Clearly - this week in generative AI

Three things that happened in the world of generative AI this week:

Generative AI is good for simple tasks, but it can’t be creative, right?

It’s pretty common to hear people who have played around with Chat GPT for a bit to say that while genAI is impressive, and clearly very good at relatively simple writing tasks, it’s definitely no good at higher level creativity. And while this may be true for certain tasks - there’s little sign of a great novel coming out of an LLM just yet -?a?group of Ivy League researchers recently compared the innovation capabilities of a group of innovation students to outputs from GPT-4. The findings were unambiguous - GPT-4 can generate ideas much faster and cheaper than students and the ideas are on average of higher quality. A word of caution before you fire your innovation team - this is a study of AI versus innovation students not AI versus innovation professionals with years of experience under their belt. Of course experience is not a guarantee of quality in innovation - the mean age of startup founders in the US is 42, so experience clearly plays some part. The study does suggest, however, that genAI can be a powerful creative tool, and that, if used correctly, can play a significant role in innovation and creation processes. ?

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4526071

SDXL launches with a high quality image model that competes with Midjourney ?

For most of the last year Midjourney has been the king of the generative image models, producing significantly more impressive results than competitors. This week, however, Stability AI released SDXL, their latest model, and it seems that it’s pretty comparable in quality for many subjects and tasks. What’s more, the free version of SDXL is easier to access and use - Midjourney is still only available via Discord , in what is a pretty unfriendly user experience. Our own experiments suggest that Midjourney is perhaps a little less prone to visual hallucinations, so perhaps not quite time to jettison the Midjourney subscription just yet. SDXL is a powerful model, and Midjourney now faces a very significant challenge, and it will be interesting to see how they respond. It seems likely that experience design and pricing models will become the next battlegrounds now underlying capabilities are at or near parity. ?

https://clipdrop.co/stable-diffusion

Inflection AI claims to be training models 100 times larger than GPT4?

Last month Inflection AI raised over $140m from Greylock , 英伟达 and others. This week CEO, Mustafa Suleyman , gave an interview to Barrons, in which he claimed to be preparing to train models that are 10 times larger than GPT-4 and then, soon after, ones that are 100 times larger than GPT-4. He claims this is going to be ‘eye-wateringly different’ and that these new models will, in the next two years,?largely eliminate hallucinations. It is possible that there is a little tech bro hyperbole going on here, but even if these new models are only 10x larger than GPT-4 this would have a very significant impact in a very short time.?

https://www.barrons.com/articles/ai-chatbot-siri-alexa-inflection-pi-fa1809f8

And a final, bonus ball for those with a technical bent and an hour or two to spare - a good, jargon-free run down of how LLMs actually work.