What I'm watching in AI in 2025

It will be a radical year

Welcome to 2025! It will be a radical year.

Literally - it’s a year of perfect squares (√2025 = 45) - a phenomenon not seen since 1936, with the next not until 2116. Profoundly auspicious for the nerds.

And figuratively - it’s the year I believe we’ll see the secular wave of AI transformation accelerate, with major progress in LLMs and other domains as well as continued transformation in software. Radical indeed.

Here are the AI trends I’m watching closely this year:

1. Code continues to make waves as one of generative AI’s killer apps. Code autocomplete tools have taken the world by storm, and we’re beginning to see the possibility for fully-automated software engineering agents like Solver. But there’s much more to come, including functional-specific coding agents automating laborious tasks like incident response. More software may also mean more problems: we’ll need new kinds of infrastructure to manage security vulnerabilities and tech debt presented by higher-velocity software development.

2. “Selling work” scales across the enterprise. We’ll see more automation of enterprise text-based work with increasingly agentic systems. The go-to-market function will be a continued area of innovation: sales, customer support, and customer experience will all continue to be transformed with LLMs. Enterprises will also begin establishing comfort with AI systems that automate more core back office functions like tax, accounting, and compliance. Multi-agent collaboration and coordination systems will be key as these functional agents scale.

3. Next-generation systems of record. The cloud transition led to “replatforming” of core business systems, producing generational companies like Salesforce and Workday. Today, there’s a unique opportunity to challenge these platforms with AI-native systems of record that are built around LLMs, build unique data advantages, and deliver differentiated customer value. This is one of the hardest roads in startups, but the prize is huge. Bring on a new wave of CRMs, EMS’, and more.

4. Voice - this time, for real. Did you live through the last breathless voice assistant wave in the late 2010s? Yeah, me too, and my Alexa still doesn’t work. This time, it’s for real, with real-time speech-to-speech models now a reality. In 2025, we’ll see voice systems interview candidates for jobs, manage customer service interactions, book reservations, and more.

5. Embodied AI brings intelligence into the real-world. LLMs have given us extraordinary verbal intelligence, but AI systems are still woefully nascent in their ability to understand and interact with the physical world. In 2025, new approaches to developing “embodied AI,” including in world models and robotic systems, will gain steam.

6. Innovation in each node in the AI value chain. Amid concerns about plateauing AI progress, we’ll see pressure on each node in the AI value chain of compute, algorithms, and data. Next-gen alternatives to the Transformer will begin to take shape, with leaders emerging across the competing camps of SSMs, liquid NNs, geometric deep learning, and more. As we hit text data ceilings, data curation, annotation, and access will become increasingly critical and strategic, as will access to and performance from advanced computing infrastructure.

7. AI transforms science. Deep learning won two Nobels in 2024 - an incredible feat. We’ll see continued impact of AI in science in 2025 in fields like biology, chemistry, and materials design. Importantly, we’ll also see progress in automating the scientific method itself. LLMs are now performing lit reviews, running virtual labs, and developing research ideas, getting us steps closer to “AI scientists.”

8. Professionalized creative tools. Since 2022, we’ve seen incredible progress in AI that can create images and video, but few have been enterprise-grade. This year, we’ll see the advent of professionalized creative platforms that go beyond controllable content generation and “close the loop” with cross-platform content distribution and optimization.

9. Inference cost continues to drop. Yes, o3 is very expensive, reportedly costing >$1k/query. In parallel, though, previous model generations (which are still very performant in terms of benchmarks like MMLU) have dropped in cost by orders of magnitude since GPT-3 became publicly accessible in 2021. We’ll see costs continue to decline in 2025 on the back of advancements in post-training, model quantization, and continued progress in open-source.

10. Small companies will have big impact. Cleverly applying AI to automate core business functions will enable more small companies to hit big numbers fast. In 2025, we’ll see the first company sub-100 employees hit $100m in annual revenue.

What do you think will be the most radical AI developments in 2025? Let me know at molly@radical.vc.