It has been 18 months since OpenAI gave the starting gun to artificial intelligence. Yes, AI has been in the industry for many years, but when ChatGPT went public, the world seemed to change forever. It was the democratization of AI.
Almost all major companies in the industry have pivoted towards some kind of AI integration, and they have gained and lost millions of dollars based on the AI prospects that these companies seem to have.
But AI has not yet changed our lives. The wild revolutions proposed in everything from data processing to search have largely driven, at least in the public sphere, only marginal improvements in functionality.
We have seen some smooth improvements in tools that we were already using, with at least some successes. So, why is AI taking so long to revolutionize the tech space and will drastic changes materialize someday?
LLM hallucinations continue to be a problem
The problem is that models can easily hallucinate, a new word that we all know since the launch of ChatGPT. Hallucinations are extrapolations. An LLM receives a question and, unlike a search engine that can apply a predefined algorithm to classify and filter relevant data, it extrapolates the question to an answer.
An hallucination occurs when a model generates nonsensical text or extrapolates data that is not included in its original training model. This extrapolation is the reason why an hallucination can be dangerous, because the incorrect response that a model gives can often be the one you expect and therefore not seem as ridiculous as it actually is.
Access to training data is becoming more expensive
Another element that has slowed down the progress of LLM in the last year is the lack of training data. One of the major impacts of AI that we are already noticing, as we mentioned before, is the abandonment of free and publicly accessible APIs.
Before ChatGPT, it was relatively easy to obtain data from user-generated content websites, such as Twitter or Reddit, to train your model. However, these two sites have blocked any API access behind costly paywalls or have directly sold training datasets, leading to mixed responses from their users. Reddit has also recently faced backlash for selling its training dataset.
This is another expense added to the already costly process of model training.
The introduction of AI products in the market has been difficult
Despite all the flashy demonstrations, companies are still struggling to bring products to market.
One of the elements of the AI wave that was recognized from the beginning, but remains true, is that the commercialization of AI products has been difficult. To date, most of the leading AI models have been limited to major technology companies, and products that rely on internet connections and online models have struggled to gain traction.
We have witnessed a wave of integrations with fine-tuned models, supported by GPT3.5/4, as well as by OpenAI’s support to first parties to fine-tune these models. Although some of the results from these narrow-focused tools have been impressive, they have often been too rough to reliably fill a gap.
Even companies as big as Google and Microsoft have had serious setbacks when it comes to introducing AI into the market. Google has introduced its AI assistant in the default search, which has already caused all kinds of problems with incorrect, silly, or directly dangerous results.
Bing Search was similarly pushed and then reversed. AI image generators have proven to be a significant problem, potentially because their biases are easily visible to the average user.
AI is changing the world
The reality of AI is that it is already changing the world. Outside the hype cycle of GPT-3, generative AI, and AGI, companies, researchers, and governments use AI daily to help them process data, find trends, make better decisions, and understand patterns.
It is used by developers to improve productivity, analysts to improve supply chain efficiency, and for better or worse, insurers to make decisions on all kinds of issues, from car insurance to healthcare.
Used to optimize drug discovery programs, which traditionally require huge supercomputers and expensive computing capabilities, and in the financial sector to model risks, detect financial crimes, and process data.
A major AI breakthrough on the horizon is weather prediction. This is an enormously expensive industry that is being rapidly revolutionized by AI. They can now take years of historical weather data and make accurate predictions about the weather, without the need for expensive supercomputer models.
Although AI has been exciting in the last 18 months, and will continue to be, it remains to be seen how it will truly change our lives.
