Almost Timely News: What’s Coming With Generative AI (2023-04-09)
Almost Timely News: What’s Coming With Generative AI (2023-04-09) ::?View in Browser
Content Authenticity Statement
100% of this newsletter was written by me, the human, with no contribution from AI except in the displayed outputs.
Watch This Newsletter On YouTube ??
What’s On My Mind: What’s Coming With Generative AI
Ann Handley posed the implicit question, “What’s coming with generative AI?” in a?recent blog post?she wrote. It’s an interesting question with a lot of different answers. Let’s tackle where the field of generative AI appears to be going and what you and I should be doing about it.
First, the big generative models like OpenAI’s GPT-4 will keep getting bigger. GPT-4 is interesting in that it didn’t really introduce any new knowledge - the model is still stuck in September 2021 as the last date it has knowledge, and boy has a lot happened since then - but it did increase its model parameters by an order of magnitude.
What does that mean? In non-technical terms, it’s able to remember more and generate more. The more parameters a model has, the bigger its pattern recognition gets. A model with 100 million parameters can remember a sentence it’s written, maybe two. Back in the early days of GPT-2, you’d notice the model started to generate nonsensical gibberish after a paragraph or two, because it had effectively forgotten what it had written.
When GPT-3 came out, it was able to remember and generate coherent paragraphs, several at a time before it started to lose its way. If you were skillful with prompts and lucky, you might even get a short page of text out of it.
GPT-4? It’s able to do multiple pages of text. How do we know? The answer is in the developer documentation. The documentation given to programmers and developers tells us how much the model can expect to ingest, and how much it can spit out.
If you look back at the documentation from the past, developers could send 1,024 tokens - word fragments or words - to GPT-2 and generate about that many (that’s about 700 words). GPT-3 permitted 2,048 or thereabouts, about 1,500 words. GPT-3.5, the model that ChatGPT uses by default, can accept and generate about 4,096 tokens/word fragments, about 3,000 words. In the API for GPT-4? That extends all the way out to 32,000 tokens, or about 24,000 words.
What that means is that the publicly available generative language models have gone from cranking out a paragraph at a time to cranking out a novella at a time. What will a future GPT-5 be able to do? Probably about 64,000 tokens, or about 45,000 words - the average length of a business book. Put a book in, and tell this futuristic model to rewrite it in Aramaic, or entirely with emoji, etc.
The tradeoff is that as each model gets bigger, it requires more and more computational power to create and to use. But for the public, and for companies that just want to get up and go, these will be the models we use and we will love them.
Think, for a moment, about every story you’ve ever started to enjoy but couldn’t complete because the work was left undone. Maybe the author stopped writing. Maybe the TV show got cancelled before it could wrap up all the loose ends - or any of the loose ends (looking at you,?Warrior Nun?and Legends of Tomorrow). Maybe the podcast just went off the air. Whatever the case may be, with today’s and future large language models, it will be feasible for the average person to feed in the story as it is and get a logical completion of the story. Would it be what the original author wrote? Maybe, maybe not. But it will exist nonetheless.
The same is true for any large text work. Suppose I fine-tuned GPT-4 on all my blog posts and books, but I really didn’t feel like writing my next book, or I felt like dictating a bunch of disconnected thoughts, then gave the random thoughts and an outline to GPT-4 as a prompt. Could it write my next book for me? Yup. Would it be as good as me hand-crafting it? Probably not as a final product, but it’d get me 95% of the way there and then all I’d have to do is edit it.
So what do you do about bigger models? Plan for their use. Start thinking about the long form content you’d like them to tackle. Start designing the extensive, multi-page prompts they’ll need to generate the results you want.
Now when it comes to generative models, bigger isn’t always better, which brings us to point two. Computationally, these multibillion or soon to be trillion parameter models are slow, costly to operate, and prone to weirdness because they have the sum total of human language in them in some capacity, and when you work with humanity as a whole, things get weird by default. The second place these models are going is… smaller.
Smaller? Why smaller? Because what we’re seeing more and more is companies creating purpose-built, custom models that are focused on a specific task. They start with a very small model, like?Eleuther.ai’s GPT-J-6B model?(which is small enough to tune and run on a gaming laptop), and then these companies fine-tune it - tech speak for retraining it - on one specific domain of expertise. The model can no longer write fanfiction or make guitar tabs out of lyrics, but it can do its functional focus very, very well - better than any of the big general models.
We saw this very recently with the?announcement of BloombergGPT, Bloomberg LP’s custom-built model programmed to do one specific thing: financial analysis. Within the Bloomberg Terminal, the venerable 41-year old software package that forms the backbone of Bloomberg LP’s business, subscribers can now ask natural-language questions of their data. For example, an analyst could ask, “Find 10 stocks that have a dividend greater than 10% that have shown positive growth in the last 5 years,” and get an accurate answer.
Within the marketing domain, our friends and colleagues at?GoCharlie?have released their CHARLIE LLM, a custom-tuned model specifically made just for content marketing. It also can’t rewrite the lyrics to a Guns ‘N’ Roses song very well, but it’s a lot more knowledgeable about creating and processing marketing content.
This is the second avenue that we’re going to see a lot more of in the coming months. Any forward-thinking company that has a lot of data that’s difficult to process for the average end user is looking at building custom models for their specific domain, because the model doesn’t have to be able to do everything - it just needs to be able to do a small number of things really, really well. And here’s a secret about custom models: the more tuned they are for a specific task, the shorter and less specific your prompts have to be, because you don’t need nearly as many guardrails in place. Its domain knowledge is so focused that there’s much less ambiguity in prompting them.
Any software company that has even a moderately complex interface is or should be looking at integrating an LLM into its interface to simplify tasks. If a company isn’t, a competitor is, and that competitor is going to eat the legacy player’s lunch.
For example, Adobe had better be looking at an LLM for tools like Photoshop. Imagine the power given to the average user to just say, “Hey, make this black and white photo in color, and remove my ex from it”, rather than needing to follow the 54 clicks and menus needed to do the same thing. Will it be perfect? No.?Will it be better than what a true Photoshop expert can do? No.?Will it be better than what an untrained user can do? Absolutely - and that’s the benefit of these kinds of models.
Large language models and generative AI rarely create something better than what a true expert can do. More often than not, they create mediocre to good results, but rarely excellent and never unique, never-before-seen results. That’s because their very nature is an averaging of what already exists, tuned to deliver above average results. Above average never beats expert, never beats world class.
But a fair amount of the time, we don’t need to beat world class on many tasks. Above average - or even mediocre - is an improvement. I’m a completely incompetent musician. I have no musical skills whatever - I know what I like in music, but I can’t create it for the life of me. Can an AI create music on my behalf? Yes. Is it good? It’s good enough for use cases where it doesn’t matter very much, like background music in a video that’s not about the music. Is it good enough to satisfy the idea I have in my head that I don’t know how to bring out of my head? Yes. Is it better than hiring a true professional musician? No.
For a company like?MarketingProfs?that has decades of text stored up? A custom-built large language model could give marketers a chat interface and a natural language solution built on highly-curated text.
The key takeaway here for you is to think about which domain-specific tasks your company does that could use a custom-built generative AI model. Perhaps you work in architecture or engineering. What would a custom-tuned model for generating or evaluating proposals look like? How much time would that save you if you had a model that did just a specific subset of repetitive tasks?
The third area which is coming very, very soon for the non-technical user is multimodal (technical users can already access this in big models like OpenAI’s GPT-4 through the API, though there’s a lengthy waitlist for access). This is the transformation of one mode of data to another. This is where we give the model an input that isn’t words, but words are the output - and in the future, we’ll see a convergence where we put in words and get out non-words output. We already have some experience with this with tools like Stable Diffusion, Midjourney, and DALL-E which take word prompts and generate images. The reverse is now possible as well - taking an image and having the model transform it to words.
Imagine taking a photograph from a conference and having the machine write up what it sees. Imagine taking a screenshot from your Google Analytics account and having the machine write a summary of your data. That’s what multimodal really is all about - the ability to transform things in and out of words.
I saw a quote on Twitter from Andrej Karpathy that encapsulates this well: “The hottest programming language in 2023 is English”. This is wholly accurate. If you can write words (in any language, not just English - I’ve had some excellent success with GPT-4 in Danish, for example), you can program the machines to do what you want them to do. I’ve fed models my Google Analytics data and had them write marketing recommendations. I’ve had models reverse engineer songs just from lyrics, transforming them into guitar tablatures that sound decent.
The key takeaway here is to start looking for non-word inputs like images that you need to perform repetitive tasks on, and plan for how you’re going to use them with large language models.
The final place these models are going is with text-like inputs that aren’t really text, but can be used as text. Take a genetic sequence. A genetic sequence can be represented like this:
AGTCATTGACATAAATCCAAGGATAATA
These are the four base pairs of DNA, written out as text. Suppose instead of making limericks, we purpose-built a model to only work with DNA and RNA, but otherwise use the same underlying technology. What could you do if you could input genetic data? What could you generate or forecast?
Novel gene therapies.
Vaccine candidates.
Understandings of protein foldings and misfoldings.
Literally the cure for cancer.
That’s what the technology behind tools like GPT-4 and ChatGPT are capable of, with enough training and specificity. They’re capable of working with text-like data and making predictions from that text-like data - and I can’t imagine the largest pharmaceutical companies in the world aren’t already doing this. If they’re not… well, they’re missing the boat. (and if you work with one of them and you’re missing the boat,?call Trust Insights.)
This is where things are going with generative AI: bigger models, custom models, multimodal, and non-text applications. If we do this well as a society and civilization, we will see enormous benefit and huge leaps in progress. If we do it badly, we’ll see dramatically worsened income inequality and scarcity battles everywhere. But good or bad, this is the direction things are going in the very near future - and I mean a year away or less.
Are you ready? Is your company?
Shameless plug: this is all offered as a?talk?that I’m doing for companies as well as events.?If you’d like me to bring it to your company, let’s chat.
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 livestream from Thursday on Marketing Mix Modeling. You’ll get a chance to see how a marketing mix model works - and whether you can trust it or not.
领英推荐
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
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 five most recent open positions, and check out the Slack group for the comprehensive list.
Advertisement: LinkedIn For Job Seekers & Personal Branding
It’s kind of rough out there with new headlines every day announcing tens of thousands of layoffs. To help a little, I put together a new edition of the Trust Insights Power Up Your LinkedIn course, totally for free.
What makes this course different? Here’s the thing about LinkedIn. Unlike other social networks, LinkedIn’s engineers regularly publish very technical papers about exactly how LinkedIn works. I read the papers, put all the clues together about the different algorithms that make LinkedIn work, and then create advice based on those technical clues. So I’m a lot more confident in suggestions about what works on LinkedIn because of that firsthand information than other social networks.
If you find it valuable, please share it with anyone who might need help tuning up their LinkedIn efforts for things like job hunting.
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: Google Analytics 4 for Marketers (UPDATED)
I heard you loud and clear. On Slack, in surveys, at events, you’ve said you want one thing more than anything else:?Google Analytics 4 training. I heard you, and I’ve got you covered. The new?Trust Insights Google Analytics 4 For Marketers Course?is the comprehensive training solution that will get you up to speed thoroughly in Google Analytics 4.
What makes this different than other training courses?
With more than 5 hours of content across 17 lessons, plus templates, spreadsheets, transcripts, and certificates of completion, you’ll master Google Analytics 4 in ways no other course can teach you.
If you already signed up for this course in the past, Chapter 8 on Google Analytics 4 configuration was JUST refreshed, so be sure to sign back in and take Chapter 8 again!
Tools, Machine Learning, and AI
Analytics, Stats, and Data Science
Dealer’s Choice : Random Stuff
Advertisement: Ukraine ???? Humanitarian Fund
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.
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:
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
Founder and Webmaster for @DataWavves | Demand Generation and Digital Marketing Pro | RevOps Marketing Professional | South Florida’s Funniest Marketing Manager | office hours 9am to 9pm EST Boca Raton
1 年Awesome ??Ann Handley !! The Legend, can’t wait to listen tomorrow