“6 New Theories About AI: Software With Superpowers § GPT-4”, Evan Armstrong2022-11-22 (GPT-4, AI economics; backlinks)⁠:

…I’ve since heard tales of multiple startups that had spent the last few years fine-tuning their own model for their own use case. Thousands of hours devoted to training, to labeling specific data sets, to curation. Then, when OpenAI released GPT-3, it was an order of magnitude better than their own version.

…For those startups that are competing on the basis of SaaS, the competitive advantage goes toward companies that already have inherent distribution or product capabilities. It would be much easier for a large company to integrate AI into their existing products than for startups to build competitive full-suite products from scratch. Microsoft, Canva, Notion, and others have already replicated some of these startups’ capabilities within their existing product suites. I expect every major software provider to integrate generative AI into their products over the next 6 months.

Note: Credible sources have told me that GPT-3’s successor GPT-4 is far beyond what people are expecting. It’s currently in testing with a variety of OpenAI’s friends and family, and it will leave most fine-tuned models in the dust.

1. Fine-tuned models win battles, foundational models win wars

2. Long-term model differentiation comes from data-generating use cases

3. open source makes AI startups into consulting shops, not SaaS companies

4. Most endpoints compete on GTM, not AI

5. AI will not disrupt the creator economy, it will only amplify existing power law dynamics. Invisible AI will be the most valuable deployment of AI

…Even something as mundane as file management can be reimagined. The typical file structure is a series of cascading folders: photos go to July photos, go to the 4^th of July, etc. In a world where AI can generate hundreds of photos every minute, what are you supposed to do? SOOT uses AI to auto-tag images by their components so you can dump them all into a visual space and sort them by their content (as well as time and variety of other attributes). For example, you can look for swimming pictures…

…and the AI will find and sort them into the ones most like your description.

Lots of companies do this, but it’s even cooler when you mix those search capabilities with generative AI. Let’s say you were looking to create a new arrow icon and had a reference image. Click on it and, in the modify field, write that you’re looking for an arrow in the style of Virgil Abloh. The AI will create a bunch of variations. Contrast this with the previous process of file management and image creation, where designers would flip back and forth between Figma, Photoshop, and Dropbox just to manage a few images.

This product is something that is only possible with AI. I don’t know what all of the startups will look like, but when you mix an invisible AI backend with AI generative capabilities, magic can happen. AI products can’t simply be model-deployed to existing use cases. AI products win when they enable entire new modalities of digital interactions.

Backlinks: