GenAI - The Behemoth of the Century
The explosion of Generative AI (GenAI) in both personal and professional context testifies to the disruptive future of the technology and the corresponding shift of power from cognitive to compute – not keeping up with GenAI already puts at risk business continuity. McKinsey reports that in 2024 already 65% of businesses use GenAI regularly, nearly double the percentage from ten months ago. The unprecedented speed of AI hackathons and events due to the low entry costs and investor funding has unleashed new possibilities and challenges for businesses. The investment in GenAI platforms to date reflects the high expectations of industry disruption – Stanford’s AI Index report reveals an eightfold rise in GenAI funding, which hit $25.2B in 2023, $10B of which was attributed to the estimated investment of Microsoft in OpenAI.
GenAI is only eighteen months old, yet it grows at an unthinkable rate and presently likely exceeds 200m users. The adoption speed of ChatGPT 3.5, which was launched in November 2002, validated beyond doubt the market demand for the AI-driven conversational tools and the vast applicability of the technology. ChatGPT gained 1 million users in under five days and 100 million in two months, setting the record for the highest growing user base in history!! In comparison, popular platforms like Netflix, Facebook and Twitter needed respectively 3.4, 0.8 and 2 years to reach 1m users.
Data Source: ChatGPT
The ’operating system’ of GenAI
LLMs are like operating systems hosting different apps. Building upon the immense success of Chat GPT, the GenAI market accommodates a number of competing players. Anthropic’s Claude and Google’s Gemini have entered the race to compete with ChatGPT to try to curtail a predicted catastrophic loss of market share to search engines and online marketplaces. Claude 3, Gemini, Llama 2 and GPT-4 are very large and advanced models but there are other players whose models are worth mentioning. Perplexity AI is an AI-powered conversational search engine that leverages different LLMs to provide targeted information - it already has a user base of 10m users in 2024 and a valuation of $1b. Meta has its own LLM called Llama and focuses on open-source AI, similarly to the Emirates’ Falcon. Companies like Hugging Face and DeepMind (Inflection) also push forth their products, making for a dynamic AI model landscape. While industry leads the way in building large GenAI models due to prohibitive training costs, UC Berkeley and Stanford have entered the space with several advanced GenAI models.
The future of GenAI
In May 2024, I saw the presentation of Jeremiah Owyang from Blitzscaling Ventures, a Stanford-born VC in the Bay Area. Jeremiah believes that the mass majority of growth in the market will be seen in the AI apps that build upon the AI models, as the large players have already defined the majority of the AI models and much further movement is not expected there. He sees the emerging category of growth in the so called AI agents – they differ from the apps, as the apps work with the user in real-time, while the agents can assign tasks to other agents, browse the internet in an automated fashion, and offer ready solutions to the end user. In his opinion, agents will need little human supervision and soon we will not know whether we are dealing with human or agents.
How does this reshape the future? Search engines like Google and online marketplace giants like Amazon will be seriously challenged by GenAI and the player landscape will experience a shift like we have not seen before. Our interactions with technology will change, the ease and precision with which we get results will take formerly unreachable dimensions and our lives will become largely dependent on AI, more so than they have already become dependent on Google and Amazon. The technology will deliver and cater to our needs better than ever before, freeing up much of our time as it advances and becomes more sophisticated. We will no longer need to search, as ‘‘someone’’ else will do the work for us based on our actual tastes and preferences. Ready for a brand new world?
Challenges
Analysts predict a strong increase in capability for cognitive power due to GenAI. Yet there are market hurdles that need to be overcome to access this power, including the following:
1. Supply hurdles
The main bottleneck to AI’s infrastructure development have been the semiconductor chips, 90% of which are produced by the Taiwan Semiconductor Manufacturing Company, subject to potential political conflicts and uncertainty – some of these chips are extremely expensive, time-consuming and difficult to get.
2. Training cost
AI models get increasingly more expensive to train – according to Stanford’s HAI report, GPT4 used $78M worth of compute to train and Google’s Gemini respectively used $191M. These high costs have made it prohibitive for governments and universities to enter the space, which is currently dominated by industry.
3. The ‘jagged technological frontier’
A study made with the cooperation of BCG involved conducting an experiment with highly-skilled consultants to assess the AI impact on productivity when knowledge-intensive, realistic and complex tasks were involved. The study defined the jagged technological frontier as the line dividing the tasks that easily fall within the capabilities of GenAI vs the tasks that fall outside the capabilities. The highly skilled workers (consultants) greatly improved their productivity and quality; among these, the consultants with generally lower than average performance benefited the most from the use of Chat GPT-4, as long as the tasks fell within the frontier. However, when the tasks were outside the frontier, consultants using AI were 19 percentage points less likely to produce correct solutions compared to those who did not use AI. This result speaks to the importance of recognizing tasks where humans would perform better on their own. The important question here is whether future data and model training will shift the frontier to the advantage of the user.
The limitations of LLMs are evident. A Telerik Academy survey of 150 professionals in Bulgaria found that 73% of businesses use LLM assistants and tools on a daily or weekly basis. At the same time, 96% of the people using LLM stated that they needed to adjust the responses, as LLM did not deliver the required result.
The conclusion is that AI presently cannot replace the amalgamation of the human brain, and so compute and cognitive power must work hand-in-hand in validating and improving results.
4. Legal and ethical issues connected to potential misuse of the technology
Regulations to address legal repercussions and copyright issues in the context of GenAI are evolving and vary across states and countries. Celebrities have vocally expressed their dissatisfaction and taken legal action for perceived use of their AI-counterparts without consent. The latest is the case of Scarlett Johansson against OpenAI, which she accuses to have violated her right of publicity as a celebrity by introducing a voice very similar to hers in its new voice assistant to ChatGPT named Sky. Professor Kristelia Garcia from Georgetown Law School reminds of a precedent from 1988 called Miller vs. Ford, where the singer Bette Midler found Ford’s appropriation of a voice similar to hers in one of their commercials to be a tort under California’s right of publicity laws. Similarly to the Scarlett Johannson case vs OpenAI, the singer Midler had turned down an offer by Ford to use her voice. There is a clear conflict between the push of regulators to assign clear limits on transparency and use of the technology versus the offsetting need to unleash creativity through models and apps in Europe vs the US. Time will show how the opposing forces will give the technology enough space to flourish without crossing important legal and ethical boundaries.
Plagiarism is another topic that falls under the potential misuse of GenAI. Now it is easier than ever to use AI to mirror someone’s style and create ‘’personal’’ works. I recently met a university student who worked on building an AI model that generated text content, mimicking styles and frameworks of the data it was trained on. To protect against misappropriation of data, OpenAI has created a way to determine whether text has been created via use of its platform. There are also platforms such as Turnitin, which detect AI-generated text. Apart from the legal repercussions of plagiarism. UC Berkeley faculty members (Camille Crittenden, Chris Hoofnagle) have accepted that the use of models like ChatGPT will be unavoidable and encouraged students to take advantage of the technology as an aiding tool to support the critical thinking process or handle cases where more tedious writing is involved. Yet it is clear that the tools must be used but not abused, as they will not make one a great writer or lawyer and human-based skills must be internalized to succeed in life.
5. Data scarcity
A model is as good as its data. As organizations realize the value of their data, they will hold on to it, utilize internally and build their own models to realize the maximum value and protect their intellectual property. This will increase the costs of acquiring data that enables robust, trustworthy AI models. In addition, there is a limit to new data that can be used to train and improve future models. Researchers have found that basing future models on synthetically generated data eventually lowers output quality, so generating synthetic data could be a solution in the short-term, but it comes with its own limitations.
6. Lack of standardized AI reporting and transparency
There is no standardized responsible AI reporting and it is difficult to compare the risks and limitations of LLMs. Fake news and deepfakes, privacy and security violations, and lack of transparency in disclosing training data are among the top areas of concern. Large model developers like Hugging Face and Anthropic have been introducing safety mechanisms to increase models’ trustworthiness. Hugging Face hosts an LLM Safety Whiteboard where, according to Stanford’s AI Index Report, in 2024 Claude-2 was rated as the safest model, while ChatGPT-4 was only 6th in safety among 10 models, as it is easier to manipulate through misleading prompts. In addition, benchmarks like Truthful QA evaluate the truthfulness of answers by the different models.
Generative AI is presently incapable of explaining conclusions. This becomes especially tricky when the data have been manipulated and with the lack of data transparency.
7. Unemployment concerns
GenAI has faced similar fears as the rest of the AI, and rightly so. A lot of jobs that require mundane, lower-skilled and repetitive tasks will be overtaken by AI. These include lower-skill, repetitive customer service tasks, which are already being handled by chatbots. The loss of jobs across the board due to AI’s advent is already a fact and this is not a diminishing trend. Many workers will be laid off in the future and will need to reorient their focus, build new skills and undertake tasks that are not easily replaceable by AI.
However, there are also professions that will not be negatively impacted by GenAI, and will even be aided by the technology. Among these are roles that require empathy and human interaction, such as teachers, leaders and doctors, knowledge-intensive jobs in science that charter new territories and solve difficult problems, and skilled handworkers whose precision and individual mastery make the difference. Creative Directors improve their ideas through image, video and text-generating AI. No matter how advanced compute power is, it cannot replace human traits, be curious nor make and substantiate new discoveries. If deployed properly as an aid, GenAI will increase productivity and quality.
We are in the midst of a new era in human-machine interaction. Generative AI is no longer a choice and in most industries and jobs it will be critical to know how to make use of the technology to flourish and ensure business continuity. Both juniors and senior employees who know how to optimize the use of the technology have a significant advantage to those who do not use it or fail to use it correctly. Strategy and technology consulting companies have shifted a considerable part of their focus to AI consulting and integrated GenAI into their skills portfolio, as they know that not keeping up with the explosive speed will take them off the competitive edge quickly and surely.
How will you harness the power of GenAI to advance your business?
Note: I generated this image on Canvas - it features a hobbit’s hut and creatures that combine characteristics of colibris and pterosaurs.