AI predictions for 2024: What prime VCs suppose

As 2023 attracts to a detailed, it’s a time of reflection on the monumental advances — and moral debates — surrounding synthetic intelligence this previous yr. The launch of chatbots like Bing Chat and Google Bard showcased spectacular pure language talents, whereas generative AI fashions like DALL-E 3 and MidJourney V6 shocked with their artistic picture era.

Nonetheless, considerations have been additionally raised about AI’s potential harms. The EU’s landmark AI Act sought to restrict sure makes use of of the expertise, and the Biden Administration issued pointers on its growth.

With fast innovation anticipated to proceed, many surprise: What’s subsequent for AI? To search out out, we surveyed main enterprise capitalists investing in synthetic intelligence startups for his or her boldest predictions on what 2024 might carry. Will we see one other “AI winter” as hype meets actuality? Or will new breakthroughs speed up adoption throughout industries? How will policymakers and the general public reply?

VCs from prime corporations together with Bain Capital Ventures (BCV), Sapphire Ventures, Madrona, Basic Catalyst and extra supplied their outlook on matters starting from the way forward for generative AI to GPU shortages, AI regulation, local weather change purposes, and extra. Whereas views differ on dangers and timelines, most agree: 2024 guarantees to be a defining yr for synthetic intelligence. Learn on for his or her boldest predictions and insights on what’s to come back in AI.

The rise and fall of generative AI startups

“Many generative AI firms will die. For those who weren’t one of many startups that raised monster rounds this yr, the long run can be unsure for you. Many [generative AI] firms will compete with one another, startups constructed on prime of OpenAI will expertise platform danger, and fundraising into these firms will dry out. At Day One, we’ve stopped these offers altogether. 

I’m enthusiastic about AI in biotech, genome, local weather and industrial purposes. AI will save lives by serving to scientists and researchers develop new remedies and diagnostics primarily based on human genomic information. Within the psychedelics {industry}, our portfolio firm Mindstate is utilizing AI to create new “states of thoughts” primarily based on their largest complete information set of journey experiences to assist deal with treatment-resistant PTSD. AI in fertility, reproductive and longevity will utterly alter the paradigm of people’ lifespan and the way we now have kids. In local weather, firms are utilizing AI to defend our ecosystems, like how Vibrant Planet is utilizing AI/ML to stop catastrophic wildfires which can be occurring globally.

AI may also be capable of learn what’s occurring in folks’s minds and venture photographs of their ideas—it’s fairly fascinating and I’m curious to see the way it’ll unlock information about human consciousness and unconsciousness.

Masha Bucher, Founder and Basic Accomplice at Day One Ventures

Convergence of information modalities in multimodal fashions

“In 2024, the convergence of information modalities—textual content, photographs, audio—into multimodal fashions will redefine AI capabilities. Startups leveraging these fashions will allow higher decision-making and improved consumer experiences, together with personalization. We are going to see novel and transformative use instances throughout industries like manufacturing, e-commerce and healthcare. On the infrastructure aspect, AI workloads will change into extra demanding, and I count on to see innovation round multimodal databases. Whereas not each use case would require multimodal fashions, first era LLM startups in lots of sectors will face new competitors and the stress to proceed to innovate and construct defensibility can be intense.”

Cathy Gao, Accomplice at Sapphire Ventures

“We’re going to see multi-modal retrieval & multi-modal inference take middle stage in AI merchandise in 2024. AI merchandise immediately are principally textual. However customers desire extra expressive software program that meets them in each modality, from voice to video to audio to code and extra. If we will get these architectures to work at scale, we may unlock software program that gives far more correct and human outcomes, from drawing the reply to creating calls in your tone and voice so you possibly can attend much less conferences to converging on the fitting end result through collaboration with different AI and human entities. To energy this, we count on ETL suppliers like Unstructured to diversify to incorporate new information sources, extra startups making use of the Section Something structure from Meta, and startups like Contextual changing into full-scale options for multi-modal retrieval.” 

Rak Garg, Vice President at Bain Capital Ventures

“We proceed seeing AI blossom into extra use instances, particularly in industries which can be each giant and rusty. In healthcare we’re enthusiastic about the potential of utilizing pc imaginative and prescient to detect most cancers, for utilizing machine studying to assist with higher diagnoses, and of utilizing generative AI to cut back paperwork. Think about we spend $4.3 trillion in healthcare within the US, which as a p.c of our GDP is nearly double of the typical of OECD, and that just about a 3rd is simply in administrative prices… And whereas Chat has been a key buzzword of 2023, firms must suppose past Chat in 2024.  Multimodal AI throughout enter, coaching, mannequin creation and output are key areas of innovation.”

Amit Garg, Managing Accomplice at Tau Ventures and Sanjay Rao, Managing Accomplice at Tau Ventures

“Multimodal fashions will make it a lot simpler to create compelling interactions with AI brokers and the standard of the AI will make it practically not possible for people to discern the distinction between a pc and a human in sure use instances. We are able to see this already in locations like Character.AI and Instagram and count on this to take maintain within the office in areas like coaching, buyer help, and advertising / gross sales. You may be constructing a relationship with a machine prior to you understand.”

Jess Leao, Accomplice at Decibel VC

Democratization of AI via open supply

“We predict that extra open-source fashions can be launched in 2024, and we’re particularly giant tech firms to be one of many main contributors. Some examples of this might embody firms like Tesla, Uber and Lyft (traditionally each huge contributors to open-source tasks), and even Snowflake. We might not be stunned if a few of these fashions spun out into firms and acquired giant funding rounds.”

Vivek Ramaswami, Accomplice at Madrona and Sabrina Wu, Investor at Madrona

“I see multimodal changing into the de facto commonplace for any giant mannequin supplier by H2’24. The marquee mannequin builders who’ve traditionally maintained proprietary fashions will start open-sourcing choose IP whereas releasing new benchmarks to reset the dialog round benchmarking for AGI.”

Chris Kauffman, Accomplice at Basic Catalyst

“There’s a race-to-the-bottom in generative AI pricing between OpenAI, Mistral, Google, and others serving open-source fashions. Most are incurring losses utilizing the prevailing {hardware} infrastructure (evidenced in per-token enter/output prices) and hoping to make it up on quantity. The crucial for generative AI firms is evident: discover pathways to profitability and scalability. Primarily based on this want, I consider VC investments will go towards creating environment friendly fashions, leveraging new AI compute {hardware}, and offering value-added companies like industry-specific mannequin fine-tuning and compliance.”

Jimmy Kan, Accomplice at Anzu Companions

GPU scarcity: A persistent downside or a brief setback?

“2024 would be the yr of actual time diffusion purposes. In 2023, we noticed some main theoretical enhancements in diffusion mannequin inference speeds — reminiscent of the unique consistency fashions paper by Track et al, and, extra just lately, LCMs. (Additionally, Adversarial Diffusion Distillation.) We’re already beginning to see tasks that use these concepts, reminiscent of Dan Wooden’s Artwork Spew (77 512×512 photographs per second, on a single 4090), Modal’s turbo.artwork (primarily based on SDXL Turbo), and fal.ai’s 30fps face swap. In 2024, we’ll see extra realtime picture, audio, and video era diffusion purposes.” 

Slater Stich, Accomplice at Bain Capital Ventures

“The GPU scarcity continues to ravage the startup ecosystem, making it onerous for brand new firms to carry their merchandise to market. There are two methods to resolve this downside; both new compute choices emerge that break freed from the Nvidia monopoly on AI, or new fashions/architectures emerge which can be extra environment friendly with compute assets. I count on to see giant quantities of funding go in direction of novel mannequin architectures that run in linear, not quadratic time, reminiscent of Mamba from Cartesia AI, along with platforms constructed round diffusion fashions and liquid neural nets as a quicker, cheaper, extra performant various to transformer-based LLMs.” 

Rak Garg, Vice President at Bain Capital Ventures

“For starters, the GPU scarcity shouldn’t be essentially as acute or definitive as everybody would possibly suppose. The larger situation is the famine of utilization for present infrastructure therein, which I consider will persist in 2024 alongside continued provide chain constraints. Fixing lower-level software program for AI would be the key to resolving the illusory GPU ‘scarcity’”’ and far more actual utilization points. Till then, the one short-term answer we now have is, merely, extra computation. That stated, I predict GPU constraints to persist in 2024, with the NVIDIAs of the world experiencing continued backlogs, whereas rivals (specifically AMD and Intel) will every achieve 1-2% of GPU market share on account of demand-side desperation.”

Chris Kauffman, Accomplice at Basic Catalyst

“A contrarian take is we’ll finally *not* have a GPU scarcity. The market will converge to a small handful of consumers and suppliers. Nvidia and others will scale as much as meet forecasted demand, and Microsoft, Google, Amazon, Fb, and lots of sovereign nations will nonetheless be giant consumers. The remainder of us will lease them from the cloud suppliers however there can be loads of leasable capability to go round. The wealthy will get richer, however high quality of life will enhance for the ‘GPU-poor’.”

Jess Leao, Accomplice at Decibel VC

“In accordance with the Taiwan Semiconductor Manufacturing Firm (TSMC) Chairman, ‘it’s not the scarcity of AI chips. It’s…our COWOS [advanced chip packaging] capability,’ and superior reminiscence and packaging capability will ramp; nevertheless, the long-term sustainability of AI in manufacturing gained’t depend on general-purpose GPUs like Nvidia H100 and AMD MI300X. Investments will shift to give attention to {hardware} specialised for inference, moderately than coaching. NPU improvements like d-Matrix and EnCharge AI, using close to/in-memory computing, are rising as cost-effective and environmentally pleasant options, appropriate for deployment each on native AI PCs and inside information facilities.”

Jimmy Kan, Accomplice at Anzu Companions

Apple and Google: Sleeping giants?

“We consider 2024 will carry some huge releases from Apple, maybe even their very own GPT.  There have been experiences of an Apple LLM identified internally as Ajax GPT. Whereas the mannequin was created for inside use, subsequent yr we may see Apple making Ajax (or associated fashions) extra public, or incorporating generative AI capabilities throughout its apps (e.g. XCode, Apple Music) and units (e.g. Mac, Imaginative and prescient Professional). And whereas this was extra machine studying than AI, simply final week Apple launched MLX – an “environment friendly machine studying framework particularly designed for Apple silicon.” Releases from Apple may have large affect over not solely present fashions however how the US approaches regulation, given Apple’s outstanding function as a shopper gadget producer.”

Vivek Ramaswami, Accomplice at Madrona and Sabrina Wu, Investor at Madrona

“If 2023 was the yr of Open AI and Microsoft owned the airwaves, subsequent yr we’ll all be speaking about Google. Google’s substantial funding in Gemini and unmatched information and compute assets will provide builders GPT-4+ capabilities in all styles and sizes, pushing the frontier for all basis mannequin suppliers. Don’t rule them out simply but.”

Jess Leao, Accomplice at Decibel VC