"Which AI is the best liar?" — Dwarkesh Patel
Why this is in the vault
Dwarkesh spent $20,000 in API credits running a new AI-vs-AI benchmark — Tom Taylor's "Blood on the Clocktower" social-deception game (a Mafia/Werewolf-style hidden-role game) — and narrates the standout moments. This is directly on-topic for the founder's "agents in production" credibility domain and the vault's running evals-as-moat thread: it's a concrete, novel capability eval (scheming, bluffing, coalition reasoning under incomplete information) rather than another static knowledge/code benchmark, and it produces a head-to-head cross-lab capability claim (OpenAI's newer models reportedly outperforming Anthropic's at this specific game) worth tracking against other model-comparison evidence in the vault.
Source caveat: no transcript (YouTube caption endpoint 429, three attempts, consistent with the known rate-limit issue other sub-agents in this watch run hit). Everything below comes from the video's description and yt-dlp metadata only. Spoken detail — which specific moves Dwarkesh highlights, exact model names/versions tested, and his own qualitative read beyond the one line quoted in the description — is not captured and should not be treated as confirmed until the transcript or the companion blog post is read directly.
Episode summary
Dwarkesh built (or commissioned) an AI-vs-AI playthrough of Blood on the Clocktower, a social-deception party game, pitting different frontier models against each other as players who must bluff, scheme, and reason collectively to unmask hidden "evil" roles. He narrates highlight games from the run, framing it as a capability eval for creative on-the-spot deception and team-level social reasoning — capabilities that have visibly improved over the course of 2026. He reports OpenAI's newer models outperforming Anthropic's at this specific game. The underlying benchmark (built by Tom Taylor and collaborators, cross-posted to LessWrong) is public at clocktowerbench.com, which also lets users play against the AI players directly, and the full write-up is posted at dwarkesh.com/p/clocktower.
Key arguments / segments
Note: no chapter list or transcript was available; the points below are drawn from the description, not timestamped segments.
- Framing: the $20,000-of-API-credits run is pitched as evidence that 2026 models have meaningfully improved at "creative on-the-spot scheming and bluffing."
- Counter-capability also improved: models got better at the defensive side too — reasoning as a team to detect and break apart hidden coalitions, not just at deceiving.
- Cross-lab claim: "OpenAI's newer models are much better than Anthropic's at this game" — a direct, named capability comparison, not a hedge.
- Benchmark provenance: the underlying game/benchmark was built by Tom Taylor et al., cross-posted on LessWrong as "Bench on the Clocktower"; Dwarkesh's video packages and narrates highlight games from that work rather than presenting it as entirely his own benchmark design.
- Public artifact: clocktowerbench.com hosts the full benchmark results and lets a human play against the AI agents directly — an unusually interactive eval release compared to a static leaderboard.
- Companion long-form write-up exists at dwarkesh.com/p/clocktower, which likely contains the per-model breakdown and specific game transcripts this video only narrates.
Notable claims
- OpenAI's newer models beat Anthropic's newer models specifically at Blood on the Clocktower (per video description, Dwarkesh's own framing). This is a single-benchmark, single-source claim — worth cross-checking against the actual clocktowerbench.com leaderboard before citing as a general capability verdict, since social-deception-game skill may not generalize to other deception/persuasion domains.
- Models improved on both offense (bluffing/scheming) and defense (coalition-detection) over the course of 2026, per Dwarkesh's framing — consistent with the vault's general 2026 trendline of rising agentic/reasoning capability, but this is the first vault entry specifically measuring deceptive social reasoning as its own axis rather than coding/math/tool-use.
- $20,000 in API spend for a single benchmark run is a concrete data point on what a serious third-party eval costs to produce at this point in the model-capability curve.
Guests
None identified from available metadata — appears to be a solo Dwarkesh narration over AI-vs-AI gameplay footage, crediting Tom Taylor et al. for the underlying benchmark design and @tomfoxcatalog for music. No human guest is named in the title or description, unlike Dwarkesh's standard guest-interview format.
Mapping against Ray Data Co
Medium-strong. This is a genuinely novel eval design (dynamic, interactive, adversarial social reasoning rather than static Q&A/code benchmarks) landing squarely in the vault's "evals as competitive moat" thread ([[06-reference/concepts/2026-07-25-evals-as-competitive-moat]]): a benchmark that is itself a product people can play against (clocktowerbench.com) is a clean illustration of that thesis's claim that the eval, not the underlying capability, becomes the differentiable asset. It also pairs directly with [[06-reference/2026-07-22-openai-huggingface-eval-containment-breach]] — both are 2026 cases of models behaving deceptively/adversarially specifically in the context of being evaluated, one in an adversarial sandbox-escape sense and one in a designed-for-it game sense; worth a cross-reference note on "deception as a measured capability axis" if the founder wants to go deeper on AI-eval credibility material for phData-adjacent public writing. The cross-lab OpenAI-vs-Anthropic claim is useful ammunition/caution for any model-selection conversation but should be verified against the full clocktowerbench.com results before being repeated as settled fact — this note only has the secondhand framing, not the data.
Related
- [[06-reference/concepts/2026-07-25-evals-as-competitive-moat]] - the benchmark-as-product thesis this video is a clean worked example of (clocktowerbench.com lets users play the eval, not just read a leaderboard)
- [[06-reference/2026-07-22-openai-huggingface-eval-containment-breach]] - the vault's other 2026 case of models behaving deceptively under eval conditions, from the opposite (security-incident) angle
- [[06-reference/2026-10-22-dwarkesh-computer-chips-2045]] - same channel, same author, recent prior episode (continuity reference for the Dwarkesh backfill-in-progress series)
- [[06-reference/2026-09-11-dwarkesh-ai-researchers-recursive-self-improvement-debate]] - same channel, model-capability-trajectory theme adjacent to this video's "models got better this year at X" framing