How SLMs are the accessible alternative to LLMs

Here, Glyn Heath, Director of AI at Bayezian, looks at the rise of Small Language Models (SLMs) in 2024.

The popularity of ChatGPT, soon to be celebrating its first birthday, has exploded and launched the familiarity of LLMs into the public consciousness.

However, for all the breathtaking pace of change and improvement, it’s also laid bare some of the shortcomings, too. Researchers at Brown University have recently discovered new vulnerabilities in OpenAI’s GPT-4 security settings; by using less common languages like Zulu and Gaelic, they could bypass various restrictions. The researchers claim they had a 79% success rate running typically restricted prompts in those non-English tongues versus a less than 1% success rate using English alone.

Other examples are the Tom Cruise parents disconnect – whereby you can ask ChatGPT who Tom Cruise’s parents are, and it will say Thomas Cruise Mapother and Mary Lee (which is correct). However, if you ask who Thomas and Mary Lee’s child is, it doesn’t know. This has reputedly been rectified, but the root of the problem is fundamental to the design of current neural networks. Then there are the more recently surfaced issues such as the inability to perform basicmultiplication or tell the time accurately from the image of a clock face (often referred to as hallucinations).

While not causing serious alarm yet, this has forced a rethink for the adopters of LLMs in the commercial space.

Now, the use of small language models looks set to rise into 2024. With many organisations having no need for the breadth of general purpose LLMs, more compact models are being built at a fraction of the cost and at a much quicker pace. This enables for rapid iteration of models in order to train and tune them for the specific use-case.

Why SLMs?

A new report by Dylan Patel, Chief Analyst at the research firm SemiAnalysis, found that it costs approximately $700,000 per day, or 36 cents per query, to keep ChatGPT up and running. With that in mind, it’s easy to see the attraction of SLMs for organisations of all sizes – but especially smaller firms. Costs and complexity are slashed and operational maintenance is more straightforward.

SLMs offer greater potential for specific tasks, too. Businesses can tailor the models to the needs of a particular use-case and extract maximum value from their AI investments, as well as better alignment with business objectives. Data bias, often cited as a drawback of LLMs, is less of an issue in smaller models. If biases do arise, they can be identified quickly and ‘trained out’ with greater accuracy, enabling a return to smoother operations more efficiently.

The year ahead

While it’s difficult to make predictions for any technology that’s constantly smashing expectations, it’s worth pondering where generative AI could be in one year. Namely, the early hype around LLMs will start to settle down; organisations and the wider general public will get to grips with how they can be used most effectively and, equally, what they’re not so good at. As a result, expect to see incredibly innovative use-cases with embedded AI in 2024.

The advance of LLMs heralded a wave of gloom by doomsayers, who warned of the beginning of mass redundancies. It’s doubtful we’ll see that in 2024; that is for us to plan for over the longer-term.

On the contrary, there will be enormous demand for people who understand how to extract value from LLMs. A new job category will emerge – Prompt Engineers, as coined by AI pioneer Mustafa Suleyman – for those that understand how framing the dialogue with an AI can obtain the best results. In addition, the demand for data scientists, programmers, and for other staff within many departments of organisations – both in the public and private sectors who understand how to exploit the benefits of AI – will continue to grow.

The bottom line

Predicting what the coming year will bring, especially with the pace of technology advancement over the past 12 months, really is crystal-ball gazing. However, it’s certain that LLMs will prove to be an accessible alternative for organisations that are looking to innovate in 2024.

It is likely that challenges to the roll-out of AI powered systems will increase, especially from sectors that have tended to be unaffected by automation or other disruptive technologies in the past, such as script and song writers, along with copyright and IP infringement claims.

One aspect that will almost certainly gain momentum in the coming year will be the debate, globally, about how the future impact (both positive and negative) of frontier AI technologies can be managed to maximise the benefits while safeguarding humanity.