State of AI丨2024 Recap & 2025 Prediction

Morning VC Brief is an expert daily curation of tech and venture news. Feel free to forward; the sign-up is here

Tech Shifts Spark Innovation and New Markets

Looking back at past tech cycles, we can see that each wave of platform-level opportunities has given rise to companies valued at tens or even hundreds of billions of dollars. The tech revolution brought about by GenAI represents a new wave of platform-level innovation opportunities that we are currently experiencing.

As a frontline observer and researcher of tech innovation and venture capital in Silicon Valley, let’s start with a quick recap of our recent observations and insights, and then look ahead to the key trends worth watching in 2025.

01. Enterprise AI Application

2024 Recap：

According to Menlo Ventures, AI-related enterprise spending reached $13.8 billion in 2024, a more than 6x increase from $2.3 billion in 2023. While spending on models still constitutes the majority, spending on applications has shown the fastest growth, rising from $600 million in 2023 to $4.6 billion this year. Clearly, GenAI is transitioning from an emerging technology to a practical tool. Among various applications, the most valuable use cases are AI coding, AI customer support, enterprise search, and AI meeting tools.

Horizontal, large-scale markets will inevitably face competition from foundational model providers and leading tech companies. Although vertical industries are more niche, they still offer substantial opportunities, as most vertical sectors evolve slowly under the dominance of traditional giants. Entrepreneurs who are experts in these verticals and deeply understand industry workflows, combined with the ever-evolving GenAI infrastructure, will have a chance to quickly capture users and build industry data moats.

2025 Prediction：

1/ Enterprise are expected to reallocate more of their traditional expenditure toward AI applications. As a result, we will likely see more GenAI startups reaching the $100 million ARR milestone.

2/ Vertical AI applications will remain a promising area for entrepreneurship. Pay Attention to traditional industries such as manufacturing, construction and home services. Sectors with low digital penetration might leapfrog the software phase and move directly into the AI phase, much like emerging markets bypassing cash and moving straight to mobile payments.

3/ The shift from Copilot to Agent in applications. We believe that in the coming years, we will witness a transition from AI empowering humans in their work, to humans supervising and guiding AI to complete tasks, and eventually to AI independently completing tasks.

02. AI Agent

2024 Recap：

When AutoGPT went viral in 2023, pioneering developers began envisioning scenarios where large models could serve as reasoning engines to help humans accomplish various complex tasks. However, it proved that, due to the limited capabilities of models at the time, the idea was more experimental than practical. As a result, our view back then was that the infrastructure for AI Agents wasn't ready, and the creation of real-world applications would require more time. It wasn't until this year, with the release of GPT-4 and o1-preview, that this experimental idea began to show potential for productization.

Sam Altman discussed the five stages in the evolution of AGI in a recent conversation with Gary Tan. He believes that we are currently in the second stage and are very close to entering the third stage. To summarize, a Level 3 Agent intelligence would be one capable of autonomously interacting with the environment, collecting information, continuously planning and executing multi-step, long-duration tasks, and collaborating with humans to solve complex problems when challenges arise.

2025 Prediction：

1/ AI Agent applications will transition from pilot products to real-world use cases, delivering tangible efficiency gains for users. Coding Agents and GTM (Go-to-Market) Agents are likely to be the two fastest-growing areas.

There are two key prerequisites for making this trend possible:

(1)The emergence of models with stronger reasoning and logical abilities. This could come in the form of the next version of o1(o1 full model just launched after this post) or open-source models that match o1 capabilities.

(2) The development of Agent infrastructure. A recent example is that /dev/Agents (Agent OS Startup) raised a seed round at $500M. Additionally, we’ve seen several startups focused on AgentOps in the latest Y Combinator F24 batch.

03. AI SWE(Software Engineer)

2024 Recap：

Whether it’s enterprise applications or the implementation of Agents, software development is the fastest and most promising direction. We’ve been focusing on this from the very beginning for three key reason:

1/ Software development inherently breaks down complex tasks into smaller, more manageable subtasks, which directly benefit from advancements in large model reasoning capabilities. Additionally, there is an abundance of data available for SWE Agent training.

2/ The results of software development and coding tasks are highly verifiable. The accuracy and reliability of outputs can be objectively and quantitatively measured.

3/ Engineers are naturally closer to the cutting edge of technology and are more willing to try and pay for tools that enhance efficiency.

As a result, we have seen the evolution from the earliest Github Copilot to the current explosion of AI tools, now represented by a logo map that no longer fits on a single screen.

2025 Prediction：

1/ SWE-bench Verified scores are likely to exceed 70% (currently 55%). We view 70% as a critical milestone because, prior to SWE-bench, HumanEval was the standard for assessing model coding capabilities. When the average performance on HumanEval surpassed 70%, it laid the technical foundation for commercially viable code completion applications like Copilot.

2/ In the longer term, AI SWE will expand the upper limit of the developer population. More professionals without coding experience will gain the ability to develop software and use this capability to create personalized productivity tools or entertainment applications. In practice, we are already seeing many companies building tools for non-developer professionals.

04. Videa Gen Model

2024 Recap：

Previously, we mainly discussed application-layer opportunities driven by the maturing infrastructure of large language models. Last year, when discussing opportunities at the model level, we believed that the early investment window for language or multimodal models like GPT-3 had already passed. Instead, we recommended focusing on areas that had yet to produce GPT-3 level models, such as video models.

Vision, as a critical medium for human interaction with the physical world, is also essential for AI's understanding of it. At the time, Runway was one of the few startups working on video generation. However, this year, with the breakthrough launch of Sora, we’ve seen a surge of video-generation startups. Companies like Luma AI and Genmo AI, which initially focused on image or 3D generation, have also shifted toward video generation this year.

2025 Prediction：

1/ Video generation models serve as engines for AI to understand the physical world, enabling users to experience real-time, interactive visual effects. A clear application scenario is gaming: using video generation models as game engines to generate responsive game content based on user inputs, making every experience unique. （Decart, backed by Sequoia Capital, has released an interactive world model called Oasis, which resembles Minecraft but is entirely driven by Video Gen models.）

2/ Beyond gaming, this will open up broader applications and imaginative possibilities in various domains. 

05. Embodied AI Foundation Model

2024 Recap：

In addition to video models, another model-level opportunity we are focusing on is the foundational models for Embodied AI. In 2023, we noticed that Google and several leading figures in academia were optimistic about the development prospects of Embodied AI, considering it as the "last mile" toward AGI. After witnessing the development trajectory of LLMs, capital and talent are likely to decisively invest in the next opportunity at the LLM level, and Embodied AI presents a field with both a high threshold and ceiling.

Looking back at 2024, we saw many leading scholars and experts joining the entrepreneurial wave of Embodied AI, with capital heavily supporting the field. This year, we also held related thematic discussions and compiled a map of 50 startups in the embodied intelligence space.

To date, the most prominent team in the field, Physical Intelligence, recently released their first-generation model, π0, which is designed for highly flexible, complex tasks requiring long-term planning. In their demo, the model was able to perform tasks such as doing laundry—removing clothes from the dryer and folding them. Physical Intelligence has made breakthroughs in both model architecture and training data, with π0 demonstrating some generalization capabilities, as it can perform tasks not seen in its pre-training data with only a short period of fine-tuning.

However, one of the paper's authors mentioned that, despite the progress made with π0, it still hasn’t reached the transformative level of GPT-2. Compared to GPT-2's training data, the robot's training dataset is still far from sufficient, lacking in quantity, diversity, and quality.

2025 Prediction：

1/ We may see breakthroughs in solutions for robot data. With the joint iteration of data and model design, the Scaling Law is expected to apply in Embodied AI.

2/ When foundational models are ready, we will also see more real-world applications for general-purpose robots. On the consumer side, we believe the home setting will progress the fastest, while on the business side, logistics, warehousing, and retail are seeing the quickest advancements.