ChatGPT is not alone... AI socio-economics impacts in the next 2 years
As an AI practitioner, in recent weeks I have been asked a lot about ChatGPT and its limitations , how to apply similar Large Language Models in real applications and in general I can feel now there is a sense AI can deliver significant change in the next few years.?
Clearly ChatGPT was indeed one of the major breakthroughs of 2022… but it wasn’t the only one with near future cascade effects.?
For that reason I wrote an article combining the latest AI development with regulations and socio economic factors (catalysts) which leads to 5 predictions for the next 2 years. I hope it is helpful.
Also in case it is helpful AI Technologies released the ‘State of AI 2023’ in January and you can download it here.
--------------------------------------
2022 was a fundamental year for advancements in AI , as well as growing awareness in the general public. This article reviews the key events of the last year together with socio-economical 'catalysts'? to give a perspective on what we should expect in the AI landscape in the future. As we describe the 'state of AI' to date, we identify five predictions that we expect to happen in the next two years.
From a research point of view it was a fructuous year: we showed that multimodality (i.e. the capacity of the same trained model to compete in different tasks) is possible (chatGPT and GATO models for example) and also large models (more than 10B parameters) showed learning behaviours that are not yet understood and 'unpredictable' (called ‘emergence’). The latter poses a safety issue in making the model useful in real applications, but it can also be an opportunity to tackle problems currently beyond reach.?
It is worth mentioning that most of the large models are likely under-trained, as we do not have enough data This is one of the main reasons why it is expected that by 2025 synthetic data will be 60% of the whole data in use in live applications (real data is not sufficient). Since models are expected to overcome the 1T parameters (currently the largest has 540B PaLM), the generation of synthetic data is becoming increasingly? necessary.
The biggest media hype was around the Google model Lamda, credited to be 'sentient' by the same Google employee who was testing it. Understandably, this claim generated a lot of media attention but there is not, from an algorithmic point of view, any evidence of it. This is a very technical discussion, but I will attempt to simplify it: any 'being' is such because its ‘objective’ is to survive i.e. anything dangerous triggers 'fears' or sentiments to avoid the (perceived) dangerous situations. Now, although it is technically possible,in principle, to set the? ‘objective’ of an algorithm to ‘stay alive’ this is not the case of Lamda (or any algorithm for that matter). Looking at the Lamda paper (page 3 for the reader who understands maths) and observing the training loss (the 'surviving mechanism' of the algorithm),? it is notable that it has been trained to mimic human speech as best as possible. In other words, Lamda has been trained in a way that it is penalised if the responses are not mimicking humans well (large loss) and rewarded otherwise (0 or low loss). If the goal was to create something even remotely close to sentient, in the case of Lamda it would be to train an algorithm to 'last a dialogue' as long as possible so that the human would keep the chatbot 'alive'. However,? this was not the case.
Data volume and cloud adoption keep growing and both are projected to grow significantly in the coming decades. This fact, combined with the decade-long low fertility rate (especially in the Western world), will create a 'shortage' of labour and it will create the required scenario for automation solutions (including AI) in the near future.?
Prediction #1: By 2025 Machine learning operations (MLOps) will be a new media adopted 'buzz world' together with 'digital twins’. Given the rate of adoption of cloud infrastructure, the automation tools and digital twins technologies would likely be one of the most talked about in the media. Digital twins is the technology that allows the replication of physical infrastructure or process (including humans) for the purpose of testing, simulating scenarios and predictive maintenance.
领英推荐
The EU already announced that by 2024 new regulations will be in place to use models in commercial applications: Four risk categories are announced depending on the level of safety. We need to wait for the actual rules to come out for clarity, but it is expected that education and health applications of AI will be considered high risks while customer service chatbots will be low risk: a model that marks exams of a student is considered high risk as it can potentially impact the whole future career of the person, while discussing a replacement kettle for your kitchen has low impact on humans life. As we now have 'data' governance, the new regulation will impose 'model' governance to ensure safety of the AI model behaviour.
Prediction #2: By 2025, while research will keep pushing the boundaries, with even more sophisticated AI models, actual complexity of AI applications will be pushed 'down' as companies would try to avoid heavy costs in governance and potential fines for an eventual sophisticated model mis-behaviour. Only simple and understandable models will be deployed in most cases.
Last year also saw the ascent of text images models (DALL-E and alike) which allow users to generate images from text and it is natural to expect soon,this year, similar text2music models will appear for general use (there are already some attempts but not widely known). In a similar way, chatGPT quickly gained users for text responses on 'almost' everything, from writing code to epistemology, and it is currently expected to produce 1B revenue by 2024. All these tools are actually within a decades long trend of 'human machine hybridization': as the smartphone was arguably first personal hardware extension (2007 saw the first smartphone, with the launch of the iPhone) and so the smartwatch we will now have a sequence of AI software extensions (assistants) which bring human-machine together.?
Prediction #3: By early 2025 in the English speaking world at least 40% of the working population will have used at least once an AI assistant. 100% of the population will have heard of AI assistants through the media.
These tools also would pose a serious threat on personal intellectual property: imagine an artist who invented his/her music, or painting, with a certain 'style'. Now the specific songs are indeed covered by copyright but the style is not: such AI could simply learn the ‘style’ and generate new songs with the same style and the artist is no longer required.?
Not to be an ego-maniac, but I already predicted that happening in? 2018 (https://www.dhirubhai.net/pulse/ai-near-future-5-things-expect-andrea-isoni/) and talked at length at Social Media Week New York 2018. I believe now we are on the verge of a public concern about it.?
Prediction #4: By 2025 specific cases of personal IP will be covered by the media, possibly involving court decisions and attempts to legislate the matter.
Finally, the research will keep advancing with new models but it is expected by some experts that the current deep learning architecture believes that it cannot bring to 'general' intelligence, hence we may be close to its actual limitations (Yann LeCun for example). There is certainly an ever increasing number of publications in AI but it is argued (https://doi.org/10.1038/s41586-022-05543-x) the actual 'disruption' of the discoveries is decreasing: an indicator of disruption is the number of ‘new’ words a language adopts over time to describe the discoveries). Also the larger the model (we expect to reach 1T parameters) the more unexpected behaviours are likely ('emergence') so less understanding of the model itself.
While certainly language historically is reflecting technical disruption, it is also true that AI models are now able to make discoveries ‘without’ deep understanding (for example, AlphaFold discovers new proteins, but the mechanism underlying it is not understood before the discovery itself). There is however a limit of what you can discover without understanding it.
Prediction #5: by 2025 R&D in AI models will slow down the race on increasing the number of parameters and focus on new architectures likely less reliant on very large numbers of parameters, and 'understanding'.