Monday, July 1, 2024
This week's edition
Happy Canada Day all! 🇨🇦
This week marks the re-launch of our newsletter!
Dose of Data Clarity will now come out on Monday mornings (ET) and it will focus on synthesising the hot data topics and discussions that are going on in the industry.
In this issue, we'll be diving into:
🤖 LLMs are amazing, but are they being applied to the right problems?
🌎 Dark data and the environmental impact of forgetting how to:
drop table d_customer_master_dev_test2;Â
🥊 OpenAI, LLaMA or Claude? The Salesforce LLM rankings are in.
😄 Your Weekly Dose of Data Hilarity
I hope you enjoy the issue!
Best regards,
Dave
🤖 LLMs are amazing, but are they being applied to the right problems?
Last week's AP article by Wyatte Grantham-Philips, which discussed the difficulties faced by McDonald's AI-driven drive-thru, generated quite a bit of discussion in the data community.
The IBM-powered solution, while impressive in many regards, fell short of its required accuracy. According to Cate Chapman from LinkedIn News, 1-in-5 orders required human intervention. It's clear that having 20% of drive-thru customers experience friction in their ordering process is not something that was acceptable to McDonalds, as the program with IBM is coming to an end this month at all test locations.
Even though the experiment didn't work for McDonalds, Grantham-Philips notes that other brands have successfully implemented similar technology. Popeyes even goes as far as to say their solution is 97% accurate at capturing orders, however, some people may have a different take on that figure. 🤣
Stories like these really underscore that understanding user needs and the delivered user experience is paramount. Without that, will anyone actually use the stuff we build? 😅
🌎 Dark data, AI and the environment
Author and data executive Sol Rashidi generated some conversation on LinkedIn with her thoughts on dark data and the energy consumption of LLMs like ChatGPT.
According to Sol, ChatGPT consumes half a litre of water for every 5-10 prompts, and an AI generated image consumes enough power to charge your smartphone to 24%! Another post suggests that a query can consume the equivalent of running a 10-watt LED light bulb for 18 minutes to 3 hours. That's quite the energy footprint!
The flip side of this is the topic of dark data. This is data that resides in your environment and is not being used for any purposes. It's collecting virtual dust. It's using up space on your servers, which equates to energy, and if this data is hooked up to pipelines its footprint is larger, especially when you consider the overhead of maintaining things that aren't being used.
Data Strategist Malcolm Hawker provides his take on the issue of dark data in his Forbes article. Malcolm offers up some good suggestions on dealing with dark data, which is becoming increasingly important these days as more companies are setting ambitious ESG goals.
I guess now is as good a time as any to take a look in our own (server) closet, figure out what's being used and what isn't, and do a little digital spring cleaning! 🌱
🥊 OpenAI, LLaMA or Claude? The Salesforce LLM rankings are in.
Last week, Salesforce AI Chief Clara Shih posted the company's LLM benchmark for CRM. Shih notes that this is an ever evolving space, with new models dropping all the time, the foundational model layer becoming increasingly commoditised.
This is absolutely correct. A quick check on Hugging Face shows over 700,000 models equipped to handle pretty much any task that you can think of.
With rapid commoditisation in the model layer, Shih points out that customers should consider using an offering (like the Salesforce AI stack) that acts as an abstraction layer and allows interchangeability between the off-the-shelf model of the day and BYO models.
The results of the Salesforce study also point out that out-of-the-box model accuracy is just one dimension when selecting your models. Enterprise users also need to consider cost and speed when making their choice.
When you put it all together, it looks like Claude is becoming a real contender! 🥊
😄 Your Weekly Dose of Data Hilarity
We've all been here at one point in our careers.... right?