As we step into 2025, I wanted to share some of the emerging themes in AI and technology that I’m watching closely. From new scaling laws to the rise of agents, this year promises to bring advancements that continues to reshape technology. Shameless plug: read more of my musings on my Substack. Okay, with that out of the way – let’s dive in.
I. Rapid Reasoning Advancements from Test-Time Compute
So far, most advances in models have been driven by scaling training compute—training larger models on more compute on larger datasets. But just as there are signs that we are hitting a wall there, we now have a new vector for scaling — test-time or inference-time compute, where models use additional compute during inference.
OpenAI’s o-1 showed that by leveraging reasoning techniques like CoT reasoning and tree search, AI can dynamically "think through" complex problems at inference time and improve performance.
OpenAI’s o-3, while not released yet, announced only 3 months after o1, showed tremendous gains even over o-1.
I expect continued progress in 2025 with rapid reasoning advancements in models (as measured by ARC-AGI eval, SWE-bench, and other evals) particularly as more research is done in this area and forms of synthetic data (reasoning traces) are used by the labs.
Sam Altman recently wrote the below, likely hinting at test-time compute and current techniques getting to AGI (with some additional tricks of course).
We are now confident we know how to build AGI as we have traditionally understood it… We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word. We love our current products, but we are here for the glorious future. With superintelligence, we can do anything else. Superintelligent tools could massively accelerate scientific discovery and innovation well beyond what we are capable of doing on our own, and in turn massively increase abundance and prosperity.
II. Multimodal AI Applications Emerge
2024 saw the rise of multimodal AI with improvements and continued advancements in audio models (Deepgram, Eleven Labs, Cartesia, Hume), Video Models (Sora, Veo2) and image models (Midjourney, Flux, etc).
However, perhaps with the exception of voice, which is being productized in various support and sales applications, the rest have largely been used for very simple generative purposes rather than truly changing core workflows and processes so far.
I believe 2025 will be the year that changes and we start seeing many more interesting multimodal AI applications. A lot of the foundational pieces are already in place today, and I expect founders to put them together in interesting ways to build vertically oriented apps that understand and process images, video and text and then carry out interesting workflows that (often) result in end outputs that are also multimodal.
This will also result in new sets of interfaces, rather than simple chat-based ones. OpenAI’s advanced voice mode is one example and Adobe’s Firefly integration into its creative suite is another example.
As multimodal capabilities improve and they begin to get productized, we’ll likely see creative, productivity, and technical workflows reimagined entirely.
III. AI Agents Start Working, Literally
Us VCs have been talking about service-as-a-software and agentic AI as representing the new thing after SaaS for over a year now, but an honest reflection on the current state is that we’re not quite there yet. Many of the applications today still largely take the form of copilots rather than true AI agents with a few exceptions in sales (11x) and support (Decagon, Sierra), often able to still do only a fraction of tasks as junior employees.
There are many reasons why they don’t work yet, but many are being solved in real-time and over the coming years: better multimodality to interpret inputs and create outputs across wider sets of tasks, better reasoning to be able to perform tasks more reliably, and better memory, orchestration and other tooling that’s necessary to have agents representing humans in organisations.
I believe 2025 will be the year of agents, when we finally start to see AI systems that act autonomously, making decisions and taking actions on behalf of users and being productive akin to human colleagues. It will be the year when we get the first glimpse of agents that really start to work, quite literally, for companies as AI colleagues.
Sam Altman recently wrote something along these lines:
We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies. We continue to believe that iteratively putting great tools in the hands of people leads to great, broadly-distributed outcomes.
Jensen Huang has shared a similar point of view on the BG2 podcast:
I’m hoping that someday NVIDIA will be a 50,000 person company with a 100M AI assistants and they’re in every single group. We'll have a whole directory of AIs that are just generally good at doing things we'll also have our inbox is going to full of directories of AIs that we work with that we know are really good specialized at our skill and so AIs will recruit other AIs to solve problems. AIs will be in Slack Channels with each other and with humans and so so we'll just be one large employee base.
There’s a lot of companies building agents across various functional and vertical roles and there’s a lot for them to figure out: from how to sell into labour budgets, how to price properly on outcomes or usage given the concept of seats may not make sense, and how to get agents to work reliably enough that discrete sets of tasks performed by humans can get delegated to agents, while still giving humans interfaces to verify and control them.
IV. Increased Consolidation in AI
In 2024, we saw the rise of the “license and hire” approach used to acquire companies, particularly at the model layer, as talent consolidated towards a few large players.
I expect consolidation to continue in AI and perhaps expand from the model layer to some of the surrounding tooling/infra layers, as leading companies continue to try to bolster talent and enter new and important critical areas adjacent to them, resulting in acquisitions of various sizes across vendors working on small model, multimodal models, inference optimisation on edge and on-devices, the RAG stack, and others.
With a new regime in place, the “license and hire” approach may not be needed, and they make take the form of regular M&A.
V. IPO Window Drifts Open
We saw a few companies go public this year, such as Reddit, ServiceTitan and Instacart, but in general, the IPO window was largely constrained to profitable and truly at scale companies. I think this year we start to see the window opening up some more with many more companies going public than the past few years.
While, I don't think we see the likes of Databricks and Stripe go out this year, we could see some interesting AI companies such as Coreweave and Cerebras and at scale companies such as Klarna and Canva go public, which may set the tune for other companies to follow.
I would also love to see some “smaller” (i.e. <$350M ARR) companies go public, but I don’t know whether we’ll quite see that yet this year, but I do expect IPO volumes to be higher than ‘23 and ‘24.
Conclusion
As these themes play out, they will shape not only the technology landscape but also how businesses and individuals interact with AI. 2025 is set to be a transformative year, and I’m excited to see how these trends evolve.
What are you keeping an eye on this year? As always, I’d love to hear your thoughts, and if you’re building in the first few areas discussed above, feel free to email me at tanay@wing.vc.
