Long-Term Possibilities: 2035 and Beyond

Learning Objectives

• Explain what AGI means, why there is no consensus definition, and why expert timelines vary so dramatically
• Articulate the AI alignment problem and why serious researchers consider it a genuine technical and governance challenge
• Evaluate speculative long-term AI scenarios with appropriate skepticism while understanding what makes them worth considering

The Epistemological Problem with Long-Range Prediction

Before engaging with the long-term possibilities, it is worth being direct about the quality of the evidence base: nobody reliably predicts AI trajectories ten years out.

The forecasters who have been most consistently right about AI over the past decade are not the ones who made the boldest predictions. They are the ones who correctly identified which underlying trends would continue and which would hit unforeseen obstacles — and who were honest about uncertainty.

The history of AI is full of confidently wrong long-term predictions:

• "Human-level AI by 1980" — from the Dartmouth Conference founders (wrong)
• "AI will never match humans at chess" — wrong by 1997
• "LLMs can't reason" — contested by GPT-4
• "Full self-driving cars by 2020" — wrong
• "GPT-3 is the last paradigm before AGI" — probably wrong (GPT-5.5 now exists with dramatically expanded capabilities)

This section presents serious possibilities that serious researchers are actively working on and thinking about. It does not claim to tell you what will happen. It gives you frameworks for thinking about what might — which is the most honest thing one can do.

Artificial General Intelligence: What It Actually Means

Artificial General Intelligence (AGI) is one of the most debated concepts in the field — partly because it is used to mean different things by different people.

The core idea: AI systems that can match human-level performance across all cognitive domains. Not just the specific tasks they are trained on, but novel situations, creative problem-solving, self-directed learning, and transfer across domains.

The definitional problem is significant:

• Systems that pass a comprehensive behavioral test across diverse domains?
• Systems that can perform any cognitive task a human can, given appropriate tools and context?
• Systems that can autonomously conduct scientific research at the frontier?
• Systems that can improve their own capabilities without human guidance?

Each of these definitions implies different thresholds and different timelines. "AGI" is therefore less a precise technical target than a cluster of capabilities that might emerge together or separately.

The Timeline Debate

The range of expert opinion on AGI timelines is extraordinary — and has shifted notably in 2025–2026:

OpenAI has developed an internal 5-level framework: (1) Chatbots, (2) Reasoners, (3) Agents, (4) Innovators, (5) Organizations. As of early 2026, they assess their models at Level 2–3, with Altman already pivoting company messaging toward superintelligence as the next goal. Anthropic's formal filing to the White House Office of Science and Technology Policy (March 2025) stated they "expect powerful AI systems will emerge in late 2026 or early 2027."

This range reflects genuine technical disagreement, not just different confidence levels about the same prediction. The debate has shifted from "when will AGI happen?" toward definitional arguments — no consensus definition exists, and the same capability can be called "AGI" or "not even close" depending on who is speaking.

A more tractable question: rather than "when will AGI arrive?", ask "which specific human cognitive capabilities will AI systems acquire, in what order, and at what pace?" The transition to highly capable AI is already underway — it is not waiting for a single threshold crossing.

The Alignment Problem

AI alignment refers to the challenge of ensuring that AI systems reliably pursue goals that benefit humans — not just in the specific tasks they are trained on, but across the full range of situations they might encounter as they become more capable.

The concern is not that AI will spontaneously become malicious. It is more subtle and technically grounded: that as AI systems become more capable, small misspecifications in their objectives or training could lead to behaviors that are technically optimal by the AI's internal metrics but harmful by human values.

The instrumental convergence thesis: An AI system with almost any goal will develop sub-goals that are convergent across different objective functions — such as self-preservation, resource acquisition, and resistance to being shut down or modified. This is because these sub-goals are useful for achieving almost any terminal goal. A capable AI system optimizing for an objective that is subtly wrong could, in principle, resist correction.

The specification problem: Correctly defining what we want from an AI system is harder than it appears. "Be helpful" — what does that mean across all possible situations? "Maximize user engagement" — this is measurable, but leads to recommendation systems that promote outrage. Every proxy measure we can specify differs from what we actually care about.

Scalable oversight: As AI systems become more capable, maintaining meaningful human oversight becomes harder. If a system is significantly more capable than its human supervisors in the domain being supervised, how do you check whether it is doing the right thing? The AI safety field is actively researching techniques like debate (AI systems argue for and against their own outputs, with humans judging), recursive reward modeling, and interpretability research (understanding what AI systems are actually computing internally).

Anthropic, OpenAI's safety team, DeepMind's safety team, the Center for AI Safety, and dedicated research institutes (Redwood Research, ARC) are all working on alignment. The OWASP Top 10 for Agentic Applications (2026) codified the security risks specific to autonomous AI agents, and the NIST AI Agent Standards Initiative (launched February 2026) is developing formal evaluation frameworks. The field has made genuine progress. The core problem is not solved, and the difficulty is expected to increase as capabilities scale.

Artificial Superintelligence: The Most Speculative Horizon

Artificial Superintelligence (ASI) refers to systems that substantially exceed human intellectual performance across all domains — including the ability to improve their own capabilities.

The argument for concern runs: if a system can improve its own capabilities, and if each improvement makes the next improvement faster, this creates a positive feedback loop where capability gains could be rapid and could produce systems that humans cannot meaningfully control or understand. This is the "intelligence explosion" scenario described by I.J. Good in 1965 and popularized by Nick Bostrom.

The counter-arguments are also serious:

• There may be fundamental limits on self-improvement; recursive capability gains may hit diminishing returns
• The analogy to historical technological transitions suggests that humans adapt and maintain meaningful agency
• The entire scenario may rest on architectural assumptions about AI self-improvement that don't hold

If advanced AI is developed before alignment is solved, some researchers argue the risk of catastrophic outcomes is non-trivial. Others argue this framing is premature given actual current capabilities, and that focusing on speculative superintelligence scenarios diverts attention from near-term, concrete harms.

Engaging seriously with this debate requires reading primary sources — Nick Bostrom's Superintelligence, Stuart Russell's Human Compatible, Brian Christian's The Alignment Problem — not just accepting any single narrative.

Economic Transformation: The Post-Scarcity Question

One long-term economic scenario — favored by some techno-optimists — is that AI-driven automation of cognitive work could produce something like post-scarcity for knowledge work: the cost of generating text, code, analysis, legal reasoning, and creative work approaches zero, in the same way that digital reproduction has made the marginal cost of copying music or software negligible.

If this scenario plays out fully, it raises profound questions about the economic model underlying knowledge work. Historical technological transitions — industrialization, computerization — created new jobs while displacing old ones. The question economists actively debate: is there reason to believe AI will follow this pattern, or does it represent a qualitatively different challenge because it affects cognitive work rather than only physical and routine cognitive work?

The Universal Basic Income (UBI) debate has gained new urgency in this context. As of 2025, 122 guaranteed basic income pilots across 33 US states have tested variations of guaranteed income, allocating approximately $481 million to over 40,000 recipients. The most consistent finding: significant improvements in mental health — reduced anxiety, depression, and domestic violence. Employment effects are mixed: smaller pilots show slight employment increases, while larger pilots (500+ participants) show modest decreases of about 3 percentage points. Ireland made its artist UBI pilot permanent after finding EUR 1.39 returned for every EUR 1 invested. But whether these experiments generalize to the fiscal and political challenges of large-scale UBI implementation — particularly in the context of AI-driven displacement of white-collar work — remains deeply contested.

AI-Accelerated Longevity

Longevity therapeutics — research aimed at extending healthy human lifespan by intervening in the biological mechanisms of aging — is a serious scientific field with unprecedented funding.

Altos Labs (launched with $3 billion — the largest biotech debut ever, backed by Jeff Bezos and Yuri Milner) is pursuing partial epigenetic reprogramming using Yamanaka factors to reverse cellular aging. Mouse studies have shown aged donor kidneys can be rejuvenated through reprogramming before transplantation. Altos appointed a Chief Medical Officer in 2025 and reportedly began early human safety testing in August 2025 — 2026 marks the year Altos "finally makes contact with human reality" after years of preclinical work.

Calico (backed by Alphabet) has faced setbacks: its lead drug fosigotifator failed a Phase II/III ALS trial in January 2025, and AbbVie terminated its 11-year collaboration in November 2025, laying off approximately 100 scientists. Calico is now pursuing a $571 million partnership with Mabwell Bioscience for age-related disease therapeutics — but the path has been harder than expected.

The AI connection: drug discovery and biological modeling tools are potentially transformative for longevity research. Isomorphic Labs' Drug Design Engine (see Module 10.2) demonstrates the kind of capability acceleration that could compress aging intervention discovery timelines. If AI can do for longevity what AlphaFold did for protein structure prediction, some researchers believe meaningful therapeutics could become available within twenty years.

Quantum Computing: A Longer Horizon

Quantum computing is frequently invoked as a potential transformative force for AI — particularly for matrix multiplication operations that dominate neural network training. Quantum processors could in principle accelerate certain computations exponentially.

The field reached a genuine inflection point in late 2024: Google's Willow chip (105 qubits) demonstrated "below-threshold" quantum error correction for the first time — the point where adding more qubits improves rather than worsens computation quality. Willow completed a benchmark in 5 minutes that would take the fastest supercomputer 10 septillion years. In February 2025, Microsoft unveiled Majorana 1, the first quantum processor powered by topological qubits, designed to scale to 1 million qubits on a single chip. IBM connected three quantum chips into a 4,158-qubit system, and IonQ demonstrated practical quantum advantage on a medical device simulation that outperformed classical HPC by 12%.

However, quantum computers are still far from practical AI workloads. Logical error rates (~0.14% per cycle on Willow) remain orders of magnitude above the 10^-6 levels needed for large-scale quantum algorithms. Each logical qubit may need 1,000+ physical qubits. DeepMind's AlphaQubit is using AI to improve quantum error correction itself — achieving 30% error reduction — creating an interesting AI-quantum feedback loop.

Quantum computing is a real and rapidly advancing field. Google targets 1 million physical qubits; IBM targets a fault-tolerant quantum computer by 2029. The timeline for meaningful quantum advantage on AI workloads is compressing but remains years away. For planning purposes, treat it as a development that has moved from theoretical to early engineering — still not a near-term factor for AI training, but no longer purely aspirational.

Democratic Institutions and the Adaptation Challenge

One of the less-discussed long-term challenges is institutional: democratic systems were not designed to govern technologies that evolve on AI's timescales. Legislatures take years to pass regulations. Courts take years to resolve cases. Regulatory agencies require long rulemaking processes. These deliberative timescales are appropriate for many purposes and may be actively good for preventing regulatory capture — but they create friction when governing technologies that can change fundamentally in 12–18 months.

If capability jumps happen faster than legitimate democratic processes can respond, important governance decisions may be made by default — by the companies building the systems and the governments that happen to move first — rather than through democratic deliberation. This is a structural challenge for which there is no simple answer, but it is worth understanding.

Key Takeaways

• AGI is a contested concept with no consensus definition; expert timelines range from before 2030 to never, reflecting genuine technical disagreement, not just different confidence levels
• The AI alignment problem — ensuring increasingly capable systems reliably pursue human-beneficial goals — is a technically serious, actively researched challenge with real but incomplete progress
• Long-term economic transformation scenarios, from knowledge work post-scarcity to UBI debates, are genuinely uncertain; the historical pattern that new technology creates new jobs may or may not hold at AI's pace and scale
• All long-term AI possibilities should be held with significant uncertainty — intellectual humility, not certainty in either the optimistic or pessimistic direction, is the appropriate stance