Honest-02: Should you build with GPT-3?
Happy Friday everyone,
Updates
This last week was filled with interesting concepts, from the update on Neuralink to discussions of the metaverse. Honest AI appeared in a call-in with Betaworks last Friday, joining panelists discussing how NLP is currently the hottest trend for startups, even if you exclude GPT-3. More improvements to the website have also been shipped, including a new article design. (love it? hate it?)
Meanwhile, the community has launched! While it is still in it's earliest form, the first brave users signed up last week. I will continue to add topics, links, and answer questions there so please join in and invite some friends or colleagues.
I encourage you to Sign up to the community for free!
Should you build with GPT-3?
In the first honest article, I break down the pros and cons of using the OpenAI API as the core part of a business. As with any new piece of technology it is important to understand the risks ahead of time, without getting too caught up in the hype.
Pricing
OpenAI debuted their pricing model for GPT-3 this week, which resulted in some pained grimaces. With pricing based per-token, it looks like many fixed-price apps will be rendered DOA. This limits the fun and creative cases along with the daily-helper apps. I am not saying they are wrong in pricing it high; If you have the best, you can charge the most. Based on my personal usage I would already be in the $400/month category. Notably this is just for the beta period, and the largest engine, davinci. We will have to wait and see for more details later. Pricing is enacted starting 10/1.
Deployment
New guidelines were also released as to which applications can actually use GPT-3. Tiered into 3 categories based in risk, they are taking an active stance to limit bad press. Notably they are banning their API's use for generating synthetic articles for SEO, pretty much anything to do with politics or medicine, and open-ended chatbots. Every usage will be hand-reviewed by the OpenAI team, which should help to ease fears that the internet will be soon overrun by intelligent bots.
It is safe to assume that with these updates the crazy hype seen over the last 3 months will start to cool off.
Interesting links this week –
If the pricing for GPT-3 has you a little disappointed (🙋) then maybe it is time to take a look at a more scaleable option. I came across Max Woolf's project aitextgen on Twitter and it looks quite interesting for specific use cases. Featuring a teeny-tiny model (1,400 times smaller than the flagship GPT-3), it should be very fast when trained for specific text generation. Just don't expect the same level of fake smarts shown by others.
Neuralink remains one of the most interesting companies in the world. While the aspirational, world-changing goals remain years or decades away, it is hard to not get caught up in the what-ifs. Machine learning is a core component of making sense of the complex waveforms emitted by the brain. If you haven't seen the video demo yet, check it out.
Do you want to understand a bit more about who designs these complex AI systems? While a few months old, this is a great conversation between Lex Fridman and Ilya Sutskever (co-founder of OpenAI) to see logical thinking in action.
Coming up –
- The first feature essay! A lengthy piece on AI, APIs, and the Metaverse. This is one you won't want to miss.
- A dive into why synthetic personas and other virtual media are going to be huge in the near future.
- More info on a sister project to Honest AI, a site that makes it very easy to use transformers (GPT2/3)!
Have a good weekend,
-Tyler