[alariccole]

ChatGPT just told me to put the turkey in my toaster oven legs facing the door, and you think it can replace school. Unless there is a massive architectural change that can be provably verified by third parties, this can never be. I’d hate for my unschooled surgeon to check an llm while I’m under.

[CamperBob2]

Just curious, not being a turkey SME, what's the downside to positioning the turkey that way?

[patrickmay]

Most turkeys of my acquaintance would not fit into a toaster oven without some percussive assistance.

[CamperBob2]

I see, I overlooked the 'toaster' part. That's a good world model benchmark question for models and a good reading comprehension question for humans. :-P

GPT 5.1 Pro made the same mistake ("Face the legs away from the door.") Claude Sonnet 4.5 agreed but added "Note: Most toaster ovens max out around 10-12 pounds for a whole turkey."

Gemini 3 acknowledged that toaster ovens are usually very compact and that the legs shouldn't be positioned where they will touch the glass door. When challenged, it hand-waved something to the effect of "Well, some toaster ovens are large countertop convection units that can hold up to a 12-pound turkey." When asked for a brand and model number of such an oven, it backtracked and admitted that no toaster oven would be large enough.

Changing the prompt to explicitly specify a 12-pound turkey yielded good answers ("A 12-pound turkey won't fit in a toaster oven - most max out at 4-6 pounds for poultry. Attempting this would be a fire hazard and result in dangerously uneven cooking," from Sonnet.)

So, progress, but not enough.

[bigstrat2003]

Don't worry, someone will put another hack on top the model to teach it to handle this specific case better. That will totally fix the problem, right? Right?

[charcircuit]

What's the alternate if someone didn't know something during a procedure? Just wing it? Getting another data point from an LLM seems beneficial to me.

[axus]

Ask a human who does. If there are no competent humans on-call before the procedure starts, reschedule the procedure.

[bigstrat2003]

A trained professional making their best guess is far more capable and trustworthy than the slop LLMs put out. So yeah, winging it is a good alternative here.