Some AI models get more accurate at maths if you ask them to respond as if they are a Star Trek character, ML engineers say::Researchers asking a chatbot to optimize its own prompts found it was best at solving grade-school math when acting like it was on Star Trek.
Its because it dosent try to answer right, but more what it thinks toy want to see/read 🤷
LLM detractors hate this one weird trick
We’ve finally figured out how to trick the computer that’s bad at math into being less bad at math.
Doh. This says to have the AI write the prompt for you, but it doesn’t give any examples of doing that.
I don’t want to get into a rabbit hole looking up examples from the wide internet.
Reverse the polarity!
AI models get more dishonest if you ask them to respond as if they were a pirate: https://www.circusscientist.com/2023/11/13/i-hired-a-pirate-to-take-orders-for-my-entertainment-business/
If you ask them to respond like a politician they answer all your questions with something completely different.
*Picks up wireless mouse* Hello, computer.
wessels
“It might just take a while to get the pitch right” https://www.youtube.com/watch?v=uyV0IVItlM4
Are you sure?!
Ah yes. Star Trek: The One With The Whales.
is this … transparent Aluminium?
They did it. The crazy son’s of bitches did it! Quite awhile ago, it’s commercially available.
There is also This transparent aluminum (linked in that same article) and it’s been used in phone/watch screens also.
Can you explain this reference for me? I do not understand.
Go watch Star Trek IV first
I’ve seen it, but it was a while ago.
Undiscovered Countries was my favourite TOS movie, because it covered the historically important Khitomer Accords.
Then go watch it second
I will, but only for Spock’s beanie.
It’s a reference to this scene from Star Trek IV: The Voyage Home (1986).
Explanation without video
Scotty, having traveled back in time to the year 1986 as part of a mission to rescue some whales, attempts to use a computer by speaking to it and then mistakenly tries to use the mouse as a microphone when the machine does not respond. He is prompted to use the keyboard instead of verbal commands and gives information on how to manufacture transparent aluminum. This material was not invented until about 150 years later according to the pre-trip history of the Star Trek future but Scotty has given it a head start.
Oh, thank you for the lengthy explainer.
All I have in return is this fairly interesting video detailing one of the ways we’ve already found transparent metals. Perhaps over the next 150 years we’ll be able to stabilise the material structure.
Thanks again for explaining.
Helping people with their work through teams has taught me that voice control is a disaster to get anything done for anything other than just dictating text.
It is only logical that an algorithm trained on the ways of a Vulcan, is precise and accurate in it operation and communication. Vastly more fascinating are the result when you ask it to behave like a human.
“Answer as if you’re a tribble.“
But then all it can do is multiply.
That’s troubling.
Did they try to ask the models to act as Euler, Gauss or Tao?
Tao said he would ask the AI model to pretend to be a colleague before asking it anything. So your suggestion works!
After a long dig: https://mathstodon.xyz/@tao/110601051375142142
This is the best summary I could come up with:
A study attempting to fine-tune prompts fed into a chatbot model found that, in one instance, asking it to speak as if it were on Star Trek dramatically improved its ability to solve grade-school-level math problems.
“It’s both surprising and irritating that trivial modifications to the prompt can exhibit such dramatic swings in performance,” the study authors Rick Battle and Teja Gollapudi at software firm VMware in California said in their paper.
“Among the myriad factors influencing the performance of language models, the concept of ‘positive thinking’ has emerged as a fascinating and surprisingly influential dimension,” Battle and Gollapudi said in their paper.
Their study found that in almost every instance, automatic optimization always surpassed hand-written attempts to nudge the AI with positive thinking, suggesting machine learning models are still better at writing prompts for themselves than humans are.
The prompt then asked the AI to include these words in its answer: “Captain’s Log, Stardate [insert date here]: We have successfully plotted a course through the turbulence and are now approaching the source of the anomaly.”
Axel Springer, Business Insider’s parent company, has a global deal to allow OpenAI to train its models on its media brands’ reporting.
The original article contains 820 words, the summary contains 198 words. Saved 76%. I’m a bot and I’m open source!
