Where is AI Actually At?

If we were to listen to two of the most influential figures in AI today, it seems that reaching a perfect, all-powerful, job-replacing algorithm is very near. Sam Altman (Altman 2024) from OpenAI and Dario Amodei (Amodei 2024) at Anthropic both agree: the real paradigm shift has been reached with the widespread adoption of some forms of AI models (think generative AI, trading algorithms). Therefore we will now see broader, higher-quality models that cover more tasks. The two seem to suggest it would suffice to follow the exponential trend to reach full automation and independent learning. However, what they don’t disclose is that many of the models that we design and engage with are limited by the fundamental structure of AI.

Current AI models are actually quite restrictive. The model remains relatively static once it has learned from a given data set, and the finished version would have to be fully retrained if there were errors in the data or rules that it runs off. An important feature of current neural networks is that most large models are ‘black-box’: engineers cannot directly map how an output has been obtained given an input. Also, models such as those by OpenAI and Anthropic are very large, which eventually has the effect of slowing computation; it is not yet known whether the rate of chip development will cancel out this issue (IBM 2023).

This type of AI is called Artificial Narrow Intelligence (ANI). It is limited to the quality of the data set, to some abstract understanding of the probability of one output compared to another, and thus to a limited and narrowly defined task (despite what Altman and Amodei argue). There are further ethical considerations to these models too. Since Large Language Models (LLMs) are so inscrutable once constructed, it can be very difficult to deconstruct the internal biases. Consider for instance ChatGPT as a form of generative AI, which often returns misinformation in the cases where the underlying data set is not complete enough. Several examples have demonstrated that these shortcomings mean the technology could not be applied to medical analysis, for example. 

Theoretically, many people believe in Altman and Amodei’s Artificial General Intelligence (AGI) – forms of intelligence that would be able to use previous learning on new tasks in completely different contexts, without the need for supervised learning (where a human provides input-output pairs for the model to study). Such models would be able to rival the flexibility of human intellectual capacities. But the scale at which AGI would be deployed in their ideal worlds is currently mere fantasy: they propose an AI capable of every task (and capable of replacing nearly every human job given the right hardware), yet we have far-from-perfected ANI. That said, OpenAI is regularly improving its models, with yet another ChatGPT update released in September using “chain-of-reasoning” to reach impressive levels of accuracy but without a specific application.

However, recent research seems to suggest that the next paradigm shift in AI will not stem from the progress from ANI to AGI, but rather involve a complete restructuring of the foundations of neural networks. This could involve training models to perform exceptionally well on specific data sets with specific aims, returning to the origins of Machine Learning. One such example was recently released by LiquidAI, an MIT-based start-up, that creates Liquid Foundation Models (LFMs). These assign equations rather than static values to define the behaviour of each neuron over time (Knight 2024). Then, when given an input, this results in an observable equation cascade; the output can be retraced on inspection, and this enhanced explainability means that the quality of predictions can be continuously improved. These models are far more situation-specific and local and aim to remove as many cloud calls as possible. Contrast this with current AI technologies, such as OpenAI, which have huge cloud integration; while allowing them to have enormous databases, it also restricts the applications of tools to web-connected instances. These LFMs could create opportunities for autonomous peer-to-peer AI systems, in remote drone networks that could identify early forest fires for instance (LiquidAI, n.d.). These systems are also far more efficient, requiring fewer parameters and less memory than common AI alternatives, ultimately meaning that the energy cost of running these LLMs would be lower – this is an important feature, considering the current environmental cost of running such algorithms.

The time of Big AI may soon be reaching its climax, with custom-made small-scale target models taking over in accuracy and efficiency. Some industries, such as financial institutions, have already been using Machine Learning tools in such restricted and specific scenarios for decades, constantly improving models for speed and quality of deep learning as research progresses. Encouraging other industries to adopt custom models at necessarily high design costs conflicts with the current global attention on large-scale open-source AI, but it may be the best prospect to truly harness the power of AI and reach some semblance of the idyllic world Altman and Co. promise.

Sources

Altman, Sam. 2024. “The Intelligence Age.” https://ia.samaltman.com/.

Amodei, Dario. 2024. “Machines of Loving Grace.” https://darioamodei.com/machines-of-loving-grace.

IBM. 2023. “Types of Artificial Intelligence.” IBM. https://www.ibm.com/think/topics/artificial-intelligence-types.

Knight, Will. 2024. “Liquid AI Is Redesigning the Neural Network.” WIRED. https://www.wired.com/story/liquid-ai-redesigning-neural-network/.

LiquidAI. n.d. “Liquid Engine.” Liquid AI. Accessed November 3, 2024. https://www.liquid.ai/liquid-engine.