AI-driven businesses will come
we haven't talked for some time. Let's catch up!
This is Rinat Abdullin, writing to you about building ML-driven products, by the way :)
There was a lot of news in the mass media about ML in the last weeks. Most of them don't have practical implications right now, so we can just skim through:
Elon Musk signed a letter asking to suspend GPT-4 training. Just a PR stunt. Nobody will do that.
Somebody suggested to airstrike data-centres that train illegal models. PR, irrelevant, and taken out of context.
Google Bard might be using data that was produced by ChatGPT, which should violate terms. Pretty irrelevant, there will probably be no outcome. At the moment, Bard is a lot worse that ChatGPT-4
Italy has banned ChatGPT-4, Germany is also considering that. Also irrelevant, since the same service is already provided by Azure OpenAI (and it is quite good for Microsoft).
Now, there are a few things that I consider to be more relevant and important.
LangChain and LlamaIndex are awesome
@Mayo wrote on twitter:
I built a GPT-4 'Warren Buffett' financial analyst to 'chat' with and analyze multiple PDF files (~1000 pages) across @elonmusk's Tesla 10-k annual reports (2020-2022)
Check out this YouTube video for more details: https://www.youtube.com/watch?v=Ix9WIZpArm0
This is important because this impressive product doesn't require much effort to implement. It stands on the shoulders of a few existing building blocks:
OpenAI APIs - used for the ChatGPT prompts and embeddings
LangChain - framework for building apps powered by language models
Vector database - essentially a knowledge DB, Pinecone in this case.
I would also add into this list LlamaIndex project (formerly GPT-index), because it can improve knowledge retrieval from documents beyond the capabilities of a vector DB with embeddings.
You can quickly build all sorts of “products” like that: code converter between the languages, landing page generators, text summarization engines, and smart bots that can tell you everything about a given company.
While exciting, there are even more interesting areas for generating value.
AI-driven businesses will come
Products above are driven by humans. Somebody gives a task, and GPT-driven code executes it. What if we reversed the situation and put AI in control?
You have probably seen messages like that: “Hey, I give 100 dollars to the ChatGPT and will do whatever it tells me to do with that in order to make money. Subscribe to follow my journey”.
The journey then follows - the model helps to identify the niche, prioritise next steps and execute them. While models (before GPT-4 with plugins) aren’t particularly good with the bookkeeping, overall they do a decent job. People like that. Journey attracts millions of views, thousands of likes and even willingness to invest in such an enterprise. If somebody hasn’t started a kickstarter campaign to have a proper AI-governed business already, it will come soon.
I’ve seen such scenario unfold in multiple media platforms and business niches already. More are likely to follow.
This is a brilliant PR stunt that can also actually work out financially in the long-term. Why? The models have inhuman focus and patience, which is important in pushing forward some businesses.
This “AI-driven business” trend is important for two reasons.
First, it shows that the current AI boom has a potential to affect a lot of people in concrete niches, right away. Cryptocurrencies, Cloud or dotcom didn’t have such an immediate impact. While it is yet to be seen if there will be long-term value generated, people are already drawn to it.
Second, note the reversal of the roles. Humans no longer drive the process: language models are making the decisions, somebody else executes them. People tend to call this form of work: “smart agents”. You can deploy thousands of smart agents in parallel, scaling up and down as needed.
Just imagine a new department in some existing business that contains a team of different agents that continuously scan for business opportunities, prioritise them, break down into projects and drive the execution using existing company assets.
While this isn’t applicable to all businesses, I have already seen calculations that compare the cost of running a GPT-4 agent to the salary of a real human. 2-10EUR per hour is not that expensive.
After finishing this newsletter I’m going to open up a code editor and try creating a small team of agents for me. I’ll give them the strategy, access to a few news outlets, my journal and latest emails, provide with memory and will ask to continuously look out for the opportunities that I have missed. It will be fun to see what happens out of it.
What are your plans for this week?
With best regards,