Reasoning AI: How Machine Minds Actually Think & Solve Problems

In partnership with

What Makes These AI Different?

I've been watching AI get smarter. It's wild how these new thinking machines work different than the old ones. Normal AI is like that kid in class who knows everything but never shows their work. You ask something, boom - answer appears. But I noticed these new reasoning models do something different. They actually think first.

I can tell because they take longer. When I give them hard math problems, they don't just spit out numbers. Instead, they work through steps like I would on paper. It's pretty amazing to watch. For instance, if I ask "What's 47 × 23?", standard AI responds "1,081" instantly. Thinking AI shows process: "I multiply 47 by 23. Breaking down: 47 × 20 = 940, then 47 × 3 = 141, so 940 + 141 = 1,081."

These models are special type of language AI. The difference? They generate thoughts before speaking. Like having a internal conversation with themselves.

Find out why 1M+ professionals read Superhuman AI daily.

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

1. Join the Superhuman AI newsletter – read by 1M+ people at top companies

2. Master AI tools, tutorials, and news in just 3 minutes a day

3. Become 10X more productive using AI

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI

How Do They Learn to Think?

The training process fascinates me. Engineers take a regular AI model - something like GPT-4. Then they teach it to have two conversations at once.

First conversation happens in it's head. That's where all the thinking occur. Second conversation is what we see - the final answer.

They use something called an "end of thinking" token. I think of it like a period that says "ok, I'm done thinking, here's my answer."

Training data includes three things: questions, the thinking process, and correct answers. Math and coding work great for this since you can check if answers are right or wrong.

The model learns through supervised training first. Then they add reinforcement learning on top. Only the final answer gets graded, not every step in between.

This helps the AI decide how long it needs to think. Simple questions might need just a few internal words. Complex problems could require hundreds of thinking tokens.

But here's the thing - it's still just predicting what comes next. The thinking part just gives it more context to make better guesses.

When Should I Use Thinking AI?

I use these models for hard stuff. Math problems that need multiple steps. Coding challenges with lots of moving parts. Strategic planning for my business.

Regular AI might tell you a train goes 180 miles in 3 hours at 60 mph. Thinking AI shows you: distance = speed × time, so 60 × 3 = 180 miles.

For business, these are game changers. I can ask them to analyze market opportunities. They'll think through different scenarios, consider risks, weigh costs and benefits.

They help with crisis management too. Product recalls, supply chain issues, financial forecasting. Instead of just giving generic advice, they actually work through your specific situation.

The key difference is interpretation. Regular search gives you articles. Thinking AI interprets your words, asks follow-ups, and reaches conclusions based on your circumstances.

The Downsides I've Found

All this thinking costs more. More tokens mean higher bills and more energy use. For simple tasks like translation or summaries, it's overkill.

Apple did a study that showed these models can overthink easy problems. Sometimes they completely fail at very hard ones too. They work best on medium-complexity stuff.

I've also noticed they don't have truly general problem-solving skills yet. And sometimes the overthinking leads to more errors or hallucinations.

But researchers see this as a new way to scale AI. Instead of just making models bigger, we can make them think longer. Early experiments show good results on math and coding benchmarks.

There are diminishing returns though. Doubling thinking time doesn't double performance. You can't help a model find answers it doesn't already know.

How I Talk to Thinking AI

My prompting strategy changed completely. With old AI, I'd say "think step by step" or give detailed plans. These new models already know how to think.

Instead, I focus on clear goals and guidelines. I give tons of context upfront - like briefing a consultant. All the data, constraints, history, exceptions they might need.

For example, instead of asking for "a marketing plan for Germany," I provide market research, budget limits, team capabilities, regulatory concerns, competitive landscape. Everything in one shot.

If I'm worried about costs, I'll add "reason briefly in five steps maximum." They seem to understand that instruction.

One thing that trips people up - some models like OpenAI hide their thinking. Others like DeepSeek show it. If the hidden reasoning is important for follow-up questions, you need to copy it into your next prompt.

Where This All Goes

These thinking models represent a big leap forward. Especially for problems that need more than memorization.

I see them as tools in a toolbox. Customer support might still work better with faster, simpler models. But complex research and multi-step coding? That's where thinking models shine.

The rule I follow: use the simplest model that solves your problem. Save money and energy when you can.

Will this lead to AGI? I don't know. But right now, if your current AI isn't smart enough, try letting it think out loud first. The results might surprise you.

The future of AI isn't just about bigger models. It's about smarter thinking. And that changes everything for how we work with artificial intelligence.