Here to stay or fade away? Generative AI in the year ahead

There are some saying we are heading towards a new (Gen)AI Winter in 2024, following the age old pattern of the hype cycle. This, however, takes a too simplistic approach to technology cycles and does not sufficiently take into account the momentum of applied innovation and the new level of innovation velocity we see in (Gen)AI.

Looking at ongoing trends, adoption patters and AI R&D there are five interlocking themes that will shape GenAI (and AI in general) in 2024 and these are setting the year up for continued progress as opposed to stagnation, let alone regression.

Theme 1: Enhancement & Adoption. In 2024 we will see a continuous increase in the capability of GenAI and its adoption beyond hype and low-hanging fruits.

Foundation Models and their subset of Large Language Models (LLMs) will continue to advance. The launch of GTP 5 by OpenAI is already hotly anticipated, stoked further by Sam Altman’s teasing at Davos, as is a new version of Anthropic's Claude. Moreover, we will continue to see more and more Open Source models entering the scene as well as big players caught out by OpenAI & Microsoft working to catch up. Google’s Gemini, that is now powering the pro version of its chat bot Bard (Bard Advanced) is one example, Amazon is also said to be working on 2 trillion parameter LLM codenamed “Olympus” and both firms have struck partnerships with GenAI Start-Ups Cohere and Anthropic in a bid to accelerate their catch-up. Furthermore, Meta just announced the start of its work on an even more powerful Llama 3 model, as part of a large investment and re-organization drive to re-orient the company around AI.

It is worth noting, however, that it is not only a race to create bigger and bigger models in 2024. There is a new bread of "Small" Language Models, such as the just released LM 2 1.6B from Stability AI, whose compactness aims to lower hardware barriers to enable more developers to participate in the Generative AI ecosystem. Also, new ways of “chinchilla” optimal scaling already give rise to a new type of models such as Bloomberg's BloombergGPT for Financial Services. Moreover, Multimodal LLMs continue to evolve as do LLM training and deployment architectures in general. This also importantly includes work on making training and deployment of Foundation Models less energy intensive and thus more sustainable.

Regarding the ongoing evolution of GenAI it is also key to consider the data underpinning Foundation Models to date. The quality and capability of current generation (Gen)AIs is impressive but limited by the data they are trained on viz. publicly available data. That means private data, which accounts for the vast majority of data in the world and in many cases is also more high quality and domain and/or task specific, has not yet been tapped into. This is the raison d'être for firms such as Scale AI and the drive behind OpenAI’s and others’ work to strike partnerships with companies to access their internal data pools. As models begin to learn from private data, this will further unlock their enhancement and evolution. Also, it will be a further accelerate the drive towards more verticalized GenAI (see Theme 2 below).

Hand in hand with more capable GenAI models we will see continuously increasing adoption. Personal adoption of Generative AI has increased massively over the last year and notwithstanding fluctuations is set to continue to do so in 2024. The usage by individuals of GenAI can be measured quite straight forward using OpenAI’s ChatGPT usage as a proxy, which reached 180 million monthly active users and 100 million weekly active ones in 2023. The new generation of AI chips in consumer PCs just showed off at CES as well as the slew of new GenAI features coming to mobile phones (see Theme 3 below) will further drive GenAI adoption and its seamless and often invisible integration into our daily lives. As will the recent push from Amazon on to integrate GenAI powered experiences into its Alexa-enabled devices, of which the company has now sold more than 500 million. We saw a sliver of this at CES where Amazon showed off the Alexa integration of Character.AI, Splash and Volley. It is furthermore impressive to see the hunger people have to better understand this new technology and work to gain the ability to use it optimally and responsibly. Digital learning platform Coursera, for instance, added a new user on average every single minute in 2023 for its Artificial Intelligence courses as CEO Jeff Maggioncalda recently shared (Link).

Gaging enterprise adoption is more challenging than in the personal space. Numbers naturally vary from survey to survey but no matter the source, the data consistently shows that a large part of organizations have adopted Generative AI. Gartner’s latest poll (Link) found 45% of firms had GenAI in piloting mode and another 10% were already in production. Importantly, the adoption is not narrow, 45% of enterprises said they are scaling Generative AI across multiple business functions. From a functional perspective (Figure 1), Sales, Marketing and Customer Services are the main areas where Generative AI is being deployed (47% of use cases).

Looking at industries (Figure 2), another survey shows Manufacturing leading the early adopters, with 20% of firms having “extensively adopted generative AI across multiple functions”, followed by Retail & Consumer Packaged Goods with roughly the same number. Among the fast followers are Financial Services, where 17% of companies state they have extensively adopted GenAI and another 20% have a moderate adoption as well as Media & Telco with 17% and 18% respectively.

Looking ahead, with company’s GenAI investments ramping up further, enterprise adoption is going to accelerate even more in 2024. Importantly, we will see more and more companies move beyond simple use cases towards deeper adoption at scale. In Europe enterprises are set to grow their GenAI investments by 115% in 2024 to $2.8B and in North America by 70% to $5.6B (Figure 3a). In fact, AI and Generative AI are budging the overall tech investment trend of "change fatigue" that Gartner sees with total 2024 IT & Tech spending forecast to grow a meagre 6.8%. AI spending, on the other hand, is set to increase massively. Between 2023 and 2027 it will grow from $124 to $297 billion with a CAGR of 19.1% and within this Generative AI is set to more than quadruple is share from 8% to 35% (Figure 3b).

If GenAI adoption is to scale effectively though, companies will need to address the challenges posed by their legacy tech infrastructure, soiled data estates as well as AI skilling & talent gaps. Furthermore, firms will have to build robust Responsible AI frameworks as without them there will always be a deficit of trust from users and without trust there is no adoption. Lastly, to generate sustainable RoI from GenAI projects, companies need to build a strong operating model and governance around it. The recent examples of Amazon selling products with AI generated names such as “I cannot fulfil this request it goes against OpenAI use policy” (Link) is testament to the pitfalls that even Big Tech companies can step into without the proper oversight surrounding GenAI.

2024 will therefore see continued advancement in Generative AI's capabilities and its adoption. It is important though not to be blinded. While it is the shiny new thing with a previously unimagined capability set, GenAI is not always the answer when Artificial Intelligence is the question. Existing "Narrow AI" models are incredibly powerful and continue to advance at pace. They are by nature highly focused and task-specific compared to generalist GenAI but in their given domain often outperform the latter. It is therefore imperative to remember this and always choose the right AI tools for any given job. Moreover, we will see more and more collaboration between different AIs both narrow and generative that will unlock yet another level of AI capability (see Theme 5 below).

Further insights

For additional reading on GenAI adoption see:

• "The State of Generative AI Adoption" by LTIMindtree's Antony Panteli, Nachiket Deshpande, Sudhir Chaturvedi, Pascal BORNET et al. https://www.ltimindtree.info/gen-ai
• "Generative AI Radar" by Infosys Knowledge Institute's Samad Masood , Kate Bevan, Harry Keir Hughes and Jeff Mosier https://www.infosys.com/services/data-ai-topaz/insights/gen-ai-radar.html
• "2023 Generative AI in the Enterprise" by O'Reilly's Mike Loukides https://ae.oreilly.com/Generative_AI_in_the_Enterprise
• "What Can Copilot’s Earliest Users Teach Us About Generative AI at Work" by Microsoft https://www.microsoft.com/en-us/worklab/work-trend-index/copilots-earliest-users-teach-us-about-generative-ai-at-work
• "5 Forces That Will Drive the Adoption of GenAI" by Harvard Business Review's and Gartner's Alexander Bant, Helen Poitevin, Nicole (Denman) Greene, and Erick Brethenoux https://hbr.org/2023/12/5-forces-that-will-drive-the-adoption-of-genai

Theme 2: Verticalization. In 2024 Generative AIs will become substantially more tailored to specific industries and tasks.

The last year was the break-out year for the technology and dominated by broad, cross-industry adoption of general-purpose LLMs. This year will see a further increase of more tailored and customized Generative AI. This will be driven by two conjoining trends.

One is the increase of domain specific LLMs, whether purpose-built or fine-tuned. These are models trained on industry or task specific data sets such as FinBERT or the abovementioned BloombergGPT for Financial Services, Google’s Med-PaLM for Healthcare or even J.P. Morgan’s recently released DocLLM for legal document and contracting tasks. Importantly, the Open Source community is highly active in this space with various datasets and tooling available on Hugging Face. Hence, while we already see initial verticalized LLMs, their number is only going to increase in 2024.

Secondly, there is an even greater verticalization of Generative AI driven by SaaS players offering GenAI solutions tailor-made for specific industries and tasks. For instance, Start-Up Harvey was built for the legal profession and famously deployed by early adopter UK law firm Allen & Overy, who now themselves released new tool co-developed with Microsoft and Harvey called “ContractMatrix” that it now offers to other law firms (Link). Similarly, Start-Ups like Unique built GenAI solutions tailor made to support bank advisors and Lucinity’s GenAI helps identify Financial Crime. Fashable brings GenAI to, you guessed it, the fashion industry and Ello and Mindsmith to Education. The breadth of GenAI Start-Ups is already impressive (Figure 4). They offer important turn-key solutions to firms where investing in DIY GenAI is either not feasible or the continuous innovation benefits of SaaS players simply outweigh those of having one's own custom Generative AI solution.

It is also impressive to see that it is not only Start-Ups leaping into this lucrative niche of verticalized GenAI. Even big players are entering the space. For instance, Moody's Corporation and LSEG (London Stock Exchange Group) have launched their own GenAI offerings for Financial Services in partnership with Microsoft, SAP is offering embedded Generative AI solutions for supply chain as well as HR and ServiceNow does so for IT Service Management and other functionalities.

Hence, whether through the growth of sector and task specific LLMs or industry tailored GenAI SaaS solutions, 2024 will see a great increase in verticalized Generative AI, further driving adoption across large and small enterprises as well as by the Start-Up community where GenAI significantly lowers the barriers of entry.

Theme 3: Extension to the Edge. In 2024 Generative AI will make its way into every pocket and even the most remote locations through edge deployable models.

Foundation Models require large amounts of compute and thus run on cloud capabilities. Given their need for compute power, they cannot run locally, for instance directly on your phone. This is set to change in 2024. The recent CES has showcased a slew of AI capable chips by Intel Corporation and AMD coming to laptops and also phone providers taking centre stage in bringing GenAI to the edge i.e. our phones.

Samsung Electronics just announced its new “AI Smartphone” Galaxy line-up and showed off its partnership with Microsoft to bring their free Generative AI Copilot to its phones and therefore into the pockets of millions. Interestingly, Apple has to date been conspicuously absent from the public GenAI race, but is in fact working intensely to bring Generative AI to their iPhones. Apple researchers released a paper mid-December called “LLM in Flash” that offers a way to run complex LLM applications directly on iPhone's memory. Moreover, it released a Machine Learning Framework (MLX) that can run Deep Learning Models on Apple Silicon, indicating Apple wants to bring AI to its iPhones in a much broader manners than “only” GenAI. Whether in 2024 we will also see the much rumoured Apple LLM codenamed "Ajax" come to light remains to be seen.

Bringing (Generative) AI to the edge will be a key step ahead for adoption in 2024. It will integrate GenAI seamlessly into all our lives, even in remote locations, without the need for costly internet connectivity. For businesses, edge GenAI will be essential for the integration of Generative AI into IoT and automation projects and for usage in isolated locations from warehouses over mining shafts to construction sites and oil rigs.

The latter examples might sound odd and leave one wondering what role GenAI has to play in these areas. In fact, we are already seeing GenAI being integrated into robots. Massachusetts Institute of Technology, for instance, used a cohort of GenAI models to enable robots to solve complex object manipulation problems more efficiently (Link). And just now, BMW Group struck a partnership to introduce humanoid general purpose robots into its South Carolina plant to automate “difficult, unsafe or tedious” manufacturing tasks (Link). These robots will come from Start-Up Figure, which is also currently working on incorporating Generative AI into robots to enhance it semantic behaviour and increase its ability to understand human commands. Making the potential of Generative AI enhanced robotics accessible at scale in such industries requires edge GenAI and thus with the advances outlined above 2024 will be an interesting year in this area.

Theme 4: Geopolitical Competition. 2024 will see a continued struggle for global AI leadership playing out across multiple dimensions, impacting the evolution of GenAI.

Nothing is ever just about one thing, and so GenAI is also not only about technology but also about which country and region will seize AI leadership. The key competitors are naturally the US and China and to a smaller extend Europe, each seeking to foster their own AI innovation ecosystem from Silicon to Application, shape impactful yet also competitive AI regulation and keep/take control of key resources along the supply chain of AI development. This race has been going on well before the rise of Generative AI but has since intensified and is set to be a key part of the political as well as technology discussion in 2024.

Taking a leading role in AI R&D to create ever more capable (Gen)AI is a key dimension of this competition. The US is ahead in this space with the highest concentration of AI talent, one of the largest data pools as well as access to vast compute power and a large business ecosystem to commercialize AI. China, meanwhile, is working to close the gap regarding AI talent and seeks to capitalized on its own vast and largely walled-off consumer base and data pools. Moreover, it is fostering its own national AI and GenAI champions from existing large Techs like Baidu, Inc. (with its Earnie Bot) to GenAI Start-Ups like 01.AI with the aim to compete with the global GenAI leaders predominantly based in the US, with the exception of French Mistral AI and UK headquartered Stability AI. In this context, it will be important to watch the real-world impact of intensifying western sanctions designed to prevent China from acquiring key AI technologies. China is reporting to be self-sufficient in semiconductors, thanks to increased local production and R&D. However, other reports suggest the sanctions are indeed hampering its ambitions to the extend that China is resorting to repurposing Nvidia gaming chips from PCs for AI (Link).

They second important competitive playing filed in 2024 will be AI Regulation. In the context of Geopolitics, regulation is not just about regulation. Regulation will be the key differentiator for countries and regions competing for AI growth either by providing a clear and detailed set of rules for AI, which is the path taken by Europe with the EU AI Act, or adopting a more hands-off pro-innovation approach, as the US and to an even greater extend Japan are doing.

There is not right or wrong way here. We have seen in other areas of emerging technology that often companies favour and actively ask for clear regulation to provide them the necessary certainty for investment and implementation at scale. This is, for instance, the case in the blockchain and crypto space. Moreover, the ask for clear regulation is also a recurring theme among consumers as a foundation for their trust in using GenAI or the products of companies leveraging it. On the other hand, we know of ample examples where too intrusive, rigid and often too uninformed regulation stops tech innovation in its tracks.

Although there are competing approaches, there is nevertheless a certain level of consensus emerging internationally. Satya Nadella highlighted in Davos that he sees a coming together around the need for global coordination and standard setting connected to (Gen)AI (Link). Indeed, despite their differing approaches, we do see policy makers are taking the OECD principles for AI as a shared global guiding benchmark.

Hopes for too cordial and effortlessly coordinated global AI regulation in 2024 are nevertheless premature. For example, China has been forthcoming and advanced in its own AI regulation and is attempting to export its approach to AI, using standard setting as a political and competitive tool. This “emphasis on ‘promote’ policies with a global dimension has helped cement China’s position in the AI value chain” of many emerging markets and is further bolstered by its Digital Silk Road infrastructure projects as Lazard's Geopolitical Advisory points out.

The competition between the different international powers along these and more dimensions will be a key underlying shaping factor for Generative AI in 2024, impacting its evolution and implementation.

Further insights

For additional reading on AI Geopolitics:

• "Geopolitics of Artificial Intelligence" by Lazard's Theodore Bunzel, Siddharth Mohandas, Paul Norwood et al. https://www.lazard.com/research-insights/the-geopolitics-of-artificial-intelligence/
• "Charting the Emerging Geography of AI" by Harvard Business Review's Bhaskar Chakravorti, Ajay Bhalla, and Ravi Shankar Chaturvedi https://hbr.org/2023/12/charting-the-emerging-geography-of-ai

For additional reading on evolving global AI Regulation:

• "The Artificial Intelligence (AI) global regulatory landscape" by EY's Nicola Morini Bianzino, Marie-Laure Delarue, Shawn Maher, Ansgar Koene, Katie Kummer, Fatima Hassan-Szlamka https://www.ey.com/en_gl/ai/how-to-navigate-global-trends-in-artificial-intelligence-regulation
• "The State of AI Regulations in 2024" by Holistic AI https://www.holisticai.com/papers/the-state-of-ai-regulations-in-2024
• "The EU's Artificial Intelligence Act" by Clifford Chance's Dessislava Savova , Alexander Kennedy, Rita Flakoll https://www.cliffordchance.com/briefings/2023/12/the-eu-s-ai-act--what-do-we-know-about-the-critical-political-de.html
• "As gen AI advances, regulators - and risk functions - rush to keep pace" by McKinsey & Company's Andreas Kremer, Angela Luget, Daniel Mikkelsen, Henning Soller, Malin Strandell-Jansson and Sheila Zingg https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/as-gen-ai-advances-regulators-and-risk-functions-rush-to-keep-pace

Theme 5: Interactive AI. In 2024 AI will continue evolving beyond being a tool and towards being able to use tools (and other AIs) itself, unlocking a new level of capability.

The biggest advance in (Gen)AI this year is set to come from its ability to move beyond “just” creation and it gaining the ability to flexibly leverage other AIs and digital tools to take actions. Such "Interactive AI" will make AI capable of tackling far more comprehensive and complex tasks than ever before. One of the most vocal person to outline this next AI evolutionary step is Deep Mind co-founder and now CEO of Inflection AI, Mustafa Suleyman in his book "The Coming Wave". In 2024 we are unlikely to see fully fledged Interactive AIs but it will be the start of the journey. Though it may sound a farfetched idea, in fact we already took the first steps towards it. For instance, Multi-Agent AI systems where several GenAIs and traditional AIs are collaborating in solving complex tasks are already possible. Microsoft Research released a LLM Framework called “AutoGen” for this in August of 2023 (Link). With such capability AI will be able to use tools whether these are computer systems, other AIs or even robots (see below). For instance, think of the AI in a bank collecting information from you, then entering it into the bank’s different systems to do a credit risk assessment, then into those for calculating interest rates and charges and finally it triggers a virtual GenAI Avatar to come back to you with a personalized loan offer and to explain the details to you.

A way of looking at this from a different angle but to the same outcome is the notion of “Embodied AI”, raised again just recently at Davos by Daphne Koller (Link). Embedding (Gen)AI into agents such as robots allows them to interact with the physical world and go beyond using just digital tools, in addition to capturing vast amounts of new "hands-on" data to learn from. Adding to the above example of MIT in this area, Google DeepMind also demonstrated the far reaching potential of Foundation Models in Robotics (Link). In a Constitutional AI approach, leveraging multiple specialised Foundation Models embedded into robots (AutoRT, RT-Trajectory and Self-Adaptive Robust Attention for Robotics Transformers (SARA-RT)), the researchers enabled these robots to learn doing household tasks in a way previously impossible. Although easy for humans, giving a robot the task “clean up the house” would currently stump even the most advanced models as it requires a deep understanding of the world, the ability to reason and leverage available tools adaptively. While household tasks may sound trivial, it showcases how the embedding of Foundation Models opens up a huge new potential for (Gen)AI enabled robotics to interact with one another and the real world.

In short, Generative AI evolving the ability to interact with other AIs as well as digital and physical tools is the next capability frontier that we are already testing and 2024 will see further exciting developments towards such Interactive AI.

Looking at these five themes it is clear that Artificial Intelligence in general and Generative AI in particular will see exciting developments in 2024. Those raising concerns about a potential (Gen)AI Winter are not wrong to point to the well-known hype cycle and its valley of disillusionment. We all had our 30 days of utter excitement when the latest GenAI milestone was reached, thinking it will single handed change the world as we know it. Then we looked at it closer and began to calm down.

At a micro level of individual Generative AI applications and models such hype cycle view is correct. On a macro level, however, this view fails to sufficiently take into account the momentum behind (Gen)AI and the completely novel velocity of innovation we witness in the field.

Generative AI entering the scene in 2023 was an eye opening moment for most people and executives. For years we knew (theoretically) about the potential of Artificial Intelligence but OpenAI managed to create and iPhone moment. It made the technology's potential tangible and made it accessible to so many that suddenly there were not more excuses not to (finally) move ahead more seriously with AI and GenAI. The momentum that stems from this watershed moment is driving the adoption and investments outlined above and will result in continued new GenAI use cases being deployed and applications being launched and each pushes against the disillusionment dynamic.

On top of this, how AI is evolving has changed. It is not only about the kind of innovation we see as outlined above but also the pace of AI innovation. It took AI from 1998 to about 2014 (16 years) to reach human parity in handwriting recognition, for AI image recognition to get to the same stage took much less time, from about 2009 to 2015 (6 years). Now, GenAI matched humans in code generation and grade school math over just the course of 2022 to 2023 i.e. in 1 year. This unprecedented speed of improvement is enabled by AI outperforming Moore's Law (compute power doubling every two years). The compute used for AI models is now doubling every 5.7 months for regular scale models and every 9.9 months for large scale models. This power provided through cloud, coupled with ever vaster amounts of data available and rapid AI scientific advances creates a velocity of innovation hitherto unknow.

Together the momentum of applied innovation and velocity of (Gen)AI advancement mean that we experience continuous strides that outdo any downwards disillusionment momentum from traditional hype cycle dynamics. For this reason and the five themes outlined it is unlikely that we are heading towards an AI Winter in 2024. Rather, we will see continuous progress and exciting new applications of GenAI entering our daily personal and work lives.