On the Edge #25
We dropped our first NFTs. It’s a present with public and private values so everyone can enjoy the public part. Check out the items on Opensea here and a small intro about the project here 🎁🌸
It was on my mind for quite a time to write a brief about some critical points of progress in AI in short post, especially the impact of transformers architecture, and recently generative models and this post from MIT Tech Review has done the same link
I learned that OpenAI planned to release GPT-4 in early 2023, but in Q4 2022 they decided to change the release time. Most people inside OpenAI were focused on and ready for GPT-4 and they didn’t know about the new decision. In November, employees receive an email that a version of GPT-3.5 will be opened to chat with the public to gather feedback. (This was the davinci-text-003 that we discussed in OtE #16, which had evident supremacy).
This decision has a management lesson that sometimes, even if you have something running in production, it’s better not to ship the bigger version but to get more feedback with the older version in a new form.
Another point is that Microsoft quickly decided to add ChatGPT to Bing, while Google was hesitant for quite a while to add LLMs to search. Microsoft Bing has less 10% of the internet search market, so they had a larger capacity to risk. So another point to be taken is not to be afraid of a huge competitor, real change is possible.
It’s also worth mentioning that in a recent study (link) which evaluates OpenAI models using Theory of Mind tests, GPT models before 2022 were showing almost no capability. The davinci-text-002 (GPT-3) that was released in Jan 2022 solved 70% of tests (comparable to 7 years old children) and the davinci-text-003 (GPT-3.5), released in Nov 2022 solved 93% (comparable to 9 years old children).
This piece from Harmony Korine, Child’s Play Blockbuster link
All-time favorite lip sync link
Detective novels
Rusted Root are one of those hippy bands that every music festival needs
Playlists