“I could do a whole talk on the question of ‘is AI dangerous.’ My response is that AI is not going to exterminate us; it's a tool that's going to empower us.”

—Oren Etzioni, founding CEO of the Allen Institute for Artificial Intelligence

Scaling Hits a Slowdown

Reasoning models like OpenAI’s o3 are the rising stars of AI. Unlike standard language models that can summarize your inbox or draft a decent email, these models are built to think. Solve problems. Make decisions. And for a while, it looked like the sky was the limit.

But a new study from Epoch AI is throwing a bit of shade on that optimism. The analysis suggests that reasoning models might soon hit some serious scaling limits.

How Reasoning Training Works

Reasoning models follow a two-step training recipe. First comes pre-training on massive amounts of human data. That’s the baseline. Then comes reasoning training, where the model is taught to solve hard problems using reinforcement learning. It’s like the difference between reading every book in the library and learning how to pass the bar exam.

Why Scaling Might Stall

The problem is simple. Reinforcement learning training is growing at a 10x rate every few months. But the total amount of computing power that AI labs can realistically throw at a model is growing at a much slower rate—about 4x per year.

This means reasoning training is on track to collide with the ceiling of available compute sometime in the next year. Once that happens, progress will slow, and dramatic jumps in AI capability will become harder to pull off.

There are other friction points too. Creating high-quality problems for these models to learn from is not easy. It takes time, human expertise, and a good chunk of change.

What to Watch

If your vendor or in-house AI tools are promising "smarter than human" reasoning, ask what models they’re using and how they’re trained. There’s a big difference between fast chatbots and true reasoning engines.

For now, the sweet spot is combining today’s capable AI with human judgment, especially in variable-rich environments like retail auto. The models are powerful, but not magic. And it looks like some of the fastest gains may be behind them—at least for now.

Objections don’t have to mean the end of the conversation. In fact, they’re often invitations to better understand your customer and build trust. When your team knows how to respond with clarity and care, those moments become a chance to move the deal forward with confidence.

• Finance & Insurance: MeasureOne
  Specializes in data verification, as well as streamlining the process of verifying income, employment, and auto insurance information for car loan applicants.
  

• Marketing & Advertising: Team Velocity
  An AI-driven platform that generates vehicle detail videos, automates ad creative, and manages cross-channel marketing by connecting inventory data, customer insights, and campaign performance.
  

• Content Creation: Spyne
  AI-driven solutions for creating high-quality automotive images and videos. Their technology allows dealerships to produce professional-grade visuals for online listings without the need for expensive equipment or photographers

In this Auto Collabs episode, Monik Pamecha (co-founder of Toma) gets real about what it takes to make AI actually helpful in a dealership.

After listening to thousands of calls (often at 3x speed), he and his team built voice AI that doesn’t just hear words—it gets what the customer is trying to do. It’s an engaging and honest chat about what to automate, what to leave alone, and how smarter tools can make things way less messy for everyone.

Catch the replay to see how AI can clean up communication without killing the human side.

• Google is bringing its Gemini AI to every vehicle that supports Android Auto in the coming months. 🚗

• TikTok’s new AI Alive tool allows users to create videos from still photos right inside the app. 📱

• OpenAI launched a safety evaluations hub to report model performance and track updates over time. 🛡️