Either AGI starts showing up in the next few years, or the whole singularity story is much farther away than most people think.

My best guess is 2028.

I think there is a real chance that we get something fair to call AGI between 2027 and 2029.

Not because every demo on Twitter is real. Not because every benchmark is profound. Not because one lab says the word “reasoning” louder than the others.

I think that because three things are now true at once:

• the models got much better
• the scaffolding around the models got much better
• the amount of useful work they can do in one shot has started to climb fast

That does not prove a singularity is near. It does mean the argument is no longer science fiction.

There is about a 60% chance that AGI-like systems arrive by the end of 2029, with 2028 as my single best guess.

If that does not start happening soon, then it probably does not happen on today’s curve ten years later.

Either the curve bites hard in the next few years, or the whole singularity story belongs to a much stranger world with different hardware, different institutions, and probably a different technical paradigm.

First, define the words

People use AGI, ASI, and singularity sloppily. So here is the plain-English version.

AGI means a system that can do most valuable cognitive work at least as well as a good human, and usually faster. If you can throw it coding, research, analysis, planning, writing, and tool use, and it holds up across the board, that is close enough for this article.

ASI means something clearly beyond the best humans at almost all of that.

Singularity is the strongest claim of the three. It means progress starts compounding so fast that normal forecasting breaks. Humans stop being the main pace-setter.

Those are not the same claim. You can believe in AGI and still be skeptical of ASI. You can believe in ASI and still think the word singularity is too dramatic.

The short history

Most AI history does not help much with this question. A few moments do.

Deep Blue beating Kasparov in 1997 was important because it showed machines could crush humans in a domain once treated as a test of intellect. But it was narrow. It was a machine for chess.

AlexNet in 2012 helped establish deep learning as the main road, not a side alley.

AlphaGo in 2016 changed the emotional tone of the field. DeepMind did not just beat a Go champion. It produced moves that looked alien to strong human players. AI was no longer just brute force plus speed. It could find ideas people missed.

Around that same period, another piece of history sits behind the current race: OpenAI was founded in 2015 partly out of a real fear that advanced AI might get concentrated inside Google after its acquisition of DeepMind. That fear now looks less naive than it did at the time. A lot of the current AI story is really about who gets to control general-purpose intelligence.

Then came the language-model phase.

The architecture break behind that phase was the Transformer. Once Attention Is All You Need landed in 2017, the field had a much better recipe for training large sequence models in parallel. A lot of what came next was that recipe getting scaled, refined, and pointed at more data.

GPT-2 in 2019 was the first time many people saw that simple next-token prediction could turn into eerily coherent text at scale. The model was still flaky, but the direction was obvious.

GPT-3 in 2020 pushed that much further. It made the “general-purpose text engine” idea feel real. You could prompt it into translation, summarization, code, Q&A, or light reasoning. A lot of people still dismissed it as autocomplete with good PR. They were not fully wrong. But they were no longer fully right either.

Then two things changed the picture more than raw scale did.

The first was RLHF and the InstructGPT line. That took models that were smart in a messy way and made them usable.

The second was ChatGPT in 2022. ChatGPT was not the first strong model. It was the first time millions of people understood, in one afternoon, that the interface to computing might change.

Then the center of gravity shifted again.

In 2024, reasoning models started to become a real product category. OpenAI o1 was the clearest public marker. “Think before answering” stopped being a research rumor and became a thing normal users could feel.

At the same time, the field moved from chat to action.

AutoGPT was an early viral moment in 2023. It was messy, overhyped, and loop-prone. But it planted the idea that you could wrap a model in tools, memory, and retries and get something agent-like.

By 2025, that idea got much more serious. Operator, deep research, and similar systems showed that agents were no longer just toys for prompt hackers. They were still unreliable. They were also useful.

And for many engineers, somewhere around November 2025, agentic coding crossed a line. Not because one product magically solved software engineering, but because the combination of better models, better tool use, and better coding harnesses made delegation feel normal. Claude Code is the cleanest symbol of that shift. It was not the first coding agent, but it was one of the clearest signs that “ask the model to edit the repo, run tests, and come back with a patch” had become a practical workflow, not just a conference demo.

Then came the more personal version of the same story.

OpenClaw belongs in this story because it made the always-on personal agent feel real. If ChatGPT turned LLMs into a consumer product, OpenClaw and similar runtimes made people ask a more dangerous question: what happens when the model is not just a chat box, but a semi-autonomous operator living in your messages, your tools, and your workflow?

That is the real through-line.

First we got narrow systems. Then general language systems. Then usable assistants. Then reasoning models. Then agents that can actually do work.

What changed in the last two years

The simple answer is that the models got better. But that is too shallow.

The deeper answer is that we are stacking several kinds of progress on top of each other now.

Training improved. Inference got more powerful. Test-time compute became a real lever. Tool use improved. Retrieval improved. Memory got less fake. Coding scaffolds got sharper. Evaluation got less academic and more task-shaped.

That shift changed the argument.

For a while, people argued past each other because one side pointed at chatbot tricks and the other side pointed at benchmark failures. Both were missing the operational middle: what happens when you give a strong model a shell, a browser, a file editor, long horizons, retries, and permissioned access to the world?

That middle is where a lot of the action is now.

This is also why software is the first domino. Code lives in text, diffs, terminals, logs, tests, and tickets. The world of software is already shaped like a language model’s natural habitat. You do not need full robotics to change software. You need long-enough task horizons, good enough judgment, and enough reliability that the review burden is still worth it.

We are getting close.

The three timeline camps

Most timeline arguments are easier to read if you stop treating them as one giant debate.

There are really three camps.

1. The fast camp

This is the AI 2027 and Situational Awareness camp.

The fast camp thinks the missing distance is small, and that once AI systems become strong enough at coding and research, they start speeding up the very process that improves them. That is the core loop.

The strongest version of this argument is not “models look cool.” It is “the systems are already eating more of the cognitive stack, and once they cross the threshold where they can materially accelerate AI R&D, the curve bends.”

AI 2027 is the most concrete public version of this view. It tries to say, in detail, what a near-term transition to superhuman coders and then much faster progress could look like. Leopold Aschenbrenner makes a related case in Situational Awareness: extrapolate compute, algorithmic efficiency, and “unhobbling,” and AGI by 2027 starts to look plausible rather than insane.

This camp is also closer to frontier-lab rhetoric than many skeptics want to admit. Top lab leaders keep describing very short timelines in slightly different language. They may avoid the word singularity, but they are not talking like people who think this is a 2045 story.

2. The medium camp

This camp is less interested in sci-fi vocabulary and more interested in massive acceleration.

Dario Amodei is a good example. In Machines of Loving Grace, he argues that powerful AI could arrive reasonably soon and then compress decades of progress in fields like biology into a few years. His framing is useful because it takes both sides seriously: intelligence helps a lot, but the physical world still imposes latency.

This camp says:

• yes, big change is close
• no, intelligence is not magic dust
• yes, software may move first
• no, wet labs, fabs, logistics, and governments will not move at model speed

I think this camp is underrated because it is less theatrical than the pure takeoff story and less comforting than the skeptic story.

It also fits a lot of current evidence better.

3. The slow or skeptical camp

This camp has several flavors.

Epoch makes the cleanest serious case for longer timelines. The argument is not “AI is fake.” The argument is that people are hand-waving bottlenecks. Compute, data, real-world experiments, deployment cost, and messy agentic work all slow things down. On that view, current progress is real but does not imply a software-only singularity in the next few years.

Gary Marcus is the stronger skeptic on the core paradigm. He thinks LLMs are not the royal road to AGI.

François Chollet is more subtle. His complaint is not that current models do nothing. It is that memorization and interpolation are being confused with general intelligence. The ARC line of critique says we should care much more about efficient adaptation to genuinely new problems.

This camp has a real point. Many bullish people talk as if a model being strong at code, math, and web tasks automatically means it is one clean scale-up away from robust autonomous research. That is not proven.

What the camps are actually disagreeing about

A lot of this debate sounds philosophical. Most of it is not.

The real disagreements are about bottlenecks.

Can AI automate AI research early enough to speed up its own improvement?

How much does real progress depend on more chips, more power, more data centers, more experiments, and more human review?

Do agentic systems keep getting better as you give them more time and tools, or do they just fail more slowly?

Is current progress mostly a smooth extension of what transformers already do well, or do we still need a real conceptual break?

And maybe the biggest one:

When a model looks bad in a bare benchmark, is that telling you something deep about its limits, or are you just measuring the wrong level of the stack?

I do not think either side fully owns that question yet.

The economic warning shots

One reason timeline talk stopped sounding academic is that people started sketching the second-order effects.

Citrini Research did this with The 2028 Global Intelligence Crisis. It is explicitly a scenario, not a prediction. That distinction is important. But the scenario still hit a nerve because it describes a world many people can now imagine: coding agents get cheap, white-collar pricing power cracks, and large parts of the economy realize too late that cognitive labor was more exposed than expected.

I would not treat that essay as prophecy. I would treat it as a useful stress test.

Once agents become good enough, the question stops being “can they reason?” and becomes “which parts of the economy were quietly built on expensive human cognition?”

My prediction

Here is my best guess.

There is about a 60% chance that we get AGI-like systems by the end of 2029.

If you want a rough split, I would put it like this:

• about 25% by the end of 2027
• about 45% by the end of 2028
• about 60% by the end of 2029

My modal year is 2028.

Why that range?

Because the strongest part of the bullish case is real. Coding, research assistance, tool use, long-horizon work, and test-time reasoning are all improving at the same time. That is exactly where you would expect the curve to steepen if AGI is near.

But I do not buy the strongest version of the fast-takeoff story. I think reliability, evaluation, deployment friction, and physical bottlenecks still loom large.

So I end up here:

2027-2029 looks live.

Much later than that, on the current story, looks less convincing.

And this is the part many people will find strange: if we get to 2030 without a real AGI break, I would update toward longer, not toward “sure, then maybe 2034.”

Why?

Because the singularity claim is not just “progress continues.” It is “progress is about to compound.”

If it does not start compounding soon, then one of two things is probably true:

• current systems are missing something important
• the real bottlenecks are stronger than the near-term bulls think

In either case, the comfortable middle story gets weaker. The idea that today’s paradigm somehow muddles along for years and then suddenly turns into a clean singularity in the mid-2030s is, to me, less believable than either a near break or a much longer detour.

That is why I think singularity is either a near-term event or a much later civilization-level event.

Soon, or not for a while.

What it means if this is right

The first big consequence is that software changes before most institutions are emotionally ready for it.

Not all software jobs disappear. That is the wrong frame. The deeper change is that the unit of work changes. One good engineer with several agents starts to look like a small team. Review, taste, problem selection, security, and system design count more. Boilerplate production counts less.

Research changes too. Not because AI replaces every scientist at once, but because literature review, hypothesis generation, experiment planning, and coding-heavy analysis all get cheaper before labs know how to price the new reality.

Security gets stranger. If you can give an agent useful authority, you can also give it dangerous authority. The same thing that makes personal agents compelling also makes them risky.

And the biggest shift is psychological.

Once a system can reliably compress a week of cognitive work into an hour of supervision, people stop arguing about whether it is “really intelligent.” They start reorganizing around it.

That is the point where the vocabulary fight counts less than the economic one.

Final view

I do not think the singularity is guaranteed.

I do think the burden of proof has changed.

Five years ago, saying “AGI by the end of the decade” made you sound reckless.

Today, saying it is impossible should require more argument than many skeptics are giving.

That does not mean the bulls are right. It means the world has moved.

My best guess is still 2028.

If I am wrong because it happens later by a year or two, fine.

If I am wrong because this whole line of progress stalls, that will be one of the most important facts of the decade.

But if I am right, then the period people will look back on as the hinge was probably not some far future date.

It was when coding agents became useful, reasoning models became normal, and the distance between “chatbot” and “co-worker” started collapsing in plain sight.