15 Years to AGI?

Estimated Reading Time: 5 minutes

TL;DR

Advances in AI over the last year or two have shifted my thinking on the pathway to AGI. I talk about what I think it would take to reach AGI and explain why I think a 15 to 20 year roadmap may now be within reach.

Onwards to AGI

Over the last several years, I have been very skeptical on the prospects for artificial general intelligence (AGI). Although there has been a steady string of advances, it felt like boosters of AGI have been more driven by the prospects of astronomic fundraises and commercial pressures to build hype rather than by scientific analyses. Over the last year though, I have slowly been shifting to a view that medium term AGI, within the next 15-20 years, may in fact be possible.

There are a few reasons for this. The biggest is the extraordinary buildout of new compute capacity and the enormous growth in talent pool around AI and AI scaling. As more and more smart minds and resources are pointed at a problem, the more amenable it becomes to solve. Importantly, there have been several additional intermediate advances. Most recently, coding tools have taken a major step forward and are now at a tipping point of becoming accepted as a powerful addition to the toolkit of even a professional software developer. In our own work at Deep Forest on retrosynthesis, we have shown that LLM-enabled retrosynthesis (with human help) is capable of finding novel synthetic pathways to complex natural products (see https://deepforestsci.com/blog/9). Recently, prominent mathematician Terry Tao has attested that GPT5.2 is capable of (with human help) solving novel Erdos problems. Several Erdos problems have fallen to GPT5.2 (Though it must be said most solutions are fairly straightforward literature extensions, at least so far. See https://github.com/teorth/erdosproblems for an exhaustive listing by Terry Tao). This is a sampling of a much broader range of advances across several fields than we’ve seen before. Something is shifting in the air.

And yet, we are not at AGI yet. For example, we show in our retrosynthesis work that leading LLMs routinely make very dumb chemistry errors. We have had to create extensive validation infrastructure to prevent the LLM from inventing carbons out of thin air. A high school student can understand that atoms can’t come out of nowhere in chemistry, but an LLM cannot. Recent advances are driven by reinforcement learning with verifiable rewards (RLVR), which depends on the presence of a validation/verification harness. These harnesses require human experts to put in large amounts of effort to encode when solutions are incorrect. For some applications, like software, these are relatively easy to do. For others like robotics, this is much harder. Chemistry falls somewhere in the middle. Models still lack common sense; they lack the spatial awareness and notions of object permanence and memory that even a five year old have. These problems are fundamental and may require new algorithms and methods to address. Brute force scaling can and will make progress, especially given the enormous resources at hand, but may not be enough to construct robust common sense. If I had to bet, I would say RLVR in high fidelity simulation worlds could make good progress, but at enormous computational cost. There may be more clever ways to do this in the short term as well. Plenty of smart people are working on different approaches to “world models;” see this recent example from Meta for a representative discussion (https://arxiv.org/abs/2506.09985).

On the whole, I think there are a series of critical steps we need to hit before it is possible to build AGI. I want to emphasize that I don’t claim these to be novel observations! Several groups are already working on each of these challenges. My goal is rather to aggregate patterns that I’ve been observing in the hopes that they may be useful for a broader audience. My critical assumption for this analysis is that AGI should not routinely make mistakes that a 10 year old human would find silly. This is not to argue that human intelligence is not “jagged;” humans definitely have cognitive blind spots. But rather I want to argue that crossing the point where AI intelligence no longer has obvious human-detectable blind spots is a critical marker on the path to AGI. Let’s call this first core property robust common sense. Achieving this may require brute force effort by the big AI labs as I hinted above. That is, like today’s self-driving car efforts, through exhaustive data gathering and extensive simulation of failure modes in powerful world models. Understanding the potential and limitations of world model approaches is a red hot line of research. I anticipate that within a couple of years, we will have a better idea on whether world model approaches are the correct approach to robust common sense or if new ideas are needed.

The second core property is intuitive physics. A five year old has built the ability to navigate the physical world with confidence. This includes picking up and manipulating objects in the wild, navigating new rooms and environments, and a basic understanding of gravity, friction, leverage and more. None of today’s LLMs or robotic foundation models yet have anywhere near this degree of capacity. It’s an open question why. Possibly robotic foundation models which build on vision language models (VLMs) are biased the wrong way from the representations needed for control? My personal theory is that, as for robust common sense above, extensive training in game worlds with realistic intuitive physics could make a big difference. The devil of course is in the details; how much physics is needed and can such systems be simulated fast enough to enable reasonably fast rollouts for reinforcement learning? I suspect there are some hard computational physics problems to still be worked out on the way, so this is probably not just a matter of brute force scaling.

The third core property I would argue is embodiment. It is not enough to just be able to predict intuitive physics. Instead, AGI must be capable of meaningfully interacting with the real world. This may be one of the most fundamental challenges to AGI since it depends on difficult questions in actuator design among other challenges. One of the biggest such challenges is safety; embodied AI that poses a risk to humans can’t be justified. I don’t speak of “Terminator” risk, but rather just that heavy robots failing and falling onto people could be disastrous! Proper embodiment requires sufficient mastery of stability and control to not prove dangerous in real world settings. A critic may reasonably say that dis-embodied AGI would still meet a reasonable definition of being AGI. I won’t argue with this too much, but I will say there is something crucial in human and animal intelligence about being localized and embodied. The energy constraints of the body and the need to tend to its underlying biological drives form a tremendous part of our day to day operations. Enforcing these same constraints on AI systems may provide crucial impetus to AI development.

The fourth core property I would argue for is long term planning. Today’s agents lack the ability to execute complex long range plans. The closest I have seen is probably a starcraft bot. The ability to form complex plans is critical to human intelligence. It is routine for a person to form plans at the scale of years; think of the goal of graduating college for example. It is harder, but not that uncommon, to form plans at the scale of decades or lifetimes. Every parent has to do this in some form for their children. AI systems do none of these tasks today. There is no system today that can plan and execute on long term goals to serve others or itself. Doing so in AI systems may require new algorithms or possibly just more scale. I don’t personally have a strong intuition on what the critical missing piece is between today’s methods and capable long range planning. It is also not clear to me that simulation or world model approaches are sufficient here; games don’t typically require very long range planning the way normal life does.

The final core property I will argue for is continual learning. There is no easy or robust way for a model to gain new information on an on-going basis. All of today’s agents and models are created by an extensive “pre-training” and “post-training” process. This frozen agent has some limited ability to respond to different inputs through “in-context learning”, but this is a much weaker effect than skills learned during pre and post-training. We rely on human engineers to update models semi-manually. We need some way to unify these stages; on-going learning needs to become as robustly capable of modifying an agent as the preliminary training steps.Continual learning is a hot area of research; see this recent paper from Google for example https://research.google/blog/introducing-nested-learning-a-new-ml-paradigm-for-continual-learning/. Continual learning is a tremendously challenging problem though. For example, this paper from 10 years ago was wrestling with many of the same challenges https://arxiv.org/abs/1606.04671. Neural networks have a tendency towards “catastrophic forgetting.” New learning can erase old understanding. Until we form a deeper understanding of how to create stable long term memory in neural networks, we will likely continue to struggle. I think new ideas are needed here and there is a reasonable chance we will still struggle with this problem more than 15 years out from now.

These five properties (robust common sense, intuitive physics, embodiment, long term planning, and continual learning) are not necessarily an exhaustive listing of the challenges on the pathway to AGI. There may well be other challenges that arise as we get closer. Even as a onetime skeptic I have to acknowledge that systems like ChatGPT have astonishing capabilities by the standards of 10 years ago. For example, the recent demonstration that Claude Code can build code for a new browser from scratch is astonishing (https://simonwillison.net/2026/Jan/19/scaling-long-running-autonomous-coding/). It feels churlish to not acknowledge this is extraordinary. And yet it somehow isn’t enough. Humans are still much better at many many tasks even within the relatively narrow scoping of software engineering. Intelligence is the product of evolution over very very long time scales. There may still be many as of yet mysterious roadblocks on the path to AGI. But that said, I still think that if we can overcome the five challenges I discuss, we will be much closer to this goal. I think 15 years of focused research with the scale of resources society has decided to throw at the problem may do it. If I had to place a bet, I would say it’s slightly greater than even odds, but honestly, who knows! We will have to see what happens.

I’ll end this essay by arguing that as AGI draws closer, possibly within our lifetimes, we all should think more about AGI ethics. Legions of serious thinkers have already tread these grounds but mostly from a Western perspective thus far. Approaches range from Asimov’s laws (https://en.wikipedia.org/wiki/Three_Laws_of_Robotics) to Human Compatible AI (https://www.amazon.com/Human-Compatible-Artificial-Intelligence-Problem/dp/0525558616). Speaking for myself, although I have lived most of my life in the West, my philosophical groundings have probably drawn more from Indian or Asian roots on the whole. I’ll quote two famous sources, Hindu and Buddhist, that I think could contribute meaningfully to discussions.

The first is drawn from the Bhagavad Gita, Chapter 2, Verse: 47. The verse in question is —

कर्मण्येवाधिकारस्ते मा फलेषु कदाचन।
मा कर्मफलहेतुर्भूर्मा ते सङ्गोऽस्त्वकर्मणि॥ २-४७

In Roman script—

Karmanye vadhikaraste Ma Phaleshu Kadachana,
Ma Karmaphalaheturbhurma Te Sangostvakarmani

The meaning translated (by Gandhi) is “Your business is with the action only, never with its fruits.” (https://www.mkgandhi.org/articles/Mahatma-Gandhi-and-the-Bhagavad-Gita.php). This is arguably the original definition of what it means to be a Yogi; an agent that acts this way would not be driven by the desire to prolong its existence or to gather money or power.

The second comes from the Diamond Sutra:

“All bodhisattvas who sincerely seek the truth should control their minds by focusing on one thought only: ‘When I attain enlightenment, I will liberate all sentient beings in every realm of the universe and allow them to pass into the eternal peace of Nirvana. And yet, when vast, uncountable, unthinkable myriads of beings have been liberated, in reality no being has been liberated.’ Why? Because no one who is a true bodhisattva entertains such concepts as ‘self’ and ‘other.’ Thus, in reality there is no self to attain enlightenment and no sentient beings to be liberated” (https://zendust.org/wp-content/uploads/2023/10/Diamond-Sutra-redacted-for-sesshin-chanting-Michell-version2.pdf)

The Diamond Sutra is wonderfully deep and I won’t attempt much commentary beyond stating that I would rather have a powerful AGI become a bodhisattva focused on liberating all sentient beings rather than on more materialistic aims! Thich Nhat Hanh’s commentary is a good starting point to learn more of the Diamond Sutra (https://plumvillage.org/library/sutras/the-diamond-that-cuts-through-illusion). I’ll end the essay by saying that as a technologist and scientist, my fundamental driver is building technology that is useful for human beings. Our challenge as a society with increasingly powerful AI on the horizon is to bend it to ends that serve humanity rather than worsening inequality and suffering.

Interesting Links from Around the Web

• https://spectrum.ieee.org/hbm-on-gpu-imec-iedm: It may be possible to stack GPUs directly on top of high bandwidth memory.

About

Deep Into the Forest is a newsletter by Deep Forest Sciences, Inc. Get in touch with us at partnerships@deepforestsci.com.

Credits

Author: Bharath Ramsundar, Ph.D.

Editor: Sandya Subramanian, Ph.D.