Almost Timely News: The Fallacy of the Single AI (2023-08-06)
Almost Timely News: The Fallacy of the Single AI (2023-08-06) ::?View in Browser
Content Authenticity Statement
100% of this newsletter was generated by me, the human. When I use AI to generate content, I will prominently label it. This is important because AI output cannot be copyrighted; this disclosure reinforces my copyright claim over my material.
Watch This Newsletter On YouTube ??
What’s On My Mind: The Fallacy of the Single AI
What’s the biggest blind spot people have about generative artificial intelligence and its capabilities? Beyond fanciful science fiction predictions that don’t really have a basis in reality, the biggest blind spot most people have about artificial intelligence, at least currently, is the fallacy of the single intelligence.
Here’s what this means. The power of generative AI becomes apparent as complexity increases. The first generative AI models were fairly limited in capabilities because they were trained on such little data. As companies like OpenAI and Google added more data, more capabilities of these models began to emerge - the ability to answer more complex questions, solve more complex problems, take on tasks that previously AI models had no hope of achieving.
This created a natural eagerness to build bigger and bigger models, something we see when people talk about AI models and how many parameters they have - 7 billion, 30 billion, 70 billion, etc. The natural assumption is that a bigger model is a better model, and to a great degree, this has proven true. GPT-3.5 was an improvement on GPT-3. GPT-4 is an improvement on GPT-3.5. The newest version of any model generally has more tokens and more parameters and more training time than its predecessor.
So it seems natural that people would assume that bigger is always better, and that whatever problem we’re facing can be solved by making models bigger. And we have ample precedent for this assumption; we saw it during the chip wars when companies like Intel and AMD kept making faster and faster CPUs, touting each chip’s speed as better than the last. I remember long ago when Intel earned massive bragging rights for reaching a consumer chip speed of 1 gigahertz back in March of 2000.
You would think, based on that example, we would have something like 23 gigahertz computers today, but we do not. In the case of CPUs, we ran into literal limits of physics that prevented us from making chips that operated much beyond 4 GHz. For example, the faster a chip goes, the more heat it generates, and the harder it is to keep it synchronized across billions of operations.
We are running into similar limitations with large language models and generative AI. The bigger a model is, the more compute power is needed to make it and operate it. Think of it like manufacturing bigger and bigger cars; not only does the assembly line need to be bigger and consume more resources, so does the final product once it’s built. Yes, a bigger model in general delivers better quality than a smaller model, but we are running into real resource limitations for bigger models. Anyone who’s logged into a system like ChatGPT has seen issues with speed and quality;?ChatGPT’s premier model, GPT-4, appears to have gotten dumber over time.
That’s why the fallacy of the single intelligence is problematic; the assumption that bigger is better will lead people to chase after bigger numbers, bigger models. The parallel assumption is that bigger models also mean more and more capabilities, more abilities for models to do ever more complicated tasks.
What happened with computers once we started hitting physical limits of computing power? The industry pivoted from making faster chips to making more of them bundled together. You started seeing companies touting that they had two CPUs - cores, then four cores, six cores, etc. Today, companies like Apple advertise having 12 core CPUs and 32 core neural engines as a way to suggest the total processing power of a system.
And that evolution - from one big chip to a bundle of chips - is what the evolution of generative AI will look like as well. In general, nature addresses complex problems with complex solutions involving many parts rather than one bigger part. One animal’s survival is not guaranteed, but having a pack of animals greatly increases the survival capabilities of every animal in the pack. We do the same thing. Think about it - can one employee power an entire company? Or do companies hire many people?
That’s where generative AI is already going; instead of trying to make a bigger, more magical single model that can do everything, we are seeing more architectures of model ensembles, collections of models chained together to accomplish tasks. Fable Studios showed this off very capably in their recent South Park generative episode; they didn’t try to force one model to do everything. Instead, they built an ensemble of different models that each did individual tasks, sewing them together with some infrastructure.
So what? Why is any of this important to the many folks who just want to use AI to solve problems? It’s important to understand because it frames how we need to think about problems and the AI solutions that we will commission to solve those problems. What we’re seeing with agent networks - ensembles of AI models - is that they’re very capable of taking a complex task, breaking it down into much smaller tasks, and then using individual models to solve those specific smaller tasks. That’s no different than how we humans solve problems; we take a big task like “achieve profitability” and break it down into smaller tasks like marketing, lead generation, deal management, etc. Until we pile up closed sales and achieve our goal.
Right now, many AI companies are little more than a skin or a wrapper on someone else’s model, usually with a pretty interface to make things easier to understand, or with specific prompts under the hood to generate desired outputs. This class of companies will go away fairly quickly. Those companies which will survive and prosper? They’ll have ensembles of models put together in very specific ways to achieve results that exceed what any individual model is capable of alone.
And that’s where the magic will happen, where companies will differentiate themselves. When you have one gigantic all-purpose super-model, there’s no real opportunity for users of the models to differentiate, to show any kind of uniqueness. If you take two people who are equally competent at prompt engineering and have them use such a system, one person will not achieve substantially greater results than another.
When you have ensembled systems? You can achieve significant differentiation, because the models you choose under the hood and the way you connect them together will dictate the results you get.
Here’s how. Suppose you were going to cook dinner and you had one super-appliance that did a bunch of things reasonably well, like every appliance manufacturer often touts. “13 appliances in one, it fries, it bakes, it broils, it makes soup!” We know those appliances can do many of those things, though not necessarily equally well. In that case, the appliance would not be a differentiator for the quality of your food because everyone would be using the same thing. (Said appliance would also be crazy expensive)
Now suppose you have a kitchen full of different appliances - a great blender, a nice induction stove, a deep fryer, etc. Instead of cooking everything in one machine, you’d use different machines for different capabilities, and in different orders. For example, you would probably not want to use a blender of steak. You would absolutely want to use a blender to mix the gravy for a roast, though. You probably would not want to use a deep fryer for soup, but you could absolutely use a deep fryer to make shrimp tempura that you put into an udon noodle soup.
The network of appliances in a kitchen creates more powerful outcomes than one single super-appliance; the network of AI models will create more powerful outcomes than one single super-model.
As end users of AI, we will probably care little what the underlying architecture of systems are, so long as we get the results we want. But as economics begins to take hold, as we begin to hit practical limits of compute power (electricity, cost, etc.), ensembles systems will be the way we continue to expand the power of generative AI - and just as we have with computer CPUs today, this change in strategy will let us keep growing the power of AI over time.
As consumers of AI - particularly if we are working with vendors that we’re paying money to - understanding what’s under the hood will help us understand the value we’re getting and how current a vendor can stay with the latest technologies. For example, if a company’s technology is predicated on a single super-model (like GPT-4), then what tends to happen is that company gets locked into that particular model. The next time something newer and better comes along, it can be challenging for the company to change the model, in the same way that it can be mechanically very challenging to rip the motor out of a car and put a totally different motor in.
If you’re built on ensembled models - particularly if you’ve been clever and smart and built an abstraction layer in - then when something new comes along, you don’t have to start from scratch. You can swap out the affected pieces while leaving the entire system intact, just like you could buy a new blender for your kitchen without it affecting your stove or microwave. That kind of agility is what we want to look for in our AI vendors, the agility to change with the ever-increasing pace of change in AI.
Whether you’re buying AI technology from someone else or building it yourself, consider how the technology is evolving and then decide whether you want to pursue or purchase systems based on one big model or a network of models. I would wager the second choice is probably the one you want to go with.
Got a Question? Hit Reply
I do actually read the replies.
Share With a Friend or Colleague
If you enjoy this newsletter and want to share it with a friend/colleague, please do. Send this URL to your friend/colleague:
ICYMI: In Case You Missed it
Besides the newly-refreshed?Google Analytics 4 course I’m relentlessly promoting (sorry not sorry), I recommend the piece on
Skill Up With Classes
These are just a few of the classes I have available over at the Trust Insights website that you can take.
Premium
Free
Advertisement: Bring My AI Talk To Your Company
领英推荐
I’ve been lecturing a lot on large language models and generative AI (think ChatGPT) lately, and inevitably, there’s far more material than time permits at a regular conference keynote. There’s a lot more value to be unlocked - and that value can be unlocked by bringing me in to speak at your company. In a customized version of my AI keynote talk, delivered either in-person or virtually, we’ll cover all the high points of the talk, but specific to your industry, and critically, offer a ton of time to answer your specific questions that you might not feel comfortable asking in a public forum.
Here’s what one participant said after a working session at one of the world’s biggest consulting firms:
“No kidding, this was the best hour of learning or knowledge-sharing I’ve had in my years at the Firm. Chris’ expertise and context-setting was super-thought provoking and perfectly delivered. I was side-slacking teammates throughout the session to share insights and ideas. Very energizing and highly practical! Thanks so much for putting it together!”
Pricing begins at US$7,500 and will vary significantly based on whether it’s in person or not, and how much time you need to get the most value from the experience.
Get Back to Work
Folks who post jobs in the free?Analytics for Marketers Slack community?may have those jobs shared here, too. If you’re looking for work, check out these recent open positions, and check out the Slack group for the comprehensive list.
What I’m Reading: Your Stuff
Let’s look at the most interesting content from around the web on topics you care about, some of which you might have even written.
Social Media Marketing
Media and Content
SEO, Google, and Paid Media
Advertisement: Business Cameos
If you’re familiar with the Cameo system - where people hire well-known folks for short video clips - then you’ll totally get Thinkers One. Created by my friend Mitch Joel, Thinkers One lets you connect with the biggest thinkers for short videos on topics you care about. I’ve got a whole slew of Thinkers One Cameo-style topics for video clips you can use at internal company meetings, events, or even just for yourself. Want me to tell your boss that you need to be paying attention to generative AI right now?
Tools, Machine Learning, and AI
Analytics, Stats, and Data Science
All Things IBM
Dealer’s Choice : Random Stuff
How to Stay in Touch
Let’s make sure we’re connected in the places it suits you best. Here’s where you can find different content:
Advertisement: Ukraine ???? Humanitarian Fund
The war to free Ukraine continues. If you’d like to support humanitarian efforts in Ukraine, the Ukrainian government has set up a special portal, United24, to help make contributing easy. The effort to free Ukraine from Russia’s illegal invasion needs our ongoing support.
Events I’ll Be At
Here’s where I’m speaking and attending. Say hi if you’re at an event also:
Events marked with a physical location may become virtual if conditions and safety warrant it.
If you’re an event organizer, let me help your event shine.?Visit my speaking page for more details.
Can’t be at an event? Stop by my private Slack group instead,?Analytics for Marketers.
Required Disclosures
Events with links have purchased sponsorships in this newsletter and as a result, I receive direct financial compensation for promoting them.
Advertisements in this newsletter have paid to be promoted, and as a result, I receive direct financial compensation for promoting them.
My company, Trust Insights, maintains business partnerships with companies including, but not limited to, IBM, Cisco Systems, Amazon, Talkwalker, MarketingProfs, MarketMuse, Agorapulse, Hubspot, Informa, Demandbase, The Marketing AI Institute, and others.?While links shared from partners are not explicit endorsements, nor do they directly financially benefit Trust Insights, a commercial relationship exists for which Trust Insights may receive indirect financial benefit, and thus I may receive indirect financial benefit from them as well.
Thank You
Thanks for subscribing and reading this far. I appreciate it. As always, thank you for your support, your attention, and your kindness.
See you next week,
Christopher S. Penn
You've put a name to the problem I had only sensed early on. And I kept thinking that maybe I needed to learn more about prompting. Or that I needed to find just the right tool. But you've nailed it--Ensembles! This is the way, Christopher Penn. Thank you.
Founder at Twibi
1 年Impressive insights! ?? Just imagine the possibilities with #generativeAI in the near future. Marking a shift in mindset will truly optimize AI's resource capabilities. ?? Can't wait to see how this shapes our interactions and creativity. #AIrevolution #limitlessgrowth
Sr. Manager, Professional Services @ Sprout Social
1 年Talissa Beall, Alice Chubb ??