Proliferation
- Counterint.AI research has vastly more open publication norms than other sensitive technologies, with 85% of breakthrough AI papers published openly compared to less than 30% for nuclear research during the Cold War.S:4.5I:4.0A:4.5
- Quant.The capability gap between frontier and open-source AI models has dramatically shrunk from 18 months to just 6 months between 2022-2024, indicating rapidly accelerating proliferation.S:4.0I:4.5A:4.0
- ClaimFine-tuning leaked foundation models to bypass safety restrictions requires less than 48 hours and minimal technical expertise, as demonstrated by the LLaMA leak incident.S:3.5I:4.5A:4.0
- QualityRated 60 but structure suggests 93 (underrated by 33 points)
- Links14 links could use <R> components
Quick Assessment
Section titled “Quick Assessment”| Dimension | Assessment | Evidence |
|---|---|---|
| Severity | High | Enables cascading risks across misuse, accidents, and governance breakdown |
| Likelihood | Very High (85-95%) | Open-source models approaching frontier parity within 6-12 months; Hugging Face hosts over 2 million models as of 2025 |
| Timeline | Ongoing | LLaMA 3.1 405B released 2024 as “first frontier-level open source model”; capability gap narrowed from 18 to 6 months (2022-2024) |
| Trend | Accelerating | Second million models on Hugging Face took only 335 days vs. 1,000+ days for first million |
| Controllability | Low (15-25%) | Open weights cannot be recalled; 97% of IT professionals prioritize AI security but only 20% test for model theft |
| Geographic Spread | Global | Qwen overtook Llama in downloads 2025; center of gravity shifting toward China |
| Intervention Tractability | Medium | Compute governance controls 75% of global AI compute; export controls reduced China’s share from 37% to 14% (2022-2025) |
AI Proliferation
Overview
Section titled “Overview”AI proliferation refers to the spread of AI capabilities from frontier labs to increasingly diverse actors—smaller companies, open-source communities, nation-states, and eventually individuals. This represents a fundamental structural risk because it’s largely determined by technological and economic forces rather than any single actor’s decisions.
The proliferation dynamic creates a critical tension in AI governance. Research from RAND Corporation↗🔗 web★★★★☆RAND CorporationResearch from RAND CorporationSource ↗Notes suggests that while concentrated AI development enables better safety oversight and prevents misuse by bad actors, it also creates risks of power abuse and stifles beneficial innovation. Conversely, distributed development democratizes benefits but makes governance exponentially harder and increases accident probability through the “weakest link” problem.
Current evidence indicates proliferation is accelerating. Meta’s LLaMA family↗🔗 web★★★★☆Meta AIMeta Llama 2 open-sourceSource ↗Notes demonstrates how quickly open-source alternatives emerge for proprietary capabilities. Within months of GPT-4’s release, open-source models achieved comparable performance on many tasks. The 2024 State of AI Report↗🔗 webState of AI Report 2025The annual State of AI Report examines key developments in AI research, industry, politics, and safety for 2025, featuring insights from a large-scale practitioner survey.Source ↗Notes found that the capability gap between frontier and open-source models decreased from ~18 months to ~6 months between 2022-2024.
Risk Assessment
Section titled “Risk Assessment”| Risk Category | Severity | Likelihood | Timeline | Trend |
|---|---|---|---|---|
| Misuse by Bad Actors | High | Medium-High | 1-3 years | Increasing |
| Governance Breakdown | Medium-High | High | 2-5 years | Increasing |
| Safety Race to Bottom | Medium | Medium | 3-7 years | Uncertain |
| State-Level Weaponization | Medium-High | Medium | 2-5 years | Increasing |
Sources: Center for Security and Emerging Technology analysis↗🔗 web★★★★☆CSET GeorgetownCenter for Security and Emerging Technology analysisSource ↗Notes, AI Safety research community surveys↗🔗 web★★★☆☆AI ImpactsAI experts show significant disagreementSource ↗Notes
Proliferation Dynamics
Section titled “Proliferation Dynamics”Key Proliferation Metrics (2022-2025)
Section titled “Key Proliferation Metrics (2022-2025)”| Metric | 2022 | 2024 | 2025 | Source |
|---|---|---|---|---|
| Hugging Face models | ≈100K | ≈1M | 2M+ | Hugging Face |
| Frontier-to-open capability gap | ≈18 months | ≈6 months | ≈3-6 months | State of AI Report |
| Mean open model size (parameters) | 827M | - | 20.8B | Red Line AI |
| US share of global AI compute | ≈60% | - | 75% | AI Frontiers |
| China share of global AI compute | 37.3% | - | 14.1% | AI Frontiers |
| AI-generated code (Python, US) | - | 30% | - | International AI Safety Report |
Drivers of Proliferation
Section titled “Drivers of Proliferation”Publication and Research Norms
Section titled “Publication and Research Norms”The AI research community has historically prioritized openness. Analysis by the Future of Humanity Institute↗🔗 web★★★★☆Future of Humanity InstituteFHI expert elicitationSource ↗Notes shows that 85% of breakthrough AI papers are published openly, compared to <30% for sensitive nuclear research during the Cold War. Major conferences like NeurIPS and ICML require code sharing for acceptance, accelerating capability diffusion.
OpenAI’s GPT research trajectory↗📄 paper★★★★☆OpenAIOpenAI: Model BehaviorSource ↗Notes illustrates the shift: GPT-1 and GPT-2 were fully open, GPT-3 was API-only, and GPT-4 remains largely proprietary. Yet open-source alternatives like Hugging Face’s BLOOM↗🔗 webHugging Face's BLOOMSource ↗Notes and EleutherAI’s models↗🔗 webEleutherAI EvaluationSource ↗Notes rapidly achieved similar capabilities.
Economic Incentives
Section titled “Economic Incentives”Commercial pressure drives proliferation through multiple channels:
- API Democratization: Companies like Anthropic↗🔗 web★★★★☆AnthropicAnthropicSource ↗Notes, OpenAI↗🔗 web★★★★☆OpenAIOpenAISource ↗Notes, and Google↗🔗 webGoogleSource ↗Notes provide powerful capabilities through accessible APIs
- Open-Source Competition: Meta’s strategy with LLaMA exemplifies using open release for ecosystem dominance
- Cloud Infrastructure: Amazon’s Bedrock↗🔗 webAmazon's BedrockSource ↗Notes, Microsoft’s Azure AI↗🔗 webMicrosoft's Azure AISource ↗Notes, and Google’s Vertex AI↗🔗 webGoogleSource ↗Notes make advanced capabilities available on-demand
Technological Factors
Section titled “Technological Factors”Inference Efficiency Improvements: Research from UC Berkeley↗🔗 webBerkeley AI Research: Detection methodsSource ↗Notes shows inference costs have dropped 10x annually for equivalent capability. Techniques like quantization, distillation, and efficient architectures make powerful models runnable on consumer hardware.
Fine-tuning and Adaptation: Stanford’s Alpaca project↗🔗 webStanford's Alpaca projectSource ↗Notes demonstrated that $600 in compute could fine-tune LLaMA to match GPT-3.5 performance on many tasks. Low-Rank Adaptation (LoRA)↗📄 paper★★★☆☆arXivLow-Rank Adaptation (LoRA)Edward J. Hu, Yelong Shen, Phillip Wallis et al. (2021)Source ↗Notes techniques further reduce fine-tuning costs.
Knowledge Transfer: The “bitter lesson” phenomenon↗🔗 web"bitter lesson" phenomenonSource ↗Notes means that fundamental algorithmic insights (attention mechanisms, scaling laws, training techniques) transfer across domains and actors.
Key Evidence and Case Studies
Section titled “Key Evidence and Case Studies”Major Open-Source Model Releases and Impact
Section titled “Major Open-Source Model Releases and Impact”| Model | Release Date | Parameters | Benchmark Performance | Impact |
|---|---|---|---|---|
| LLaMA 1 | Feb 2023 | 7B-65B | MMLU ≈65% (65B) | Leaked within 7 days; sparked open-source explosion |
| LLaMA 2 | Jul 2023 | 7B-70B | MMLU ≈68% (70B) | Official open release; 1.2M downloads in first week |
| Mistral 7B | Sep 2023 | 7B | Outperformed LLaMA 2 13B | Proved efficiency gains possible |
| Mixtral 8x7B | Dec 2023 | 46.7B (12.9B active) | Matched GPT-3.5 | Demonstrated MoE effectiveness |
| LLaMA 3.1 | Jul 2024 | 8B-405B | Matched GPT-4 on several benchmarks | First “frontier-level” open model per Meta |
| DeepSeek-R1 | Jan 2025 | 685B (37B active) | Matched OpenAI o1 on AIME 2024 (79.8% vs 79.2%) | First open reasoning model; 2.5M+ derivative downloads |
| Qwen-2.5 | 2024-2025 | Various | Competitive with frontier | Overtook LLaMA in total downloads by mid-2025 |
| LLaMA 4 | Apr 2025 | Scout 109B, Maverick 400B | 10M context window (Scout) | Extended multimodal capabilities |
The LLaMA Leak (March 2023)
Section titled “The LLaMA Leak (March 2023)”Meta’s LLaMA model weights were leaked on 4chan↗🔗 webLLaMA leakSource ↗Notes, leading to immediate proliferation. Within just seven days of Meta’s controlled release, a complete copy appeared on 4chan and spread across GitHub and BitTorrent networks. Within weeks, the community created:
- “Uncensored” variants that bypassed safety restrictions
- Specialized fine-tunes for specific domains (code, creative writing, roleplay)
- Smaller efficient versions that ran on consumer GPUs
Analysis by Anthropic researchers↗🔗 web★★★★☆AnthropicAnthropic researchSource ↗Notes found that removing safety measures from leaked models required <48 hours and minimal technical expertise, demonstrating the difficulty of maintaining restrictions post-release.
State-Level Adoption Patterns
Section titled “State-Level Adoption Patterns”China’s AI Strategy: CSET analysis↗🔗 web★★★★☆CSET GeorgetownCSET analysisSource ↗Notes shows China increasingly relies on open-source foundations (LLaMA, Stable Diffusion) to reduce dependence on U.S. companies while building domestic capabilities.
Military Applications: RAND’s assessment↗🔗 web★★★★☆RAND CorporationRAND's assessmentSource ↗Notes of defense AI adoption found that 15+ countries now use open-source AI for intelligence analysis, with several developing autonomous weapons systems based on publicly available models.
SB-1047 and Regulatory Attempts
Section titled “SB-1047 and Regulatory Attempts”California’s Senate Bill 1047↗🏛️ governmentSenate Bill 1047Source ↗Notes would have required safety testing for models above compute thresholds. Industry opposition cited proliferation concerns: restrictions would push development overseas and harm beneficial open-source innovation. Governor Newsom’s veto statement↗🏛️ governmentveto statementSource ↗Notes highlighted the enforcement challenges posed by proliferation.
Current State and Trajectory
Section titled “Current State and Trajectory”Capability Gaps Are Shrinking
Section titled “Capability Gaps Are Shrinking”Epoch AI’s tracking↗🔗 web★★★★☆Epoch AIEpoch AIEpoch AI provides comprehensive data and insights on AI model scaling, tracking computational performance, training compute, and model developments across various domains.Source ↗Notes shows the performance gap between frontier and open-source models decreased from ~18 months in 2022 to ~6 months by late 2024, with the gap narrowing to just 1.7% on some benchmarks by 2025. Key factors:
- Architectural innovations diffuse rapidly through papers; 85% of breakthrough AI papers published openly
- Training recipes become standardized; 30% of Python code written by US open-source contributors was AI-generated in 2024
- Compute costs continue declining (~2x annually); inference costs dropped 10x annually for equivalent capability
- Data availability increases through web scraping and synthetic generation
- Model size growth: Mean downloaded model size increased from 827M to 20.8B parameters (2023-2025)
Open-Source Ecosystem Maturity
Section titled “Open-Source Ecosystem Maturity”The open-source AI ecosystem has professionalized significantly, with Hugging Face reaching $130 million revenue in 2024 (up from $10 million in 2023) and a $1.5 billion valuation:
- Hugging Face hosts 2 million+ models with professional tooling; 28.81 million monthly visits
- Together AI and Anyscale provide commercial open-source model hosting
- MLX (Apple), vLLM, and llama.cpp optimize inference for various hardware
- Over 10,000 companies use Hugging Face including Intel, Pfizer, Bloomberg, and eBay
Emerging Control Points
Section titled “Emerging Control Points”Export Controls Timeline and Effectiveness
Section titled “Export Controls Timeline and Effectiveness”| Date | Action | Impact |
|---|---|---|
| Oct 2022 | Initial BIS export controls on advanced AI chips | Began restricting China’s access to frontier AI hardware |
| 2024 | BIS expands FDPR; adds HBM, DRAM controls | 16 PRC entities added; advanced packaging restricted |
| Dec 2024 | 24 equipment types + 140 entities added | Most comprehensive expansion to date |
| Jan 2025 | Biden AI Diffusion Rule: 3-tier global framework | Tier 1 (19 allies): unrestricted; Tier 2 (~150 countries): quantity limits; Tier 3 (≈25 countries): prohibited |
| May 2025 | Trump administration rescinds AI Diffusion Rule | Criticized as “overly bureaucratic”; 65 new Chinese entities added instead |
| Aug 2025 | Nvidia/AMD allowed to sell H20/MI308 to China | US receives 15% of revenue; partial reversal of April freeze |
Compute Governance Results: US controls 75% of worldwide AI compute capacity as of March 2025, while China’s share dropped from 37.3% (2022) to 14.1% (2025). However, despite operating with ~5x less compute, Chinese models narrowed the performance gap from double digits to near parity.
Production Gap: Huawei will produce only 200,000 AI chips in 2025, while Nvidia produces 4-5 million—a 20-25x difference. Yet Chinese labs have innovated around hardware constraints through algorithmic efficiency.
Model Weight Security: Research from Anthropic↗🔗 web★★★★☆AnthropicResearch from AnthropicSource ↗Notes and Google DeepMind↗🔗 web★★★★☆Google DeepMindDeepMindSource ↗Notes explores technical measures for preventing unauthorized model access. RAND’s 2024 report identified multiple attack vectors: insider threats, supply chain compromises, phishing, and physical breaches. A single stolen frontier model may be worth hundreds of millions on the black market.
Key Uncertainties and Cruxes
Section titled “Key Uncertainties and Cruxes”Will Compute Governance Be Effective?
Section titled “Will Compute Governance Be Effective?”Optimistic View: CNAS analysis↗🔗 web★★★★☆CNASCNAS analysisSource ↗Notes suggests that because frontier training requires massive, concentrated compute resources, export controls and facility monitoring could meaningfully slow proliferation.
Pessimistic View: MIT researchers argue↗🔗 web★★★★☆MIT Technology ReviewMIT Technology Review: Deepfake CoverageSource ↗Notes that algorithmic efficiency gains, alternative hardware (edge TPUs, neuromorphic chips), and distributed training techniques will circumvent compute controls.
Key Crux: How quickly will inference efficiency and training efficiency improve? Scaling laws research↗📄 paper★★★☆☆arXivKaplan et al. (2020)Jared Kaplan, Sam McCandlish, Tom Henighan et al. (2020)Source ↗Notes suggests continued rapid progress, but fundamental physical limits may intervene.
Open Source: Net Positive or Negative?
Section titled “Open Source: Net Positive or Negative?”| Argument | For Open Source | Against Open Source |
|---|---|---|
| Power Concentration | Prevents monopolization by 3-5 tech giants | Enables bad actors to match frontier capabilities |
| Safety Research | Allows independent auditing; transparency | Safety fine-tuning can be removed with modest compute |
| Innovation | 10,000+ companies use Hugging Face; democratizes access | Accelerates dangerous capability development |
| Enforcement | Community can identify and patch vulnerabilities | Stanford HAI: “not possible to stop third parties from removing safeguards” |
| Empirical Evidence | RAND, OpenAI studies found no significant uplift vs. internet access for bioweapons | DeepSeek R1 generated CBRN info “that can’t be found on Google” per Anthropic testing |
Key Empirical Findings:
- Open-weight models closed the performance gap from 8% to 1.7% on some benchmarks in a single year
- AI-related incidents rose 56.4% to 233 in 2024—a record high
- China’s AI Safety Governance Framework 2.0 (Sep 2024) represents first Chinese government discussion of open-weights risks
The Core Tradeoff: Ongoing research↗📄 paper★★★☆☆arXivOngoing researchWathela Alhassan, T. Bulik, M. Suchenek (2023)Source ↗Notes attempts to quantify whether open-source accelerates misuse more than defense, but the empirical picture remains contested.
Is Restriction Futile?
Section titled “Is Restriction Futile?”“Futility Thesis”: Some researchers argue that because AI knowledge spreads inevitably through publications, talent mobility, and reverse engineering, governance should focus on defense rather than restriction.
“Strategic Intervention Thesis”: Others contend that targeting specific chokepoints (advanced semiconductors, model weights, specialized knowledge) can meaningfully slow proliferation even if it can’t stop it.
The nuclear proliferation analogy↗🔗 web★★★★☆RAND Corporationnuclear proliferation analogySource ↗Notes suggests both are partially correct: proliferation was slowed but not prevented, buying time for defensive measures and international coordination.
Policy Responses and Interventions
Section titled “Policy Responses and Interventions”Publication Norms Evolution
Section titled “Publication Norms Evolution”Responsible Disclosure Movement: Growing adoption of staged release practices, inspired by cybersecurity norms. Partnership on AI guidelines↗🔗 webPartnership on AIA nonprofit organization focused on responsible AI development by convening technology companies, civil society, and academic institutions. PAI develops guidelines and framework...Source ↗Notes recommend capability evaluation before publication.
Differential Development: Future of Humanity Institute proposals↗🔗 web★★★★☆Future of Humanity Institute**Future of Humanity Institute**Source ↗Notes for accelerating safety-relevant research while slowing dangerous capabilities research.
International Coordination Efforts
Section titled “International Coordination Efforts”UK AI Safety Institute: Established 2024↗🏛️ government★★★★☆UK GovernmentUK AISISource ↗Notes to coordinate international AI safety standards and evaluations.
EU AI Act Implementation: Comprehensive regulation↗🔗 webEU AI ActThe EU AI Act introduces the world's first comprehensive AI regulation, classifying AI applications into risk categories and establishing legal frameworks for AI development and...Source ↗Notes affecting model development and deployment, though enforcement across borders remains challenging.
G7 AI Governance Principles: Hiroshima AI Process↗🔗 webHiroshima AI ProcessSource ↗Notes developing shared standards for AI development and deployment.
Technical Mitigation Research
Section titled “Technical Mitigation Research”Capability Evaluation Frameworks: METR↗🔗 web★★★★☆METRmetr.orgSource ↗Notes, UK AISI↗🏛️ government★★★★☆UK GovernmentUK AISISource ↗Notes, and US AISI↗🏛️ government★★★★★NISTUS AI Safety InstituteSource ↗Notes developing standardized dangerous capability assessments.
Model Weight Protection: Research on cryptographic techniques, secure enclaves, and other methods for preventing unauthorized model access while allowing legitimate use.
Red Team Coordination: Anthropic’s Constitutional AI↗🔗 web★★★★☆AnthropicAnthropic'sSource ↗Notes and similar approaches for systematically identifying and mitigating model capabilities that could enable harm.
Future Scenarios (2025-2030)
Section titled “Future Scenarios (2025-2030)”| Scenario | Probability | Key Drivers | Proliferation Rate | Safety Implications |
|---|---|---|---|---|
| Effective Governance | 20-30% | Strong international coordination; compute controls hold; publication norms shift | Slow (24-36 month frontier lag) | High standards mature; open-source has guardrails |
| Proliferation Acceleration | 35-45% | Algorithmic efficiency gains (10x/year); DeepSeek-style innovations; compute governance circumvented | Very Fast (less than 3 month lag) | Misuse incidents increase 2-5x; “weakest link” problem dominates |
| Bifurcated Ecosystem | 25-35% | Frontier labs coordinate; open-source proliferates separately; China-based models diverge on safety | Mixed (regulated vs. unregulated) | Two parallel ecosystems; defensive measures become critical |
Scenario Details
Section titled “Scenario Details”Scenario 1: Effective Governance Strong international coordination on compute controls and publication norms successfully slows proliferation of most dangerous capabilities. US maintains 75%+ compute advantage; export controls remain effective. Safety standards mature and become widely adopted. Open-source development continues but with better evaluation and safeguards.
Scenario 2: Proliferation Acceleration Algorithmic breakthroughs dramatically reduce compute requirements—DeepSeek demonstrated frontier performance at ~5x less compute cost. Open-source models match frontier performance within months. Governance efforts fail due to international competition and enforcement challenges. Misuse incidents increase but remain manageable.
Scenario 3: Bifurcated Ecosystem Legitimate actors coordinate on safety standards while bad actors increasingly rely on leaked/stolen models. China’s AI Safety Framework diverges from Western approaches. Two parallel AI ecosystems emerge: regulated and unregulated. Defensive measures become crucial.
Cross-Links and Related Concepts
Section titled “Cross-Links and Related Concepts”- Compute Governance - Key technical control point for proliferation
- Dual Use - Technologies that enable both beneficial and harmful applications
- AI ControlSafety AgendaAI ControlAI Control is a defensive safety approach that maintains control over potentially misaligned AI through monitoring, containment, and redundancy, offering 40-60% catastrophic risk reduction if align...Quality: 75/100 - Technical approaches for maintaining oversight as capabilities spread
- SchemingRiskSchemingScheming—strategic AI deception during training—has transitioned from theoretical concern to observed behavior across all major frontier models (o1: 37% alignment faking, Claude: 14% harmful compli...Quality: 74/100 - How proliferation affects our ability to detect deceptive AI behavior
- International Coordination - Global governance approaches to proliferation challenges
- Open Source AI - Key vector for capability diffusion
- Publication Norms - Research community practices affecting proliferation speed
Sources and Resources
Section titled “Sources and Resources”Academic Research
Section titled “Academic Research”- AI and the Future of Warfare - CSET↗🔗 web★★★★☆CSET GeorgetownCenter for Security and Emerging Technology analysisSource ↗Notes
- The Malicious Use of AI - Future of Humanity Institute↗📄 paper★★★☆☆arXivThe Malicious Use of AI - Future of Humanity InstituteMiles Brundage, Shahar Avin, Jack Clark et al. (2018)Source ↗Notes
- Training Compute-Optimal Large Language Models - DeepMind↗📄 paper★★★☆☆arXivHoffmann et al. (2022)Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch et al. (2022)Source ↗Notes
- Constitutional AI: Harmlessness from AI Feedback - Anthropic↗📄 paper★★★☆☆arXivConstitutional AI: Harmlessness from AI FeedbackBai, Yuntao, Kadavath, Saurav, Kundu, Sandipan et al. (2022)Source ↗Notes
Policy and Governance
Section titled “Policy and Governance”- Executive Order on AI - White House↗🏛️ government★★★★☆White HouseBiden Administration AI Executive Order 14110Source ↗Notes
- EU Artificial Intelligence Act↗🔗 webEU AI ActThe EU AI Act introduces the world's first comprehensive AI regulation, classifying AI applications into risk categories and establishing legal frameworks for AI development and...Source ↗Notes
- UK AI Safety Institute↗🏛️ government★★★★☆UK GovernmentUK AISISource ↗Notes
- NIST AI Risk Management Framework↗🏛️ government★★★★★NISTNIST AI Risk Management FrameworkSource ↗Notes
Industry and Technical
Section titled “Industry and Technical”- Meta AI Research on LLaMA↗🔗 web★★★★☆Meta AIMeta Llama 2 open-sourceSource ↗Notes
- OpenAI GPT-4 System Card↗🔗 webOpenAISource ↗Notes
- Anthropic Model Card and Evaluations↗🔗 web★★★★☆Anthropicobservations of strategic reasoningSource ↗Notes
- Hugging Face Open Source AI↗🔗 webHugging Face Open Source AISource ↗Notes
Analysis and Commentary
Section titled “Analysis and Commentary”- State of AI Report 2024↗🔗 webState of AI Report 2025The annual State of AI Report examines key developments in AI research, industry, politics, and safety for 2025, featuring insights from a large-scale practitioner survey.Source ↗Notes
- AI Index Report - Stanford HAI↗🔗 webAI Index ReportStanford HAI's AI Index is a globally recognized annual report tracking and analyzing AI developments across research, policy, economy, and social domains. It offers rigorous, o...Source ↗Notes
- RAND Corporation AI Research↗🔗 web★★★★☆RAND CorporationRAND: AI and National SecuritySource ↗Notes
- Center for Security and Emerging Technology↗🔗 web★★★★☆CSET GeorgetownCSET: AI Market DynamicsI apologize, but the provided content appears to be a fragmentary collection of references or headlines rather than a substantive document that can be comprehensively analyzed. ...Source ↗Notes