Implementing AI during a worldwide talent shortage

C-suite demands for the proliferation of AI throughout the enterprise are often complicated by the lack of available talent and the requisite skills to endeavor on such deployments. Budget is rarely the limiting factor — especially for larger organizations. What’s missing is the people with the knowledge and hands-on skills to test and institute AI throughout an organization.

When the right machine learning (ML) models are combined with the right use cases, AI can augment customer service, perform administrative tasks, analyze huge data sets, and perform many more organizational functions in enormous volume and with low error rates. Business leaders know this. Yet they’re being held back from acting on that knowledge.

New research by SambaNova Systems has shown that, globally, only 18% of organizations are rolling out AI as a large-scale, enterprise-scale initiative. Similarly, 59% of IT managers in the UK report that they have the budget to hire additional resources for their AI teams, but 82% said that actually hiring into these teams is a challenge.

Every hour of repetitive tasks that can be cut by automating or augmenting with AI is an hour that employees can spend deriving value through higher-order, lateral thinking tasks. Firms are watching their competitors find a competitive edge when they test, iterate, and roll out wide-scale AI programs, casting about for whatever AI and ML expertise they can attract in the meanwhile.

This skills crisis is not new, nor surprising, nor easily solved. It’s been an issue across the tech sector as a whole for years, if not decades. In 2011, a PwC study found that more 56% of CEOs were concerned about a lack of talent to fit digital roles. And more than a decade later, 54% of tech leaders ranked talent acquisition and retention as the number one threat to business growth.

The era of AI has made this problem more acute — the pace of change is outpacing what’s come before.

The skills crisis is exacerbated by the rapid pace of change in AI models

The challenge for anyone working in AI who wants to keep their skills up to date is two-fold. Firstly, the pace of change is breathtaking, and seemingly getting faster all the time. Secondly, as models become bigger, they become less accessible for software engineers to train, as large models need big budgets to run.

The hottest topic in AI is probably large language models (LLMs). The first Generative Pre-trained Transformer (GPT) model was launched by OpenAI in 2018 — which, as a general purpose learner, is not specifically trained to do the tasks it’s good at. The model leverages deep learning and is able to carry out tasks such as summarizing text, answering questions, and generating text output — and doing so on a human-like level. The first model came out four years ago, but it only leveraged 150 million parameters (a dataset of less than a million web pages). The breakthrough for GPT and large language models came with GPT-3, which launched in 2020 and had 175 billion parameters, more than a thousand times the number of the first GPT model.

Since this first large language GPT model from OpenAI (which has significant investment from Microsoft), others have been released from Google, Meta and Aleph Alpha. It’s no coincidence that these huge tech companies are behind large LLMs: They require huge amounts of experience to train and run. GPT-3 was trained on 45 terabytes of data and likely cost millions of dollars in computing to create the model. Even the recently-released open-source LLM by BigScience, BLOOM, took the combined efforts of more than 1,000 volunteer researchers, $7 million in grants, and access to the Jean Zay supercomputer near Paris.

Although the concepts are accessible, it’s much more difficult for a typical software engineer to get hands-on experience with the models because of the expense of running them.

The challenge of building a team

SambaNova research found only one in eight IT leaders have fully resourced teams with enough skilled workers to deliver on what the C-suite is asking. A further one in three are struggling to meet the demands placed on them. The rest (over half) are unable to deliver on the C-suite’s vision with the people they have.

IT leaders have the budget to hire, but recruitment and retention can often prove to be a hugely complex and difficult process. Technology companies aren’t in a race for hardware or resources so much as they’re in a race for the best minds. As a consequence, those minds have become a valuable resource in and of themselves.

Issues to do with supply shortages are multifarious, often difficult to isolate and overlap. One of the key obstacles facing teams that want to hire new talent for their AI initiatives, and the cause of that dearth, is that, as a practical discipline, AI is relatively new. It has been studied in theory and practice for as long as we’ve had the computing and technical know-how to achieve it, but formal, academic education has only just become widespread. This doesn’t help the organizations that need a fully-formed, comprehensive talent pool now.

Faculty staff with experience and training in AI — both in theory and in practice — are hard to come by for universities. Despite speculation about the pull from the tech sector, many researchers remain interested in academia. However, the enormous demand for courses and a relatively short history of graduates from such a new discipline all depress the number of available professors and constrict the talent pipeline.

So, not only will organizations struggle to hire the AI skills that they need, but that those looking for an education in AI to acquire these skills will struggle too. That’s why organizations need to look for alternative ways to achieve their AI/ML goals.

How upskilling can help promote talent from within

There are ways for engineers to upskill and extend their knowledge in AI. There are a number of open source projects such as TensorFlow (open sourced from Google) and Pytorch (open sourced from Meta).

Upskilling, as a practice and a workplace policy, is as good for the employee as it is for the company. The organization gets a future-proof workforce with broader skills and interdisciplinary AI capabilities, working with the latest approaches and research to improve its knowledge base. For the employee, they ensure that their skillset is aligned with current sector trends and they can future-proof their own careers, setting themselves up for longevity in the industry.

By investing in learning programs, companies can help to ameliorate some of the more severe impacts of the skills crisis. These programs can bridge the gap between the talent organizations already have and the talent they need to implement models and ML programs that can create additional value. This means having a very clear view of where upskilling programs begin and end: The skills they’d like their workforce to have and how they can promote them from within.

Therefore, when top talent does become available, they act as an augmentation to an already-functioning AI team, rather than the foundation to a project that is waiting for them to arrive.

When is outsourcing the right option?

There is, of course, another option. Outsourcing. Having an outside start-up or expert AI company partner with an enterprise can help them get access to the value and cost-savings of AI. However, this comes with a whole host of issues and considerations. It will be the right option in some cases, but there are drawbacks that need to be taken seriously.

The integration of start-ups and other enterprises within a corporate structure does not always work smoothly: The startup culture of ‘move fast, break things’ can clash with a more considered, bureaucratic approach. The difference between short-term thinking and long-term thinking may also emerge, depending on the dynamics of the partnership. As a rule, these implementation projects are either long-term or short-term investments, and it’s vital to get on the same page early so that timetables and priorities are clear.

For smaller enterprises looking wistfully at the gravitational pull that companies like Google and Meta have to build star-studded AI initiatives, outsourcing is a way to fast-track their own development. Much like a small start-up hiring a freelancer to do its web design, copywriting, or financials, SMEs can use outsourcing to implement the right AI models quickly and without huge up-front costs — alongside assurances of return on investment.

On that note, enterprise leaders should consider the technical effectiveness of any outsourcing partners and their specific metrics for success. If a partner is able to clearly define and show how effective its models and algorithms are, how much it can do with the data, and how long the training process might take, this shows that there is some common basis, and expectation, for what success looks like.

Ultimately, given a historic shortage of AI talent, enterprises and team leaders need to make the decisions that are right for them. The costs of going in-house and constructing your own team from the ground up, at a time when Big Tech firms like Google, Meta, and others are engaged in a tug-of-war for experienced employees, may be hugely costly and inefficient. But no two projects, or companies, are made equal, and only those with the data at their fingertips can say whether they need outside help or not.

What’s the next step for under-resourced AI teams?

Enterprises and smaller organizations are coming to the realization that the small models that have been deployed around the company for various purposes have become unmanageable; they are fragmented, siloed, and frequently incomprehensible to everybody but their creator.

As staff leave for better offers, more favorable working conditions, or merely a change, entire processes and systems are being left behind. Companies aren’t sure if these huge amounts of AI models and their uses can be audited, and often these departures freezes models in time. Like an archaeological finding, nobody wants to touch them lest they break.

The benefits, present and future, of AI are all around us. We see the statistics daily: Billions of dollars of value added, thousands of hours saved in administrative tasks, and the disruption of entire industries. However, the gap between what C-suite level executives want and what they can have is unfortunately large — and that begins with their struggle to hire the right people.

The UK government has recently set out proposals for a new rulebook on AI, on top of existing funding allocations, to truly establish the UK as a global AI hub. To realize that potential, more must be done. This begins at the university level: Feeding huge demand with top-class courses, experienced lecturers, and hands-on, practical experience with the models.

But businesses can’t always afford to wait such a long time to reap the benefits of AI, and with the host of options available to them in the short term, they may not have to.

Marshall Choy is SVP of product at SambaNova Systems