Key Moments
The State of AI Startups in 2024 [LS Live @ NeurIPS]
Latent Space PodcastKey Moments
AI startups in 2024: More competition, open source gains, cheaper models, new modalities, and evolving investment.
Key Insights
The foundation model landscape is more competitive, with OpenAI no longer having a dominant lead; open source models are increasingly viable.
The cost of AI intelligence has dramatically decreased, making advanced capabilities more accessible and affordable for startups.
New modalities like video generation (Sora), advanced voice synthesis, and AI in biology are emerging as significant areas for innovation.
Despite concerns about an AI bubble, funding for startups (excluding major foundation model labs) appears more rational and sensible.
Successful AI startups are focusing on service automation, enhanced search, democratizing creative/technical skills, and enabling infrastructure.
The "GPT wrapper" narrative is fading, as startups can capture value by effectively leveraging existing models and focusing on product innovation.
AI presents opportunities in traditionally difficult markets for venture capital, such as legal, healthcare, and education, by offering significant cost reductions.
Incumbents face an innovator's dilemma with AI; their traditional advantages like distribution may be challenged by new AI-driven UIs and business models.
EVOLVING FOUNDATION MODEL LANDSCAPE
In 2024, the foundation model ecosystem has become significantly more competitive. Unlike late 2023, where OpenAI models dominated, current evaluations show a closer race among various proprietary and open-source models. This increased competition is also reflected in enterprise spending, with OpenAI's share decreasing from approximately 90% to around 60%, indicating a willingness to explore and switch between different AI providers. Open-source models, in particular, have shown remarkable advancement, becoming competitive in crucial areas like math, instruction following, and robustness, challenging the notion that only large, proprietary models can deliver top-tier performance.
DECREASE IN INTELLIGENCE COSTS AND NEW MODALITIES
The cost of AI intelligence has plummeted, with flagship model API costs seeing an 80-85% reduction over the last year and a half. This dramatic price decrease makes advanced AI capabilities economically feasible for a much wider range of applications and startups. Concurrently, new modalities are rapidly maturing. We see AI making strides in biology, with models outperforming established tools like AlphaFold. Advancements in low-latency voice enable new interaction experiences, and nascent use cases are emerging in execution, with companies like Dev.ai showcasing autonomous agent capabilities. Video generation, exemplified by Sora, also represents a significant leap forward.
THE MYTH OF THE "GPT WRAPPER" AND STARTUP VALUE CAPTURE
The prevailing narrative that startups could only be "GPT wrappers" with little intrinsic value is being debunked. The reality is that a diverse ecosystem of innovation is flourishing due to multiple competitive model options, price competition, and the rise of open-source alternatives. Startups can now leverage these models effectively to create unique customer value. Test-time compute scaling allows companies to better match user value with compute spend, shifting the financial burden from upfront pre-training capex to customer-paid inference. This paradigm shift empowers startups to build valuable products without needing to develop their own foundational models.
EMERGING STARTUP PATTERNS AND SUCCESSFUL STRATEGIES
Several patterns are emerging among successful AI startups. Service automation is a key area, addressing tasks too expensive or complex for human labor, from scribing to advanced customer support. Enhanced search is another, with tools like Perplexity and Glean redefining how users find and interact with information. The democratization of creative and technical skills continues to drive growth, enabling a broader user base to generate content and code. Investing in enabling infrastructure, particularly compute and specialized data, remains critical as the demand for data with reasoning traces and expert knowledge grows.
OPPORTUNITIES IN UNDERSERVED MARKETS AND INCUMBENT CHALLENGES
AI is unlocking substantial opportunities in markets traditionally considered difficult for venture capital, such as legal, healthcare, and education. By offering capabilities that are orders of magnitude cheaper or entirely novel, startups can reshape buying patterns and market structures. Incumbents, though possessing distribution, face an innovator's dilemma. Their business models, often based on per-seat licenses, can be threatened by AI solutions that automate workflows and potentially reduce the need for human seats. Furthermore, incumbents often lack the specific, granular data (like reasoning traces) that startups require to build higher-quality, specialized AI products.
THE NEW PARADIGM: ADAPTING TO SOFTWARE 3.0
The current era is characterized as "Software 3.0," representing a full-stack rethinking that favors startups. The speed of change and the inherent difficulty for large incumbents to pivot quickly create advantages for agile startups. Market opportunities are expanding beyond simple software replacements to entirely new paradigms. Business models are evolving, with a growing focus on outcome-based pricing and managing gross margins effectively, especially concerning compute costs. Developing creative products, understanding the non-deterministic nature of AI outputs, and re-engaging with infrastructure management are crucial for success in this dynamic landscape.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
In 2024, the race for foundation models has become much closer, with OpenAI no longer holding a dominant lead. Open-source models are increasingly competitive, and the gap between state-of-the-art and smaller models is shrinking.
Topics
Mentioned in this video
AI model that launched image upload functionality in October 2023, previously only handling text input/output.
A company demonstrating the effectiveness of new search paradigms using AI.
Open-source language model that is highly competitive, ranking among the top evaluated models.
A smaller language model that was state-of-the-art in late 2023, now surpassed by newer models.
A company demonstrating the effectiveness of new search paradigms using AI.
AI agent that offers code execution capabilities for a monthly fee.
Customer Relationship Management software, used as an example of how new UX and code generation can challenge incumbents.
A newer, small form-factor model that significantly outperforms previous small models on evaluations.
An open-source biology model developed by Chai Discovery that surpasses AlphaFold 3.
Google's AI model, announced shortly before NeurIPS in 2023.
A previous benchmark for biology models, now outperformed by Chai 1.
A pioneering company in AI image generation, illustrating the demand for creative tools.
Previous AI model from Google, noted by the speaker as being forgotten.
Platform for users to rate and compare language model generations, used to show the narrowing gap between AI models.
AI model for video generation, which the speaker had early access to.
A consumer company noted for its large user base and engagement, representative of successful platforms.
A successful company in text-based AI applications.
A consumer company noted for its large user base and engagement, representative of successful platforms.
Company whose AI model spend has increased, showing diversification away from OpenAI.
Venture capital firm where Sarah Guo previously worked.
Leading AI research lab, whose models were previously dominant but now face increased competition.
A company known for its voice models, highlighting the importance of low-latency voice applications.
A portfolio company that offers AI-powered lip-syncing and dubbing for speeches.
A successful company in text-based AI applications.
Collaboration on Chai 1, an open-source biology model that outperforms AlphaFold 3.
More from Latent Space
View all 134 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free