🤖 AI 2027 Is Starting to Look Scarily Accurate. But One Big Prediction Is Missing

Introduction

When the AI 2027 Paper came out in 2025, many predictions sounded pretty crazy. AI agents would do real work, coding would become more automated, and AI companies would use AI to build even better models.

At that time, I thought the timeline looked a bit like a sci-fi movie with an Excel sheet.

But in 2026, it’s getting harder to laugh at it.

I checked the AI 2027 Paper against the AI 2027 Tracker, new benchmarks, data from major AI labs, and recent research. Some predictions are getting much closer to reality than I expected.

There are a few parts I agree with, some parts I think are too early to call, and one prediction matters much more than the rest.

Okay, let’s get into it. This is how I see the AI 2027 Paper after checking what has actually happened so far.

I. What AI 2027 Expected to Happen by Now

Before judging whether the AI 2027 Paper was right or wrong, I want to set the baseline first. The paper didn’t just say models would get smarter. It expected AI to start doing more real work, while AI itself would help speed up AI development.

Key points

• AI 2027 expected AI to move into real work and AI development by 2026.

• It also predicted rapid growth in compute and AI infrastructure.

By 2025 and 2026, the timeline expected several big changes:

• AI agents would become useful for real work, even if they still made mistakes on long or complex tasks.

• Coding agents would create enough value that developers would use them in daily work.

• AI labs would start using their own models more for coding, experiments, and AI research. This is also a key part of the paper’s takeoff forecast.

• Compute and AI infrastructure would grow very fast. The compute forecast expected bigger training runs and around $600 billion in AI data center spending by 2026.

• AI capabilities would also start having a bigger effect on jobs, cybersecurity, and the global AI race.

In simple terms, the AI 2027 Paper expected 2026 to be the point when AI started moving beyond “a chatbot that gives good answers” and became much more involved in real work.

That was a very aggressive timeline. The good part is that we now have enough real data to check how much of it actually happened.

II. What the AI 2027 Paper Got Right

After comparing the original predictions with what is happening now, I think the AI 2027 Paper got the overall direction surprisingly right.

Key points

• AI agents are already doing real tasks, but reliability is still a problem.

• Coding agents are becoming part of daily developer work.

• AI is already helping build better AI, though the 50% R&D speedup is not proven.

• The AI infrastructure race is clearly happening, with massive spending on compute and data centers.

1. AI Agents Developed Almost as Expected

In the “Stumbling Agents” section, AI 2027 predicted that computer-use agents would handle real tasks like ordering food or working with spreadsheets.

The paper also expected them to stay unreliable. By early 2026, it still expected agents to struggle with long tasks. You can read the original prediction in the AI 2027 scenario.

That prediction looks much closer to reality now.

In July 2025, OpenAI launched ChatGPT agent, which could browse websites, fill out forms, edit spreadsheets, run code, and complete multi-step tasks from start to finish.

The reliability problem is still here. METR’s latest research measures both 50% and 80% time horizons because completing a long task sometimes is very different from completing it reliably.

METR also warns that its time horizon doesn’t mean an agent can handle every real-world task of the same length.

So I think “stumbling agents” is still a fair name.

Agents can do much more useful work now, but I still wouldn’t hand one a messy project and expect everything to be perfect the next morning.

2. Coding Agents Became Part of Real Work

In the “Stumbling Agents” section, AI 2027 predicted that coding systems would start acting more like autonomous agents than simple assistants.

They could take instructions, make major code changes on their own, and sometimes save developers hours or even days of work.

That prediction now looks pretty strong:

• Claude Code: Anthropic found that coding-agent activity across GitHub projects more than doubled since late 2025, while Claude Code users now spend about 20 hours per week using it.

• Codex: OpenAI reported more than 5 million weekly users by June 2026.

• Cursor: More than 70% of Fortune 500 companies use Cursor to deploy and manage coding agents.

For me, this is one of the clearest wins for the AI 2027 Paper.

Coding agents have moved past the “wow, it can write code” stage. Developers and companies are now giving them real parts of the job.

3. AI Is Already Helping Build Better AI

This is the prediction I care about most.

AI 2027 expected frontier labs to use their own models to speed up AI research. By early 2026, it predicted that AI assistants could make algorithmic progress about 50% faster.

We can already see part of that happening. As of May 2026, Claude wrote more than 80% of the code merged into Anthropic’s codebase. Anthropic also says Claude can match or outperform skilled humans in some well-defined research experiments.

I think AI 2027 got the direction right.

But the 50% speedup still isn’t proven. AI is clearly helping build better AI, but we don’t yet have enough public evidence to say the whole research process is moving 1.5x faster.

4. The AI Infrastructure Race Became Real

AI 2027 expected compute demand and data center spending to rise quickly. Its compute forecast projected around $600 billion in AI data center spending by 2026.

That race is already easy to see. Alphabet now expects $195 billion to $205 billion in total capital spending for 2026, as it expands capacity for growing AI and cloud demand.

I don’t think we need ten more numbers here. When one company plans to spend close to $200 billion in a year and is still trying to add capacity, the infrastructure race already looks very real.

So far, the AI 2027 Paper got the direction surprisingly right. The harder question is whether it also got the speed right.

III. Where the AI 2027 Paper Is Falling Behind

So far, I think AI 2027 got the direction mostly right. But when I check the timeline and exact numbers, the paper starts to look less accurate.

1. AI Research Acceleration Is Still the Biggest Missing Piece

The biggest question is still the R&D speedup.

A July 2026 study gave frontier agents up to 6 days to work on two open-ended AI research problems. The agents handled much of the engineering work, but they still didn’t make major progress on the main research questions.

So I still wouldn’t give AI 2027 full credit here. AI can already help with coding and experiments, but we still don’t have strong public evidence that it can speed up the whole AI research process at the level the paper expected.

2. Some Exact Numbers Are Still Behind the Timeline

AI 2027 also made some very specific predictions. The scenario expected the stock market to rise about 30% in 2026.

But by September 2, 2026, the S&P 500 was up only about 12% for the year.

I wouldn’t call this prediction wrong yet because 2026 isn’t over. But right now, AI 2027 looks better at predicting where things are going than exactly when they will get there.

IV. It Still Depends on One Big Assumption

This is the most important part for me. Many predictions that look right so far are still part of the setup.

AI 2027 only reaches its extreme scenario if one loop starts to speed up: AI helps build better AI, and the new AI then helps research move even faster.

The paper expects several big steps:

• AI research becomes 2x–4x faster: AI starts speeding up the whole research process, not just helping with small tasks.

• Coding becomes almost fully automated: AI can handle most coding work with much less human help.

• A superhuman coder appears: AI 2027 expects Agent-3 to become better at coding than the best human programmers.

• AI can copy and maintain itself: An agent could find compute, run new copies, and keep working on its own.

• Huge numbers of agents work together: The paper describes around 200,000 Agent-3 copies working in parallel.

We can already see small signs of this. In March 2026, Marinka Zitnik shared an experiment where 15 AI agents worked together on one research problem for 48 hours, making more than 574 edits to a shared research board.

Around the same time, SkyPilot gave Claude Code access to 16 GPUs. The agent submitted about 910 experiments in eight hours, around 9x faster than the estimated sequential setup.

These examples are impressive, but they are still far from 200,000 agents or a proven 2x–4x speedup across an entire AI lab.

For me, this is the real test of the AI 2027 Paper. The setup is starting to appear. The part that could create the fast takeoff is still ahead.

Conclusion

After checking the predictions one by one, I don’t think the AI 2027 Paper is something we should ignore.

It got several big directions right. AI agents are doing real work, coding agents are becoming more common, AI is helping build better AI, and the compute race is getting much bigger.

But I still wouldn’t say AI 2027 is fully right. The most important part hasn’t happened yet: AI still hasn’t clearly shown that it can speed up the whole AI research process enough to create a fast self-improving loop.

So this is how I see the paper right now: the setup looks surprisingly close, but the takeoff still isn’t proven.

That’s why I think AI 2027 is still worth watching. If AI R&D starts speeding up much faster from here, this paper will become a lot more important.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• ChatGPT Just Got a New Superpower? (Computer History & More Updates)

• Grok Bot Just Killed 1,000 Startups? An Insanely Easy AI Agent Anyone Can Use 24/7

• I Tested & Ranked Every Personal AI Agent in 2026 From Beginner-Friendly to Hardest*

• This OpenRouter Trick Literally Lets Anyone Run Claude Code for Free or Nearly Free (Full Guide)*

• How to Build Your Agent Skills Once, Run It Almost Everywhere: Claude Code, Codex, Cursor*

  *indicates a premium content, if any