Domain Collapse Timelines: AI Surplus Capability in Chemistry, Biology, and Physics

Executive Summary

• Physics and materials science are closest to collapse (2026–2030 for computational subdomains), driven by tight mathematical feedback loops, abundant simulation data, and demonstrated breakthroughs like GNoME's 2.2 million crystal predictions and ML force fields achieving 10,000× DFT speedups — but theoretical physics (new laws, unified theories) remains a distant holdout likely beyond 2035.

• Biology is bifurcated: protein structure prediction has already collapsed (AlphaFold 2/3, 2022–2024), but full drug discovery collapse is constrained by an irreducible clinical trial bottleneck — the ~90% historical failure rate has not yet improved with AI, and no AI-discovered drug has achieved FDA approval as of early 2026. Expect in-silico biology collapse by 2028–2030, in-vivo/clinical collapse by 2033–2038.

• Chemistry's collapse is most imminent for routine tasks (retrosynthesis, property prediction, yield optimization: 2027–2029), but novel reaction discovery and fully autonomous synthesis face a "physical world bottleneck" — self-driving laboratories are the critical enabling technology, and their maturation rate will determine whether the 2027 or 2032 scenario materializes.

• The single most important variable across all three domains is the self-driving laboratory (SDL): the integration of AI planning with robotic execution and closed-loop feedback. SDL maturation compresses collapse timelines by 3–5 years; without it, AI remains a powerful but incomplete tool requiring human experimental execution.

• "Domain collapse" is granular, not binary: every domain will collapse sub-field by sub-field (structure prediction before function design; property screening before synthesis; simulation before theory), and researchers and entrepreneurs should map their specific sub-domain rather than treating each field as monolithic.

Cross-Provider Consensus

1. Protein Structure Prediction Has Already Collapsed

Providers: Anthropic, OpenAI, Grok, Gemini, Gemini-Lite, Perplexity (all six) Confidence: HIGH

All providers independently confirm that AlphaFold 2 (2020–2022) and AlphaFold 3 (2024) effectively solved the protein structure prediction problem, predicting >200 million protein structures at near-experimental accuracy. This is the clearest existing example of domain collapse in the experimental sciences. AlphaFold 3 extends to DNA/RNA/ligand complexes, and open-source competitors (Boltz-2, RoseTTAFold, Chai-1) have democratized the capability. The Nobel Prize in Chemistry 2024 awarded to Hassabis, Jumper, and Baker formally recognized this collapse.

2. The Physical World Bottleneck Is the Primary Decelerant

Providers: Anthropic, Perplexity, Gemini-Lite, Grok Confidence: HIGH

Multiple providers independently identified that the experimental sciences face a fundamental constraint absent in mathematics and coding: AI cannot yet autonomously execute physical experiments. Anthropic explicitly states: "AI agent systems struggle to independently collect data that requires hands-on experimentation." Perplexity frames this as "data abundance and feedback loop speed" being the true determinants of collapse timing. Gemini-Lite notes that physics "requires adherence to fundamental, immutable laws" beyond pattern recognition. This consensus finding is the most important structural insight for timeline estimation.

3. Chemistry Collapse Is Estimated at 2027–2030

Providers: Grok (2027–2029), OpenAI (~2028), Anthropic (2029–2032), Gemini (2027–2030), Perplexity (2027–2029) Confidence: MEDIUM-HIGH

Five of six providers converge on a 2027–2030 window for chemistry domain collapse, with retrosynthesis and property prediction collapsing first (2027–2028) and novel reaction discovery/full autonomous synthesis later (2031–2034). The primary disagreement is on the upper bound.

4. Biology's Full Collapse Is Constrained by Clinical Trial Biology

Providers: Anthropic, Gemini, OpenAI, Perplexity, Grok Confidence: HIGH

All providers note that while computational biology is advancing rapidly, the ~90% clinical trial failure rate has not meaningfully improved with AI. Gemini specifically cites failed Phase II/III trials for AI-designed drugs (Fosigotifator for ALS, navacaprant for depression). Anthropic notes "no AI-discovered drug has achieved FDA approval as of December 2025." This creates a hard floor on biology collapse timelines regardless of in-silico progress.

5. Self-Driving Laboratories Are the Critical Enabling Technology

Providers: Grok, Anthropic, OpenAI, Gemini, Gemini-Lite Confidence: HIGH

Five providers independently identify autonomous/self-driving laboratories as the key variable that will determine whether collapse timelines compress or extend. Anthropic cites Argonne's 6,000 battery experiments in 5 months. Grok describes SDLs as "the wet-lab GPU." OpenAI cites Carnegie Mellon's Coscientist executing autonomous chemistry experiments. The convergence on this point is strong.

6. Mathematics and Coding Are in Late-Stage Partial Collapse

Providers: All six providers Confidence: HIGH

Universal agreement that math and coding represent the reference case. Anthropic provides the most nuanced view: "late-stage partial collapse for routine tasks but pre-collapse for frontier research." FrontierMath went from <2% AI solve rate to >40% within a short period. IMO gold-medal performance achieved. However, generating genuinely novel mathematical knowledge remains pre-collapse.

7. Physics Materials/Simulation Subdomains Are Collapsing Faster Than Theoretical Physics

Providers: Grok, OpenAI, Anthropic, Gemini, Perplexity Confidence: HIGH

All providers distinguish between computational/materials physics (near-collapse, 2026–2031) and theoretical physics (far from collapse, 2035+). GNoME's 2.2 million crystal predictions, ML force fields, and fusion plasma control via RL are cited as evidence of materials/simulation collapse. Theoretical physics — deriving new fundamental laws — is universally treated as the most distant target.

Unique Insights by Provider

Grok

• "Knowledge collapse" as a second-order risk: Grok uniquely flags the MIT economics paper on "AI, Human Cognition and Knowledge Collapse" — the risk that AI surplus capability erodes human learning incentives, creating long-term epistemic fragility. This is distinct from model collapse (AI training on AI data) and represents a societal-level risk that other providers ignore. Matters because it suggests domain collapse has negative externalities beyond the obvious disruption narrative.
• Pearl model benchmark: Grok cites the Pearl model achieving 40% better performance than AlphaFold 3 on drug benchmarks, suggesting the AlphaFold 3 baseline is already being surpassed by specialized successors — a leading indicator of accelerating capability.

OpenAI

• Halicin and Abaucin as collapse precursors: OpenAI provides the most detailed treatment of specific AI-discovered antibiotics (Halicin 2020, Abaucin 2023) as concrete proof-of-concept events. The framing of these as "early collapse events" — not just impressive results — is analytically useful for calibrating what collapse looks like in practice before it becomes systemic.
• Insilico Medicine's 18-month drug candidate: The specific case of an AI-designed fibrosis drug entering human trials in 18 months (vs. typical 5 years) is cited as a concrete timeline compression metric that other providers don't quantify as precisely.

Anthropic

• The "physical world bottleneck" as a formal framework: Anthropic provides the most rigorous structural analysis of why experimental sciences collapse slower than math/code, presenting a comparative table across six dimensions (verifiability, feedback loop speed, data availability, emergent complexity, regulatory barriers, physical world coupling). This framework is the most analytically useful contribution for understanding differential collapse rates.
• Boltz-2 binding affinity prediction: Anthropic uniquely highlights Boltz-2's ability to co-fold protein-ligand pairs and output binding affinity estimates in ~20 seconds at gold-standard accuracy — collapsing what previously required 6–12 hours of free-energy perturbation calculations. This is a specific, verifiable capability milestone.
• Evo 2 genomic foundation model: The 9.3 trillion nucleotide training dataset across 100,000 species, capable of designing bacterial-length genomes and identifying disease-causing mutations, is highlighted as a landmark in genomic AI that other providers underemphasize.

Gemini

• FDA/EMA joint "Good AI Practice" guidelines (January 2026): Gemini uniquely identifies the regulatory institutionalization of AI in drug development as a structural accelerant. The 10-principle framework from FDA/EMA signals that regulators are building the scaffolding for AI-validated drug approvals — a prerequisite for biology collapse that other providers treat as a barrier rather than an evolving constraint.
• Energy-GNoME and rare earth disruption: Gemini highlights the Politecnico di Torino spin-off targeting renewable energy materials specifically to challenge geopolitical "rare earth" dependencies — a strategic implication of materials AI that other providers miss.
• EDEN (Basecamp Research) and Trillion Gene Atlas: Gemini provides the most detailed treatment of the 28-billion parameter EDEN model trained on 10 billion novel genes from 1 million uncatalogued species, and the subsequent Trillion Gene Atlas initiative targeting 100 million species. This represents the most ambitious genomic data infrastructure project currently underway.
• Claude Opus 4.6 physics benchmark performance: Gemini reports 91.3% on GPQA Diamond and 78.3% on LAB-Bench FigQA (surpassing 77% human expert baseline) — specific benchmark data that grounds the "physics collapse imminent" claim in measurable capability.

Gemini-Lite

• Concise constraint hierarchy: While less detailed than other providers, Gemini-Lite's clean three-row comparative table (Biology: transformers/geometric DL, bottleneck biological noise, 2026–2028; Chemistry: generative models/GNNs, bottleneck data scarcity, 2027–2030; Physics: PINNs/neural operators, bottleneck theoretical consistency, 2030+) provides the clearest at-a-glance framework for non-specialist readers. The framing of physics as requiring "adherence to fundamental, immutable laws" rather than just pattern recognition is a useful conceptual distinction.

Perplexity

• Quantitative capability percentage estimates: Perplexity uniquely provides specific "percentage toward collapse threshold" estimates (Chemistry: 70–80%, Biology: 50–60%, Physics: 30–40%) with annual improvement rates (Chemistry: 3–4%/year, Biology: 2–3%/year, Physics: 1.5–2.5%/year). While these numbers are necessarily approximate, they provide a quantitative framework absent from other providers.
• Negative data scarcity in biology: Perplexity specifically identifies the lack of large-scale negative data (toxicants, off-targets) as a distinct bottleneck in biology — not just data scarcity generally, but the asymmetric availability of positive vs. negative experimental outcomes. This is a technically precise insight that matters for understanding why predictive toxicology lags.
• Inverse problem non-uniqueness in physics: Perplexity uniquely frames physics' difficulty as the "non-uniqueness of inverse problems" — multiple physical mechanisms producing identical observations — which is a more precise characterization than simply "data scarcity" or "complexity."

Contradictions and Disagreements

Contradiction 1: Physics Collapse Timeline — 2026–2027 vs. 2032–2040

Gemini argues physics collapse is "imminent" with an estimated date of 2026–2027, citing Claude Opus 4.6's GPQA Diamond performance and GNoME as evidence of near-surplus capability. OpenAI similarly suggests "within ~18 months" (from early 2026) for parts of physics.

Anthropic provides a sharply different view: condensed matter theory 2034–2040, theoretical physics 2040+, with even crystal structure prediction not collapsing until 2028–2030. Perplexity estimates physics at only 30–40% of collapse threshold with a median collapse date of Q3 2034.

The core disagreement: Gemini and OpenAI appear to conflate materials screening capability with physics domain collapse broadly, while Anthropic and Perplexity maintain a stricter definition requiring theoretical physics to be included. This is a definitional dispute as much as an empirical one. Readers should note: if "physics collapse" means materials property screening and simulation speedup, 2027–2029 is defensible. If it means AI deriving new physical laws or solving condensed matter theory, 2035+ is more credible.

Contradiction 2: GNoME's Significance — Landmark vs. Overhyped

Gemini and OpenAI treat GNoME's 2.2 million crystal predictions as a landmark achievement representing near-collapse in materials discovery, with the A-Lab's 71% synthesis success rate as validation.

Anthropic directly cites critics who found "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility" and notes that "a number of the so-called novel materials predicted by DeepMind appeared to be well-ordered versions of disordered ones that were already known." Anthropic concludes: "GNoME was simply another reminder of how challenging it is to capture physical realities in virtual simulations."

This is a genuine empirical dispute with published criticism of the GNoME paper. Readers should investigate the specific critiques in the materials science literature before treating GNoME as definitive evidence of materials collapse.

Contradiction 3: Biology Collapse Timeline — 2026–2028 vs. 2033–2038

Gemini estimates in-silico biology collapse at 2026–2028 and treats this as nearly achieved. Gemini-Lite similarly estimates 2026–2028 for biology broadly.

Anthropic provides a much more granular and conservative view: protein-ligand binding prediction 2027–2028, genomic variant interpretation 2028–2030, de novo protein/antibody design 2028–2031, preclinical drug candidate generation 2029–2031, and full drug discovery (target to approval) 2033–2038. Perplexity estimates biology at 50–60% of collapse threshold with a median collapse date of Q2 2031.

The resolution: Gemini appears to be referring specifically to structural biology and in-silico tasks, while Anthropic and Perplexity are using a broader definition that includes clinical translation. Both can be correct simultaneously if the definition is clarified.

Contradiction 4: AI Drug Discovery Clinical Success Rates

Grok states "first fully AI-designed drugs reached Phase II trials in 2025 with positive results" and cites "80–90% success rate in Phase I trials" for AI-discovered molecules.

Anthropic and Gemini directly contradict this optimism: Anthropic states "no AI-discovered drug has achieved FDA approval as of December 2025" and notes the ~90% clinical failure rate "has not yet improved with AI." Gemini cites specific Phase II/III failures (Fosigotifator, navacaprant). Perplexity notes "predictive toxicity remains unreliable."

This is a significant factual tension. The 80–90% Phase I success rate cited by Grok (sourced from Anthropic's report) refers to safety trials (Phase I), not efficacy (Phase II/III) — a critical distinction. Phase I success does not predict Phase II/III success, where the ~90% failure rate persists. Readers should not interpret Phase I safety success as evidence of clinical collapse.

Contradiction 5: Chemistry Collapse Upper Bound — 2028 vs. 2032+

OpenAI and Perplexity estimate chemistry collapse around 2027–2029. Anthropic provides a more conservative upper bound: novel reaction discovery 2031–2034, full autonomous synthesis 2032–2035. The disagreement centers on how to weight out-of-distribution generalization failures — Anthropic emphasizes that "common benchmarks test models in an in-distribution setting, whereas many real-world uses are in out-of-distribution settings," suggesting current benchmark performance overstates real-world capability.

Detailed Synthesis

The Concept of Domain Collapse: Definitional Clarity

"Domain collapse" as used across providers refers to the inflection point where AI achieves surplus capability — not merely human-level performance, but the ability to industrialize discovery such that progress becomes compute-constrained rather than human-genius-constrained [Grok, OpenAI, Gemini]. The analogy is instructive: just as the Industrial Revolution made physical labor scalable by decoupling output from individual human effort, domain collapse makes cognitive labor in a scientific field scalable by decoupling discovery from individual human insight [Gemini].

Critically, collapse is not binary. Every provider who engaged seriously with the question concluded that collapse proceeds sub-field by sub-field [Anthropic, Perplexity, Grok]. Protein structure prediction collapsed in 2022; protein function design has not. Molecular property prediction is near-collapse; novel reaction discovery is not. Crystal structure screening is approaching collapse; condensed matter theory is not. This granularity is essential for practical decision-making.

The Reference Case: Mathematics and Coding

Mathematics and coding provide the clearest evidence of what collapse looks like in practice. In mathematics, FrontierMath benchmark performance went from <2% to >40% solve rate within a short period [Anthropic]. DeepMind's AlphaProof solved 4/6 IMO problems at silver-medalist level [OpenAI]. Google DeepMind's Aletheia agent solved approximately 6/10 research-level problems [Anthropic]. However, Anthropic provides the most nuanced assessment: "late-stage partial collapse for routine tasks but pre-collapse for frontier research." The ability to solve competition problems does not yet translate to generating genuinely novel mathematical knowledge — a distinction that will recur across all three experimental domains.

In coding, 53% of senior developers believe AI already outperforms most human programmers [OpenAI], and AI coding assistants are now used daily by a large majority of developers. The key structural feature enabling rapid collapse in math and coding is formal verifiability — outputs can be checked against ground truth instantly, creating tight feedback loops that accelerate AI improvement [Anthropic, Perplexity]. This feature is largely absent in experimental sciences, which is the fundamental reason collapse timelines are longer.

Chemistry: The Nearest Collapse in Experimental Science

Chemistry sits at approximately 60–80% of collapse threshold [Perplexity, Anthropic], with the range reflecting genuine uncertainty about out-of-distribution generalization. The most advanced capabilities are in molecular property prediction and retrosynthesis planning, where transformer models trained on USPTO databases achieve 90%+ accuracy on forward reaction prediction and 50–70% on multi-step retrosynthesis [Perplexity].

The most significant recent advance is MIT's FlowER model, which integrates the bond-electron matrix (tracking all atoms, bonds, and lone electron pairs throughout reactions) to enforce conservation of mass and electrons — moving beyond statistical pattern matching to physically grounded reaction prediction [Gemini]. This addresses the core criticism of earlier LLM-based chemistry: that models treated atoms as linguistic tokens and could "hallucinate" impossible chemistry [Gemini]. Similarly, UCLA/Utah researchers have demonstrated AI prediction of stereochemistry — the 3D arrangement of atoms — reducing months of empirical work to days [Gemini].

The critical enabling technology for chemistry collapse is the self-driving laboratory. Argonne National Laboratory demonstrated 6,000 battery chemistry experiments in 5 months via AI+robotics [Anthropic]. Carnegie Mellon's Coscientist autonomously designed, executed, and analyzed chemical experiments in 2023 [OpenAI]. The Acceleration Consortium is building SDL infrastructure at scale [Grok]. A 2026 breakthrough by the Sigman and Doyle labs demonstrated that models pre-trained on massive datasets can be fine-tuned on as few as 5–10 experiments, with Active Learning loops optimizing reactions to >90% yield in half the time of human experts [Anthropic].

The primary remaining bottleneck is out-of-distribution generalization. Anthropic's analysis is the most rigorous here: "despite compelling performance on popular benchmark tasks, strange and erroneous predictions sometimes ensue when using these models in practice" because "common benchmarks test models in an in-distribution setting, whereas many real-world uses are in out-of-distribution settings." This is the key reason why the 2027–2028 optimistic timeline and the 2031–2034 conservative timeline for novel reaction discovery are both defensible.

Collapse timeline: Retrosynthetic planning and molecular property prediction: 2027–2028 [Anthropic, Perplexity, Grok]. Reaction yield optimization and drug-like molecule design: 2028–2030 [Anthropic, Gemini]. Novel reaction discovery: 2031–2034 [Anthropic]. Full autonomous synthesis: 2032–2035 [Anthropic].

Biology: The Bifurcated Domain

Biology presents the starkest internal bifurcation of any domain. Protein structure prediction has already collapsed — this is the unanimous view of all six providers. AlphaFold 2's achievement of near-experimental accuracy across ~200 million proteins, followed by AlphaFold 3's extension to DNA/RNA/ligand complexes, represents the clearest case of domain collapse in experimental science [all providers]. Boltz-2 can now co-fold protein-ligand pairs and output binding affinity estimates in ~20 seconds at gold-standard accuracy, collapsing what previously required 6–12 hours of free-energy perturbation calculations [Anthropic].

Genomics is advancing rapidly. Evo 2, trained on 9.3 trillion nucleotides across 100,000 species, can identify disease-causing mutations and design bacterial-length genomes [Anthropic]. Basecamp Research's EDEN (28 billion parameters, 10 billion novel genes from 1 million uncatalogued species) represents the most ambitious genomic foundation model currently deployed [Gemini]. DeepMind's AlphaMissense evaluated 71 million possible missense mutations and predicted 32% as likely pathogenic with 90% precision [OpenAI] — a task that would have been "astronomically time-consuming without AI."

Drug discovery presents the most sobering picture. Over 173 AI-designed programs are in clinical development as of 2026 [Anthropic], but no AI-discovered drug has achieved FDA approval [Anthropic, Gemini]. The ~90% clinical trial failure rate persists [Anthropic, Gemini, Perplexity]. Specific high-profile failures — Fosigotifator (ALS) and navacaprant (depression) — demonstrate that AI's ability to compress preclinical timelines does not translate to improved clinical success rates [Gemini]. The economic structure of AI drug discovery partnerships reflects this skepticism: while >$15 billion in partnerships were announced in 2025, actual upfront payments averaged only 2% of headline values [Gemini].

The January 2026 FDA/EMA joint "Good AI Practice in Drug Development" guidelines represent a structural shift — regulators are building the framework for AI-validated drug approvals rather than treating AI as purely a research tool [Gemini]. This is a prerequisite for biology collapse that other providers underemphasize.

Collapse timeline: Protein structure prediction: already collapsed (2022–2024) [all providers]. Protein-ligand binding prediction: 2027–2028 [Anthropic]. Genomic variant interpretation: 2028–2030 [Anthropic, Perplexity]. De novo protein/antibody design: 2028–2031 [Anthropic]. Preclinical drug candidate generation: 2029–2031 [Anthropic]. Full drug discovery (target to approval): 2033–2038 [Anthropic, Gemini]. Systems biology/virtual cell: 2035+ [Anthropic].

Physics: The Most Stratified Domain

Physics presents the widest range of collapse timelines across sub-domains of any field analyzed. The disagreement between providers (2026–2027 per Gemini vs. 2033–2038 per Anthropic/Perplexity) is largely definitional: materials screening and simulation are near-collapse, while theoretical physics is far from it.

In materials science, GNoME's prediction of 2.2 million crystal structures (380,000 stable) represents a near-tenfold expansion of known stable inorganic crystals [Gemini, OpenAI]. The A-Lab's 71% synthesis success rate on GNoME-predicted compounds over 17 days validates that in-silico predictions translate to physical reality at meaningful rates [Gemini]. However, Anthropic cites published criticism that many "novel" materials were well-ordered versions of already-known disordered compounds, and that thermodynamic stability does not guarantee synthetic feasibility — a genuine empirical dispute [Anthropic].

ML force fields represent perhaps the most concrete near-term collapse signal: 10,000× speedup over DFT atomistic simulations expected by 2026 [Grok], with Allegro-FM enabling simulations 1,000× larger than previous models [Anthropic]. The THOR framework computes thermodynamic properties hundreds of times faster than supercomputer simulations while preserving accuracy [Anthropic]. These are not incremental improvements — they represent qualitative capability shifts.

Theoretical physics shows the most nascent but intriguing AI capability. GPT-5.2 Pro identified a closed-form formula for gluon scattering amplitudes by recognizing patterns from specific cases [Anthropic]. Gemini found a novel solution using Gegenbauer polynomials for cosmic string physics [Anthropic]. Physics foundation models (Walrus, AION-1) demonstrate cross-domain transfer from fluid dynamics to astronomy [Grok]. Claude Opus 4.6 scored 91.3% on GPQA Diamond (PhD-level physics/chemistry/biology) and 78.3% on LAB-Bench FigQA, surpassing the 77% human expert baseline [Gemini].

Perplexity provides the most precise characterization of physics' fundamental difficulty: the "non-uniqueness of inverse problems" — multiple physical mechanisms can produce identical observations, making it impossible to uniquely infer underlying physics from data alone. This is a deeper obstacle than data scarcity and explains why theoretical physics collapse is genuinely distant.

Collapse timeline: Crystal structure prediction: 2028–2030 [Anthropic]. Materials property screening: 2029–2031 [Anthropic, Perplexity]. Atomistic simulation (replacing DFT): 2029–2032 [Anthropic]. Particle physics data analysis: 2030–2033 [Anthropic]. Condensed matter theory: 2034–2040 [Anthropic]. Theoretical physics (new laws): 2040+ [Anthropic, Perplexity].

Cross-Domain Structural Insights

The most important structural insight across all three domains is the feedback loop speed hierarchy [Perplexity, Anthropic]: chemistry has faster feedback loops than biology (hours vs. weeks/years), and both have faster loops than physics experiments (though physics simulation loops are fast). This hierarchy directly predicts collapse order.

The SDL convergence thesis [Anthropic] holds that self-driving laboratories are the single most important variable for compressing collapse timelines. When AI planning integrates with robotic execution and closed-loop feedback, the physical world bottleneck shrinks dramatically. The Argonne 6,000-experiment result [Anthropic] and the A-Lab 17-day synthesis run [Gemini] are early demonstrations of what mature SDLs will enable.

The proprietary data moat [Anthropic, Perplexity, Gemini-Lite] is emerging as the primary competitive advantage in all three domains. As foundation models commoditize, the value shifts to proprietary experimental datasets that enable fine-tuning. Pharma companies using federated learning to fine-tune models on proprietary structural data without sharing raw IP [Anthropic] represent an early version of this dynamic.