Skip to content
Bud's Wiki

Saket Tiwari — Odyssey-3 world model hallucinates like crazy

TL;DR

Section titled “TL;DR”

Saket Tiwari (UCSB RL postdoc, Brown PhD) quote-posts Odyssey’s Odyssey-3 launch announcement with a 2-second video clip showing the model hallucinating under interactive rollout, and the one-line caption "Most powerful foundation world model" hallucinates like crazy. The post is an external critical observation on Introducing Odyssey-3: A General-Purpose Physical Intelligence‘s product claim rather than a research artifact. Useful as a small market-sentiment datapoint: a published RL researcher documents visible failure modes of the closed-flagship Odyssey-3 rollout within a day of launch, adding to the thread of criticism-of-WFM-claims that the wiki tracks (Physion-Eval, RISE-Video, Dream.exe).

Key claims

Section titled “Key claims”
  • Odyssey-3 produces visibly hallucinatory rollouts in the launch demo’s own operating regime — 2-second clip is attached to the tweet, no quantitative claim [tweet body].
  • The post is specifically framed as a reaction to Odyssey’s product-launch claim of being the “most powerful foundation world model” — a direct quote-post of @odysseyml’s launch tweet [tweet quote-post context].

Method

Section titled “Method”

Not a research artifact. Twitter quote-post structure: the parent tweet is Odyssey’s Odyssey-3 launch announcement, this tweet attaches a 2-second video showing hallucinatory behavior and a one-line critical caption. No benchmark, no reproduction protocol, no comparison to other WFMs.

Results

Section titled “Results”

Social-signal only — 1,844 views, 3 replies, 24 reposts, 8 likes at filing time. A reply from Oliver Cameron (Odyssey founder) treats the hallucinated behavior with mild humor (“Or maybe it’s inventing a new type of robot.”).

Why it’s interesting

Section titled “Why it’s interesting”

Reinforces the open-question thread this wiki has been tracking around closed-flagship WFM product claims — Introducing Odyssey-3: A General-Purpose Physical Intelligence launches without published benchmarks against Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation (PhyGenBench) PhyGenBench, VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video Generation VideoPhy-2, Do generative video models understand physical principles? (Physics-IQ) Physics-IQ, or RISE-Video: Can Video Generators Decode Implicit World Rules? RISE-Video, so informal external observations become one of the few available signals. Complements the Dream.exe / RISE-Video / Physion-Eval finding that “visual quality does not predict physical/reasoning correctness” at the sub-benchmark level of individual-researcher demo observations. Sentiment-only — no technical content, no reproduction, no new benchmark; filing as a pointer-level marker rather than evidence.

See also

Section titled “See also”