high school, I learned how to type on a keyboard.

?

My sister is in high school now, and she's learning how to code and use 3D printers.

?

What will we be teaching high schoolers in 15 years? AI development.

?

If technology continues to advance at this breakneck pace, I wouldn't be surprised if in 15 years high school students are learning how to build LLMS and develop application layers for AI.

Naturally there needs to be significant development of this technology to bring it to the masses; and as a venture investor I'm keen on investing into platforms that will enable all 8 billion people on the planet to build and interact with AI.

So what needs to happen to create this reality?

Cost Compression of Model Training

AI models are not expensive to use after they've been put into production, but are incredibly expensive to develop and train. GPT-4 cost over $100M to build - a gargantuan amount to a small startup, let alone an individual. A single run of a model of the scale and complexity of GPT-4 can cost anywhere from $500K - $5M. Training costs may come down over time as NVIDIA and other AI-enabled semiconductor manufacturers develop novel silicon that have enhanced processing power, but that technology is still years away from commercialization.

In the same way that SpaceX's launch systems have decreased the cost to reach orbit and have spurred a wave of innovation in the space economy, AI needs a platform that can decrease the cost to train models and push them into production. AI development is a black box and is difficult to understand the true cost and ROI of creating such models. Some companies like Ternary inform business leaders of the underlying cost of developing a new model and platforms like Granica can help identify areas for cost reduction and improved efficiency in model development - but there needs to be better tools that can decrease the cost of AI development and inform business leaders on the cost and ROI of building this technology.

Model Governance and Ethics

One of the many fallacies around AI is that models will decide to end humanity. This is obviously wrong and is driven by a lack of understanding of AI, however there are biases within models that can lead to negative outcomes. In the same way humans can lose their train of thought, models drift from desired outcomes when live data doesn't match the ontology of their training data. Companies like Fiddler AI, Monitaur, and Credo AI help monitor models and alert AI practitioners to deficiencies in their models - but governance is subjective. We need a base level of AI governance akin to the international bill of rights developed by the UN.

No-code / Low-code AI Dev Tools

Publicly traded tech companies have long regarded their model development and deployment frameworks as core intellectual property. Many of those internal projects are now open source (e.g Uber's Michelangelo, Netflix's Metaflow, etc) but were created as the backbone for AI/ML models that informed Uber's route optimization and Netflix's recommendation engine, respectively. For consumers, there's no need for proprietary AI development frameworks. Character.AI is perfect for creating fictional characters. Stability AI 's Stable Diffusion or Midjourney are great platforms for text to image generation. API keys are easily accessible through all of these platforms but still require a degree of technical ability to stitch them all together. A solution like Monster API is great for developers, but not for the everyday consumer. Where are the tools for non-technical users?

Let’s Chat!

By investing in these areas we can propel a new era of AI development and make this technology accessible to enterprises and consumers alike. If you’re an early stage founder building AI infrastructure and dev tools, I’d love?to hear?from you!