10 Trends in LLM Dev Business Owners Need To Watch in 2024

Artificial intelligence (AI) may have been around for years, but its large language models (LLMs) have captured the attention of business owners and AI professionals.

In fact, according to McKinsey, one-third of organizations are using generative AI in at least one business function. This means a growing need for skilled AI and ML engineers to satisfy market demand.

Knowing the key LLM trends will help businesses make informed decisions on what models they might consider building their new projects, as well as allow AI developers to stay in the loop and update their skill sets accordingly.

But how are LLMs evolving exactly? In this article, we will look at the future of large language models, including the movement toward multimodal input to the growth of the open-source market and the increasing cost-effectiveness of contemporary language models.

Table of Contents Table of Contents

1. Key Takeaways
2. Top 10 LLM Trends to Keep an Eye on
3. The Bottom Line
4. FAQs
5. References

Show Full Guide

Table of Contents

1. Key Takeaways
2. Top 10 LLM Trends to Keep an Eye on
3. The Bottom Line
Show Full Guide
4. FAQs
5. References

Top 10 LLM Trends to Keep an Eye on

1. LLMs Will Become Increasingly Multimodal

This was one of the key themes in Sam Altman’s recent conversation with Bill Gates, where Altman noted that “multimodality will definitely be important” in the future. This is further highlighted by OpenAI’s launch of GPT-4V last year, which enabled users to use image inputs in ChatGPT.

Likewise, Google has also attempted to shift toward this approach by creating a family of multimodal LLMs called Gemini, a model that Demis Hassabis, CEO and co-founder of Google DeepMind says was “built from the ground up to be multimodal,” supporting text, code audio, image, and video input.

2. The Gap Between Open and Closed-Source Models will Continue to Close

Just recently, Meta announced that Code Llama 70B, a fine-tuned version of Llama 2 designed for writing and editing code, had outperformed GPT 3.5’s on the HumanEval benchmark (53% compared to 48.1%) while coming close to GPT-4’s performance (67%).

In addition, Mistral AI recently announced the release of 8x7B, a language model with 46.7B total parameters, which it says offers 6x faster inference than Llama 2 70B and “matches or outperforms GPT 3.5 on most standard benchmarks.”

Although there’s still a while to go before these tools are at the point where they can challenge top performers like GPT-4, there is a steadily growing ecosystem of viable LLMs for enterprises to choose from.

3. The Rise of Small Language Models

The high cost of training and running AI models has been a barrier to adoption for many organizations. By some estimates, training an LLM such as GPT 3.5 could cost over $4 million, a significant investment for any organization.

As of 2024, there are numerous examples of these models in the wild, including Stability AI’s newly released Stable LM, a language model with 1.6 billion parameters, which was trained on 2 trillion tokens, including multilingual data in English, Spanish, German, Italian, French, Portuguese, and Dutch.

There’s also Microsoft’s Phi-2, a 2.7 billion parameter model released in December 2023, which features outstanding reasoning and language understanding capabilities to the point where it can outperform models up to 25x larger due to its highly curated dataset.

These releases are just several examples of an increasing number of models that are designed to function more efficiently than LLMs.

4. Language Models Are Going to Become Less Expensive

At the same time, there is a concerted effort among vendors to cut the cost of training and running LLMs.

This is shown by the fact that less than a month ago, OpenAI announced it would be lowering prices on its GPT 3.5 Turbo model, with input prices reducing by 50% to $0.0005 /1K tokens and output prices reducing by 25% to $0.0015 /1K tokens.

However, it’s not just OpenAI that is looking to cut costs. Recently, Anthropic also cut costs for its popular proprietary LLM, Claude 2.

5. More Will Experiment with Direct Preference Optimization as an Alternative to RLHF

For years, reinforcement learning from human feedback (RLHF) has been used as a technique to help train machine learning algorithms to align with the preferences of human users.

However, Stanford researchers recently discovered a compelling alternative – direct preference optimization (DPO), which is likely to see much more use among LLM vendors.

Under RLF, a developer would need to build a reward model based on human feedback, which would help fine-tune the model based on human preferences. In contrast, Stanford’s method provides an alternative technique for training language models with preferences without time-consuming reinforcement learning.

“DPO identifies a mapping between language model policies and reward functions that enables training a language model to satisfy human preferences directly, with a simple cross-entropy loss, without reinforcement learning or loss of generality. With virtually no tuning of hyper parameters, DPO performs similarly or better than existing RLHF algorithms,” the study said.

6. The Move to Autonomous Agents

Autonomous agents are another component of LLM development that’s growing significantly. Last year, autonomous agents like AutoGPT gathered a lot of attention due to their ability to interact with language models like GPT 3.5 and GPT-4 and perform tasks independent of human input.

For instance, the Center for AI Safety warns that malicious actors could create rogue autonomous agents, citing the incident where a developer used GPT-4 to create ChaosGPT, an AI agent instructed to “destroy humanity.”? While the organization notes it didn’t get far, it does demonstrate how such tools can be weaponized.

7. Robotics-Focused Vision-Language Action Models Will Pick Up Speed

AI has been a staple of robotics development for years, playing an integral part in the development of advanced humanoid robots like Hanson Robotics’ Sophia, but it’s becoming increasingly clear that more AI vendors are looking to invest in this research area.

According to Business Insider, as of January 2024, Microsoft and OpenAI are considering investing $500 million into the rapidly rising robotics startup Figure AI.

Last year, we also saw the release of Google DeepMind’s Robotics Transformer 2 (RT-2), a vision-language action (VLA) model designed to help robots understand and perform actions.

Essentially, RT-2 uses an LLM to generate motion controls and gives robots the ability to interpret commands. This includes placing an object onto a particular number or icon, picking up the smallest/largest object, or picking up the object closest to another object.

8. More AI Vendors Will Offer Users Custom Chatbots

As AI engineering continues to mature, there is a move in the marketplace toward customization. More specifically, more vendors are offering customizable chat assistants.

This can be seen most obviously with the launch of OpenAI’s GPTs in 2023 – essentially custom versions of ChatGPT that can be shared with other users via the newly released GPT Store.

Hugging Face is also now offering users the option to make their own custom chatbots in the Hugging Chat Assistant, choosing between any open LLM and assigning a name, avatar, and description.

Given that other organizations like Bytedance are also evaluating custom chatbots as a potential solution, we can expect more vendors to follow suit.

9. Generative AI Will Be Used in More Consumer Apps

Aim Research anticipates that 40% of enterprise applications will have embedded conversational AI in 2024, with real-time outputs in 70% of practical applications by 2030.

As of today, we can see generative AI used in a variety of popular products, including Grammarly, which added generative AI to its proofreading solution in June 2023 to give users the ability to produce content on demand.

Likewise, HubSpot has also added AI-driven tools to HubSpot CRM, including a Content Assistant, which can generate blog titles, outlines, and content such as blog posts, landing pages, webpages, and outreach emails.

10. Retrieval Augmented Generation Will Make LLMs Smarter

Finally, in an effort to improve the performance of language models, we’re seeing more and more researchers experimenting with retrieval augmented generation (RAG).

Research from Pinecone shows that using GPT-4 with RAG improved the quality of answers by 13%, even with regard to information that the LLM was trained on. This means that the increase in quality would be more pronounced if the questions were related to private data.

The Bottom Line

LLMs as a technology may be young, but their capabilities are evolving fast. With multimodality becoming more popular and LLMs potentially becoming more computationally efficient and cost-effective, barriers to AI adoption are decreasing.

Although these solutions are a long way off artificial general intelligence (AGI), they are getting better, and we can expect to see adoption increase throughout the course of 2024 as more use cases emerge.

FAQs