Friday Squid Blogging: Bobtail Squid – Source: www.schneier.com

Clive Robinson • August 23, 2025 8:30 PM

@ Ismar, ALL,

With regards the ACM paper,

https://www.schneier.com/blog/archives/2024/06/new-blog-moderation-policy.html

You’ve linked to, it should carry a “health warning”.

Because it’s not just an up-hill struggle to read. Due in part to the fact it’s written in what reads like “Domain experts” trying to be more “non expert friendly” and comming off like parents trying to answer their childs “Why is the sky blue?” Question.

Thus you get things like,

“such as top-p or top-k sampling with nonzero temperature.”

And no further explanation. Thus a non domain reader knows not if it is a throw away statment or something of significant relevance. With further no way of knowing what relevance the writers ascribe to it…

Thus for non domain experts the paper suffers one of four fates,

1, It imparts confusion not clarity.
2, It gets put down / ignored.
3, The reader consults a traditional search engine.
4, The reader asks an AI to expand / explain.

Some will realise that the best option for those with expertise in other domains is at the moment still ‘3 – consult a traditional search engine’.

But AI slop is polluting the internet, and not only does that have a bad feedback effect on “Hover it all up AI”, but less well publicised it effects traditional search engines as well. Which additionally are also suffering from Corry Doctrow’s “enshitification” problem as well as a double whammy. Thus combined making them increasingly worse than useless or a significant grueling challenge to use, that at times feels like wading through a turbulent stream of diarrhoea…

Thus of the four options available to non domain experts the two (3/4) that might have helped them forward are being polluted with noise and bias and becoming in effect useless.

Leaving only the first two options (1/2) where a non domain experts gets at best no value in the paper and have at the very least wasted their time, if not worse.

I do not know if this is because the authors are just poor writers, or they were trying to put to much into too little assigned space… But what is clear is,

“They have not achieved meaningful communications.”

Thus few can judge the merit of the message they were attempting to convey.

But even now a traditional search has been polluted… To see why I say

“that at times feels like wading through a turbulent stream of diarrhoea…”

Lets “walk it through”,

If you “duck-duck” with AI turned off you get nine poor examples before you get to,

https://codefinity.com/blog/Understanding-Temperature%2C-Top-k%2C-and-Top-p-Sampling-in-Generative-Models

How many people would go that far?

And unfortunately whilst it sort of explains Top-K and Top-P to a level most can start productively thinking about in terms of probability they were taught in school… Temperature remains a mystery behind the “door-stop” of the technical term “logits” that do not get explained in a meaningful way…

And another search gives a page,

https://peterchng.com/blog/2023/05/02/token-selection-strategies-top-k-top-p-and-temperature/

Does a less well explanation of Top-K and Top-P but… explains Temperature via Softmax in a way few would understand, with only the pictured graphs giving a clue.

But of “logits” there is no sign…

To save the pain of looking through many pages with eye aching formula, in a “Digital Neural Network”(DNN) neuron there are a large number of inputs from tokens or previous layer neurons. These inputs are multiplied by the “weights” these are summed. and thus an output range between “plus and minus infinity” is potentially possible. This needs to be brought into a usable range –ie normalised– before it can be used in the next layer.

This normalisation is sometimes done via a “Sigmoid function” which has certain desirable characteristics for this. Because it is a mathematical function with a graph that has a characteristic S-shaped or “sigmoid” curve, that reversibly maps any real input via the curve to an output in the ranges from minus one to one. This curve can be done via the exponential or “log”(ln) function so in this case is called a “logistic function” And this is where the statistical term “logit” comes from as it’s “the inverse transform of the logistic function”. Or more readily it’s “the logarithm of the odds p/(1-p)” so,

logit(p) = ln(p/(1-p))

However…In the world of AI and artificial neural networks Sigmoid and Logistic get used as synonyms. Worse they use “logits” as a shorthand for the raw “not yet normalised” input to the final layer normalisation function (SoftMax).

Also remember as far as we can tell natural neurons have different equivalent output curves that are not reversible so not Sigmoids, and in effect are the integration of multiple variable frequency pulse train inputs so not really linear sums of the inputs either.

The question of speed and required accuracy of the Sigmoid curve normalisation arises and a crude three straight line linear approximation has been used successfully. With a closer approximation done faster by a look-up table. Thus giving a “memory time trade off”.

But what of “Temperature” this is a function carried out on the logits vector prior to the Softmax Sigmoid function normalisation and can be thought of as a multiplicative or scaling rather than additive bias. However to the eye on the output of the Softmax functions it looks somewhat like an additive shift or bias. The result is it effects the degree of “randomness” that some see incorrectly see as “creativity”.

Which brings me back to my point about the ACM paper and the four options. As can be seen from my partial demonstration above, option 3 of consulting a traditional search engine can be grueling work for those without sufficient domain knowledge, but also it has low relevance to the argument being made…

Which is a shame because the message in the paper is getting lost in the message for all but a few, who would in effect be “part of choir/congregation”.