Hands-on on Google PaLM API and MakerSuite: Summary: vastly behind GPT-4; more of a competitor to the Llamas. Details: of the 4/5 foundation models discussed at Google I/O, only the “Bison” model is available not available in Europe; Google emphasizes in various places that this is a preview, is experimental, … and its external or producti... Read more 16 May 2023 - 1 minute read
Gartner has published their guide “How Should CMOs Respond to ChatGPT Today?”. PDF metadata confirms Creation Date: April 19. But it seems to written around “ChatGPT-on-GPT-3.5”, so basically technology from 2020: (by Yang et al) Points like Numerical challenges. Large language models like ChatGPT converse by predicting what’s next in a t... Read more 11 May 2023 - 1 minute read
As someone who’s spent quite some money on Focus@Will, I was excited to get an invite to Google MusicLM (under the “AI Test Kitchen” umbrella). Unfortunately, it’s so bad: prompts take little effect, it’s more like a random generator. Could not even reproduce the example from the abstract of the paper. very poor audio quality, like on a 90i... Read more 11 May 2023 - less than 1 minute read
One colleague recently shared that ChatGPT knows the inside of a (rare?) book which is not available in fulltext online, concluding that they must have scanned works. This study takes a systematic approach to survey this: An Archaeology of Books Known to ChatGPT/GPT-4. Code and data here: gpt4-books. Furthermore, this other study presents proof ... Read more 06 May 2023 - 1 minute read
I have published my content filtering/scoring system for RSS feeds on Github: FilterRSS. Of general interest may be how GPT-4 can be used to self-instruct, to build the actual filters from a few examples: FilterRSS#system-prompt - or at least give you some ideas for starters. My use-cases include this leading researcher on Twitter who unfortuna... Read more 04 May 2023 - less than 1 minute read