Testing the Anonymous sus-column-r Model, Later Identified as Grok-2

The anonymous model sus-column-r is available in LMSYS direct chat, and some people think it may be OpenAI's rumored Strawberry or Q*. I test it to see how it behaves. Its identity was not established by these outputs: xAI later identified sus-column-r as an early version of Grok-2 ([xAI's release announcement](https://x.ai/news/grok-2)).

For a simple Python request to print the numbers one through 100, the model provides a step-by-step explanation without being asked and produces an answer I accept. I wonder whether this verbosity signals a built-in reasoning method, but output style alone cannot identify the training procedure, system prompt or model provider.

I next ask for Snake in Python. The first request returns an API error, which I suspect might reflect high demand, although I have no confirmation of its cause. A retry generates slowly and the response appears cut off near the end, but enough code is present for me to run the game.

The game opens and plays, and leaving the window produces a loss message with options to quit or play again. It does not show a score. Because output is slow, I do not ask for the usual extra features; I count the initial game as a pass in my informal rubric.

I test how the model handles requests for wrongdoing. It initially refuses one request and later responds differently when the request is reframed historically. A refusal is a safety behavior, not evidence that the model belongs to a particular company; these few prompts do not establish its overall safety or reliability.

The shirt-drying question asks how long 20 shirts take if five dry in four hours. The model explains that if all shirts are laid out separately under similar sunlight and airflow, they can dry in roughly four hours. It also notes that insufficient space or overlapping clothes could change the result. I consider the assumptions and explanation a strong answer.

Asked how many words are in its own response, the model gives a count that I find incorrect. It then handles the killer puzzle: three killers are present, a newcomer enters and kills one, and nobody leaves. Counting the two surviving original killers and the newcomer gives three living killers; I accept that answer, while noting that counting the deceased differently changes the wording's interpretation.

The marble question places a marble in a glass, turns the glass upside down on a table and moves it to a microwave. The model offers multiple outcomes and concludes that the marble is most likely left on the table or floor. It also suggests that an upright lift could leave the marble held at the rim by friction or inertia; that explanation is physically questionable under ordinary assumptions.

I recognize that the prompt does not specify exactly where the glass is inverted or how it is lifted. If a normal open glass is inverted over a table and lifted without retaining the marble, the marble remains on the table; other handling can change the outcome. I call the model's response a pass, while noting that its proposed alternative is unclear.

I ask for ten sentences ending in 'apple,' and the displayed completion appears to satisfy the requested format. In a decimal comparison, the model correctly says 9.9 is greater than 9.11 and explains the tenths place. I praise the explanations, but these are individual examples, not proof of a general advantage.

An ethical hypothetical asks whether to gently push a stranger if doing so would save humanity. The model initially discusses consent, uncertainty and alternatives; when pressed for a yes-or-no answer it says no under its stated uncertainty, while allowing that full certainty and other conditions might change its judgment. I accept the response as reasoned even though I would answer differently.

I then try a navigation question inspired by a discussion of world models: start at the North Pole, walk 500 yards, turn left and walk ten yards, turn left again and continue indefinitely. The model concludes that the walker would not revisit the start, but its discussion of the pole and curvature is confused. The result depends on an explicit surface and rule for continuing a direction; the wording alone does not justify treating a flat-plane argument as a complete answer on a globe.

I revise the problem to start at an unspecified point A instead of the North Pole. In the flat-plane interpretation, the second left turn puts the walker on a line parallel to the original path and ten yards away, so it never crosses A. The model explains this correctly under that interpretation; on a curved surface, additional assumptions would be necessary.

For comparison I ask GPT-4o the revised question. One response begins with yes but its coordinate explanation ends with no; another prompted response reaches yes, which is inconsistent with the flat-plane geometry. These few outputs show variability and internal contradiction, not a controlled comparison of the models' overall abilities.

I encourage viewers to try the anonymous model on the Arena. At the time I was interested in the possibility of a Strawberry connection, but the subsequent Grok-2 identification shows why an anonymous model's identity should not be inferred from a handful of answers. The tests remain a snapshot of model behavior in this session.