"If it’s easy to make a good-enough app out of the box, the barriers to entry are mercilessly low."

"If GPT-3 is so easy to adopt and build products with, incumbents will do it too."

"If the baseline GPT-3 performance cannot be substantially improved to create a substantial (10x) proprietary edge, the competition will shift away from technology to other dimensions of competition—particularly in marketing and distribution."

"Meanwhile, the profits will accrue to the true beneficiaries: 1) the algorithm owners,  OpenAI (and, by extension, Azure), 2) to marketing platforms, particularly Google and Facebook."

"Unfortunately, most of the products built on GPT-3 will be identical to each other, and few will have a meaningful edge."

"When Google first came out, it delivered such better search results than other engines—possibly by a subjective factor of 10 or more—that there was little reason to use any competitor."

"I argue: The better GPT-3 works out of the box, the harder it will be for any single company to build meaningful differentiation. Frankly, early demos already look very good."

"Before GPT-3, building anything close to the Good Enough AI was really hard. A company would have needed to invest many millions into its own algorithms and data cleanup to get anywhere close to the human user bar."

"There are many more competitors who can offer a “good enough” product."

"The range of competition is compressed into a narrower band."

"Because the companies don’t own the core technology behind GPT-3, their ability to improve beyond the baseline performance is limited."

"Then imagine how much worse this will get when GPT-4 comes out. And again with GPT-5."

"Plus, it’s likely that any proprietary progress you make on GPT-3 will be totally wiped out by GPT-4; the same way your tweaks on GPT-2 would have been wiped out by GPT-3."

"Yes, the algorithm isn’t everything. Products can still differentiate with their user experience, product design, customer support, adding on human services, and so on."

"How much can a GPT-3 user improve on the base algorithm experience in a proprietary way?"

"To get to the sweet spot of that improvement curve (80-20 rule), how much investment do you need to make?"

"That because GPT-3 already constitutes such a large dataset and heavy training, marginal improvements are hard to come by and will look log-linear"

"How about companies built on GPT-3? While the beta API is free, OpenAI will charge for API access. Every API call will cost you money. As you get more users and users use your product more, you will pay OpenAI more money. This sounds more like a Spotify situation than an Amazon situation."

"The simplistic reflex answer here is that AI companies can achieve this with data, but whether this effect even exists in AI is questionable. If the marginal improvements to more data plateau quickly, then companies with more data won’t have an advantage."

"In fact, since the most obvious uses of GPT-3 are human-to-computer interactions, it might be anti-network building."

"But with GPT-3, it’s likely this effect quickly taper off, particularly because GPT-3 only has a limited “working memory” (currently limited to around 1000 words). If you watch someone play AI Dungeon for a few minutes, this becomes apparent—while it stays fairly coherent in a narrow time window, it forgets characters and events that came just 5 minutes earlier."

"In Clayton Christensen’s terms, GPT-3 becomes a sustaining innovation that benefits the existing players, rather than a disruptive innovation that benefits new entrants."

"Sustaining innovations offer better performance or cost savings to existing products. They most benefit market leaders, who use sustaining innovations to augment their existing products."

"Disruptive innovations offer significantly worse performance than existing products; they at first look like toys, nothing that any professional user would rely on."

"A tricky but important point: within the same industry, an innovation can be disruptive to some incumbents and sustaining to other incumbents."

"Incumbents can easily use GPT-3 because it works right out of the box. Capital expenses are low, and building a GPT-3-powered product wouldn’t require a drastic overhaul of the company’s business model."

"Peloton, Lululemon, and Strava will build AI fitness coaches"