Looking for help forecasting inventory. You may want to take a close look at this. Bedrock is no doubt the best name in the generative AI space today.
I’m far from convinced that “generative” AI (Large language models and also image generation) is an effective approach for forecasting sales (and other stock movements) and hence necessary inventory cover. I feel that long standing and more proven big data machine learning models remain the best approach, but need careful integration to business processes, careful data management/cleansing (garbage in, garbage out), and a strong and proactive data science team.
Generative has emerging strengths in merchandising (textual and visual), and may have useful predictive power in product ideation and ranging, and a wide range of models help personalisation (display and pricing) and conversion/upselling/crosselling.
But less so in “execution”.
At this stage I am in agreement with you @Miles_Thomas. I wanted to highlight the inventory aspect because Amazon pushed it forward and I thought it was a bit interesting if nothing else.
Offering AWS customers access to generative artificial intelligence is for me, like the early solution for certain use cases. Use cases such as forecasting etc., feel like a way off - and almost need a niche solution rather than a generalist solution.
Yes this is a hyped sector, but AWS democratizing should concern Microsoft, Meta, Salesforce and others.
Having generative AI “do the math” doesn’t make a ton of sense to me, BUT, having generative AI provide explanations around decisions and/or to engage with a user for natural language requests makes some sense to me. “How many more units of chlorine tablets will I sell next week if I offered a 20% off coupon to all visitors? what’s the break-even point for gross margin $?”
Yep that’s a fair point, to democratise access to underlying ML models.
That many of these systems don’t understand math and the way their LLM is generating results would have me concerned that the data coming out of them would be unreliable if asking something like “How many more units of chlorine tablets will I sell next week if I offered a 20% off coupon to all visitors?” I think that’s a very valuable thing to be able to trend and know, but I’m not sure that any of the systems that I’ve seen out there so far can really understand that question and give a useful answer. What they give will seem useful, but I personally wouldn’t have a lot of confidence in it without backing explanations, which also need to not be hallucinations.
The right combination for many use cases will be a blend of more traditional ML, with LLM making it easier to interact with and control the ML layer.
Your sales uplift example is a perfect case in point. The ML layer can be trusted to give consistent answers for the uplift, but LLM makes it easier to get to that answer.
Because, math.
(a phrase I like to use with people who try to come up with something absurd where the fundamental calculations are incorrect.)