AI Tsunami— Making Some Sense of it All
"Though this be madness, yet there is method in't." - Hamlet, Act II, Scene 2.
If you’re reading this post then you probably know the feeling of going to Twitter and getting your daily overdose of AI-related news. Whether it makes you feel anxious or ecstatic, it’s really hard to make some sense of everything that’s coming at us.
I’m writing this content piece to distill the constant barrage into a somewhat more clear “prediction”. Note that I’m not inventing anything new here because everything I’m going to talk about has already been sighted in the wild, but writing it down helps me better understand the macro of the AI revolution.
P.s. I’m not affiliated in any way to any of the people/projects that I mention here, besides my own (promptchainer.io). If you don’t know us, you should definitely check us out.
AI Hardware
AI-oriented hardware is set to become more crucial and popular as the demand for processing power continues to rise:
Graphics cards are already a mainstay in the AI department and I don’t think this will change. Keep an eye on established players like Nvidia and AMD.
Dedicated AI chips produced by both big companies and small startups are going to be big. Graphics card dominance in AI may not last when faced with purpose-built AI chips.
P.S. AI models that design new hardware? Coming soon?
Models
Multi-Modal: In the realm of AI models, the future seems to be steering towards an increasingly multi-modal direction. This means that models will be developed to integrate different types of data inputs, such as language (LLMs), graphics, and sound, into a unified model. The result is an AI that can understand and generate a more comprehensive range of data types. There are plenty of multi-modal models already out there (GPT4 included), but none are “commercially” available or easily obtainable (correct me if I’m wrong!)
Efficiency: Another significant trend is the development of smaller, more efficient models. These AI models will be lightweight enough to run on a variety of devices, from high-end servers to everyday smartphones. Examples: Fine-tune a model on a commercial GPU. GPT4All open source LLM (sub-GPT 3.5 performance, BUT YOU CAN RUN IT LOCALLY!).
Continuous Learning: We are also likely to see models that are continuously learning and updating. Instead of being periodically trained on new data, these models will incorporate new information in real-time, making them more adaptable and accurate.
Context is Key: AI’s capability to comprehend context will also expand dramatically. Experimental models are already capable of understanding up to 100,000 tokens at a time (check out the 100k Claude model on Poe), and this figure could increase to a million tokens in the near future. Moreover, the evolution of models like AutoGPT, which can generate their prompts, signifies a leap towards more autonomous and sophisticated AI systems.
Open Source: Open-source AI models will keep pushing the envelope in performance, often lagging just 3–6 months behind their commercial counterparts. This democratization of AI technology will lead to a more dynamic and competitive AI landscape. This is helped by the fact that simpler models can be trained on the outputs of a larger, more capable one.
Big Players: OpenAI. Google. Meta. Amazon. Anthropic. Elon Musk? These will keep dominating the game. Hmm, when will Apple integrate AI?
AI Infrastructure
Self-Hosting: As AI models become more sophisticated, the infrastructure supporting these models will also need to evolve. The ability to host and fine-tune a model on your data and deploy it with ease will be paramount (Like Izzy with Robo-boys). Model hosting is likely to emerge as a significant business, with companies providing comprehensive services to manage AI applications.
Monitoring: Tools that track how many calls are made, which prompts are used, and the costs associated with these operations. Advanced software could even offer recommendations on how to reduce these costs, making AI more accessible and cost-effective. There are already many such tools. Look at Aimstack, Usagepanda, Portkey.
Caching and Compression: Technologies like caching will be employed to reduce the number of calls to AI models, ensuring no double calls happen, while “prompt compression” techniques will optimize the delivery of prompts.
Hallucinations, Security, and Alignment: Ways to enhance fact-checking, prevent hallucinations, and improve alignment and security will also be explored. Nvidia Guardrails comes to mind.
AI Cloud: Big tech players like Azure, AWS, and Google Cloud will fully integrate AI into their services. AI models will be able to access all resources on your cloud with ease. AutoGPT-like cloud-based agents will serve as “always-on” helpers, connected to your apps, waiting for tasks, and executing them with speed and efficiency.
Full Auto: BabyAGI, AutoGPT. All of these were invented last month but are already ubiquitous and enjoy a vast following, and it’s only going to get better and more advanced with each passing day. With that said, with GPT4 level performance AutoGPT is (IMO) currently a novel curiosity. This will change once more powerful LLMs are available.
New “Languages”
As we become more adept at utilizing AI models, their usage will evolve into both an art and a science. Users will utilize a variety of techniques to harness the full potential of AI models. One such technique is the use of chain prompts, which are chains of smaller prompts interlinked with traditional code and logic. This approach allows developers to achieve complex goals by breaking them down into manageable tasks which help keep the result more predictable. Langchain is a big player in this area (though IMO they’re not so much about chains as about connecting prompt and context). That’s also what we do at promptchainer.io, check us out if you didn’t already!
Another intriguing concept is the use of “super prompts.” These are single-shot, long-form prompts designed to accomplish more complex tasks like discovering new scientific theories, writing books, developing applications, and more. The sophistication of these prompts could lead to AI models generating unprecedented outcomes. The big question here is “How do we come up with super prompts?” Hint: What if we use a more abstract, higher-level language and compile it into a super-prompt?
The development and use of these new techniques will necessitate the creation of innovative coding languages. Some of these languages will be text-based, similar to the programming languages we use today. However, as AI becomes more multi-modal and interactive, we might also see the development of visual “prompting” languages. These languages will allow prompt engineers to construct prompts in a more intuitive and visual manner, making AI more accessible to a wider audience.
Yes, prompt engineers are going to be a BIG thing.
Conclusion
Hang on to your seats.