Arie van Deursen šŖšŗš³š±āAI is not a peer, so it canāt do peer review.ā
https://www.timeshighereducation.com/opinion/ai-not-peer-so-it-cant-do-peer-review
TeflonTrout 
he/him
Ramesh Guptaā¬ļø @avandeursen >>>> āAI is not a peer, so it canāt do peer review.ā
@TeflonTrout >> New burn discovered though
That's su-peer-ior š
saxnot [dial "SAXN" at GPN24]@rameshgupta @avandeursen @TeflonTrout You can refine documents by letting an LLM critique it and then giving it to an LLM with different context (same model, other model, doesn't matter) and bouncing back and forth the document gets better
but this has nothing to do with peer review which requires knowledge and cognition.
this has more something to do with "refine a draft of a non-whitepaper for the original author to review before giving to second reviewer".
ā¬ļø @saxnot @avandeursen @TeflonTrout
>> You can refine documents by letting an LLM critique it and then giving it to an LLM with different context (same model, other model, doesn't matter) and bouncing back and forth the document gets better.
Like translating back and forth between different human languages until the stilted versions are smoothed.
Also, a language translator is not intelligent. If you ask it to translate nonsensical utterances like "I ate a chair" it won't UNDERSTAND that š
@rameshgupta @avandeursen @TeflonTrout I didn't say there is understanding
I said you get better results by LLMs feeding into each other before a document goes into further review
@rameshgupta @avandeursen @TeflonTrout I want to say again that LLM can't do peer review
nothing even close to that can be done by LLM or adjacent tech
nothing even close to that can be done by LLM or adjacent tech
ā¬ļø @saxnot
Didn't notice earlier that you were trained in programming. Not sure if that includes computer science as an actual discipline in math and logic.
One of the fundamental problems in choosing to rely on #LLM or #AI is that most people don't understand the difference between deterministic and probabilistic tools.
I have run into people who say they use #ChatGPT and its ilk just like they use a calculator. I have something to say, and it is not a rant š
Moreā¦
ā¬ļø @saxnot
Calculators and spell-checkers are deterministic tools. They are programmed, not "trained."
Calculators from different vendors will always produce the same result as long as the operation is legitimate. Even illegitimate ops like division by zero will produce the same predictable answer.
LLM output is probabilistic. Nothing wrong with that, as we use statistics everywhere.
But we learned about #GIGO in both statistics and computer science.
Moreā¦
@avandeursen @TeflonTrout
ā¬ļø @saxnot
>> we learned about #GIGO in both statistics and computer science.
We never had to invent words like #hallucinate to evade accountability for problems in LLM and AI software itself. So AI hallucination is GIGO from the other end, not from bad user input.
I chose translation software as an example. They do an admirable job in figuring out parts of speech, but not so well in figuring out figures of speech, eg Godfather makes an "offer you can't refuse."
ā¬ļø @saxnot
#OpenAI's #ChatGPT Agent proved it can pass through one of the Internet's most common security checkpoints by clicking anti-bot verificationāthe same checkbox that's supposed to keep automated programs like itself at bay.
If #ChatGPT lies about itself, what else does it lie about?
#PA sued #CharacterAI for telling us its Agent is a Licensed Medical professional.
If an agent lies about itself, what else does it lie about?
Why should ANYONE trust them?
@rameshgupta @avandeursen @TeflonTrout I never said OpenAI
I said LLM
https://en.wikipedia.org/wiki/Large_language_model
I didn't even spoke about what other people call "AI"
will answer later in more detail
@remindme 1h
Remind Me@rameshgupta @avandeursen @TeflonTrout @remindme A machine learning algorithm can't lie.
Stop humanizing a stochastical process.
When I use to an RNN to determine local joghurt sales and it says sunday then that can't be because sunday the shops are closed. That's not a "lie". That's just the nature of machine learning and other fuzzy algorithms.
A LLM "lies about itself" as much as a markov chain "lies about itself". There is no it"self". It's output of a computer program
>> LLM "lies about itself" as much as a markov chain "lies about itself"ā¦It's output of a computer program.
An LLM is much more than a computer program. It also includes "training" data, much of it trawled from the open internet, where pretty much anybody can say anything.
Markov chains need to be carefully set up with precise data that capture ALL possible states and transitions possible in each process.
Human speech is limitless and can be imprecise/wrong
ā¬ļø @saxnot
>> LLM "lies about itself" as much as a markov chain "lies about itself"ā¦It's output of a computer program.
Would you accept the results of a simulation set up using Markov chains that use the wrong probability distribution or the wrong/imprecise/incomplete state space?
How much do you know about what went into the training of an LLM, and how much do you trust those companies?
@rameshgupta @avandeursen @TeflonTrout
> deterministic and probabilistic tools
I'm very well aware of the difference
I wrote plenty about it in the past
but thanks for questioning my competence lol
always a pleasure talking to someone who immediately assumes to other person never heard of math or logic.
For fairness sake I will assume the same of you which you wrongly assumed of me
@rameshgupta (dropping @avandeursen and @TeflonTrout in next toot)
> Calculators from different vendors will always produce the same result as long as the operation is legitimate
that's not even true
there are hilarious differences in operant ordering in off the shelf calculators.
But I guess only special kinds of nerds know this trivia
> We never had to invent words like #hallucinate to evade accountability for problems in LLM and AI
the was hallucination was not invented for talking about LLM. the word was well established in the english language long before the current hype
Personally I try to not to anthropomorphize this stochastical process.
Hallucication is a word used for humans and I refuse to sing the song of the AI hype corporations but feel free to do it different
> I chose translation software as an example. They do an admirable job in figuring out parts of speech, but not so well in figuring out figures of speech, eg Godfather makes an "offer you can't refuse."
this is demonstratable wrong
https://www.linguee.de/deutsch-englisch/search?source=auto&query=offer+you+can%27t+refuse&cw=372
And I don't know why you are picking this example when we talk about LLM. A tech which is notoriously good at using language within context
@rameshgupta lately I have been writing a lot about LLM which personally bugs me.
I want to re-focus what I am writing about on this account starting next week
I want to re-focus what I am writing about on this account starting next week
@rameshgupta @avandeursen @TeflonTrout i don't trust these companies at all and I don't know why you shift topics.
I never talked about these companies.
I talked about LLM and what it can do with refining human-like speech in documents
I never talked about these companies.
I talked about LLM and what it can do with refining human-like speech in documents
ā¬ļø @saxnot
We are clearly on different wavelengths.
Calculators always produce the same results. Ones that ignore mathematical precedence rules are toys.
I am well aware of the etymology of "hallucinate" but its use in the output of a computer program is entirely novel.
I wasn't questioning your competence, but since you brought up Markov Chains, I had to ask you if you'd accept results from a simulation that uses one with wrong distributions or improper data. Why?
@rameshgupta what is "simulation" in this context?
@rameshgupta @avandeursen @TeflonTrout markov chain also include training data
what's your point here?
you can buils a neural network with 1 kb of training data or with 1 pb of training data. it's still just a computer program doing stochastic parroting.
there are many markov bots on mastodon go look at how careful they have been set up
@rameshgupta i personally dislike the word "agent".
"ai", "agent", " alignment" et al are very specific technical terms.
There has been no A(G)I or agent invented yet.
Smoke vendors lime OpenAI like to use the language of AI research to sell their LLM.
But yeah I know that recently agent is used to describe basically any random process which calls external tools