Artificial General Intelligence: Inevitability or Industry Jargon?
Tom Everill | 6 May 2024
It is undeniable that Artificial Intelligence (AI), as an emerging general-purpose technology, will profoundly change our societies and our species. Over the past year, in response to this narrative, global equity markets have seen massive gains as corporations worldwide scramble to implement AI tools into their businesses before competitors, and investors place bets on future industry leaders. Given the rapid rate of progress we have observed between the public release of OpenAI’s GPT-3.5 model in November 2022 and today, it is easy to imagine that such progress will continue exponentially and without end, and perhaps it will, but, what if it does not?
Some industry CEOs, like Sam Altman, believe that the ultimate goal of their work is to develop Artificial General Intelligence (AGI), which his company OpenAI claims is a ‘benefit [to] all of humanity.’ Altman himself has described AGI as an AI system that is generally more intelligent than humans in all ways. He not only views the technology as an inevitability, but foresees its arrival in the near future. However, potential bottlenecks exist, especially regarding energy and physical computing infrastructure, and market hype cycles often fog reality (Dotcom Bubble and recent cryptocurrency ‘bull runs’). It is also important to recognise that continued exponential technological advancement, a framework popularised by Moore's Law, does not always align with reality due to external factors, including geopolitical insecurity, resource scarcity, financial crises and pandemics.
Development Prerequisites
Last month, McKinsey published a report entitled What is Artificial General Intelligence? which highlights eight capabilities of AI that need to be ‘master[ed]’ before AGI can be achieved:
Visual Perception
Audio Perception
Fine Motor Skills
Natural Language Processing
Problem Solving
Navigation
Creativity
Social and Emotional Engagement
To summarise, the report argues that the sensory perception capacity of AI remains far from the human level, limited by both software and hardware constraints, noting that many autonomous cars are still fooled by minor vandalism on road signs and that even best-in-class systems cannot interpret, particularly sound, at a human level. In terms of interacting with the physical world, it argues that AI-powered robots still lack human-level dexterity for precise tasks such as surgery, but that solid progress has been made, such as OpenAI’s single-arm robot, which solved a Rubik’s Cube in under 4 mins in 2019.
Another challenge identified is natural language processing, most notably the ability to understand the non-verbal and implicit aspects of human communication, something challenging to derive simply from statistical correlations in datasets. In terms of problem-solving, the report suggests that AGI systems should be able to learn from their environment, adjusting to new situations in the absence of human guidance. It also notes that whilst autonomous navigation capabilities have progressed, systems are far from being able to navigate independently without human priming. Furthermore, the author identifies a need for AI systems to be able to rewrite their own code to match or exceed human creativity levels, something which first requires an intuitive understanding of vast amounts of human-written code. Finally, it cites the need for high-level social and emotional engagement capabilities for humans to want to interact regularly with AGI. While some systems can already mimic human emotions to a limited extent, humans themselves often struggle to read emotions, and the current state-of-the-art remains far from achieving even parity with our abilities.
Hallucinations
One major issue users have found with generative AI chatbots like ChatGPT, Anthropic’s Claude and Google’s Gemini has been a tendency to ‘hallucinate.’ The term describes instances whereby generative AI models generate false information presented as fact. This can be due to erroneous connections made between concepts within the models’ training data or simply a lack of information on a specific topic.
During a Q&A session last month, Nvidia Corp. CEO Jensen Huang was asked how hallucinations should be managed. Huang responded that the issue was easily solvable using a retrieval-augmented generation (RAG) technique, where answers derived from training datasets are cross-referenced with relevant, up-to-date, real-world information from external sources to improve reliability. Despite Huang’s confidence, generative AI companies and research organisations are already employing RAG and similar techniques, yet these issues persist. Part of the issue is the complex nature of ‘truth’. There is inherent subjectivity to many human concepts and, as such, AI systems must be provided with frameworks for identifying and implementing nuance in their outputs.
Flawed Industry Trajectory?
Another potential issue is how much of the recent, rapid generative AI progress has been made. In 2017, a group of Google engineers released a paper, ‘Attention is All You Need', introducing the groundbreaking ‘transformer’ neural network architecture. As computing power was scaled, transformer models demonstrated the ability to recognise patterns in increasingly large datasets, resulting in broad performance improvements. When developers began to see the materialisation of unpredictable yet desirable emergent properties in generative AI models simply by scaling compute resources, the industry’s philosophy shifted accordingly. This new approach, based on the empirical observation that larger models tend to perform better (including on tasks for which they were not designed), has faced criticism from some in the field for a few reasons.
Firstly, as models grow, the associated costs and energy required to train them increase exponentially, raising concerns surrounding the sustainability of such an approach. This is especially relevant within our contemporary political context, where environmentalism is highly prioritised in most developed countries, characterised by a shift away from efficient but polluting energy sources. AI data centres are notoriously energy intensive, requiring more than double the electricity input of regular data centres. Some, like Oregon-based nuclear reactor firm NuScale, believe the future of AI data centres lies in energy input stemming from in-house Small Modular Reactors (SMRs). This builds on a blueprint developed by Imperial College London which operated its own reactor for teaching and training purposes between 1965 and 2010. There are no SMRs in commercial operation today, however, China is in the process of building the world’s first.
Another flaw of this approach is that, as models grow more powerful and their outputs more sophisticated, they become prone to more complex errors such as more subtle and hard-to-detect forms of hallucination and unintended bias. Resolving these complex failures requires more resources, adding cost and exacerbating the aforementioned sustainability issues. The process of rectifying such issues is made more challenging by an incomplete understanding of why the models act as they do.
Moreover, the expansion of generative AI capabilities in this framework relies on the ability to increase the size of training datasets. There is a possibility that the additional data required does not exist or does not exist in a form usable in generative AI training. It is also possible that new regulations regarding how data can be used in the training process will be introduced, handicapping the ability of developers to access vital data or to do so profitably.
For example, OpenAI is currently embroiled in a legal battle with the New York Times, which alleges that the organisation violated copyright laws by using millions of its articles to train its GPT models without permission. The NYT argues that its data was used to build products that could substitute its content, potentially negatively impacting its revenue. If the lawsuit is successful, Microsoft and OpenAI could face severe consequences to progress made on their models and underlying business strategies if forced to shut down or significantly alter models trained on NYT data.
Additionally, a verdict favouring the NYT could set an industry precedent, potentially resulting in slowed development progress. A future framework regarding fair use of content in training generative models could mirror a recent USD 60 million deal between Google and social media platform Reddit for the right to use its user-generated content. The deal provides Google access to Reddit’s Data Application Programming Interface (API) allowing the tech giant to better understand and utilise Reddit’s data. The deal is not exclusive, meaning that Reddit is free to make other similar deals. This would increase costs associated with developing generative models while creating significant value for companies possessing vast quantities of usable data.
Finally, an approach centred around constant investment in and development of expensive, manpower and resource-intensive infrastructure naturally raises barriers to entry, likely insurmountable for smaller firms. While this may not directly inhibit AGI development, the reduction of competitiveness in the industry could stunt innovation. Furthermore, the centralisation of power in AI development, as with all monopolies, opens the door to potential abuse, particularly concerning in this case, given the immense power of the technology in question.
How will we know?
A less technical challenge to the development of AGI concerns definitions. The lack of a universally accepted definition makes predicting if and when the AGI will arrive more difficult. Given the extent to which AGI has already entered the public consciousness and become industry jargon, it’s puzzling (although understandable given the currently theoretical nature of the issue) that a consensus definition has not emerged.
This is particularly abstruse in the case of Microsoft and OpenAI’s USD 10 billion deal which contains a clause stating that Microsoft loses the right to the developer’s IP after AGI (as defined by OpenAI’s board) has been achieved. It is surprising that the world’s most valuable company would enter into such a high-stakes agreement with largely undefined parameters, especially while deferring definitional power to its smaller partner. This illustrates the immense value it sees in OpenAI’s work.
Likewise, OpenAI’s charter contains a clause stating that if another entity reaches ‘late-stage AGI development’ before it, instead of competing with this entity, it will seek to assist it. The intention is to prevent a development race in which quick deployment is prioritised at the expense of safety. The lack of a clear definition for ‘late-stage AGI development’ or AGI generally diminishes the reassurance OpenAI seeks to provide in its charter. One can imagine a scenario where it refuses to acknowledge that a competitor’s system represents AGI to avoid honouring this clause before continuing to prioritise profit and market share growth. Depending on the extent of Microsoft’s board’s influence on OpenAI’s decision-making process, this may not be a far-fetched scenario.
The motivations of OpenAI, which claims to be committed to the principles of broadly distributed benefits, long-term safety, technical leadership, and cooperative orientation, have been questioned as of late. This is partly a result of its deal with Microsoft, which seemingly contradicts its mission statement of operating as a non-profit seeking to develop AGI ‘to benefit humanity’ given that its technology now provides financial gain to Microsoft, a for-profit company. This was the fundamental motivation for a lawsuit launched against the organisation earlier this year by billionaire Elon Musk who claims it has become a ‘closed-source de facto subsidiary’ of Microsoft.
Forecast
Short-term
Market hype will continue as new AI breakthroughs are made, especially if US and global interest rates are lowered. These breakthroughs will likely continue, at least in the short term, but perhaps also on a longer time horizon.
Medium-term
As legacy media struggles to remain profitable and related jobs are threatened by generative AI, more legal actions against developers are expected. Entities with vast collections of data will seek to monetise it by selling it to developers.
Long-term
Data’s value as a commodity is expected to increase further. Those who discover novel ways of collecting and synthesising data about the world to be used in training AI models will profit immensely.