The Agentic AI Reasoning Era Begins : Recently, Sequoia Capital published an insightful post about generative AI. They mentioned a significant shift happening right now. We are moving from “thinking fast” to “thinking slow.”
What does that mean? It means AI is evolving. In the past, AI generated quick responses. Think of it like a parrot repeating phrases. Now, we’re entering an era where AI can take a moment to reason and think, much like humans do.
This evolution is being driven by big players like Microsoft with OpenAI, AWS with Anthropic, and Google with DeepMind. They are stabilizing their technologies and now focusing on building real reasoning capabilities into AI. It’s an exciting time!
In the past, AI focused on generating quick replies based on patterns it learned. But now, companies like Microsoft, OpenAI, Amazon with Anthropic, and Google with DeepMind are making strides in stabilizing this technology. The race is now on to add true reasoning capabilities to AI.
For example, OpenAI has introduced a new model called “Strawberry,” which was formerly known as Q*. This model can pause and think before providing a response. It’s similar to how AlphaGo, the famous Go-playing AI, operated back in 2016.
The Significance of Inference-Time Compute
So, what does inference-time compute mean for AI? It means that AI can “stop and think” for more thoughtful responses. Instead of rushing to an answer, AI will analyze its options before deciding.
That’s a game-changer! We’re not just looking at AI that can mimic human responses. We’re moving toward AI that can reason through complex problems. It’s like moving from instinctual thinking, known as System 1 thinking, to more deliberate reasoning, called System 2 thinking.
System 1 is all about quick, automatic responses. Think of it like when you quickly recall the capital of Bhutan. You either know it or you don’t. But with System 2, you pause to consider more complex issues, like planning a vacation or solving a tricky math problem.
The AlphaGo Example
Let’s take a moment to talk about AlphaGo. Back in March 2016, AlphaGo played against the legendary Go master, Lee Sedol. It was a landmark moment in AI history because AlphaGo showed the world that AI could think, not just mimic.
AlphaGo wasn’t just programmed with responses. It could analyze countless potential future scenarios and choose the best one. The more time it had to think, the better it performed.
This capability is what we’re now seeing in models like Strawberry. The model can stop and think, evaluating different possible responses before deciding on the best one.
From Mimicry to True Reasoning
But the challenge with current models lies in how to score the responses. When AlphaGo played, it could simulate the entire game and see which move would lead to victory. In AI applications, though, how do you evaluate a response, like the first draft of an essay?
That’s a tricky question! For now, Strawberry excels in areas like coding and math but struggles with more open-ended tasks like creative writing.
This limitation is important to understand. Even though AI is advancing, there’s still a gap when it comes to subjective tasks that require human-like creativity and reasoning.
System 1 vs. System 2 Thinking
Let’s dive deeper into the concept of System 1 and System 2 thinking. As we mentioned earlier, System 1 is about quick responses, while System 2 is about deeper, deliberate reasoning.
This shift is crucial for AI. For simple tasks, System 1 might be enough. But when it comes to solving complex problems, like breakthroughs in science, System 2 thinking is necessary.
AI must learn to take its time when faced with tough questions. Imagine if AI could think for hours, days, or even longer. Would we finally crack complex mathematical problems?
It’s exciting to think about the possibilities! As AI moves toward inference clouds—where it can scale its computing power based on the task’s complexity—we’ll see much more sophisticated applications.
The Race for Reasoning Capabilities
Now, let’s talk about the competition among big players like OpenAI, Anthropic, Google, and Meta. As they develop more powerful reasoning capabilities, will there be one model to rule them all?
That’s an interesting question! Initially, many believed one company would dominate the market. However, we’ve seen fierce competition at the model level.
The models keep improving, and we’re witnessing rapid advancements. The landscape is not as straightforward as we once thought. The application layer, where these models get used in real-world scenarios, is where the real competition lies.
The Importance of Cognitive Architectures
Speaking of the application layer, let’s discuss cognitive architectures. These determine how AI systems think and respond. Each system has its own way of interpreting user input and generating a response.
For instance, take the AI product Factory. Each “droid” they create mimics human thought processes to solve specific tasks. This means they break down tasks into manageable parts, just like a human would.
This approach allows AI to operate more efficiently and effectively in various fields, whether it’s software engineering or customer support.
The Future of AI Applications
So, what does this mean for aspiring AI entrepreneurs? If someone wants to start a business in AI, where should they focus?
That’s a crucial question! Competing in infrastructure is tough because of giants like NVIDIA. Competing at the model level is equally challenging with companies like OpenAI leading the pack.
However, focusing on the application layer may be more feasible. Companies are now building sophisticated cognitive architectures that combine multiple models and tools to deliver tailored solutions.
These new applications aren’t just simple wrappers around existing models. They incorporate complex logic and reasoning capabilities that can handle real-world complexities.
Service-as-a-Software
Let’s touch on a transformative shift happening in the market—service-as-a-software. This model allows companies to turn labor-intensive tasks into automated services.
We’re witnessing a shift from selling software by the seat to selling work based on outcomes. For example, Sierra is an AI that resolves customer issues on websites. They get paid per resolution instead of per seat.
This model could redefine how businesses approach AI. It creates an opportunity for companies to provide targeted services rather than just selling software licenses.
New Agentic Applications
With this shift, we’re starting to see a new generation of agentic applications. These are AI systems capable of performing complex tasks with minimal human intervention.
We’re seeing applications like Harvey, an AI lawyer, and Glean, an AI work assistant, emerging in various sectors. Each of these applications brings unique capabilities to the table.
Another example is XBOW, which is building an AI pentester. This tool can perform penetration tests to assess a company’s security, offering the same quality as highly skilled human pentesters but at a fraction of the cost.
The SaaS Universe and AI
Now, let’s address a pressing question many investors have: Will the rise of AI disrupt existing SaaS companies?
Initially, many thought the answer was no. After all, established companies have data and distribution advantages. But with the complexity of cognitive architectures and the engineering required to harness AI’s capabilities, things might change.
It’s possible that we underestimated what it means to be “AI native.” Just like the transition from on-premise software to SaaS reshaped businesses, AI could drive another major transformation.
Investment Opportunities
So, where are investors focusing their attention in this evolving landscape?
There are several key areas:
- Infrastructure: This remains a domain for hyperscalers, making it less appealing for venture capitalists.
- Models: This is also dominated by large companies. Investors are drawn to cutting-edge research, but it’s a competitive space.
- Developer Tools: These are seen as valuable by venture capitalists, especially given the success stories in the cloud transition.
- Applications: This layer is the most promising for venture capital. Historically, many successful application companies emerged during the cloud and mobile transitions, and AI seems poised for similar growth.
Closing Thoughts
As we move into the next act of generative AI, we expect to see reasoning capabilities enhance applications significantly. The potential for agentic applications is immense!
With continued research in reasoning and inference-time compute, we’re on the brink of witnessing AI systems that can think and act in ways that feel almost human-like.
Exciting times are ahead! As we wrap up, it’s important to remember that while AI is advancing rapidly, there are still challenges to overcome. The interplay between cognitive architectures and real-world applications will be crucial in shaping the future.