As 2023 comes to an end, it’s a good time to look back at the significant advancements and ethical discussions around artificial intelligence this year. The introduction of chatbots like Bing Chat and Google Bard demonstrated impressive natural language capabilities, while generative AI models like DALL-E 3 and MidJourney V6 amazed us with their creative image generation.
However, there were also concerns about the potential dangers of AI. The EU’s AI Act aimed to restrict certain uses of the technology, and the Biden Administration issued guidelines on its development.
With rapid innovation expected to continue, many are curious about the future of AI. To get some insights, we asked leading venture capitalists who invest in AI startups for their boldest predictions for 2024. Will we see another “AI winter” as the hype meets reality, or will new breakthroughs speed up adoption across industries? How will policymakers and the public react?
VCs from top firms like Bain Capital Ventures, Sapphire Ventures, Madrona, General Catalyst, and more shared their views on topics ranging from the future of generative AI to GPU shortages, AI regulation, climate change applications, and more. While opinions differ on risks and timelines, most agree that 2024 will be a pivotal year for artificial intelligence. Here are their boldest predictions and insights on what’s to come in AI.
The rise and fall of generative AI startups
Many generative AI companies will struggle. If you weren’t one of the startups that raised significant funding this year, the future is uncertain. Many generative AI companies will compete with each other, and those built on top of OpenAI will face platform risks, leading to a decline in fundraising. At Day One, we’ve stopped considering these deals altogether.
I’m excited about AI in biotech, genome, climate, and industrial applications. AI will save lives by helping scientists and researchers develop new treatments and diagnostics based on human genomic data. In the psychedelics industry, our portfolio company Mindstate is using AI to create new “states of mind” based on their extensive data set of trip reports to help treat treatment-resistant PTSD. AI in fertility, reproductive health, and longevity will completely change how we view human lifespan and reproduction. In climate, companies are using AI to protect our ecosystems, like how Vibrant Planet is using AI/ML to prevent catastrophic wildfires globally.
AI will also be able to understand what’s happening in people’s minds and project images of their thoughts. It’s fascinating, and I’m curious to see how it will unlock knowledge about human consciousness and unconsciousness.
Convergence of data modalities in multimodal models
In 2024, the convergence of data modalities—text, images, audio—into multimodal models will redefine AI capabilities. Startups using these models will enable better decision-making and improved user experiences, including personalization. We will see new and transformative use cases across industries like manufacturing, e-commerce, and healthcare. On the infrastructure side, AI workloads will become more demanding, and I expect innovation around multimodal databases. While not every use case will require multimodal models, first-generation LLM startups in many sectors will face new competition and intense pressure to innovate and build defensibility.
We’re going to see multi-modal retrieval and multi-modal inference take center stage in AI products in 2024. AI products today are mostly textual, but users prefer more expressive software that meets them in every modality, from voice to video to audio to code and more. If we can get these architectures to work at scale, we could unlock software that provides much more accurate and human results, from drawing the answer to making calls in your tone and voice so you can attend fewer meetings to collaborating with other AI and human entities to achieve the right outcome. To support this, we expect ETL providers like Unstructured to diversify to include new data sources, more startups using the Segment Anything architecture from Meta, and startups like Contextual becoming full-scale solutions for multi-modal retrieval.
We continue to see AI expand into more use cases, especially in large and outdated industries. In healthcare, we are excited about the potential of using computer vision to detect cancer, machine learning to improve diagnoses, and generative AI to reduce paperwork. Considering we spend $4.3 trillion on healthcare in the US, which is almost double the average of OECD countries, and that almost a third is just administrative costs, there’s a lot of room for improvement. While Chat has been a key buzzword of 2023, companies need to think beyond Chat in 2024. Multimodal AI across input, training, model creation, and output are key areas of innovation.
Multimodal models will make it much easier to create compelling interactions with AI agents, and the quality of the AI will make it nearly impossible for humans to distinguish between a computer and a human in certain use cases. We can already see this in places like Character.AI and Instagram, and we expect this to become more common in the workplace in areas like training, customer support, and marketing/sales. You will be building a relationship with a machine sooner than you think.
Democratization of AI through open source
We predict that more open-source models will be released in 2024, and we expect large tech companies to be major contributors. Companies like Tesla, Uber, and Lyft (historically big contributors to open-source projects), and even Snowflake, could be involved. We wouldn’t be surprised if some of these models spun out into companies and received large funding rounds.
I see multimodal becoming the standard for any large model provider by the second half of 2024. The major model builders who have historically maintained proprietary models will begin open-sourcing select IP while releasing new benchmarks to reset the conversation around benchmarking for AGI.
There is a race-to-the-bottom in generative AI pricing between OpenAI, Mistral, Google, and others serving open-source models. Most are incurring losses using the existing hardware infrastructure and hoping to make it up on volume. The imperative for generative AI companies is clear: find pathways to profitability and scalability. Based on this need, I believe VC investments will go toward developing efficient models, leveraging new AI compute hardware, and providing value-added services like industry-specific model fine-tuning and compliance.
GPU shortage: A persistent problem or a temporary setback?
2024 will be the year of real-time diffusion applications. In 2023, we saw major theoretical improvements in diffusion model inference speeds, such as the original consistency models paper by Song et al., and more recently, LCMs. We’re already starting to see projects that use these ideas, such as Dan Wood’s Art Spew, Modal’s turbo.art, and fal.ai’s 30fps face swap. In 2024, we’ll see more real-time image, audio, and video generation diffusion applications.
The GPU shortage continues to impact the startup ecosystem, making it difficult for new companies to bring their products to market. There are two ways to solve this problem: either new compute options emerge that break free of the Nvidia monopoly on AI, or new models/architectures emerge that are more efficient with compute resources. I expect to see significant funding go towards novel model architectures that run in linear, not quadratic time, such as Mamba from Cartesia AI, in addition to platforms built around diffusion models and liquid neural nets as a faster, cheaper, more performant alternative to transformer-based LLMs.
The GPU shortage is not necessarily as acute or definitive as everyone might think. The bigger issue is the underutilization of existing infrastructure, which I believe will persist in 2024 alongside continued supply chain constraints. Fixing lower-level software for AI will be key to resolving the perceived GPU “shortage” and more real utilization issues. Until then, the only short-term solution is more computation. That said, I predict GPU constraints to persist in 2024, with companies like NVIDIA experiencing continued backlogs, while competitors like AMD and Intel will each gain 1-2% of GPU market share due to demand-side desperation.
A contrarian view is that we will eventually not have a GPU shortage. The market will converge to a small number of buyers and suppliers. Nvidia and others will scale up to meet forecasted demand, and companies like Microsoft, Google, Amazon, Facebook, and many sovereign nations will still be large buyers. The rest of us will rent GPUs from cloud providers, but there will be plenty of leasable capacity. The rich will get richer, but the quality of life will improve for the “GPU-poor.”
According to the Taiwan Semiconductor Manufacturing Company (TSMC) Chairman, “it’s not the shortage of AI chips. It’s our COWOS [advanced chip packaging] capacity,” and advanced memory and packaging capacity will ramp up. However, the long-term sustainability of AI in production won’t rely on general-purpose GPUs like Nvidia H100 and AMD MI300X. Investments will shift to focus on hardware specialized for inference, rather than training. NPU innovations like d-Matrix and EnCharge AI, utilizing near/in-memory computing, are emerging as cost-effective and environmentally friendly solutions, suitable for deployment both on local AI PCs and within data centers.
Apple and Google: Sleeping giants?
We believe 2024 will bring some big releases from Apple, perhaps even their own GPT. There have been reports of an Apple LLM known internally as Ajax GPT. While the model was created for internal use, next year we could see Apple making Ajax (or related models) more public, or incorporating generative AI capabilities across its apps (e.g., XCode, Apple Music) and devices (e.g., Mac, Vision Pro). And while this was more machine learning than AI, just last week Apple released MLX – an “efficient machine learning framework specifically designed for Apple silicon.” Releases from Apple could have a massive influence over not only existing models but also how the US approaches regulation, given Apple’s prominent role as a consumer device manufacturer.
If 2023 was the year of OpenAI and Microsoft dominated the headlines, next year we will all be talking about Google. Google’s substantial investment in Gemini and unrivaled data and compute resources will offer developers GPT-4+ capabilities in all shapes and sizes, pushing the frontier for all foundation model providers. Don’t rule them out just yet.
Preparing for the long-term AI shift
Everyone who jumped into AI this past year will exit 2024 knowing what chiplets are. As we continue to grapple with the limits of Moore’s Law, we will also see new architectural paradigms come into play—not only with new core semiconductor architectures like chiplets but also with advanced packaging and interconnect.
Edge-to-cloud or “hybrid AI,” integrating both cloud and edge devices like smartphones, laptops, vehicles, and IoT devices, offers benefits in performance, personalization, privacy, and security. As generative AI models shrink and on-device capabilities improve, this approach will become increasingly feasible and essential for scaling AI to meet global enterprise and consumer needs in the long term.