Greetings, fellow humans. 👋

This is Not A Bot - the newsletter about AI that was definitely not written by AI. I’m Haroon, founder of AI For Anyone, and today, I'm excited to share another installment of Not A Bot's Expert Q&A series.

Today's guest is May Habib, CEO of Writer, an AI-powered content creation startup that just released a trio of large language models to power their enterprise copy assistant.

During our chat, May and I cover a range of topics, including:

• How companies can leverage Writer's foundational models

• Why Writer created their own models instead of training on top of generic ones

• The future of content marketing

• More

This was such an enlightening conversation, and I hope you enjoy it as much as I did!

Let's jump into it...

May Habib discusses Writer's foundational models and the future of content marketing

Not A Bot Expert Q&A Series: May Habib

Haroon: [00:00:00] May, so you had a big announcement just yesterday about Palmyra and Writer.com is releasing a set of foundational models. They're three different sizes and it's really exciting because I always thought that this was the future of foundational models.

We started with the generic ones. We had GPT-3, we have, a whole host of other LLMs and then inevitably they were gonna become more focused on particular use cases. And so do you wanna describe a little bit about the inception story of Palmyra, the motivation behind it, and a little bit more about how exactly it would work at a company?

May: Yeah, totally. We started experimenting with the OpenAI APIs maybe more than 18 months ago. And we're super early on Bert. Super early on T5. We knew we wanted to be both an encoder and decoder company as a result of just where our customers use cases were going.

And I think, even before [00:01:00] our customers got sophisticated enough, honestly to be telling us the things they were telling us, and for kind of all of us together to be seeing the future together it's never felt right that we would ever be able to build our whole business on APIs that were available to everybody, like I felt like such a non-defense play. Even before it became literally mission critical, we were already building Palmyra. And so it has been amazing just how quickly our customers' use cases have developed and their sophistication and maturity has accelerated, as well as our own team's ability to, just like really learn the technology.

Inference here is huge. We are a product that is built into realtime processes and the generations and the results have to be blazing fast and very good. And those two things are quite hard. And so it took us a while to really assemble the, the killer team that you need to be able to do [00:02:00] and build your own model and put it into production.

It's not a research project. It's not like it works once and it's okay or it works in a test environment, like it, has to work. And so yeah, that was, that was really a journey that started 18 months ago that led to the announcement we made yesterday.

Haroon: Can you walk me through the thought process of what eventually led you to deciding on going down this route? Because obviously there's other companies that are a pproaching it a bit differently, but I'm curious why you landed on the, building your own models and letting companies fine tune that route instead.

May: Yeah. When we were raising our A we told people we were going to be doing this. And I don't think most people believed us . I think the best investors like literally two sentences into the pitch we're like, oh, and then you're gonna build your own. We're like, yeah, duh. And it wasn't It wasn't even a thought process as much as like it was always going to happen.

And I think for us, like our early journey into Transformers just gave us a ton of confidence. We told people we were gonna be spending a few million bucks to [00:03:00] do this. This is use of proceeds. I don't think that most people believed us, honestly. Obviously the folks that invested did. And the thought process came down to us wanting to be the company that defines the way that professionals and AI work together.

And enterprise problems have always been what me and Wasim are most excited about. Literally what are the things we do at work and how can those things be made more interesting, more exciting, faster. I don't wanna spend 16 hours at my desk and most knowledge workers don't. And I think bringing AI into those workflows has always been our motivator. So it's just gonna be impossible for us to build our vision out without owning the entire stack ourselves.

Haroon: The approach that Writer's taking is more of a privacy first approach where companies will own the model and then train them on their own data rather than all of the training and all the personalization living in the cloud somewhere. [00:04:00] Why this approach? I'm curious like the quick pros, cons, breakdown of why this approach.

May: Yeah, I think in the enterprise accuracy matters because it is not about, write anything or decide anything, it is like super specific and a very high bar. Like we are talking four nines, right? In terms of what is expected and without getting a lot of proprietary information and decisions that you've made previously into the models, you're not gonna get four nine and you're not gonna get that data if everybody else is gonna get it as a result of giving it to an LLM.

And it's really the only logical conclusion to that kind of conundrum.

Haroon: I have a ML engineering background and then eventually I ended up on the growth side of things and I did a little bit of product marketing in between. And one of the challenges that we have in the growth side of things, on a product marketing side of things is just generating enough content and generating enough collateral to equip the sales team and things of that [00:05:00] sort.

And with tools like Writer.com and other content creation tools, that's become a lot easier. The next progression was always gonna be personalization, like personalizing content to the context of the specific company so that it's not just the generic stuff that everybody's gonna be able to get from a ChatGPT or any other generic tool. How does this next sort of phase of Writer.com with these fine-tuneable foundational models, how does that help companies with creating their own brand voice and then augmenting on top of that brand voice rather than the more generic models that most people are using?

May: Yeah, I would add to that even that the space is moving so fast, I don't wanna say too much. But soon we're not even gonna be needing to fine-tune anymore. And so a lot of these processes built on LLMs are way stations while the, like the underlying architectures [00:06:00] get innovated on.

We've got more to say on that soon. But, let's take the fine-tuning concept at face value. It is, it's remarkably different. It's content that feels oh wait, did I write this? You're confused. You're really not sure. If maybe you actually wrote it before, right?

It sounds so much like something you would've written. And we sell to professional writers. We don't sell to teams where everybody sucks at writing. Or it doesn't matter that much. It's not perfect, right? We sell to folks where. The bar is already quite high. And they wanna just do more of that high bar.

So the fine tuning is really important. The being able to stack fine-tuning models, right? That's really where we are right now. And you can't do that with the other commercial models. And they might be able to add that. Sure. But even that, only get so far. The title being the right length and the CTA starting with non-gerund verbs and there being the right number of lines for every [00:07:00] blurb on your landing page, like for it to be helpful beyond a like surface level. Oh, this broke my writer's block. To really be able to power processes and change how teams do things to get the like tens of millions, hundreds of million of savings a year requires just a deeper level of integration of AI into a business process. And so that's, we're in the business of that. We're in the business of uncovering that pain.

We're in the business of solving those solutions. And there will always be a massive market growing and massive for personal productivity augmentation, right? Like your ChatGPT Pro we're building the team based AI augmentation product.

Haroon: And in terms of stacking fine-tuning models, I'm curious to learn more about that. What are the use cases or what are the applications that you've seen, whether it's Writer.com customers or outside of that where companies have leveraged stacked fine-tune models [00:08:00] for augmenting their writing process.

May: Yeah, I can give you the example of UnitedHealthcare. They're one of our largest customers. And so you start with Palmayra. Then there is a smaller brain on top of that is healthcare. There is a smaller brain on top of that brain that is Optum. And then within that there wellness. There is communication, there is thought leadership, right? There is internal comms and internal operations use cases. You're pulling from a foundation model but you are really able to do quite a bit more of just specification as you go through those stats, and then there's post-processing as well in terms of just making sure things are coming out, you know exactly how a person would've written it or analyzed it.

Haroon: Interesting. So it's almost like profiles for the different types of writing that you do, because a blog post might be different than a newsletter that somebody at a given [00:09:00] company might write.

May: Yeah and we are, generation is almost a content generation is one aspect of what large language models to do can do. Sometimes customers are using Writer to generate an answer that gets fed into a process. And so what you're training on is previous examples of the analysis, if that makes sense. And we'll have a lot more to say about that too soon.

Haroon: Okay, so I want to take a step back and I want to get your thoughts on what content creation at companies is gonna look like in the future. A lot of folks nowadays are very unsure. They don't really have a POV on this, with everyone having access to powerful tools, what is the future gonna look like?

Are we just gonna see the internet being spammed by generic content? Are we just gonna see so much content that people are gonna start to ignore it? What do you think the future of content marketing looks like and how companies can stand out in this future that's moving at breakneck speeds.

May: Yeah [00:10:00] I certainly think that the way that we index information isn't going away. The way that users process that index is certainly changing. I think is funny because it is going to be content that is created by AI, yes, supervised by humans to be read by bots that then answer the question for humans. So I do think the consumer's search experience is like irrevocably.

Changed for the better. And SEO as a practice goes nowhere. It's gonna continue to be incredibly important because, we have to index information in databases like the algorithms for figuring out what is true, what is good. That is a service that Google will be providing till we all die.

That's my point of view. The way a consumer digests that is certainly changing. So from a company perspective, content marketing, top of funnel, I don't think [00:11:00] there's a big change. And the faster you can adopt to having coverage for everything that matters for your brand, the better it is and certainly doing it like the way that Google asks you to do it. Which, all of our customers get enormous SEO uplift using Writer including us. We are literally one person in content, two and a half million sessions a month. So that is a lot of AI. Visitors actually in all sessions. Sessions is higher. US content marketing in general I do think as a result of top of funnel being bots talking to bots, people talking to people will take on even more importance. And the more that a brand can infuse its voice and its unique points of view and like what it stands for as a brand in those human to human conversations, the more that it's going to stand out.

So I do think there is gonna be just this kind of lopsided barbell. Two the human interaction it's not going to mean that top funnel isn't as important as it is today. But I do think if you [00:12:00] lose the human piece, it won't matter that you won the bot.

Haroon: With Palmyra, how can companies get started with it how can companies leverage it today?

May: So using the two open source models is as simple as getting 'em on Hugging Face and you can actually try both models on Hugging Face before figuring out how to implement 'em into your applications. And we already have customers who have just one off use cases as well as kind of test use cases already using them for app production use.

The 20 billion parameter model, Palmyra Large, isn't open sourced. But the API is coming to the self-service plan which we are excited about. So that wasn't part of the announcement. But that is coming in a couple of weeks. and the the same access that our enterprise customers have to both their custom apps as well as the out-of-the-box apps and the underlying model our self-service customers will have [00:13:00] as well. So all of the same security and privacy. The self-service plan doesn't have the custom, fine-tuned experience that enterprise customers have but it is a really good model.

And for our B2B use cases, beats Da Vinci for all of the benchmarking tests, et cetera.

Haroon: What are you most excited about in this space and Writer's role in this space moving forward?

May: Oh, I could not be more excited. It is so fun to have a space we've been working on for a decade be what everyone is talking about and is excited about and we have so many champions and power users that have been in content their whole lives too. And they're the center of the innovation conversation.

It's not social and mobile and Web3 and things that like, didn't really impact our lives. So I am really excited to bring everything that content folks have gotten smart on over the past year plus to like mostly every other function [00:14:00] too. So deepening what we do for content people and then really bringing the power of this generation of AI tools across the enterprise, across all the teams on the enterprise. Yeah, it is, it's gonna be a really fun next few years. We're hiring everywhere and these are real problems, real customers, lots of user facing roles at Writer from NLP and ML, all the way to user education and success.

So it's a really fun time to build something huge in this space.

And that does it for today's interview. Huge thanks to May for joining us and sharing her wisdom.

For more Q&As from leaders in AI, like Mark Cuban, CEO of Runway ML, and co-founder of Hugging Face (coming soon) plus daily AI news, be sure to subscribe to Not A Bot, the world's most subscribed daily AI newsletter. 👇

As always, thanks for reading, and see you next time. ✌️

- Haroon - (definitely) Not A Robot and @haroonchoudery on Twitter