What I’m excited for in 2025: more models, more use cases, more action
Last year, I wrote about the rise of multimodal AI and why I was excited for modalities beyond text — including image, speech, and video. With various foundation model companies releasing new multimodal models and hundreds of infrastructure and application layer companies emerging across voice and video, my excitement has proven to be warranted.
This year, beyond the multimodal revolution, I’m most excited about two developments — higher accuracy models unlocking new use cases, and agentic solutions becoming more truly agentic — moving beyond singular steps and into fluid sequences that can actually take actions on behalf of the user.
Higher accuracy models unlocking new use cases
Given continued improvements in accuracy across models (e.g. OpenAI’s o1, Anthropic’s Claude 3.5 models, Google’s Gemini 2.0*), we’re entering a phase where applications can address problems that require greater precision, accuracy, and reliability.
Today, many early AI applications operate at accuracy levels in the 50–80% range (compared to the target/desired human quality). In certain contexts, this has been sufficient. For example, a call center might use AI to handle 60% of all calls and route the remaining calls to an existing low-cost human call center. For these use cases where the quality or accuracy gap can be filled by relatively low-cost human labor (either on the provider side or on the customer side), the ROI remains positive. Yet in other use cases, the effort required to bring the accuracy up to the necessary level renders the value negligible or simply creates too much mental burden for the customer to want to change their way of doing things.
This is changing. As models continue to improve, applications will begin to address use cases that were previously uncompelling or unviable. Often times, though not always, these use cases are more analytical than creative in nature and/or in more regulated industries. I’m excited for applications to address these use cases that benefit from and/or require higher accuracy levels, including but not limited to complex financial analysis, compliance monitoring, pharmaceutical research, supply chain optimization, content moderation, etc.
Agentic solutions will become more truly agentic
(A.K.A. history rhymes — from point solutions to full stack solutions)
Over the last year, there’s been many uses of the word “agentic”, and the birth of a dozen “AI agent” startups for every imaginable job function. So what exactly is an agentic solution? I liked Andrew Ng’s explanation — unlike non-agentic workflows (zero-shot), agentic AI involves iterative processes of reflection (the model critiquing and improving itself), tool use (API calls), planning, and multi-agent collaboration**.
2025 isn’t the birth of agentic AI, so what’s different now? Over the last few years and especially in 2024, we’ve seen significant developments across AI infrastructure — across the data, model, and agent layers. Within data infrastructure, we’ve seen more solutions across data storage, transformations, and orchestration. Many of these solutions make it easier for data from one application to interact with data from another application. Within model infrastructure, we’ve seen solutions tackling model training, deployment, monitoring and observability (and more). Within agent infrastructure, we’ve seen the emergence of agent-specific infrastructure around storage, hosting, and evaluation as well as foundation models releasing human-agent interaction capabilities (e.g. Project Mariner from Google or Computer Use from Anthropic). All of these providers make it easier for companies to build more complex, multi-step, and autonomous AI workflow products.
On top of these new infrastructure providers, we’ve seen improvements in reasoning***. Debates continue around whether LLMs are truly reasoning or merely mimicking it through memorization and pattern recognition. However, the results are noteworthy — models are performing better on meaningful benchmarks and demonstrating greater contextual understanding (e.g. OpenAI’s o1).
This combination of improving infrastructure and reasoning means that applications will be able to communicate and collaborate more smoothly with each other, unlocking true multi-step and/or autonomous applications. For example, a travel planning app might move beyond generating itineraries and actually make intelligent bookings on your behalf. A notetaking app might move beyond transcription and actually generate accurate follow-up tasks and calendar invites in the apps of your choice. A RFP app might move beyond drafting the RFP and actually find the right opportunities for the business.
These advancements should create a new category of action-oriented platforms that deliver tangible, end-to-end value to users.
In sum, improved model performance, more robust reasoning, and maturing infrastructure are setting the stage for the next wave of AI innovation. Higher accuracy will unlock applications in previously unfitting use cases, and agentic solutions will continue to create more action-oriented and magical user experiences. If you’re building in these areas, feel free to reach out on LinkedIn or Twitter!
Footnotes
*Footnote 1: a nice (non-exhaustive) visualization of 2024’s Open and API model releases can be found here - https://huggingface.co/spaces/reach-vb/2024-ai-timeline.
**Footnote 2: More on Andrew Ng’s four design patterns that characterize agentic reasoning and workflows here and here.
***Footnote 3: What is reasoning? I like this definition from Melanie Mitchell — “The word ‘reasoning’ is an umbrella term that includes abilities for deduction, induction, abduction, analogy, common sense, and other ‘rational’ or systematic methods for solving problems. Reasoning is often a process that involves composing multiple steps of inference. Reasoning is typically thought to require abstraction — that is, the capacity to reason is not limited to a particular example, but is more general. If I can reason about addition, I can not only solve 23+37, but any addition problem that comes my way.” I also enjoyed this talk by Nathan Lambert on reasoning.
****Footenote 4: It’s hard not to be optimistic and excited about the promise of AI. It’s equally important to emphasize the need for building responsible and safe AI, issues such as data privacy, algorithmic bias, and the unintended consequences of autonomous systems require careful consideration.
