> On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.
Seriously though, at some point AI does actually need an input from the user, if we are imagining ourselves still as the users.
If instead we just let AI synthesise everything, including our demands and desires, then why bother being involved at all. Take the humans out of the loop. Their experience of the outputs of the AI is actually a huge bottleneck, if we skip showing the humans we can exist so much faster.
And have been conditioned to accept LLM responses as reality.
Enter 2+2. Receive 4. That's the right answer. If you enter 1+1, will you still receive 4? It's easy to make a machine that always says 4.
Again, in general sense. Software engineers are too used to computers being fragile wrt. inputs. Miss a semicolon, program won't compile (or worse, if it's JavaScript). But this level of strictness wrt. inputs is a choice in program design.
Error correction was just one example anyway. Programmers may be afraid of garbage in their code, but for everyone else, a lot of software is meant to sift through garbage, identifying and amplifying desired signals in noisy inputs. In other words, they're producing right figures in the output out of wrong ones in the input.
And also:
> they need to supply ones specifically engineered to be uncorrelated with the right figures.
I assume most people will understand this way (including me) when it's said to "input wrong figures".
This does not refute the concept GIGO nor does it have anything to do with it. You appear to have missed the point of Babbage's statement. I encourage you to meditate upon it more thoroughly. It has nothing to do with the statistical correlation of inputs to outputs, and nothing to do with information theory. If Babbage were around today, he would still tell you the same thing, because nothing has changed regarding his statement, because nothing can change, because it is a fundamental observation about the limitations of logic.
I do usually know what the point of any comment quoting that Babbage's statement is, and in such cases, including this one, I almost always find it wanting.
--
Exactly, it's 5. You just have to correct the error in the input.
You can't deal with humans without constantly trying to guess what they mean and use it to error-correct what they say.
(This is a big part of what makes LLMs work so well on wide variety of tasks, where previous NLP attempts failed.)
Case in point, I recently had ChatGPT point out, mid-conversation, that I'm incorrectly using "disposable income" to mean "discretionary income", and correctly state this must be the source of my confusion. It did not guess that from my initial prompt; it took my "wrong figures" at face value and produced answers that I countered with some reasoning of my own; only then, it explicitly stated that I'm using the wrong term because what I'm saying is correct/reasonable if I used "discretionary" in place of "disposable", and proceeded to address both versions.
IDK, but one mistake I see people keep making even today, is telling the models to be succinct, concise, or otherwise minimize the length of their answer. For LLMs, that directly cuts into their "compute budget", making them dumber. Incidentally, that could be also why one would see the model make more assumptions silently - these are one of the first things to go when one's trying to write concisely. "Reasoning" models are more resistant to this, fortunately, as the space between the "<thinking> tags" is literally the "fuck off user, this is my scratchpad, I shall be as verbose as I like" space, so one can get their succinct answers without compromising the model too badly.