Conjecture
Page Status
Quality:37 (Draft)⚠️
Importance:28 (Peripheral)
Last edited:2026-01-29 (3 days ago)
Words:1.6k
Structure:
📊 18📈 0🔗 35📚 0•15%Score: 10/15
LLM Summary:Conjecture is a 30-40 person London-based AI safety org founded 2021, pursuing Cognitive Emulation (CoEm) - building interpretable AI from ground-up rather than aligning LLMs - with $30M+ Series A funding. Founded by Connor Leahy (EleutherAI), they face high uncertainty about CoEm competitiveness (3-5 year timeline) and commercial pressure risks.
Issues (1):
- QualityRated 37 but structure suggests 67 (underrated by 30 points)
Research Lab
Conjecture
Importance28
Websiteconjecture.dev
Related
Organizations
Researchers
Safety Approachs
Overview
Section titled “Overview”Conjecture is an AI safety research organization founded in 2021 by Connor LeahyResearcherConnor LeahyBiography of Connor Leahy, CEO of Conjecture AI safety company, who transitioned from co-founding EleutherAI (open-source LLMs) to focusing on interpretability-first alignment. He advocates for ver...Quality: 19/100 and a team of researchers concerned about existential risks from advanced AI. The organization pursues a distinctive technical approach centered on “Cognitive Emulation” (CoEm) - building interpretable AI systems based on human cognition principles rather than aligning existing large language models.
Based in London with a team of 30-40 researchers, Conjecture raised over $10M in Series A funding in 2023. Their research agenda emphasizes mechanistic interpretability and understanding neural network internals, representing a fundamental alternative to mainstream prosaic alignment approachesArgumentWhy Alignment Might Be HardComprehensive synthesis of why AI alignment is fundamentally difficult, covering specification problems (value complexity, Goodhart's Law), inner alignment failures (mesa-optimization, deceptive al...Quality: 61/100 pursued by organizations like AnthropicLabAnthropicComprehensive profile of Anthropic tracking its rapid commercial growth (from $1B to $7B annualized revenue in 2025, 42% enterprise coding market share) alongside safety research (Constitutional AI...Quality: 51/100 and OpenAILabOpenAIComprehensive organizational profile of OpenAI documenting evolution from 2015 non-profit to commercial AGI developer, with detailed analysis of governance crisis, safety researcher exodus (75% of ...Quality: 46/100.
| Aspect | Assessment | Evidence | Source |
|---|---|---|---|
| Technical Innovation | High | Novel CoEm research agenda | Conjecture Blog↗🔗 webConjecture BlogSource ↗Notes |
| Funding Security | Strong | $30M+ Series A (2023) | TechCrunch Reports↗🔗 web★★★☆☆TechCrunchTechCrunch ReportsSource ↗Notes |
| Research Output | Moderate | Selective publication strategy | Research Publications↗🔗 webResearch PublicationsSource ↗Notes |
| Influence | Growing | European AI policy engagement | UK AISI↗🏛️ government★★★★☆UK GovernmentUK AISISource ↗Notes |
Risk Assessment
Section titled “Risk Assessment”| Risk Category | Severity | Likelihood | Timeline | Trend |
|---|---|---|---|---|
| CoEm Uncompetitive | High | Moderate | 3-5 years | Uncertain |
| Commercial Pressure Compromise | Medium | High | 2-3 years | Worsening |
| Research Insularity | Low | Moderate | Ongoing | Stable |
| Funding Sustainability | Medium | Low | 5+ years | Improving |
Founding and Evolution
Section titled “Founding and Evolution”Origins (2021)
Section titled “Origins (2021)”Conjecture emerged from the EleutherAI collective, an open-source AI research group that successfully recreated GPT-3 as open-source models (GPT-J, GPT-NeoX). Key founding factors:
| Factor | Impact | Details |
|---|---|---|
| EleutherAI Experience | High | Demonstrated capability replication feasibility |
| Safety Concerns | High | Recognition of risks from capability proliferation |
| European Gap | Medium | Limited AI safety ecosystem outside Bay Area |
| Funding Availability | Medium | Growing investor interest in AI safety |
Philosophical Evolution: The transition from EleutherAI’s “democratize AI” mission to Conjecture’s safety-focused approach represents a significant shift in thinking about AI development and publication strategies.
Funding Trajectory
Section titled “Funding Trajectory”| Year | Funding Stage | Amount | Impact |
|---|---|---|---|
| 2021 | Seed | Undisclosed | Initial team of ≈15 researchers |
| 2023 | Series A | $30M+ | Scaled to 30-40 researchers |
| 2024 | Operating | Ongoing | Sustained research operations |
Cognitive Emulation (CoEm) Research Agenda
Section titled “Cognitive Emulation (CoEm) Research Agenda”Core Philosophy
Section titled “Core Philosophy”Conjecture’s signature approach contrasts sharply with mainstream AI development:
| Approach | Philosophy | Methods | Evaluation |
|---|---|---|---|
| Prosaic Alignment | Train powerful LLMs, align post-hoc | RLHF, Constitutional AI | Behavioral testing |
| Cognitive Emulation | Build interpretable systems from ground up | Human cognition principles | Mechanistic understanding |
Key Research Components
Section titled “Key Research Components”Mechanistic Interpretability
- Circuit discovery in neural networks
- Feature attribution and visualization
- Scaling interpretability to larger models
- Interpretability researchSafety AgendaInterpretabilityMechanistic interpretability has extracted 34M+ interpretable features from Claude 3 Sonnet with 90% automated labeling accuracy and demonstrated 75-85% success in causal validation, though less th...Quality: 66/100 collaboration
Architecture Design
- Modular systems for better control
- Interpretability-first design choices
- Trading capabilities for understanding
- Novel training methodologies
Model Organisms
- Smaller, interpretable test systems
- Alignment property verification
- Deception detection research
- Goal representation analysis
Key Personnel
Section titled “Key Personnel”Leadership Team
CL
Connor Leahy
CEO and Co-founder
SB
Sid Black
Co-founder
GA
Gabriel Alfour
CTO
Connor Leahy Profile
Section titled “Connor Leahy Profile”| Aspect | Details |
|---|---|
| Background | EleutherAI collective member, GPT-J contributor |
| Evolution | From open-source advocacy to safety-focused research |
| Public Role | Active AI policy engagement, podcast appearances |
| Views | Short AI timelines, high P(doom), interpretability-necessary |
Timeline Estimates: Leahy has consistently expressed short AI timelineAgi TimelineComprehensive synthesis of AGI timeline forecasts showing dramatic acceleration: expert median dropped from 2061 (2018) to 2047 (2023), Metaculus from 50 years to 5 years since 2020, with current p...Quality: 59/100 views, suggesting AGI within years rather than decades.
Research Focus Areas
Section titled “Research Focus Areas”Mechanistic Interpretability
Section titled “Mechanistic Interpretability”| Research Area | Status | Key Questions |
|---|---|---|
| Circuit Analysis | Active | How do transformers implement reasoning? |
| Feature Extraction | Ongoing | What representations emerge in training? |
| Scaling Methods | Development | Can interpretability scale to AGI-level systems? |
| Goal Detection | Early | How can we detect goal-directedness mechanistically? |
Comparative Advantages
Section titled “Comparative Advantages”| Organization | Primary Focus | Interpretability Approach |
|---|---|---|
| Conjecture | CoEm, ground-up interpretability | Design-time interpretability |
| AnthropicLabAnthropicComprehensive profile of Anthropic tracking its rapid commercial growth (from $1B to $7B annualized revenue in 2025, 42% enterprise coding market share) alongside safety research (Constitutional AI...Quality: 51/100 | Frontier models + interpretability | Post-hoc analysis of LLMs |
| ARCOrganizationARCComprehensive overview of ARC's dual structure (theory research on Eliciting Latent Knowledge problem and systematic dangerous capability evaluations of frontier AI models), documenting their high ...Quality: 43/100 | Theoretical alignment | Evaluation and ELK research |
| RedwoodOrganizationRedwood ResearchRedwood Research is an AI safety lab founded in 2021 that has made significant contributions to mechanistic interpretability and, more recently, pioneered the "AI control" research agenda. | AI control | Interpretability for control |
Strategic Position
Section titled “Strategic Position”Theory of Change
Section titled “Theory of Change”Conjecture’s pathway to AI safety impact:
- Develop scalable interpretability techniques for powerful AI systems
- Demonstrate CoEm viability as competitive alternative to black-box scaling
- Influence field direction toward interpretability-first development
- Inform governance with technical feasibility insights
- Build safe systems using CoEm principles if successful
European AI Safety Hub
Section titled “European AI Safety Hub”| Role | Impact | Examples |
|---|---|---|
| Geographic Diversity | High | Alternative to Bay Area concentration |
| Policy Engagement | Growing | UK AISIOrganizationUK AI Safety InstituteThe UK AI Safety Institute (renamed AI Security Institute in Feb 2025) operates with ~30 technical staff and 50M GBP annual budget, conducting frontier model evaluations using its open-source Inspe...Quality: 52/100 consultation |
| Talent Development | Moderate | European researcher recruitment |
| Community Building | Early | Workshops and collaborations |
Challenges and Criticisms
Section titled “Challenges and Criticisms”Technical Feasibility
Section titled “Technical Feasibility”| Challenge | Severity | Status |
|---|---|---|
| CoEm Competitiveness | High | Unresolved - early stage |
| Interpretability Scaling | High | Active research question |
| Human Cognition Complexity | Medium | Ongoing investigation |
| Timeline Alignment | High | Critical if AGI timelinesAgi TimelineComprehensive synthesis of AGI timeline forecasts showing dramatic acceleration: expert median dropped from 2061 (2018) to 2047 (2023), Metaculus from 50 years to 5 years since 2020, with current p...Quality: 59/100 short |
Organizational Tensions
Section titled “Organizational Tensions”Commercial Pressure vs Safety Mission
- VC funding creates return expectations
- Potential future deployment pressure
- Comparison to Anthropic’s commercialization pathLabAnthropicComprehensive profile of Anthropic tracking its rapid commercial growth (from $1B to $7B annualized revenue in 2025, 42% enterprise coding market share) alongside safety research (Constitutional AI...Quality: 51/100
Publication Strategy Criticism
- Shift from EleutherAI’s radical openness
- Selective research sharing decisions
- Balance between transparency and safety
Current Research Outputs
Section titled “Current Research Outputs”Published Work
Section titled “Published Work”| Type | Focus | Impact |
|---|---|---|
| Technical Papers | Interpretability methods | Research community |
| Blog Posts | CoEm explanations | Public understanding |
| Policy Contributions | Technical feasibility | Governance decisions |
| Open Source Tools | Interpretability software | Research ecosystem |
Research Questions
Section titled “Research Questions”Key Questions (6)
- Can CoEm produce AI systems competitive with scaled LLMs?
- Is mechanistic interpretability sufficient for AGI safety verification?
- How will commercial pressures affect Conjecture's research direction?
- What role should interpretability play in AI governance frameworks?
- Can cognitive emulation bridge neuroscience and AI safety research?
- How does CoEm relate to other alignment approaches like Constitutional AI?
Timeline and Risk Estimates
Section titled “Timeline and Risk Estimates”Leadership Risk Assessments
Section titled “Leadership Risk Assessments”Conjecture’s leadership has articulated clear views on AI timelines and safety approaches, which fundamentally motivate their Cognitive Emulation research agenda and organizational strategy:
| Expert/Source | Estimate | Reasoning |
|---|---|---|
| Connor Leahy | AGI: 2-10 years | Leahy has consistently expressed short AI timeline views across multiple public statements and podcasts from 2023-2024, suggesting transformative AI systems could emerge within years rather than decades. These short timelines create urgency for developing interpretability-first approaches before AGI arrives. |
| Connor Leahy | P(doom): High without major changes | Leahy has expressed significant concern about the default trajectory of AI development in 2023 statements, arguing that prosaic alignment approaches pursued by frontier labs are insufficient to ensure safety. This pessimism about conventional alignment motivates Conjecture’s alternative CoEm approach. |
| Conjecture Research | Prosaic alignment: Insufficient | The organization’s core research direction reflects a fundamental assessment that post-hoc alignment of large language models through techniques like RLHF and Constitutional AI cannot provide adequate safety guarantees. This view, maintained since founding, drives their pursuit of interpretability-first system design. |
| Organization | Interpretability: Necessary for safety | Conjecture’s founding premise holds that mechanistic interpretability is not merely useful but necessary for AI safety verification. This fundamental research assumption distinguishes them from organizations pursuing behavioral safety approaches and shapes their entire technical agenda. |
Future Scenarios
Section titled “Future Scenarios”Research Trajectory Projections
Section titled “Research Trajectory Projections”| Timeline | Optimistic | Realistic | Pessimistic |
|---|---|---|---|
| 2-3 years | CoEm demonstrations, policy influence | Continued interpretability advances | Commercial pressure compromises |
| 3-5 years | Competitive interpretable systems | Mixed results, partial success | Research agenda stagnates |
| 5+ years | Field adoption of CoEm principles | Portfolio contribution to safety | Marginalized approach |
Critical Dependencies
Section titled “Critical Dependencies”| Factor | Importance | Uncertainty |
|---|---|---|
| Technical Feasibility | Critical | High - unproven at scale |
| Funding Continuity | High | Medium - VC expectations |
| AGI Timeline | Critical | High - if very short, insufficient time |
| Field Receptivity | Medium | Medium - depends on results |
Relationships and Collaborations
Section titled “Relationships and Collaborations”Within AI Safety Ecosystem
Section titled “Within AI Safety Ecosystem”| Organization | Relationship | Collaboration Type |
|---|---|---|
| AnthropicLabAnthropicComprehensive profile of Anthropic tracking its rapid commercial growth (from $1B to $7B annualized revenue in 2025, 42% enterprise coding market share) alongside safety research (Constitutional AI...Quality: 51/100 | Friendly competition | Interpretability research sharing |
| ARCOrganizationARCComprehensive overview of ARC's dual structure (theory research on Eliciting Latent Knowledge problem and systematic dangerous capability evaluations of frontier AI models), documenting their high ...Quality: 43/100 | Complementary | Different technical approaches |
| MIRIOrganizationMIRIComprehensive organizational history documenting MIRI's trajectory from pioneering AI safety research (2000-2020) to policy advocacy after acknowledging research failure, with detailed financial da...Quality: 50/100 | Aligned concerns | Skepticism of prosaic alignment |
| Academic Labs | Collaborative | Interpretability technique development |
Policy and Governance
Section titled “Policy and Governance”UK Engagement
- UK AI Safety InstituteOrganizationUK AI Safety InstituteThe UK AI Safety Institute (renamed AI Security Institute in Feb 2025) operates with ~30 technical staff and 50M GBP annual budget, conducting frontier model evaluations using its open-source Inspe...Quality: 52/100 consultation
- Technical feasibility assessments
- European AI Act discussions
International Influence
- Growing presence in global AI safety discussions
- Alternative perspective to US-dominated discourse
- Technical grounding for governance approaches
Sources & Resources
Section titled “Sources & Resources”Primary Sources
Section titled “Primary Sources”| Type | Source | Description |
|---|---|---|
| Official Website | Conjecture.dev↗🔗 webConjecture BlogSource ↗Notes | Research updates, team information |
| Research Papers | Google Scholar↗🔗 web★★★★☆Google ScholarGoogle ScholarSource ↗Notes | Technical publications |
| Blog Posts | Conjecture Blog↗🔗 webConjecture BlogSource ↗Notes | Research explanations, philosophy |
| Interviews | Connor Leahy Talks↗🎤 talkConnor Leahy TalksSource ↗Notes | Leadership perspectives |
Secondary Analysis
Section titled “Secondary Analysis”| Type | Source | Focus |
|---|---|---|
| AI Safety Analysis | LessWrong Posts↗✏️ blog★★★☆☆LessWrongLessWrong PostsSource ↗Notes | Community discussion |
| Technical Reviews | Alignment Forum↗✏️ blog★★★☆☆Alignment ForumAI Alignment ForumSource ↗Notes | Research evaluation |
| Policy Reports | GovAI Analysis↗🏛️ government★★★★☆Centre for the Governance of AIGovAIA research organization focused on understanding AI's societal impacts, governance challenges, and policy implications across various domains like workforce, infrastructure, and...Source ↗Notes | Governance implications |
| Funding News | TechCrunch Coverage↗🔗 web★★★☆☆TechCrunchTechCrunch CoverageSource ↗Notes | Business developments |
Related Resources
Section titled “Related Resources”| Topic | Internal Links | External Resources |
|---|---|---|
| Interpretability | Technical InterpretabilitySafety AgendaInterpretabilityMechanistic interpretability has extracted 34M+ interpretable features from Claude 3 Sonnet with 90% automated labeling accuracy and demonstrated 75-85% success in causal validation, though less th...Quality: 66/100 | Anthropic Interpretability↗🔗 web★★★★☆Transformer CircuitsMechanistic InterpretabilitySource ↗Notes |
| Alignment Approaches | Why Alignment is HardArgumentWhy Alignment Might Be HardComprehensive synthesis of why AI alignment is fundamentally difficult, covering specification problems (value complexity, Goodhart's Law), inner alignment failures (mesa-optimization, deceptive al...Quality: 61/100 | AI Alignment Forum↗✏️ blog★★★☆☆Alignment ForumAI Alignment ForumSource ↗Notes |
| European AI Policy | UK AISIOrganizationUK AI Safety InstituteThe UK AI Safety Institute (renamed AI Security Institute in Feb 2025) operates with ~30 technical staff and 50M GBP annual budget, conducting frontier model evaluations using its open-source Inspe...Quality: 52/100 | EU AI Office↗🔗 web★★★★☆European Union**EU AI Office**Source ↗Notes |
| Related Orgs | Safety Organizations | AI Safety Community↗🔗 webAI Safety CommunitySource ↗Notes |