The AI Landscape of 2024

In 2023, generative AI went mainstream, with everyone amazed at the capabilities provided by transformer language models like OpenAI’s GPT model and the image generation capabilities of models like stable diffusion. The hype around this technology has likely reached the peak of expectations, and while there may be some disappointment forthcoming as defined by the Gartner technology adoption cycle, the potential of the technology is truly exciting and we will likely see companies adopting the technology at an ever-increasing rate moving from Proof of Concepts to full-featured product launches.

Source: Wikipedia

2024 will feature both significant breakthroughs in models and products but also society catching up with the implications of what these models can do and the government introducing regulations and laws around these models.

The Tech

On the technical side, 2023 was pivotal in breakthroughs in transformer scaling, exemplified by GPT-4 from OpenAI. Claude 2 from Anthropic and Llama2 from Meta. These models showcased an unprecedented ability to generate coherent and contextually relevant text, revolutionizing industries from content creation to customer service. These models will continue to amaze and inspire in 2024 as language models expand their capabilities with things like ever-longer context windows, native retrieval augmented generation support, and better complex task following.

While GPT-4 is said to have about 2 trillion parameters in a mixture of experts architecture, the expectation is that the next generation of frontier models will have parameter counts in the tens of trillions. At these scales, we are pushing the boundaries of rich human-generated text data to train these language models on, and so in the next year, significant work will be done on synthetic data creation for language model training. While 2023 made RLHF (Reinforcement Learning with Human Feedback) a household word, expect further advancements in software and algorithms to also enable ever more complex language models. Most firms are currently incorporating a promising technique from Stanford called DPO (Direct Preference Optimization) into their model training. The first sets of these models are expected to be released in 2024.

Additionally, models will not just be limited to text data; as the saying goes, “a picture is worth a thousand words.” Other modalities like audio, video, and images will also be integrated into these language models. In the latter part of 2023, we’ve seen the limited release of Google’s Gemini models and Open AI’s Vision-enhanced GPT-4.

In 2024, these capabilities will expand even further. Meta has already released ImageBind (ImageBind: Holistic AI Learning across Six Modalities), a multimodal embedding along with Ego-Exo-4d (Introducing Ego-Exo4d), a dataset that provides both first and multiple third-person video of tasks being accomplished along with audio narration and expert critiques.

Source: Ego-Exo 4D, Meta

Expect companies to train on this and other similar datasets to create ever more capable multimodal models that are much more comprehensive, trained not only on text, but on audio, video, and text across multiple viewpoints simultaneously, enabling much more intuitive AI applications that cater across a broad spectrum of human to computer interactions.

To power all of these ever increasingly complex models, companies will require ever larger compute from ever more powerful machines. Facebook’s original Llama model (released in February of 2023) with only 65 billion parameters required 61000 GPU days on the most advanced available Nvidia A100 GPUs at the time. The most advanced current GPU, the Nvidia H100, would have reduced that training time by 75%.

2023 also saw the emergence of the GPU rich and GPU poor (Patel, Dylan, and Daniel Nishball.) based on your company's access to Nvidia hardware. Nvidia GPUs have become the workhorses of AI/ML over the last several years. Nvidia software + hardware were critical to creating the original transformer architecture for the first generation of language models. As everyone wanted to train AI models for their own purposes last year, a common refrain became the inability to access these specialized chips. The expected wait time to order the most advanced Nvidia H100 server is somewhere between 9 to 15 months.

This shortage doesn’t show any signs of abating over the next year, but there is a renewed effort to break the Nvidia monopoly on AI. AMD will make concerted efforts to improve the software ecosystem around their MI series chips to enable an alternative to Nvidia in AI workloads. Not content to wait months for more capacity, the large cloud providers have taken matters into their own hands. The lack of availability of these chips has pushed the major cloud players into developing their own chips. AWS has heavily pushed its Trainium/Inferentia chip line, Google has refreshed its TPU lineup (gen 5), and Microsoft has developed its own in-house chip, MAIA (the Microsoft AI Accelerator).

Nvidia will not rest on its laurels, however, and will be releasing an even more powerful AI chip in March 2024 (the B100). Not to be left out, the open source community has heavily invested in the ability to run these models locally on consumer-grade GPU infrastructure so a high-end gaming PC from the last 3 years can also run customized variants of these ever-improving models. By the end of 2024, expect to see capable small language models that will be able to run locally on devices like smartphones.

The Uses

2024 will also be the year that AI starts to deliver gains in efficiency and productivity for businesses. In 2023, we saw companies trying to deploy generative AI in numerous proofs of concepts and experimentation, but there was no real consensus on what types of products could be built.

But most critically, it won’t be possible for a company to exist without an AI strategy in 2024. In 2023, outright bans on generative AI products like ChatGPT in the workplace just led to workers using it without permission. A survey in early 2023 indicated that over 43% of professionals were using ChatGPT and that 70% of workers used it without telling their firms (Graham, Becky). Now, firms are trying to maximize productivity by enabling their employees to use these AI tools in a secure and compliant manner.

Microsoft has fully integrated its AI Copilot offerings into its entire product ecosystem from Outlook to Office, while Google has integrated its Duet offering into the entirety of Google Workspace. While 2023 had a boom of smaller companies providing AI-assisted notetakers, these tools have been slowly integrated into the more mainstream video conferencing and meeting tools. Zoom has integrated these directly into their platform (Zoom), and Google has teased a feature that will not only summarize meetings but attend them on your behalf using AI to ask questions and provide feedback that you would have asked (Peters, Jay.)

Source: The Verge

Expect the next email from a co-worker summarizing deliverables from the status meeting to be at least AI assisted if not entirely AI written.

As AI gets integrated directly into workflows, guardrails will be critical in these early stages of deployment. It is still relatively easy for actively malicious users to push these tools off course. A Chevrolet dealership in California that deployed a customer service chatbot went viral after a user convinced the chatbot to sell a brand-new Tahoe for $1.

Source: Chris Bakke X.com

While humorous, the wider internet quickly discovered that they could also use this as a gateway to access the most advanced models from OpenAI for free, incurring additional significant expenses for the dealership.

Deploying AI tech into business workflows will not be without these types of speed bumps; expect numerous embarrassing incidents like this to continue as firms roll out the technology for everything from customer service to content moderation.

One key trend for 2024 will be the movement of AI tools beyond single-use applications into task-accomplishing agents. Key research in 2023 indicated that AI models were capable of planning complex tasks, and companies like LlamaIndex and LangChain quickly created agent frameworks to provide tools for these models to interact with the larger world through both APIs both in consuming data as well as taking actions on that data. Companies offering relatively simple agents through techniques like retrieval augmented generation have found significant success in answering questions about specific documents; this has even become a core feature of OpenAI’s GPTs product (Introducing Gpts.).

As of December, Olivia Moore of Andreessen Horowitz tracked nearly 60 separate companies offering agent platforms across a wide spectrum of use cases.

Source: Olivia Moore X.com

Next year, we will see an explosion of these companies. If 2023 was the year of the AI Chatbot, then 2024 will be the year of the AI Agent, and these tools will become more complex, capable, and contentious. Once agents start interacting with data systems at scale, we will need robust mechanisms to check these outputs. One key risk that prevents significant adoption is the risk that a deployed AI agent can cause either reputational damage through inappropriate responses like Tay (Tay (Chatbot)) or actual monetary damage, perhaps by refunding a customer for a hallucinated order they never made.

If an AI agent really can sell a car, then the incentives to break the agent become that much more rewarding. 2024 will see significant effort dedicated to improving these guardrails and managing hallucinations from language models.

We believe the governance of AI outputs will be a continuing (and growing) need and have built BrandGuard with this in mind. In the Marketing space, having an automated system like BrandGuard that understands your brand style and guidelines and can check for violations across it all in a matter of seconds mitigates the risk of working with AI, makes the whole process faster and creates a more consistent brand experience for your customers.

The Impact

These reputational risks are not just limited to companies. The rapid improvements in AI technology in 2023 caught society off guard. Deepfakes became a part of our lexicon, with warnings about AI voice cloning (Evans, Carter, and Analisa Novak) proliferating across both traditional and social media. The improvements in image generation models meant that almost anyone could generate consistently realistic images like that of Pope Francis wearing a white jacket (Dolan, Leah).

Source: CNN

In 2024, this technology will improve quality and expand to things like video. While current video generation models create short, poor-quality videos, significant research is in progress to create high-quality synthetic videos. The next viral video you see on TikTok may already be AI-generated.

Governments have also been paying attention. The EU passed its AI Act in early December and the US government issued an executive order directing the creation of regulations around AI at the end of October. While this is a first pass, expect new regulations to come out from other regulatory bodies at both an industry and geographical level.

A key theme that will play out through 2024 and beyond will be AI models and their relationship with copyright laws. Because there is some copyrighted data in most languages and image model training data, rights holders are actively litigating the image training data for these image generation models (Getty Images v. Stability AI) as well as the text data that goes into language models. Multiple authors (Italie, Hillel) and the New York Times (Michael) are currently suing OpenAI for copyright infringement. The legal framework governing these situations is still evolving, with debates focusing on authorship, originality, and the extent of transformation required to mitigate copyright infringement risks.

As we move through 2024, the AI landscape is characterized by rapid innovation, increasingly advanced AI Agent product integrations, and growing societal and ethical considerations. While we can’t quite predict the future, we are excited to see what we will be writing about in January 2025.

For more information on BrandGuard or to sign up for a demo, please email contact@brandguard.ai.

WORKS CITED

Dolan, Leah. “Look of the Week: What Pope Francis’ Ai Puffer Coat Says about the Future of Fashion.” CNN, Cable News Network, 29 Mar. 2023,

www.cnn.com/style/article/pope-francis-puffer-coat-ai-fashion-lotw/index.html.  

Evans, Carter, and Analisa Novak. “Scammers Use AI to Mimic Voices of Loved Ones in Distress.” CBS News, CBS Interactive, 19 July 2023,

www.cbsnews.com/news/scammers-ai-mimic-voices-loved-ones-in-distress/.  

Graham, Becky. “70-Percent-of-Workers-Using-CHATGPT-at-Work-Are-Not-Telling-Their-Boss.” Fishbowl Insights, 2 Feb. 2023,

www.fishbowlapp.com/insights/70-percent-of-workers-using-chatgpt-at-work-are-not-telling-their-boss/.  

“ImageBind: Holistic AI Learning across Six Modalities.” AI at Meta,

www.ai.meta.com/blog/imagebind-six-modalities-binding-ai/.  

“Introducing Ego-Exo4d: A Foundational Dataset for Research on Video Learning and Multimodal Perception.” AI at Meta,

www.ai.meta.com/blog/ego-exo4d-video-learning-perception/.  

“Introducing Gpts.” Introducing GPTs,

www.openai.com/blog/introducing-gpts.  

Italie, Hillel. “‘Game of Thrones’ Creator and Other Authors Sue Chatgpt-Maker Openai for Copyright Infringement.” AP News, AP News, 21 Sept. 2023,

www.apnews.com/article/openai-lawsuit-authors-grisham-george-rr-martin-37f9073ab67ab25b7e6b2975b2a63bfe.  

Michael. “The Times Sues OpenAI and Microsoft over A.I. Use of Copyrighted Work.” The New York Times, The New York Times, 27 Dec. 2023,

www.nytimes.com/2023/12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html.

Patel, Dylan, and Daniel Nishball. “Google Gemini Eats the World – Gemini Smashes GPT-4 by 5x, the GPU-Poors.” SemiAnalysis, SemiAnalysis, 28 Aug. 2023,

www.semianalysis.com/p/google-gemini-eats-the-world-gemini.  

Peters, Jay. “Google Meet’s New AI Will Be Able to Go to Meetings for You.” The Verge, The Verge, 29 Aug. 2023,  

www.theverge.com/2023/8/29/23849056/google-meet-ai-duet-attend-for-me.  

“Tay (Chatbot).” Wikipedia, Wikimedia Foundation, 25 Dec. 2023,

www.en.wikipedia.org/wiki/Tay_(chatbot).  

Zoom. “Meet Zoom Ai Companion, Your New AI Assistant! Unlock the Benefits with a Paid Zoom Account.” Zoom, 19 Oct. 2023,

www.zoom.com/en/blog/zoom-ai-companion/.  

‍