It has been an incredible year in text-based AI. There have been a lot of advances and you only realize how much the industry has changed when you take a step back and look at how things were a year ago and how they are now.
2021 was the year of the writing assistant tools. Lots and lots of these tools popped up as more and more people got access to OpenAI and GPT-3. The fact that OpenAI had this in private beta and had a long waitlist ensured that getting access was to be a part of an exclusive club and by leveraging this technology when nobody else could access it, you could make money.
If we circle back to March / April time 2021, you could incorporate OpenAI into your product but there wasn't really much else for you to work with. This meant that you were somewhat shackled and at the mercy of the rules and regulations that OpenAI put on startups to use their technology. Rules which weren't always fair.
Building with OpenAI at this time in the first half of 2021 meant a pretty in-depth approval process for your app where you had to make videos for each and every prompt that you wanted to include in your application. You needed to show how these worked, what the inputs were, and what they looked like in your application. You had to show you had thought of how to prevent misuse and you needed to divulge your monetization and pricing strategies.
You were unable to launch without the green light from OpenAI and could get your API key revoked if you did so. There were certain annoyances I found particularly frustrating back then in terms of content length, creating long-form content, and giving users the ability to run the approved prompts in their own systems.
One of the more frustrating elements of OpenAI's terms was that they were not fair for all. They had approved certain startups to be able to produce long-form content whereas they weren't approving it for anyone else. These companies had an unmatched advantage in terms of being able to produce this content and monopolized it. If you were running a startup at the time, it was mightily frustrating to hear from users with feedback saying; 'why can't you do it like X'. It was a fair question and one you could not really answer.
With this selective policy, certain startups could get ahead and others could not compete. Fortunes were made off the back of this policy choice. I remember asking OpenAI at the time, if we bought a company with legacy approval, do we get that legacy approval and they said yes. What happened less than a week later? A flurry of acquisitions of these legacy approved companies ensured the playing field remained unlevel, a further monopolization of the legacy approved features and fortunes for the founders who were lucky to get access to OpenAI prior to everyone else.
Being in a closed ecosystem where you don't have the freedom to experiment and build completely openly was something I was not a fan of and I was desperate to find alternatives to OpenAI at this point. The technology is great, but the restrictions were infuriating when you are trying to move fast in a startup environment. I looked at a lot of the other technology available which was available via open-source, mainly GPT-2 and Neo 2.7B (which was released in March 2021).
Both of these were prohibitive to what I wanted to use in terms of quality or ease of hosting on my own system but they gave a glimpse of what was to come. A gold rush in AI was on the horizon and being restricted to just one provider was going to be a thing of the past. I ended up stumbling across Inferkit at this time. A hosted version of Megatron 13B which you could access via an API with no restrictions.
The last part is super important to emphasize. No restrictions here. This was incredibly liberating! You could apply the technology from Inferkit as you liked. Sure, the larger models from OpenAI were of better quality but for the sake of moving fast, Inferkit was something I played with more and more and ended up building the first version of our long-form writing assistant at Content Villain with it.
The next major development was the introduction of GPT-J. On the 4th of June 2021, it was announced in a blog post. This was a gamechanger in terms of an open-source model performing so well on benchmark tests and being able to produce quality outputs. The only issue was, how to host it? What resources would you need and how much would this cost? Fairly prohibitively expensive for a startup to do without large funding or scale.
It was great to see a model perform on par with something from OpenAI without restrictions. The future was certainly looking a lot brighter with the evolution of AI models and it seemed that the space was heating up with more interest and more development.
We didn't have long to wait for the next big development in the space. I got a tip-off via LinkedIn from a friend on 10th August 2021 that I might be interested in a link and boy was I! This was AI21 Labs. The company behind Wordtune and an Israeli startup with an impressive founding team and models of 7.5B and 178B in size. True competition in terms of sizing to anything that OpenAI had created. I signed up instantly and loved to compare what my prompts looked like in AI21 compared to existing technologies and was pleasantly surprised.
I looked through the terms and didn't see anything as restrictive as OpenAI's terms and thought I need to put this into my product as soon as possible! I took a few calls with their team where I was excited and wanted to see where their vision was and if they planned any restrictions down the road and they seemed quite open to users applying the technology as they see fit.
It wasn't long before I ended up replacing Inferkit in my own app with AI21 as it provided a higher quality output and the cost made sense to do that.
A few days after my initial experimentation with AI21, I received an email from Cohere. I had signed up on their waiting list a few months previously and knew they were cooking up some large language models but had no idea when they were going to make these available for use. This email was an invitation to an onboarding call to get access to their beta. I was very excited and after so many months of being restricted to one company, I now had options and that felt great.
The onboarding call with Cohere was awesome, very exciting to talk with anyone in this space as everybody is so energized to drive it forward and make great things and I really bought into the approach that Cohere was taking and the way they wanted to make finetuning as simple as possible for everyone.
Being able to experiment with multiple large language models and having the freedom to build a new prompt or use case and instantly apply it to an app or business was a liberating experience. No lengthy approval processes were required for most of the technologies now available and the old startup point of view of moving fast and breaking things was possible again. I was a happy bunny.
It was around this time also that a new crop of businesses started to enter the ecosystem. You saw the arrival of Language Models-as-a-service where businesses would host some of the more resource-intensive open-sourced models and give you the ability to pay per API request or tokens used like AI21, OpenAI, or Cohere worked. This was a pricing strategy that we could get on board with and having this access made it easier to experiment across the whole spectrum of AI models available at the time.
Notable companies who were pushing the boundaries on this were Banana, Forefront, NLPCloud, and a newcomer to the space in Goose. They offer a valuable service in making the open-source models available to non-technical folk and have plans scaling to fit the needs you have.
For the rest of 2021, AI21 and Cohere both went through phases of closed beta taking suggestions from users and then releasing them publicly for anyone to signup. Conveniently, OpenAI also announced they would be doing away with their waitlist and that anyone could sign up for GPT-3 access without going through the long wait that was common before. Text-based artificial intelligence was truly starting to be something that anyone with a spare 5 minutes could access and play with. That was great to see!
At the start of February 2022, EleutherAI announced NeoX 20B, a massive model that was open source and performed well on a lot of benchmarks. It is an incredible achievement to train such a large model without the commercial backing that some of the other companies in the space have access to. The model was immediately available for people to play with via Goose.
Fast forward two more months to April 2022 and it was the turn of commercial European models to throw their hats into the arena with the launch of Muse API from Lighton and the Luminous models from Aleph Alpha bringing native models trained in Spanish, French, German, Italian and English. Now we're just spoiled with choice!
When you look back on the year like this, you realize how far things have come. From really only being able to play with the 4 models from OpenAI to now being able to play with over 30 high-quality models is awesome. Competition breeds innovation and choice means that the consumer has the power to vote with their feet in terms of what technology they want to support.
I understand how having such a large amount of models can also be daunting to someone new to the space. Where do you start with testing, experimenting, and building? Who should you sign up with? Which is the best? It is difficult to know!
The advancement has definitely led to fragmentation when building out prompts and ideas. No easy way to store your prompts, no easy way to see which model works best for your needs. No easy way to compare and contrast. This is partly why Riku.AI was born. We become the engine room of your AI creations and a vault to store and save with others.
We are very confident that there will be further developments this year and will keep our fingers on the pulse in providing you with the latest news and advancements in this incredible field.