Novel / Unknown Approaches

Capability

Novel / Unknown Approaches

Analyzes probability (1-15%) of novel AI paradigms emerging before transformative AI, systematically reviewing historical prediction failures (expert AGI timelines shifted 43 years in 4 years, 13 years in one survey cycle) and comparing alternative approaches like neuro-symbolic (8-15% probability), SSMs (5-12%), and NAS (15-30%). Concludes current paradigm faces quantified limits (data exhaustion ~2028, compute costs approaching economic constraints) but near-term timelines favor incumbent approaches.

3.3k words · 1 backlinks

Key Links

Source	Link
Official Website	stockholmresilience.org
Wikipedia	en.wikipedia.org

Overview

This category represents the probability mass we should assign to approaches not yet discovered or not included in our current taxonomy. History shows that transformative technologies often come from unexpected directions, and intellectual humility requires acknowledging this. The field of AI has undergone cyclical periods of growth and decline, known as AI summers and winters, with each cycle bringing unexpected architectural innovations. We are currently in the third AI summer, characterized by the transformer paradigm, but historical patterns suggest eventual disruption.

The challenge of forecasting AI development is well-documented. According to 80,000 Hours' analysis of expert forecasts, mean estimates on Metaculus for when AGI will be developed plummeted from 50 years to 5 years between 2020 and 2024. The AI Impacts 2023 survey found machine learning researchers expected AGI by 2047, compared to 2060 in the 2022 survey. This 13-year shift in a single year demonstrates the difficulty of prediction in this domain.

Beyond the "known unknowns" such as scaling limits and alignment challenges, we face a vast terrain of "unknown unknowns": emergent capabilities, unforeseen risks, and transformative shifts that defy prediction. The technology itself is evolving so rapidly that even experts struggle to predict its capabilities 6 months ahead.

Estimated probability of being dominant at transformative AI: 1-15% (range reflects timeline uncertainty; shorter timelines favor current paradigms, longer timelines favor novel approaches)

Why Include This Category

Loading diagram...

Arguments for Allocating Probability Here

Argument	Explanation	Historical Evidence
Historical track record	Major breakthroughs often unexpected	Transformer attention mechanism existed since 2014; breakout came in 2017
Epistemic humility	We don't know what we don't know	Expert AI timeline estimates shifted 13 years in one survey cycle
Active research	Many smart people working on new ideas	63% of neuro-symbolic papers focus on learning/inference innovation
Combinatorial space	Possible architectures vastly exceed explored	NAS tools discovering architectures matching human-designed ones
Scaling approaching limits	Current paradigm may hit ceiling	Epoch AI predicts high-quality text data exhausted by 2028

Arguments Against High Probability

Argument	Explanation	Supporting Evidence
Current approaches working	Transformers haven't hit hard ceiling	Training compute grew 5x/year 2020-2024
Incremental progress	Breakthroughs usually build on existing work	Gen AI built on cloud, which built on internet
Selection effects	Best ideas tend to be discovered early	Attention, backprop, deep networks all pre-2000 concepts
Time constraints	Limited years until TAI (if near)	Median expert estimate: AGI by 2047
Investment momentum	$109B US AI investment in 2024	Massive resources dedicated to current paradigm

Historical Precedents

The history of technology provides crucial context for estimating the probability of paradigm shifts. As documented by research on technological paradigm shifts, notable figures consistently fail to predict transformative changes. Wilbur Wright famously said in 1901 that "man would not fly for 50 years"; two years later, he and his brother achieved flight.

Past Paradigm Shifts in AI

Shift	Year	From	To	Lead Time	Was It Predicted?	Impact
Neural network revival	2012	Symbolic AI	Deep learning	30+ years	Partially (by few)	AlexNet: 15% error reduction on ImageNet
Attention/transformers	2017	RNNs/CNNs	Transformers	3 years (attention existed 2014)	Somewhat surprising	Enabled 100B+ parameter models
Scaling laws	2020	"Need new ideas"	"Just scale"	N/A	Surprising to many	Kaplan et al. showed predictable improvement
In-context learning	2020	Fine-tuning	Prompting	N/A	Not predicted	GPT-3 few-shot emerged unexpectedly
RLHF effectiveness	2022	Supervised only	RLHF	5 years	Somewhat expected	ChatGPT achieved 100M users in 2 months
Reasoning models	2024	Pre-training focus	Post-training scaling	N/A	Not predicted	Novel RL techniques changed compute allocation

Expert Forecasting Track Record

Forecast Source	Year Made	Prediction	Actual Outcome	Error
Metaculus AGI median	2020	≈2070	Now estimate ≈2027	43 years shift
AI Impacts survey	2022	AGI by 2060	Updated to 2047 (2023)	13 years shift
LEAP panel superforecasters	2024	MATH benchmark 14% by 2026	GPT-5.2 achieved 33% in 2025	2.4x underestimate
FrontierMath experts	2024	31% accuracy by end 2025	29% achieved Aug 2025	Roughly accurate

Lessons from History

Lesson	Implication	Quantified Example
Old ideas revive	Attention was known; transformers made it work	3-year gap between attention (2014) and transformers (2017)
Combinations matter	Transformer = attention + layernorm + scale	Multiple paradigms combine to create breakthroughs
Empirical surprises	In-context learning emerged unexpectedly	Zero capability below ≈1B params, then emergent
Scaling surprises	Scaling laws weren't obvious a priori	5x/year compute growth 2020-2024
Experts underestimate	Specialists often wrong about own field	Wilbur Wright: "50 years", achieved in 2

Potential Sources of Novelty

Paradigm Shift Candidates Comparison

The following table compares the most promising alternative paradigms based on current research momentum and potential impact.

Paradigm	Maturity	Research Momentum	Key Advantage	Key Limitation	Est. Probability of Dominance by 2040
Neuro-Symbolic AI	Growing	63% of papers focus on learning/inference	Combines reasoning + learning	Scalability/joint-training remains "holy grail"	8-15%
State Space Models	Early	Mamba, RWKV active development	Linear complexity vs quadratic attention	Haven't matched transformer performance at scale	5-12%
Neural Architecture Search	Maturing	NASNet, EfficientNet production-ready	AI-designed architectures	Often optimizes within existing paradigms	3-8%
Neuromorphic Computing	Early	Intel Loihi, IBM TrueNorth	1000x energy efficiency	Software ecosystem immature	2-5%
Quantum ML	Nascent	NISQ-era experiments	Exponential state space	Coherence, error correction unsolved	1-3%
World Models	Growing	Video prediction, robotics	Causal understanding	Data requirements unclear	5-10%
True Unknown	N/A	N/A	Cannot be characterized	Cannot be characterized	1-5%

Areas Where Breakthroughs Might Emerge

Area	Potential	Current Status	Key Research Groups	Timeline Estimate
Learning algorithms	Beyond backprop/SGD	Active research	DeepMind, Anthropic	3-7 years
Architectures	Beyond attention	SSMs gaining traction	Mamba team, RWKV	2-5 years
Objective functions	Beyond token prediction	Minimal progress	Academic labs	5-10 years
Training paradigms	Beyond supervised/RL	Post-training scaling emerging	OpenAI, Anthropic	1-3 years
Hardware-software co-design	Novel compute substrates	Neuromorphic, analog	Intel, IBM, startups	5-15 years
AI-for-AI	AI designing AI	AutoML/NAS advancing	Google, Microsoft	2-5 years

Specific Speculative Directions

Direction	Description	Current Evidence	Probability of Major Impact	Key Uncertainties
Algorithmic breakthroughs	New training methods beyond gradient descent	Forward-forward algorithm (Hinton 2022)	10-25%	Whether alternatives can match scale
Physics-based computing	Quantum, analog, optical	Google quantum supremacy claims	3-8%	Error correction, coherence
Biological insights	From neuroscience	Sparse coding, predictive processing	5-15%	Translation to algorithms
Emergent capabilities	Unexpected abilities at scale	In-context learning, chain-of-thought	Ongoing (certain)	Which capabilities next
AI-discovered AI	AI designs better architectures	NAS matches human designs	15-30%	Search space definition
Causal/world models	Move beyond correlation	Causal AI research growing	10-20%	Scalable causal inference

Paradigm Evolution Dynamics

The following diagram illustrates potential pathways for paradigm evolution, including both incremental improvements and discontinuous shifts.

Loading diagram...

What Novel Approaches Might Look Like

Possible Characteristics

Characteristic	Explanation	Current Paradigm Comparison	Historical Precedent
More efficient	Orders of magnitude less compute	GPT-4: ≈10^25 FLOP training	DeepSeek: 95% fewer resources claimed for similar performance
Different training	Not gradient descent	Backprop since 1986	Forward-forward algorithm (Hinton 2022)
Different objectives	Not next-token prediction	Autoregressive LLMs dominant	World models, energy-based models
Different hardware	Not GPUs	NVIDIA dominates	Neuromorphic: 1000x energy efficiency potential
Different capabilities	Strong at what transformers struggle with	Reasoning, planning, efficiency	Neuro-symbolic: explicit reasoning

Current Paradigm Constraints (Drivers of Potential Shift)

According to Epoch AI's scaling analysis, the current paradigm faces several quantifiable constraints:

Constraint	Current Status	Projected Exhaustion	Implication
Training Data	High-quality text near exhaustion	2028 median estimate	New data sources or paradigms needed
Compute Costs	$7 trillion infrastructure proposal (Altman 2024)	Investors prefer 10x increments	Economic limits approaching
Energy	Data centers need 32% yearly growth	Grid capacity constraints	Physical infrastructure bottleneck
RL Scaling	Labs report 1-2 year sustainability	Compute infrastructure limits	Post-training gains may plateau
Model Size	GPT-4: ≈1.8 trillion params (estimated)	Diminishing returns observed	Architecture efficiency matters more

Warning Signs We Might Miss Something

Sign	What It Suggests	Quantified Evidence
Fundamental capability ceilings	Current approaches hitting limits	Reasoning models required novel techniques beyond scaling
Efficiency gaps with biology	Brains use far less energy	Human brain: ~20W; GPT-4 inference: ≈100kW
Certain tasks remain hard	Reasoning, planning, learning efficiency	Neuro-symbolic needed for explicit reasoning
Theoretical gaps	Don't understand why current methods work	Only 5% of neuro-symbolic papers address meta-cognition
Benchmark saturation	Easy benchmarks solved	GPT-5.2 hit 33% on LiveCodeBench Pro

Safety Implications

A paradigm shift in AI development would have profound implications for AI safety research. The Stanford HAI AI Index 2025 notes that safety research investment trails capability investment by approximately 10:1. A novel paradigm could either invalidate existing safety research or provide new opportunities for alignment.

Why Novel Approaches Are Concerning

Concern	Explanation	Risk Level	Mitigation Difficulty
Unpredictability	Can't prepare for unknown risks	High	Very High
Rapid capability jumps	New paradigm might be much more capable	Very High	High
Different failure modes	Safety research might not transfer	High	Medium
Misplaced confidence	We might assume current understanding applies	Medium	Low
Compressed timelines	Less time to develop safety measures	Very High	Very High
Open-source proliferation	Novel techniques spread faster than safety measures	High	High

Why They Might Be Better

Potential Benefit	Explanation	Probability	Example
Designed for safety	New approaches could prioritize interpretability	15-25%	Neuro-symbolic: 28% papers address explainability
Different incentives	Might emerge from safety-focused research	10-20%	Interpretability-first architectures
Better understanding	New paradigms might be more theoretically grounded	20-30%	Causal AI provides formal guarantees
Natural alignment	Could have built-in alignment properties	5-15%	Symbolic reasoning more auditable
Efficiency enables safety	More compute for alignment research	25-35%	If 10x more efficient, more safety testing possible

Safety Research Transferability by Paradigm

Current Safety Research Area	Neuro-Symbolic	SSMs	Neuromorphic	Unknown
Interpretability	High transfer	Medium	Low	Unknown
RLHF/Constitutional AI	Medium	High	Low	Unknown
Formal verification	Very High	Medium	Medium	Unknown
Scalable oversight	Medium	High	Low	Unknown
Deceptive alignment detection	Low	Medium	Low	Unknown

Research Questions

What Should We Monitor?

Area	What to Watch	Key Indicators	Monitoring Frequency
Academic ML	Novel architectures, theoretical results	ArXiv papers, NeurIPS/ICML proceedings	Weekly
Industry labs	Unpublished breakthroughs	Hiring patterns, patent filings, leaked benchmarks	Monthly
Interdisciplinary	Physics, neuroscience, mathematics	Cross-disciplinary conferences, Nature/Science publications	Quarterly
AI-for-AI	AI systems discovering new AI methods	NAS/AutoML progress, AI-generated code quality	Monthly
Hardware developments	Novel compute substrates	Chip announcements, energy efficiency benchmarks	Quarterly
Scaling signals	Evidence of plateaus or breakthroughs	Epoch AI tracking, benchmark progress	Continuous

How to Prepare for Unknown Unknowns?

Strategy	Rationale	Investment Level	Priority
General safety research	Focus on principles that transfer	High	Critical
Monitoring infrastructure	Track developments broadly	Medium	High
Paradigm-agnostic alignment	Don't overfit to transformer-specific approaches	High	Critical
Worst-case planning	Assume capabilities might jump unexpectedly	Medium	High
Rapid response capacity	Ability to pivot safety research quickly	Medium	Medium
Diverse research portfolio	Fund safety research across multiple paradigms	High	High

Key Monitoring Organizations

Organization	Focus	Update Frequency	URL
Epoch AI	Compute trends, scaling analysis	Weekly	epoch.ai
LEAP Panel	Expert forecasts on AI development	Monthly	forecastingresearch.org
AI Index (Stanford HAI)	Comprehensive AI metrics	Annual	hai.stanford.edu
Metaculus	Prediction markets on AI timelines	Continuous	metaculus.com
80,000 Hours	AI safety career/research priorities	Quarterly	80000hours.org

Bayesian Reasoning

How to Update

Observation	Update Direction	Magnitude	Current Signal (2025)
Transformers continue scaling	Novel approaches less likely near-term	-3 to -5%	5x/year growth continuing
Hard ceiling hit	Novel approaches more likely	+10 to +20%	Not yet observed
Data exhaustion	Novel approaches more likely	+5 to +10%	2028 estimate approaching
Theoretical breakthrough	Pay attention to specific direction	Variable	Neuro-symbolic momentum
AI discovers better architecture	Accelerates unknown-unknown risk	+5 to +15%	NAS producing competitive models
Major lab pivots to new approach	Strong signal	+15 to +25%	Not observed

Probability Estimate by Timeline

Timeframe	Probability of Novel Paradigm Dominance	Key Assumptions	Confidence
By 2027	1-3%	Current scaling continues; no major breakthroughs	Medium
By 2030	5-12%	Data/compute limits start binding; research progresses	Medium
By 2035	10-20%	Current paradigm hits fundamental limits	Low
By 2040	15-30%	Long timeline allows paradigm maturation	Low
By 2050+	25-45%	Historical base rate of paradigm shifts	Very Low

Why 1-15% Range Is Reasonable

The range reflects uncertainty about timelines and paradigm persistence:

Lower bound (1%): If transformative AI arrives within 3-5 years via current paradigm scaling, novel approaches have insufficient time to mature. The median Metaculus estimate of AGI by ~2027 supports this scenario.

Upper bound (15%): If current paradigm hits hard limits (data exhaustion, scaling saturation) before transformative AI, alternative approaches become necessary. Epoch AI projections of 2028 data exhaustion support this possibility.

Central estimate (5-8%): Accounts for historical base rate of paradigm shifts (~1 per decade in computing), current research momentum in alternatives, and uncertainty in both timelines and scaling projections.

Critical Questions

Uncertainty	Scenarios	Current Evidence	Resolution Timeline
How locked-in is the current paradigm?	Fundamental (like the wheel) vs. Transitional (like vacuum tubes)	Transformer dominance 7+ years suggests maturity	2-5 years
How much does understanding matter?	Empirical scaling sufficient vs. Theory needed for next leap	Deep learning theory still immature	Unclear
Will AI-discovered AI come before TAI?	Yes (accelerates) vs. No (current paradigm dominates)	NAS producing competitive models	2-4 years
How would we recognize a breakthrough?	Clear benchmark jump vs. Gradual realization	Historical: transformers looked incremental initially	Retroactive
What are the true scaling limits?	Near current frontier vs. Orders of magnitude remaining	Epoch: 2e29 FLOP feasible by 2030	3-5 years
Will safety concerns force paradigm change?	Interpretability needs drive alternatives vs. Current approaches adapted	28% of neuro-symbolic papers address explainability	Ongoing

Scenario Analysis

Scenario	Probability	Key Trigger	Implications for Safety
Transformer dominance continues	55-70%	Scaling continues working; no hard limits	Current safety research remains relevant
Hybrid integration (Transformer + Neuro-symbolic)	15-25%	Reasoning limitations drive integration	Safety approaches must span paradigms
Gradual SSM/alternative transition	5-12%	Efficiency requirements dominate	Moderate adaptation of safety research
Discontinuous breakthrough	3-8%	Fundamentally new approach discovered	Major safety research pivot required
AI-designed paradigm	5-10%	NAS/AutoML produces novel architecture	Accelerated timeline; compressed safety window

Sources & Resources

Primary Research

Source	Type	Key Finding	Year
Epoch AI: Can AI Scaling Continue?	Analysis	2e29 FLOP runs feasible by 2030; data exhaustion ≈2028	2024
Neuro-Symbolic AI 2024 Systematic Review	Survey	63% papers on learning/inference; 5% on meta-cognition	2024
LEAP Expert Panel	Forecasts	Experts underestimate AI progress on benchmarks	2024
80,000 Hours: AGI Timeline Review	Analysis	Metaculus median shifted from 50 years to 5 years (2020-2024)	2025
NAS Systematic Review	Survey	NAS producing architectures matching human designs	2024

Paradigm Shift Analysis

Source	Focus	Relevance
Paradigm Shifts in Tech (Medium)	Historical patterns	Technologies build upon predecessors
AI Paradigm Analysis (Taylor & Francis)	AI as paradigm shift	Pattern similarity to historical tech revolutions
Neuro-Symbolic AI Overview	Third AI wave	Hybrid approaches as potential successor
AllianceBernstein: AI Paradigm Shift	Investment perspective	Paradigm shift timing uncertainty

Forecasting and Uncertainty

Source	Focus	Key Insight
Our World in Data: AI Timelines	Expert surveys	13-year shift in AGI estimates (2022-2023 surveys)
The Problem with AGI Predictions	Prediction failures	Experts often wrong about own field
Clearerthinking: AI Disagreement	Methodology	Sources of forecasting disagreement
Science: AI and Unknown Unknowns	Uncertainty	Even experts struggle to predict 6 months ahead

Technical Directions

Source	Focus	Status
Neural Architecture Search Advances (NSR)	AutoML/NAS	AI designing AI architectures
Google 2025 Research Breakthroughs	Industry progress	Quantum, weather, scientific applications
FTI Consulting: AI Frontiers 2025	Research directions	Agentic AI, multimodal, reasoning
Neuro-symbolic for Robustness (Springer)	Hybrid approaches	Interpretability, uncertainty quantification

References

180,000 Hours AGI Timelines Review80,000 Hours·Benjamin Todd·2025▸

A comprehensive review of expert predictions on Artificial General Intelligence (AGI) from multiple groups, showing converging views that AGI could arrive before 2030. Different expert groups, including AI company leaders, researchers, and forecasters, show shortened and increasingly similar estimates.

★★★☆☆

80000hours.org

2Our World in Data: AI TimelinesOur World in Data▸

★★★★☆

ourworldindata.org

3Systematic review on neural architecture searchSpringer (peer-reviewed)▸

★★★★☆

link.springer.com

4Epoch AIEpoch AI▸

★★★★☆

epoch.ai

5Advances in neural architecture searchacademic.oup.com▸

academic.oup.com

6Google's 2025 Research BreakthroughsGoogle AI▸

★★★★☆

blog.google

7NAS Overviewautoml.org▸

automl.org

8Epoch AIEpoch AI▸

Epoch AI is a research organization collecting and analyzing data on AI model training compute, computational performance, and technological trends in artificial intelligence.

★★★★☆

epoch.ai

9MetaculusMetaculus▸

Metaculus is an online forecasting platform that allows users to predict future events and trends across areas like AI, biosecurity, and climate change. It provides probabilistic forecasts on a wide range of complex global questions.

★★★☆☆

metaculus.com

1080,000 Hours methodology80,000 Hours▸

★★★☆☆

80000hours.org

Novel / Unknown Approaches