Marketing Insight: The LLMs behind ChatGPT-4, Google Gemini and other leading AI chatbots
Large Language Models (LLMs) have taken the AI world by storm, particularly in the arena of chatbots and virtual assistants. This blog post is designed to provide marketing leaders with an in-depth understanding of LLMs, their significance in the market, and how to responsibly move forward using them.
What Are Large Language Models (LLMs)?
Think of LLMs as supercharged AI systems trained on massive amounts of text data. While it might sound simple, these models are trained to predict the next word. If you were given the sentence “I see her once in a blue…” to complete it you would likely choose the word “moon”. LLMs do the same thing by looking for patterns based on the data they were trained on.
This is what gives them their powerful capability to generate content and responses that mirror the quality of human language. These models are behind breakthroughs like ChatGPT-4, Jasper, and Google Gemini, which are perfect illustrations of LLMs in action.
The booming market of Large Language Models
The market for LLMs is witnessing exponential growth and is expected to skyrocket from $6.4 billion in 2024 to a whopping $36.1 billion by 2030 (Source: MarketsandMarkets).
However, the field is not just about sheer size; there are variances in performance influenced by the data on which each LLM is trained. The below comparison chart is a snapshot of some of the leading general purpose LLMs in the market and their variances in performance as of 26th March 2024.
Key insights comparing leading LLMs
Release Date: Newer releases consistently outperform their predecessors, highlighting the rapid rate of development in this space. For example, in December 2023 Google Gemini replaced PaLM2 (the LLM that Google Bard was built on) and in February 2024 Google Gemini Ultra was made available. Whatever models you choose to work with today will likely have a new release within the next year or a new player that outperforms existing ones.
MMLU: MMLU is benchmark that checks how well language models understand and solve a wide range of tasks and problems. While Google Gemini Ultra 1.0 is the highest performing with a 90% MMLU, its predecessor lagged behind ChatGPT 4.0, which has an 86.4% MMLU until recently. Claude 3 Opus was released in March 2024 by Anthropic and has an 86.8% MMLU. This technology is getting better at interpreting our requests and producing the outcomes we want.
Big Tech Backers: It’s no surprise that each of these is backed by big tech players. Microsoft has recently invested €15m in Mistral AI and has also invested a whopping €11.72bn in OpenAI, while partnering with Meta on Llama 2. Microsoft Copilot uses Open AI’s ChatGPT-4 LLM, while their cloud platform Azure leverages Llama 2. Amazon is investing up to €3.75bn in Anthropic, while Google continuously invests in the development of Gemini. The race is definitely on and Microsoft is hedging its bets.
Creativity: This characteristic refers to how original and engaging the LLM's content can be. Can it come up with fresh ideas for marketing campaigns, product descriptions, or social media posts? While this requires testing and the quality of your prompt can enhance your results, from my own experience I find Google Gemini to be more creative than ChatGPT-4. That said, you can engineer your prompts to get more creative outputs. Gemini Advanced is Google’s AI chatbot that uses their highest performing LLM, Gemini Ultra 1.0, and you can sign up for a free trial for two months. Early reports also suggest Claude 3 Opus is superior to ChatGPT-4 for creative writing.
Multilingual Capabilities: This may be important if your marketing efforts target a global audience. How well can the LLM handle different languages and cultural nuances? Many LLMs have been predominantly trained on English data and as a result, their comprehension of other languages will not be as strong. Claude 3 Opus and Mistral Large are more advanced in their comprehension of other languages in comparison with Google Gemini and ChatGPT-4.
Data Security & Privacy: To ensure LLMs uphold the highest data security and privacy standards, it's crucial to consider their training data. Remember, the more data they train on, the more proficient or biased they can become. The prompts you input in each LLM might shape future models unless you explicitly disable this feature. ChatGPT-4 offers the choice to opt out of training future models, whereas Google Gemini utilises your prompts or 'conversational data' for training. This significantly impacts the potential applications of LLMs. You wouldn't want to input sensitive business information into an LLM that could potentially disclose it as a response to a user query in the future.
Meet Poe: Your answer to testing AI chatbots simultaneously
Choosing an LLM or AI chatbot is largely dependent on what you want to use it for. Poe is a helpful service that lets you try chatbots powered by most of the major AI models in one place, including OpenAI’s GPT-4, Google’s Gemini Pro, Anthropic’s Claude 3, Meta’s Llama 70B and more.
It’s likely a more reliable source of performance than this blog post alone as by now you should appreciate how fast this space is changing.
What are LLMs saying about your brand?
Many teams are focussed on the content that they can generate with LLM powered AI chatbots, however, as marketers it’s also important for us to consider how we want LLMs to perceive our brand and compare it to competitors.
This is particularly true as consumers go to AI chatbots to find answers to almost anything they can think of. Taking steps to analyse how LLMs perceive your content can help you strategically shape your messaging and position yourself against competitors.
Integrating LLMs responsibly in marketing
With LLMs at the forefront of AI technology, the onus falls upon marketing leaders to integrate them into their strategies with care. Here are a few places to start:
Integrating LLMs with business goals
An LLM is only as effective as the goals it is intended to serve. Whether it's content creation, audience engagement, or market analysis, aligning the use of LLMs with specific business objectives and testing it is key.
Analyse how LLMs perceive your content
Consider analysing what LLMs are they telling you about how well your brand positioning, creative and messaging are working. How can you elevate your brand to appear as the answer or leading voice on certain topics?
Ensuring ethical use of LLMs
Biases in training data can translate into discriminations in AI-generated content. Marketers must be vigilant in identifying and rectifying these biases for fair and inclusive practices. Transparency in the use of AI, especially LLMs, is critical for maintaining consumer trust and compliance with data privacy regulations. Don’t shy away from disclaimers that highlight where you are using AI.
Balancing Innovation with risk mitigation
Marketing leaders must strike a balance between leveraging the latest technology and safeguarding their operations against unforeseen challenges. This includes understanding and mitigating the potential legal, brand, and operational risks associated with LLMs.
Strategically collaborate with IT and security teams
While marketing leaders provide strategic vision, those with technical expertise can offer valuable insights into selecting, implementing, and securing LLMs. Collaborative efforts enhance the effectiveness and safety of AI strategies.
If you’ve gotten this far, hopefully you’re more confident in your understanding of LLMs and know your Claude 3 Opus from your Llama 2 70b and are ready to sign the petition with me for names that are easier on the eye!
If your team is interested in getting ahead of generative AI and the human skills needed for the changing workplace, see here for more details.