With generative AI threatening businesses and side hustles, how you can find space? ChatGPT was the chatbot that has taken the world by storm. At first, it left the world in awe of its capabilities, this feeling of disbelief before a technical prodigy is followed by a wave of fear. In a sense, the hype about its theoretical capabilities has turned into the fear that ChatGPT will replace humans in their work.
We looked at the cloud and AI with the hope of opportunity, are they instead storm clouds? Will LLMs kill my business? On the one hand, Elon Musk and Sam Altman have stoked this fear:
“We’ve got to be cautious here,” Altman told ABC News on Thursday. “I think it doesn’t work to do all this in a lab. You have got to get these products out into the world and make contact with reality. Make the mistakes when stakes are low. But all of that said, I think people should be happy that we are a little bit scared of this.”
But is that really the case? Did we just create Multivac?
Meanwhile, Musk is calling for a moratorium because on the artificial intelligence side, he is behind and today he is forced to chase. While Altman has exploited hype and fear for marketing purposes. On the other hand, as LeCun stated these models are not capable of thinking, but nonetheless they can be dangerous.
Science fiction scenarios aside, we can say that ChatGPT has already started killing business.
For example, it would seem Chegg (a company that offers homework assistance) is the first victim of ChatGPT, with a loss of more than 40 % value of its shares.
“However, since March we have seen a significant spike in student interest in ChatGPT. We now believe it’s having an impact on our new customer growth rate.”
And it is clear that now every partner is thinking about how to integrate ChatGPT into their business. In fact, Expedia plans to do so, and IBM and Snapchat have done so. And we, Mercury has integrated to our SEO engine, Blog platform and customer services chatbot.
In any case, it appears that ChatGPT is integrated into a number of businesses and many more will make use of it. So the question arises does it make sense to make a generative AI start-up? or to create a start-up whose business may be destroyed by the next ChatGPT upgrade? Or even just a side hustle that costs the customer less to subscribe to ChatGPT Plus?
After all, a company or freelancer doing content writing, graphic design, web development, or course creation, why should a company hire or pay for the service?
We have No Moat and we probably will never have
A leaked Google document states, “We Have No Moat…. And neither does OpenAI.”
So far we have talked about ChatGPT leaving out Google. After all, Google has a great tradition of creating sophisticated LLMs. Why have they wasted time in countering ChatGPT? and what do they mean when they claim that neither company is winning the arms race?
Google noted how a product that did not meet the highest standards could damage its reputation (see what happened when Bard was launched). On the other hand, Google has noticed that the companies are having their stable diffusion moment.
Open-source seems to be winning the AI arms race.
Google has realized that the gap between Open-source and ChatGPT models is closing rapidly. In addition, open-source models are much more customizable, and private and are more capable in some aspects (a specialized model beats a generalist model on a certain task).
Vicuna cost only $100 dollars and can do the same things as GPT-4 at slightly lower quality. The training of GPT-4 must have cost probably millions. People will probably not pay an expensive subscription for a restricted model when there is a free alternative. Nor will companies.
The future of LLMs is kind of the future of DALL-E. Stable Diffusion is now the standard, a model on which product integration, user interfaces, marketplaces, and so many other models are now based. While OpenAI DALL-E on the other hand is not as widely used and is losing ground.
Can I build a generalist AI as a business?
Less than a year ago we had not heard of either MidJourney or Stability AI, and now almost every one of us uses one of their products. In a sense, StableDiffusion is a model that was built from DALL-E as a base. Why then not create its own new generalist model?
OpenAI has been working on building LLM for years since the first GPT has an experienced team on building the model, training it, and deploying it on the market. Also, the infrastructure is not cheap especially if you want fast service and a user-friendly interface.
While it is true that training a model in a self-supervised manner means using only a huge amount of text, quality matters a lot. As LLaMA and Google showed with PalM-2, the more curated the data, the better the model performs. OpenAI has a team dedicated to this.
Also, ChatGPT and many of OpenAI’s products can be used by just about any user, and creating a similar user experience is not at all easy. For many people that is what determines their choice of use, not performance.
Furthermore, OpenAI has the additional advantage of having created a strong and recognizable brand over the years. OpenAI is synonymous with AI; to compete, a technical report showing an alleged superiority of the model is not enough. Users will prefer ChatGPT and also companies will prefer to use a product that the average audience knows. Many companies and users choose what they now consider the default.
Ultimately, OpenAI has accumulated another advantage: 100 million users and most importantly their data. Data that are used to make the model better and better.
How can you create a start-up or side hustle in the age of LLMs
More quality than quantity
This applies to any business or idea. For example, content writing can no longer be based on the number of articles or words. ChatGPT even in its free version writes much faster and with more than appreciable quality.
But LLMs are not capable of critical thinking (techniques like the chain of thoughts can only mitigate this problem) and often hallucinate. So you can only compete with reasoned articles with a defined opinion. Anyone who has used ChatGPT knows that the model does not take a position, and an article with a bold opinion supported by clear data and sources cannot be produced by ChatGPT.
Similarly, ChatGPT does not understand customer needs. On the other hand, one study showed how patients perceive ChatGPT’s responses as more empathetic and quality. Therefore, it will again need to focus on the client’s needs.
Domain knowledge
A generalist model is good on everything, but the best on nothing. Knowledge of a specific domain is what allows you to be able to win.
The very alignment of a chatbot leads to performance degradation in some tasks. Also, as huggingGPT and similar models show an LLM is not capable of everything but needs to interface with “expert” models to accomplish some tasks.
These models can only be trained with data that are proprietary. In addition, they often require knowledge of the specific domain. So businesses that have access to knowledge and data that are not publicly available have an industry advantage.
Don’t fight artificial intelligence
AI will not replace developers or data scientists, but it will replace those who will not use an AI assistant.
The code produced by ChatGPT needs to be tested and needs knowledge on how to use and integrate it. In any case, many more developer and machine learning engineer positions will open up that are currently threatened.
Court cases and regulations will reshape the field
The EU and other institutions have been passively watching social networks, and at least the EU does not want to make the same mistake with AI.
“The European Parliament must enter the trilogue with the strongest possible position to protect the rights of all people inside and entering the EU,” said Caterina Rodelli, EU policy analyst. The new EU AI act may be much more stringent than expected. As seen with GDPR violating the rules leads to costly penalties. Also, larger companies are more exposed to reputational damage.
Recently ChatGPT was banned for privacy concerns in Italy (now reinstated). The Italian authority asked for information about OpenAI’s data collection, questions however initially answered. On the same idea, other regulatory agencies have also moved. OpenAI in any case from now the possibility to opt-out of data collection.
In addition, there are already lawsuits against GitHub Copilot and MidJourney and their products by programmers and artists.
How the regulations evolve and the outcome of the lawsuits will decide the scope of the big companies. There will be niches that larger companies will consider of little interest, risky to their business, or simply a generalist product that cannot satisfy.
Small companies and freelancers will be able to cover these spaces more easily, with less risk both legal and of being put out of business by the big companies.