Conjecture

Stated Theory of Change

Conjecture has had three stated theories of change, each superseding the last:

ToC 1 (March-Nov 2022): Alignment requires a breakthrough nobody has found yet. Maximize the probability of finding it by running many decorrelated research bets in parallel. "If you need to roll high, roll many dice."

ToC 2 (Feb 2023-present): Cognitive Emulation (CoEm). Build safe AI by treating it as a software engineering problem. Instead of building opaque, end-to-end black-box systems, compose modular, auditable building blocks that emulate human-like cognition. This keeps AI in the "human regime" where existing institutions can govern it, while also being commercially valuable (enterprises want reliable AI, not "a quirky brain in a jar with a personality disorder").

ToC 3 (Nov 2023-present, via ControlAI): The AI safety community has failed to inform the public and lawmakers about extinction risks. Direct institutional engagement -- briefing lawmakers, getting public commitments to binding regulation -- is the bottleneck, not technical research or persuasion. International treaties on AI development are needed, starting with non-superpowers and creating incentives for superpowers to join.

Revealed Theory of Change

Actions tell a different story than mission statements.

Revealed priorities by resource allocation:

• ~85% of effort appears directed at product development (Axiom, Cubed, tactics.dev, CTAC language, AWS infrastructure)
• ~10% at public communication and policy (Leahy's media appearances, ControlAI coordination)
• ~5% or less at alignment research (no published results since late 2022 from researchers still at Conjecture)

Revealed priorities by personnel:

• Every named alignment researcher from the 2022 era has departed
• Current team appears product-focused (VP of Product & Commercial, AWS infrastructure partnership)
• ControlAI has absorbed the policy/governance talent (Miotti, Shimi)

Revealed priorities by leadership attention:

• Leahy's public appearances increasingly focus on policy and x-risk communication
• Alfour's most recent interview (April 2025) is about international treaties, not CoEm
• The Cognitive Software Roadmap reads more as a philosophical manifesto than a technical progress report

The gap between stated and revealed theory of change is significant. Conjecture says it is building safe AI through Cognitive Emulation. What it appears to be doing is (a) building AI products that use safety as a brand differentiator, while (b) its leadership contributes to AI policy through ControlAI and public advocacy. Neither of these is bad, but neither is what the stated theory of change promises.

Key Assumptions

Assumption 1: CoEm is a viable approach to safe AI

• Evidence for: The software engineering analogy is compelling. Type systems, formal methods, and modular design genuinely do make software more reliable. The Wright Brothers analogy resonates.
• Evidence against: No published empirical results in 3 years. The researchers who might have validated it have left. Phase 5 (actual CoEm) of the roadmap is described as "speculative." The proposal itself is labeled "naive" by its own authors. Current LLM capabilities may not decompose cleanly into the primitive building blocks CoEm requires.
• Testable: Yes. Does a CoEm system actually outperform standard approaches on reliability/auditability benchmarks? This test has not been conducted or published.
• If wrong: Conjecture is building ordinary AI products with a safety brand, not a fundamentally different approach to AI.

Assumption 2: The world will choose not to build unbounded ASI, making CoEm relevant

• Evidence for: Growing policy momentum (ControlAI's 112 lawmakers, House of Lords debates). Public concern about AI risks is increasing.
• Evidence against: Every major lab is racing to build the most capable systems possible. DeepSeek, OpenAI, Anthropic, Google DeepMind are all scaling aggressively. No country has implemented binding restrictions on AI capabilities.
• Testable: Yes, by policy outcomes over the next 2-5 years.
• If wrong: CoEm is "a nice idea that gets steamrolled by GPT-N." Conjecture's products compete on normal business merits, not on being safer.

Assumption 3: Leahy's public profile is a durable asset rather than a liability

• Evidence for: Genuine attention-getting. Builds awareness of AI risks. Attracts talent and investors.
• Evidence against: "World's second most famous AI doomer" persona may alienate mainstream ML researchers and enterprise customers. P(doom) 90% creates credibility tension with running a product company.
• If wrong: Conjecture is perceived as an ideological org rather than a serious technical effort.

Strengths

1. Genuine intellectual honesty. The 8-month retrospective is the most candid self-assessment by any AI safety org. Publishing "most of our efforts have not made meaningful progress" requires real epistemic integrity. This culture of honest self-assessment, if it persists, is extremely valuable.

2. ControlAI's DIP approach is demonstrably effective. 112 UK lawmakers supporting their campaign in ~1 year, House of Lords debates on AI extinction risk -- these are concrete policy outcomes. Alfour's "Spectre" critique of the safety community may be self-serving but contains real insight: most safety orgs have not been direct enough about extinction risks with policymakers.

3. Financial independence from frontier labs. Conjecture does not depend on Anthropic, OpenAI, or Google for funding, compute, or career pipeline. This gives them genuine freedom to criticize, which they exercise.

4. The cognitive software vision is intellectually serious. Even if CoEm is unproven, the Cognitive Software Roadmap's critique of current AI development practices is substantive. The Oracle Database analogy for current AI development is apt. The programming language approach draws on deep software engineering principles (type systems, formal methods, Unix philosophy).

5. Early-mover advantage on sparse autoencoders. The Sharkey et al. SAE work preceded and influenced Anthropic's much more famous work. This demonstrates the team had genuine technical insight, even if those researchers have since left.

Weaknesses and Risks

1. The talent exodus is severe. Virtually every named researcher from the founding era has left. The people who produced the most cited work are gone. What remains is a CEO/communicator, a CTO/architect, and what appears to be a product team. An alignment research org without alignment researchers is a product company with an alignment brand.

2. No empirical validation of CoEm after 3 years. The most important thing Conjecture could do to establish credibility -- publish benchmarks showing CoEm systems are actually more reliable and auditable than standard approaches -- has not been done. This absence undermines every claim about the technical approach.

3. Serial pivots erode credibility. Interpretability -> Simulators -> Epistemology -> CoEm -> Products. Each pivot was individually defensible, but the pattern reads as a team that keeps looking for product-market fit rather than a research org methodically working a problem.

4. Governance concentration risk. Leahy is sole director and sole person with significant control. No board, no external oversight. If Leahy's judgment fails, there is no check.

5. The CoEm thesis may be irrelevant if the world doesn't slow down. CoEm explicitly depends on the world making "the sensible choice to not go straight for ASI." If frontier labs keep building unbounded systems, CoEm-bounded systems may be too weak to matter commercially or for safety.

6. The Christiano critique is unresolved. Alfour's own arguments suggest narrow technical alignment is insufficient because aligned systems become parts of larger unaligned systems. If true, this undermines Conjecture's own technical work in favor of ControlAI's policy work. The intellectual case for CoEm is weakest from Conjecture's own CTO.

Cross-References

• ControlAI is the most important entity in Conjecture's orbit. It may end up being the lasting contribution of the Conjecture ecosystem, even if Conjecture itself dissolves (as Manifold's 77% probability suggests).
• EleutherAI is the origin community. Conjecture represents the "alignment branch" of EleutherAI, while Stability AI represents the "capabilities branch." Both emerged from the same community with very different outcomes.
• Anthropic had its sparse autoencoders work preceded by Conjecture's SAE interim report. This is the clearest evidence of Conjecture contributing to the broader technical alignment field.
• ARC (Alignment Research Center) -- the Christiano-Alfour debate is a key intellectual document. Christiano's "buck-passing" critique is the strongest technical argument against Conjecture's trajectory.
• PauseAI and AI advocacy orgs -- ControlAI is described as the "inside game" strategy (direct lawmaker engagement) vs. PauseAI's "outside game" (public demonstrations). They are complementary, not competing.

What Would Change This Assessment

• Published CoEm benchmarks showing genuine improvement over standard approaches would be the single biggest positive update. If CoEm systems demonstrably outperform standard LLM approaches on reliability/auditability in controlled evaluations, the entire assessment shifts.
• Significant new funding round or revenue milestone would indicate external validation of the product direction.
• Successful international AI treaty with ControlAI involvement would validate the policy theory of change.
• Return of alignment researchers or significant new technical hires would indicate the research direction is alive, not abandoned.
• Conjecture dissolving with ControlAI continuing would confirm the Manifold thesis that the policy arm is the lasting contribution.

Self-Critique

What sources should I have checked but didn't:

• The Compendium (100+ page PDF, not fetched). Contains the most complete argument for extinction risk from the Conjecture team.
• Full debate transcripts (Leahy vs. Beff Jezos, Leahy vs. George Hotz) -- only available as video/audio.
• The original Conjecture AMA on Alignment Forum (founding Q&A).
• Full text of the LessWrong critiques post (only search summary available).

Where is this analysis potentially biased:

• I may be reading the talent exodus too negatively. In a fast-moving field, researchers move frequently, and departures don't necessarily indicate organizational failure.
• The absence of published CoEm results may reflect legitimate infohazard concerns rather than lack of progress -- Conjecture's stated policy is to keep things secret by default.
• I may be underweighting ControlAI's importance by framing it as a "spinoff" rather than as the primary vehicle through which the Conjecture ecosystem creates value.

What would a thoughtful person who disagrees say: "Conjecture correctly identified that the bottleneck is not technical alignment research but policy and governance. They pivoted accordingly -- ControlAI for policy, Conjecture for products that demonstrate what safe AI looks like in practice. The researcher departures reflect normal career mobility, not organizational failure. CoEm is unpublished because it's commercially sensitive, not because it doesn't work. Leahy and Alfour are doing exactly the right thing by building both a business case and a policy case for safe AI, rather than publishing papers that get cited but change nothing."

My single weakest claim: That the talent exodus indicates organizational dysfunction rather than normal career mobility. I have no inside knowledge and no statements from departed staff. The inference is pattern-based, not evidence-based.

What information would most change my view: Published, peer-reviewed benchmarks showing CoEm systems meaningfully outperforming standard approaches on reliability and auditability metrics. This would resolve the central uncertainty about whether Conjecture is building something genuinely new or repackaging standard AI engineering with safety branding.