Page 133 - The Black Box Book

The current way of developing programming tools tends to prioritize Final thoughts about the AI curator: In this project, Machine Learning
4
the developer’s time by means of making it as easy as possible to use. (ML) was used to sort visual content based on its similarities. From my
8
The demand for a fast development process leads towards detaching point of view – like any type of ML of this kind, including current popstars
the praxes of programming from its core (hardware) functionality. One like DALL-E / Imagen or GPT, this process was developed, handpicked,
thing is trans-compiled to another several times, layer after layer before adjusted, and supervised by a person and is, therefore, deterministic and
the final transcription gets compiled and the final production code is built. traceable to the last byte. I think we should approach it as such. I often
This is due to the tendency to make programming faster, universally op- encounter starry-eyed approaches to ML results. Treating the process
timized, secure and maintained by approaching natural language syntax itself as a “black box” where we cannot actually see what is happening
– bringing the first layer of some programming tools almost to the realm should not make it more divine than pressing the enter key on a keyboard,
of written speech. This method may result in lengthier computation times, being oblivious to the process behind the wonders it just set in motion.
but more crucially, it places modern programming tools within the black Thus, getting used to the enter button performing so many things seemed
boxes of superior compilers, even for developers. This is inevitable as amazing back in the 90s; this is also bound to be the case concerning
technology becomes increasingly complex and layered; specialization is ML generating new sets of novel or creative combinations in the 2020s.
normal as programming becomes more widespread. In addition, this does
not imply that the final code is ineffective; on the contrary, it is standard-
ized and efficient. It is just a process that is indirect and straightforward.
In the end, there is a machine-level effective code. 8 I prefer to refer to current AI models as Machine Learning because,
in my experience, AI does not exist yet, and it is more of a sci-fi concept,
Using the Tensorflow code library to compute the ML results in this a catchphrase if you will; current mechanisms are better described as
5
project can serve as an example. To use it, a developer writes a couple ML. I believe this saves some thought because it does not sound so out
of hand, as we are talking about node-pulling vectors.
of dozens of lines of code where the content is selected and prepared to
be processed and where the core library is set up and called to action.
While the real code computing the result is millions of lines long, written
and maintained by thousands of contributors, the programmer just uses
it. In the end, there could be a GUI (Graphic User Interface, the common
way users interact with applications – buttons, images, textareas, etc.)
to set up TensorFlow, so that we would not need Python code to run it.
6
The programmer would select the images, tweak a few sliders, and wait
for a result. It is also possible to set it up using voice commands, con-
7
tributing to about a quarter of queries in HCI . With OpenAI’s DALL-E2
or ChatGPT it becomes apparent that the process of creation in terms
of HCI is strictly dependent on the ability of the user to describe their
thoughts with words or sentences, as that is their fundamental method
– using keywords. Following the same trend – to quantify the means of
creation that mirror the user’s native language.

4 Programming tools could be, for example, APIs, frameworks or pro-
gramming languages.

5 Tensorflow is a framework that allows developers to build and deploy
ML models.

6 Python – programming language

7 HCI – Human-Computer Interaction, could be writing text using
a keyboard, clicking on a button, or speaking to a microphone (e.g. “Hey
Siri” commands).

132 133