The Unsolved Frontier: The Next Few Decades

October 15, 2026·10 min read

It’s October of 2026. Artificial Intelligence is an integral part of our lives now; from discussing personal feelings and emotions to getting work done in a more effective way, AI has revolutionized how we operate as a society. Interestingly, though, it does come with an unsettling feeling about who we are and what is valuable. For the longest time, the identity of a human existence has been about how we think and what we do with the thoughts that come to our mind. This philosophy has shaped pedagogy and academia in defining ways: an engineer thinks differently about a problem than a doctor, and a doctor thinks differently than a screenwriter. The essence of higher education has been a pursuit to shape the way we think. That still holds value in how we create effective AI-native systems. However, given the rapid pace of development of the technology, with major AI labs competing to create the next best model: some of which have been used to solve milestone math and biology problems¹, the question “Is there a moat in how we think?” is getting less obvious with time. People have likened the advent of AI to the Industrial Revolution: a surge in technology-driven automation led to efficiency gains in production chains. While it is true of the AI revolution, I think this is something much more fundamental, closer to discovering fire. If you think about it, we have really found a way to make sand think: we have passed electrons through silica and made it behaviorally simulate human thought. This is also driving the question, “What would our species look like if we didn’t have to think?” and in the absence of any right or wrong answers, the responses fall largely into two buckets: one where this is an existential crisis, and one where this is an existential opportunity. Much before any of those eventualities become obvious, however, we are likely to experience societal shifts in terms of how new technology evolves around AI, in the next few decades.

Robots everywhere!

Over the last few years, we have made tremendous progress in robotics: robots have gone from fumbling basic movement like traversing any complex terrain to dancing and volleying across obstacle courses². A lot of their growth has been amplified by the advent of Generative AI. Key challenges in Reinforcement Learning training of robots have been mitigated with grounded, synthetic, AI-generated environments: Waymo vehicles have benefited from being able to train in AI-generated versions of cities to mimic real world scenarios^{3, 4}, such as a snowy San Francisco (last time it actually snowed in San Francisco was in 1976!⁵), or unusual edge cases making up for the long tail of traffic possibilities hard to come by in real world captures, but essential for ensuring traffic safety. After all, Cruise had their license suspended based on one single incident originating from a human-driver accident⁶, a scenario extremely unlikely to take place, and something very hard to train for. Initial traction in robotics targeted manufacturing and packaging facilities, where the state space is more bounded. Increasingly, there is a shift towards consumer robotics: robots that can help small businesses perform menial human tasks such as making coffee or folding laundry. The robotics movement ultimately is gaining momentum in two different directions from here: military operations and consumer household help.

Whether autonomous armed combat is a desired outcome is a philosophical question, but I think the same question is finding representation in other areas as well. Asking “Who is responsible?” when a self driving car causes a loss of life is not entirely different from the same question posed about an autonomous militant pulling the trigger on a civilian or friendly troops. This is a vast oversimplification, of course: the dynamics of a battlefield, especially in civilian-dense areas are vastly more ethically nuanced than a car aiming to prevent any harm to any party involved. But the core question of responsibility is being defined more thoroughly. Even in a relatively milquetoast corporate setting, where AI agents are taking action on behalf of their users, similar questions almost always find representation.

Over the next few decades, we will see robots engaging more in households: watching children when parents are away, entertaining them with activities more conducive to their development than the current go-to for occupied parents: having them consume content on YouTube. With the advent of consumer self-driving cars (as opposed to the current transportation-as-a-service model by the likes of Waymo), we might find consumer robots engaging in the outer society even more naturally: picking up laundry from the laundromat or getting groceries and arranging them into the refrigerator. When AI first started gaining traction, a lot of people dreamed of a world where their chores would be taken care of by AI, and they would be free to spend time with arts and intellectual pursuits. Thus far, AI has had quite the opposite effect: it automated tasks that are intellectually stimulating, while leaving the chores for humans to do. With consumer robotics, this promise is more tractable, and a better version of that future seems to be achievable. But as AI moves from generating information to taking actions in the physical world, the question of whether it can do something increasingly gives way to a harder one: whether we can trust it to do so, and who is responsible when it gets something wrong.

Trust redefined

The concept of trust in the data we consume has materially shifted in the last few years. Digital media generated by AI are harder to distinguish from real ones than ever before. Soon, if not already, AI content will be virtually indistinguishable from human-generated content. This has raised serious questions about what will count as evidence for a claim. Think of grassroots journalism, for instance, reporting on-ground incidents under an authoritarian government cracking down on dissidents while denying any such allegations. In a pre-AI world, we counted raw video footage as a piece of evidence for those incidents to have happened. In a post-AI world, all such videos and recordings could be dismissed merely by the technology being able to generate “deepfakes”. In that world, our notions of trust need careful evaluation: do we only trust established international journalists for verified coverage? If we do, what happens when a government bans those select few journalists from accessing the grounds? In this scenario, it’s easier for governments to stifle dissent, and control the narrative much more strongly than they have been able to so far. Even if an investigative journalist from reputable press manages to get on the ground and record footage, that footage can also be dismissed by leadership under pressure from the said regime. This raises a question about how we create a world where we trust the information we consume.

Some companies are approaching this in a different way. Apple recently released a feature on its newest iPhones, called Reference Image^{7, 8}, where captured pictures are cryptographically signed to establish proof of validity. Earlier, in February 2021, a coalition of companies namely Adobe, Microsoft, Truepic, Arm, Intel, and BBC launched a standard called C2PA, meant to cryptographically verify provenance for digital content⁹. The first camera to adopt that standard was the Leica M11-P, which was released in October, 2023^{10, 11}. This has since been adopted by various popular manufacturers like Sony, Canon and Nikon¹². While cryptographic provenance authenticates a particular file, it does not necessarily survive every way in which we consume and redistribute media. A screenshot of a verified photograph, for instance, is itself a new and unverifiable image. As content gets copied, compressed, reposted, or captured from one medium into another, the authenticity chain can break. C2PA attempts to mitigate this through secondary ‘soft binding’ mechanisms, either by embedding an imperceptible watermark or by computing perceptual fingerprints. Over time, with LLMs being trained on watermarked content however, it is possible LLMs also learn to generate images with said signatures, making soft-binding based verification less effective. A more robust long-term solution may therefore depend on cooperation among the platforms through which media is distributed. A social network receiving a cryptographically authenticated image could verify it at the point of upload and cryptographically attest that any resized, compressed, or otherwise transformed version it produces was derived from that verified original. If such provenance is preserved as content moves between platforms, the chain of authenticity could survive ordinary internet distribution without relying exclusively on soft verification mechanisms. Critically, however, this does not entirely solve the issue with journalistic verification: if capture-level authentication remains limited to relatively expensive or export-controlled devices from a select set of manufacturers, it may not benefit less affluent populations relying on locally accessible equipment without such verification systems.

The issue of trust is also increasingly present in open-source contributions. Traditionally, a contributor who is confident in their contribution would make a Pull Request — the mechanism to contribute to a repository on GitHub — and the maintainers of the repository would have to review the contribution, and assess its validity. In a world where AI writes most of the codebase, writing code is easier and faster than ever before, leading to a much higher volume of potential Pull Requests on open-source repositories¹³. In this world, the developer’s role has shifted from writing code to reading code and making sure it makes sense, and the task of reading Pull Requests and telling valid contributions from otherwise can seem like a contributor offloading their task onto the repository maintainer. One way to address it would be for open-source repositories using specialized agents themselves to check the validity of contributions, but this effort adds cost overhead that may raise questions about who pays for it for open-source projects, and the possibility of biases introduced by code reviewed by the same LLM as the one that wrote it, even when the review system itself is capable enough.

In the end, how we trust in the future is anyone’s guess, but we need to get it right. If we fail to do so, we risk creating a paranoid, chronically suspicious, and uncomfortable society. This problem becomes even more important as AI systems gain greater agency. An assistant that merely answers a question requires relatively little trust. One that acts on our behalf, manages our homes, moves through the physical world, or makes decisions using intimate knowledge of our lives requires considerably more. Thus, the question of trust is not only about distinguishing human content from AI-generated content. It is also about whom we are willing to give access to the information these systems need in order to be useful.

A new Home Appliance: compute clusters

For many decades leading up to now, we have not had the widespread introduction of a new home appliance. In the 1930s we had the refrigerator, in the 50’s, we had the washing machine, in the 70’s we had the dishwasher, and the microwave in the decade after. All of these served to make our life at home easier, more convenient, and did so in a transformative way. We finally are on the cusp of a new Home Appliance: a compute cluster.

Recently, there has been some traction with consumer agentic products: AI systems that can do tasks for the consumer. It started with OpenClaw, an open-source system that took the world by storm in 2025¹⁴. Since then, we have seen Meta release Muse¹⁵, OpenAI release Dots¹⁶, and other less major players with their own offerings. There is a fundamental flaw in this model: an agent that truly works for an individual needs access to sensitive information that the large SaaS platforms should not be entrusted with. For such an agentic consumer product to be truly effective, it would need access to our financial statements, home security systems, home IoT systems, private conversations, none of which, I would posit, is suitable to be handed over to a research lab for inference.

One plausible answer is to move the intelligence, and much of the data it needs, into the home itself. This machine would be like any computer: a cluster of CPU, GPU and memory, but unlike most computers, serving inference to other devices over the internet in an encrypted manner. This cluster could be powerful enough to run capable open-source models (or eventually licensed closed source models), to the tune of 500 - 700B parameters at a reasonable token rate, and household members (and robots) could connect to it over the internet for LLM inference from their computers, phones, smart glasses, and so on. As of today, this model remains cost prohibitive for most households: the memory shortage has driven memory costs through the roof, forcing even companies like Apple to raise their prices on consumer hardware offerings^{17, 18}. However, in the next few decades, as memory costs stabilize, smaller model performance improves, and open-source models become competitive against closed source offerings by research labs, this reality is not far fetched. When this happens at scale, like the appliances that transformed our society in the decades past, this too shall transform our daily lives, potentially at a far greater scale than the ones before. The power to leverage private AI for work, convenience and entertainment would enable our society to be far more productive and contribute substantially to the economy at large.

Conclusion

I have always been an optimist, placing my confidence in the ability of our resilient species to figure out a way to solve whatever problems come its way, and this does not change with the advent of AI. These next few decades bring a lot of promise, but perhaps like a problematic +1 that joins the party, also invite the potential for tremendous harm. As AI becomes capable not only of thinking with us but acting for us, the questions of responsibility, trust, and control will become increasingly difficult to separate from the technology itself. Robots may give intelligence a physical presence in our lives, the erosion of reliable evidence may force us to rethink how trust is established, and increasingly capable personal agents may push us toward owning more of the infrastructure on which that intelligence runs. With AI, we have the potential to empower ourselves at a far greater capacity than anything else before, and we as a species must capitalize on this existential opportunity.