Big Think Clips · Published 2026-09-03

The paradox at the heart of AI and science | Terence Tao

privateReview: unreviewedConfidence: high
artificial intelligencemathematicsproofsparadoxscienceproof-indigestionverification-bottleneckcurationprovenancetraining-pipelineserendipityexplainabilityoverfitting
Open on YouTube ↗

Summary

Overview

  • Speaker: Terence Tao
  • Channel: Big Think
  • Main topic: How AI is changing math and science
  • Purpose: To analyze the impact of AI on mathematics and science, weighing its immense problem-solving capabilities against the loss of foundational human intuition and process. Mathematician Terence Tao explores the paradox at the intersection of artificial intelligence, mathematics, and scientific discovery. While AI tools function like helicopters—transporting researchers directly to goals without teaching them the path—they simultaneously revolutionize research by solving complex problems, generating proofs, and automating calculations. However, this efficiency creates 'proof indigestion,' requiring human curation, editing, and fundamental paradigm shifts in how we approach scientific inquiry and education.

Topic Map

The Hiking Analogy of Science and AI

  • Explanation: Comparing traditional scientific exploration (hiking a trail with a map, discovering unexpected phenomena along the way) with AI-assisted research (flying in a helicopter directly to a waterfall).
  • Key claims:
    • Traditional science involves a process of discovery, getting lost, and finding unintended phenomena.
    • AI tools act like helicopters, flying researchers directly to a specific goal without teaching them how to get there.
    • While AI achieves goals efficiently, it risks losing the broader process and peripheral insights.
  • Examples:
    • Going on a hike, finding a waterfall, making a map, and discovering unexpected sights along the way.
    • Using AI to jump straight to a specific scientific conclusion.
  • Terminology:
    • hiking analogy
    • paradigm
    • efficiency
  • Why it matters: It highlights the tension between efficient problem-solving and deep, foundational understanding in scientific research.

The Evolution of Scientific Paradigms

  • Explanation: Tracing the evolution of science from theory and experiment, to simulation, big data, and now artificial intelligence.
  • Key claims:
    • Science traditionally relied on two paradigms: theory and experiment.
    • Simulation and big data introduced computational and data-driven scientific methods.
    • AI represents a new paradigm that transforms all previous modes of doing science.
  • Examples:
    • Kepler and Newton creating theories of planetary motion and gravity.
    • Gauss computing the first 100,000 prime numbers to predict the prime number theorem.
    • Simulating hurricanes in supercomputers.
  • Terminology:
    • theory
    • experiment
    • simulation
    • big data
    • prime number theorem
  • Why it matters: Understanding how AI fits into the historical continuum of scientific methodology.

Large Language Models in Mathematics

  • Explanation: How LLMs process language and predict tokens, and how this seemingly simple statistical mechanism translates into mathematical problem-solving.
  • Key claims:
    • LLMs predict the next word in a sentence based on massive multidimensional data spaces.
    • Language contains unspoken, unwritten patterns and rules that LLMs absorb.
    • With enough computing power, data, and iteration, LLMs begin to mimic intelligent reasoning and mathematical proof generation.
  • Examples:
    • 'Roses are red, violets are blue.'
    • Solving 2 + 3 = 5.
    • Autocompleting mathematical proofs.
  • Terminology:
    • large language models (LLMs)
    • regression
    • tokens
    • multidimensional space
    • curve fitting
  • Why it matters: Explains the underlying mechanism of AI models and why they can perform complex tasks despite not possessing conscious understanding.

The Paradox of AI in Science: Power vs. Understanding

  • Explanation: Examining the contradiction that AI is becoming vastly more powerful while human scientists do not get better at science just by using it.
  • Key claims:
    • AI is becoming more powerful, capable, and makes fewer mistakes.
    • AI operates without human scientists understanding the underlying steps or being able to explain why a discovery is interesting.
    • We may need to redesign our perception of what science is and what we want out of it.
  • Examples:
    • Debugging AI simulations without knowing how the solution was derived.
    • Getting proofs from AI that lack intuitive human explanation.
  • Terminology:
    • paradox
    • explainability
    • optimization
    • brute force
  • Why it matters: Raises fundamental questions about the nature of scientific progress, attribution, and human expertise.

Proof Indigestion and the Future of Curation

  • Explanation: The current state of mathematics where AI generates an overwhelming volume of proofs, leading to 'proof indigestion' and a heavy burden on human curation.
  • Key claims:
    • AI is generating an unprecedented volume of proposed solutions and proofs.
    • Human experts are experiencing 'proof indigestion' due to the sheer quantity of outputs to verify, triage, and curate.
    • The final textbook form of a proof still requires human editing, organization, and editorial polish.
  • Examples:
    • OpenAI solving a famous open question posed by Paul Erdős in 1946 regarding the planar unit distance problem.
    • Sorting through hundreds of AI-generated proofs of varying quality.
  • Terminology:
    • proof indigestion
    • triage
    • curation
    • planar unit distance problem
    • Erdős
  • Why it matters: Highlights the practical bottleneck in modern scientific publishing and peer review when AI accelerates output beyond human processing capacity.

Key Points

AI acts as a powerful problem-solving accelerator rather than a replacement for human curiosity.

  • Explanation: AI can rapidly test hypotheses, generate proofs, and solve difficult problems at scale, outperforming humans in raw speed and breadth.
  • Evidence: Recent OpenAI models solving longstanding mathematical problems like Erdős's planar unit distance problem.
  • Practical implication: Scientists must learn to integrate AI tools to accelerate routine tasks while preserving fundamental inquiry and critical thinking.

The risk of overfitting and lack of deep mechanistic understanding.

  • Explanation: AI models excel at pattern matching within training data but may fail to extrapolate or understand underlying causal mechanisms.
  • Evidence: Historical examples like Kepler moving from Platonic solids to elliptical orbits because data precision revealed flaws in rigid theories.
  • Practical implication: Human oversight is vital to ensure AI-generated solutions are generalizable, correct, and meaningful.

The training pipeline crisis for future generations of scientists.

  • Explanation: If AI replaces graduate students in performing initial exploratory research and solving introductory problems, the pipeline for training elite human scientists may break.
  • Evidence: Graduate students learn by struggling through difficult problems; automating this away removes essential career experience.
  • Practical implication: Educational institutions must adapt training paradigms to ensure human mathematicians and scientists develop deep problem-solving skills.

Frameworks, Models & Processes

The Four Paradigms of Science

  • How it works: Classification of scientific methodology from empirical observation to AI-driven discovery.
  • Components:
    • Theory (Kepler, Newton)
    • Experiment
    • Simulation (Supercomputers)
    • Big Data and AI (Machine Learning)
  • When to use: To understand the historical progression and current state of scientific research methodologies.

Examples & Case Studies

Kepler attempted to fit planetary orbits to the five Platonic solids before realizing Tycho Brahe's precise observational data disproved it, eventually leading to elliptical orbits.

  • Illustrates: The interplay between elegant theory and precise empirical data, and how initial models often require revision.
  • Lesson: Theory and experiment must constantly correct one another; data alone does not guarantee truth.

An OpenAI model recently disproved the 80-year-old belief regarding the planar unit distance problem posed by Paul Erdős in 1946.

  • Illustrates: Autonomous AI solving prominent open mathematical problems previously out of reach for human consensus.
  • Lesson: AI can discover non-obvious constructions and bypass human cognitive biases.

Actionable Takeaways

  • Immediate:
    • Use AI tools to accelerate secondary tasks such as literature searches, code writing, and proof-reading.
    • Recognize the limitations of current LLMs regarding reasoning depth and overfitting.
  • Strategic:
    • Redesign scientific education and training pipelines to account for AI assistance.
    • Improve curation, verification, and triage mechanisms to combat proof indigestion.
  • Questions to investigate:
    • How replicable are AI-driven scientific breakthroughs across different domains?
    • How can we train the next generation of scientists when foundational problem-solving is automated?

Claims Worth Verifying

  • An OpenAI model autonomously solved the planar unit distance problem posed by Paul Erdős in 1946. (factual)
  • Gauss computed the first 100,000 prime numbers by hand to form the prime number theorem. (historical)

Notable Quotes

"Science and mathematics has changed a lot over the last centuries." (at 2:17) "We are at a somewhat risky point in in sort of the the structure of of funding the scientific enterprise in general." (at 45:17)

Compressed Summary

  • AI accelerates scientific and mathematical research by automating complex computations and proof generation.
  • A significant paradox exists: AI provides powerful answers without deep human understanding or explainability.
  • Researchers face 'proof indigestion' due to the overwhelming volume of AI-generated scientific outputs.
  • Human curation, education paradigms, and rigorous peer review remain essential to prevent overfitting and ensure real progress.
  • Keywords: artificial intelligence, mathematics, proofs, paradox, science
  • Core insight: AI acts as a powerful computational helicopter in science, rapidly delivering solutions while challenging human understanding, curation, and the traditional training pipeline.

Core insights

7
Tradeoffhigh noveltymoderate evidence

When generation becomes cheap, verification and curation become the rate-limiting step: AI produces more candidate proofs than human experts can triage, edit, and polish, so the pipeline bottleneck moves from producing solutions to absorbing and organizing them.

Why it matters

Agentic and research systems that scale generation without scaling verification will accumulate unverified, unranked output. Throughput gains are capped by downstream human review capacity, not model capability.

Generalization

Any AI pipeline whose output requires human adjudication (proofs, code patches, candidate hypotheses, medical findings) will shift its constraint to triage, ranking, and curation; architecture should budget explicitly for these stages.

Human experts are experiencing 'proof indigestion' due to the sheer quantity of outputs to verify, triage, and curate.
Open source video
The final textbook form of a proof still requires human editing, organization, and editorial polish.
Open source video
Mental Modelhigh noveltymoderate evidence

Capability and understanding are decoupled: AI systems get more powerful and make fewer mistakes while the humans using them do not become better scientists, because the model produces results without exposing the intermediate reasoning or explaining why a discovery is interesting.

Why it matters

Teams often assume that tooling improvements compound into operator expertise. If capability is delivered as opaque endpoint results, the human skill base stagnates even as outputs improve, which weakens debugging, review, and error recovery.

Generalization

Delegating the intermediate reasoning to a model trades short-term output quality for long-term loss of the operator's ability to inspect, explain, or repair the system.

AI operates without human scientists understanding the underlying steps or being able to explain why a discovery is interesting.
Open source video
We may need to redesign our perception of what science is and what we want out of it.
Open source video
Failure Modehigh noveltymoderate evidence

Removing the process of struggle removes the training signal: if AI absorbs initial exploratory research and introductory problem solving, the pipeline that produces future elite human experts — who historically learned by struggling through hard problems — may break.

Why it matters

This is a slow, deferred failure: systems look more productive now while degrading the human capability pool needed to supervise and advance them later. It applies to engineering orgs that automate junior/entry-level work.

Generalization

Any apprenticeship or skill-acquisition pipeline that is optimized away in the name of efficiency destroys the dataset of experience future experts are trained on.

If AI replaces graduate students in performing initial exploratory research and solving introductory problems, the pipeline for training elite human scientists may break.
Open source video
Graduate students learn by struggling through difficult problems; automating this away removes essential career experience.
Open source video
Tradeoffmedium noveltymoderate evidence

Goal-directed AI acts like a helicopter: it delivers the researcher directly to the target without teaching the route, discarding the detours, unexpected observations, and peripheral findings that traditional exploratory process generates.

Why it matters

Optimizing hard for a specified objective can systematically suppress the incidental discoveries that often carry the most value. Systems that only return the answer lose the exploratory byproducts.

Generalization

Objective-maximizing agents are blind to off-target signal; architectures that preserve and surface intermediate artifacts, dead ends, and side observations recover some of that value.

AI tools act like helicopters, flying researchers directly to a specific goal without teaching them how to get there.
Open source video
While AI achieves goals efficiently, it risks losing the broader process and peripheral insights.
Open source video
Failure Modemedium noveltymoderate evidence

AI-generated artifacts can be correct in outcome yet unexplainable, producing operational fragility: debugging simulations or trusting proofs without knowing how the solution was derived leaves no mechanistic handle when the model is wrong.

Why it matters

Without an inspectable derivation, failure diagnosis, regression detection, and trust calibration become guesswork, and errors may only surface at the point of real-world use.

Generalization

Correctness without provenance is an observability gap; capturing intermediate reasoning at generation time is the only reliable way to debug or audit later.

Debugging AI simulations without knowing how the solution was derived.
Open source video
Getting proofs from AI that lack intuitive human explanation.
Open source video
Empirical Resultmedium noveltyweak evidence

AI models excel at pattern matching within their training distribution but may fail to extrapolate or grasp causal mechanism; historical precedent (Kepler abandoning Platonic-solid orbits once data precision exposed the flaw) shows that tighter data can overturn a model that fits comfortably.

Why it matters

Models that fit existing data well can be structurally wrong about underlying mechanisms; human oversight is framed as necessary to ensure AI solutions are generalizable and meaningful, not just well-fit.

Generalization

In-distribution fit is not evidence of mechanism; evaluation should probe extrapolation and causal validity rather than held-out fit alone.

AI models excel at pattern matching within training data but may fail to extrapolate or understand underlying causal mechanisms.
Open source video
Historical examples like Kepler moving from Platonic solids to elliptical orbits because data precision revealed flaws in rigid theories.
Open source video
Predictionmedium noveltyweak evidence

AI is positioned not as an additional method but as a paradigm that transforms all prior modes of science (theory, experiment, simulation, big data) rather than sitting alongside them.

Why it matters

If AI restructures rather than supplements prior methodologies, then interfaces, tooling, and validation norms of the earlier paradigms need re-engineering around AI, not incremental bolting-on.

Generalization

New general-purpose infrastructure tends to recompose the layers beneath it rather than occupy a single slot, which argues for redesign over integration.

AI represents a new paradigm that transforms all previous modes of doing science.
Open source video
Science traditionally relied on two paradigms: theory and experiment.
Open source video

Deep dives

4

Verification and curation as the bottleneck in AI-augmented research

Research question

How can triage and curation of AI-generated mathematical proofs be scaled to match generation throughput without reintroducing opacity?

Why

Generation is cheap but verification is rate-limiting; end-to-end system throughput is capped by human curation capacity, not model capability.

However, this efficiency creates 'proof indigestion,' requiring human curation, editing, and fundamental paradigm shifts in how we approach scientific inquiry and education.
Open source video
Source video

Provenance and observability in AI-generated derivations

Research question

Can capturing intermediate reasoning at generation time make AI-generated proofs debuggable and auditable?

Why

Correctness without provenance leaves no mechanistic handle for debugging, regression detection, or trust calibration.

While AI tools function like helicopters—transporting researchers directly to goals without teaching them the path—they simultaneously revolutionize research by solving complex problems, generating proofs, and automating calculations.
Open source video
Source video

Pattern matching vs. causal mechanism in LLM mathematical reasoning

Research question

To what extent do LLMs' mathematical proofs rely on in-distribution pattern matching versus genuine causal/mechanistic reasoning?

Why

Good fit on training data does not imply generalizability; evaluation should probe extrapolation and causal validity.

Language contains unspoken, unwritten patterns and rules that LLMs absorb.
Open source video
With enough computing power, data, and iteration, LLMs begin to mimic intelligent reasoning and mathematical proof generation.
Open source video
Source video

Goal-directed AI and the loss of serendipitous discovery

Research question

What architectural changes preserve peripheral findings and unexpected observations when AI agents optimize for a specified goal?

Why

Objective-maximizing agents discard the detours and unexpected observations that often produce the most valuable scientific insights.

While AI achieves goals efficiently, it risks losing the broader process and peripheral insights.
Open source video
Traditional science involves a process of discovery, getting lost, and finding unintended phenomena.
Open source video
Source video

Article ideas

4

Proof Indigestion: Why AI's Generation Speed Is Not the Bottleneck

The limiting factor in AI-augmented research is not model capability but human verification and curation capacity; architectures must budget for triage, ranking, and deduplication as first-class stages.

Angle

Engineering and research operations perspective.

Source video

The Helicopter Problem: How Goal-Directed AI Erases the Path to Discovery

Optimizing AI for goal attainment systematically discards the exploratory byproducts—detours, dead ends, unexpected observations—that produce the most valuable scientific insights; we need agents that preserve and surface intermediate artifacts.

Angle

Science of science and AI agent design.

Source video

AI as a New Scientific Paradigm: Why Bolting It On Will Fail

AI does not add a fifth method to theory, experiment, simulation, and big data; it transforms each of them, forcing a redesign of validation, attribution, and peer review rather than incremental integration.

Angle

Metascience and infrastructure.

Source video

Debugging the Unexplainable: Provenance as a First-Class Artifact in AI-Generated Proofs

Correctness without provenance is an observability gap; capturing intermediate reasoning at generation time is the only reliable way to debug, audit, and calibrate trust in AI outputs.

Angle

Systems engineering for AI.

Source video

Project ideas

4

Proof Triage Engine

beyond-evals

An automated triage layer (novelty scoring, deduplication, correctness pre-checks, quality ranking) can increase human curation throughput of AI-generated mathematical proofs by at least 3x compared to unranked output, without increasing the rate of missed critical errors.

Proof of concept

Build a pipeline that ingests AI-generated proof candidates, applies automated novelty and correctness heuristics, ranks and clusters them, and presents them to human curators in a prioritized queue. Measure curation time and error detection rate.

Measurement

Time to curate a fixed set of proofs, number of critical errors missed, human curator satisfaction.

Source video

Provenance Capture for AI Derivations

new

Capturing intermediate reasoning steps (e.g., chain-of-thought, tool calls, intermediate lemmas) as structured provenance at generation time enables debugging of incorrect AI-generated proofs that would otherwise be opaque, reducing time-to-diagnose by at least 50%.

Proof of concept

Instrument an LLM proof generator to log all intermediate steps, store them as a provenance graph, and build a debugger interface. Compare diagnosis time for erroneous proofs with and without provenance.

Measurement

Time to identify the faulty step in a proof with provenance vs. without, on a set of known-buggy generated proofs.

Source video

Serendipity-Preserving Agent

movement-lab

An AI research agent that explicitly records and surfaces off-target observations and dead ends can produce a higher rate of unexpected but valuable findings (measured by expert novelty ratings) than a purely goal-directed agent, without unacceptable loss in goal attainment speed.

Proof of concept

Develop two agents: one goal-only, one goal-plus-peripheral-capture. Run both on a set of open-ended scientific exploration tasks. Collect all outputs, have experts rate unexpectedness and value.

Measurement

Number of expert-rated 'unexpected valuable findings' per task, goal completion time.

Source video

Extrapolation Stress Test for Mathematical LLMs

beyond-evals

Current LLM proof generators fail on out-of-distribution mathematical problems at a significantly higher rate than on in-distribution problems, revealing reliance on pattern matching rather than mechanistic reasoning.

Proof of concept

Create a benchmark of mathematical problems that are structurally novel but solvable with known techniques. Evaluate several LLMs. Compare success rates to standard benchmarks.

Measurement

Accuracy on OOD vs. ID benchmarks, with correlation to problem similarity to training data.

Source video

Architectural implications

3

The scarce resource moves from answer generation to answer verification and curation.

Before

Pipeline is designed around maximizing generation throughput and raw model capability.

After

Pipeline includes explicit triage, ranking, deduplication, and curation stages with their own budget and capacity model.

Consequence

Verification capacity, not model capability, determines end-to-end system throughput; triage automation becomes a first-class component.

Source video

AI delivers results without exposing the path that produced them.

Before

Reasoning traces are assumed recoverable or reconstructable after the fact.

After

Intermediate artifacts, derivations, and provenance are captured and stored at generation time as part of the artifact.

Consequence

Debugging, audit, and trust calibration depend on generation-time capture; retrofitting provenance later is not possible.

Source video

Automating exploratory and introductory work removes the experience that produces expert supervisors.

Before

Junior/entry-level tasks are treated as low-value work suitable for full automation.

After

Some struggle-based work is deliberately preserved or scaffolded as part of the human development pipeline.

Consequence

Organizations must treat human capability production as an explicit pipeline with its own metrics, not as a byproduct of production work.

Source video

Tradeoffs and failure modes

5

Efficiency vs. serendipitous discovery

Benefit

AI flies directly to the goal, achieving outcomes far faster than exploratory process.

Cost or risk

The detours and peripheral observations that produce unexpected findings are discarded along with the path.

While AI achieves goals efficiently, it risks losing the broader process and peripheral insights.
Open source video
Source video

Generation throughput vs. human verification capacity

Benefit

AI generates an unprecedented volume of proposed solutions and proofs.

Cost or risk

Experts face proof indigestion — more output than can be verified, triaged, and curated, and the final form still needs human editing and polish.

Human experts are experiencing 'proof indigestion' due to the sheer quantity of outputs to verify, triage, and curate.
Open source video
Source video

Capability gains vs. operator understanding

Benefit

AI is becoming more powerful, capable, and makes fewer mistakes.

Cost or risk

Users do not get better at the underlying discipline, cannot explain why a result is interesting, and cannot debug derivations they did not follow.

AI operates without human scientists understanding the underlying steps or being able to explain why a discovery is interesting.
Open source video
Source video

Automating exploratory work vs. the expert training pipeline

Benefit

Automation of introductory and exploratory problem solving removes tedious, low-skill labor.

Cost or risk

The struggle that builds elite human problem-solving skill disappears, potentially breaking future expert supply.

Graduate students learn by struggling through difficult problems; automating this away removes essential career experience.
Open source video
Source video

In-distribution pattern fitting vs. mechanistic validity

Benefit

Models fit patterns well within their training data and can mimic reasoning and proof generation.

Cost or risk

They may fail to extrapolate or capture causal mechanism, so good fit does not imply a correct or generalizable answer.

AI models excel at pattern matching within training data but may fail to extrapolate or understand underlying causal mechanisms.
Open source video
Source video

Open questions

5

How can triage and curation of AI-generated artifacts be automated enough to keep pace with generation, without reintroducing the same opacity problem?

Why unresolved

The summary identifies proof indigestion as a bottleneck but offers no mechanism for scaling verification; it only notes that human editing and polish are still required.

Research direction

Build and benchmark automated triage layers (novelty scoring, dedup, correctness pre-checks, quality ranking) and measure human curation throughput with and without them.

Source video

Can provenance be captured well enough at generation time to make otherwise opaque AI derivations debuggable?

Why unresolved

The summary frames derivation opacity as a hazard but does not describe any trace-capture or explanation mechanism.

Research direction

Prototype systems that log intermediate derivations and reasoning steps as first-class artifacts and test whether they improve debugging and regression detection.

Source video

Does delegating exploratory and introductory work actually degrade downstream expert capability, and can that degradation be measured?

Why unresolved

The pipeline crisis is presented as a risk ('may break') rather than an observed outcome; no measurement is reported.

Research direction

Longitudinal studies comparing skill development of cohorts whose early work is AI-automated versus not, with scaffolded-struggle interventions.

Source video

Is in-distribution fit on large corpora sufficient evidence of generalizable mechanism, or does evaluation need adversarial extrapolation tests?

Why unresolved

The overfitting concern is supported only by a historical analogy (Kepler), not by direct measurements of model extrapolation failure.

Research direction

Design extrapolation and out-of-distribution benchmarks that distinguish pattern matching from causal/mechanistic generalization.

Source video

If AI is a paradigm that transforms all prior scientific modes rather than adding one, what re-engineering of validation and attribution does that force?

Why unresolved

The claim is asserted at a high level; the summary gives no concrete change to validation, peer review, or attribution practice.

Research direction

Specify which parts of theory, experiment, simulation, and big-data workflows change once AI is embedded in them, and what new validation primitives are needed.

Source video

Key claims

7
causalVerification needed

AI-generated proof and solution volume exceeds human capacity to verify, triage, and curate it, producing 'proof indigestion'.

Evidence

Human experts are experiencing 'proof indigestion' due to the sheer quantity of outputs to verify, triage, and curate.

Question

What is the measured ratio of AI-generated candidate proofs to expert verification capacity in current mathematical practice?

Source video
factualVerification needed

AI can operate without humans understanding the underlying steps or being able to explain why a discovery matters.

Evidence

AI operates without human scientists understanding the underlying steps or being able to explain why a discovery is interesting.

Question

In documented AI-assisted discoveries, how often are the intermediate derivation steps unavailable or unintelligible to the human collaborators?

Source video
predictionVerification needed

Automating graduate-student exploratory research may break the pipeline that trains elite human scientists.

Evidence

If AI replaces graduate students in performing initial exploratory research and solving introductory problems, the pipeline for training elite human scientists may break.

Question

Does early-career exposure to AI-automated exploratory work correlate with reduced later research independence or problem-solving skill?

Source video
causalVerification needed

AI models excel at pattern matching within training data but may fail to extrapolate or understand causal mechanisms.

Evidence

AI models excel at pattern matching within training data but may fail to extrapolate or understand underlying causal mechanisms.

Question

On which classes of mathematical problems do LLM proof generators fail to extrapolate beyond the training distribution?

Source video
opinionVerification not requested

AI represents a new paradigm that transforms all previous modes of doing science rather than adding a separate one.

Evidence

AI represents a new paradigm that transforms all previous modes of doing science.

Source video
causalVerification needed

With enough compute, data, and iteration, LLMs begin to mimic intelligent reasoning and mathematical proof generation.

Evidence

With enough computing power, data, and iteration, LLMs begin to mimic intelligent reasoning and mathematical proof generation.

Question

At what scale of compute and data does token prediction transition from shallow continuation to verifiable multi-step proof generation?

Source video
factualVerification needed

OpenAI models solved a famous open question posed by Paul Erdős in 1946, the planar unit distance problem.

Evidence

OpenAI solving a famous open question posed by Paul Erdős in 1946 regarding the planar unit distance problem.

Question

Was the Erdős planar unit distance problem result independently verified and peer-reviewed, and what was the extent of human contribution?

Source video

Connections

5