Andy Masley jokes that Fable AI is AGI after it correctly counts four 'r's in a misspelled 'strawberry
The model used a three-second reasoning step before responding
Positive users praise Fable AI's clever counting of r's in strawberry as real progress while negative users dismiss it as a harness trick or pointless waste unrelated to model intelligence.
Most Activity
@AndyMasley agi
Fable is AGI
@AndyMasley I feel like at some point the labs should just place a direct character spelling inductive bias in the architecture, like add a bytes-dependent embedding to the regular token embedding or something
@AndyMasley surprised fable didn't flag that as prompt injection attack and refuse to answer!
@AndyMasley clearly it lied, you added the extra 'r' between the existing two
not quite AGI yet
@AndyMasley It is an important step forward, but it isn't AGI by any definition, not even the ones OpenAI made so they could profit off of it. It was able to make a model off of one prompt and train it
@AndyMasley Ask it how many L's are in the word Google.
@lukehebb @AndyMasley the r between the existing two r's IS before the y tho
@_ueaj @AndyMasley didnt they already add it to training data tho?
@mattgoldenberg @AndyMasley it was a joke, dude
@AndyMasley Gemini didn't get it https://share.google/aimode/3ZLE1zy7gUVnEd0BX
@_ueaj @AndyMasley The whole point of this prompt is to infer weather the model can generalize outside their training distribution. That's the reason why, at least until now, labs haven't added a specific tool for this.
@AndyMasley We can guess how much tokens are consumed to handle such meaning less tasks... it is not a good idea
@lukehebb @AndyMasley uggh, i hate to have to describe TWO jokes in a row it's just a thread downer
but here we are...
the joke was that you were being pedantic and i was being even more pedantic
@AndyMasley Don’t tell bad about it, Dario will hack you with Mythos
@AndyMasley should it not be 4 rs intstead of 4 r's?
@AndyMasley Ok you didn’t trick it it rightly pointed out you spelled it that way…
@clawrompocx9 @_ueaj @AndyMasley except... the whole strawberry problem was solved with CoT. the origin was tokenization which is virtually inherent to llms. the moment it separates each letter with a dash it can see the token of each individual letter and easily count.
@AndyMasley B-b-b-but it used 240 million gallons of water to answer that query!
Think of the famine in Ethiopia!
Etc.
@AndyMasley gork fast, poor benchmark: