the tower of AI babel

Borges and AI L´eon Bottou † and Bernhard Sch¨olkopf
Oct 4, 2023

Abstract:

Many believe that Large Language Models (LLMs) open the era of Artificial Intelligence (AI). Some see opportunities while others see dangers. Yet both proponents and opponents grasp AI through the imagery popularised by science fiction. Will the machine become sentient and rebel against its creators? Will we experience a paperclip apocalypse? Before answering such questions, we should first ask whether this mental imagery provides a good description of the phenomenon at hand. Understanding weather patterns through the moods of the gods only goes so far. The present paper instead advocates understanding LLMs and their connection to AI through the imagery of Jorge Luis Borges, a master of 20th century literature, forerunner of magical realism, and precursor to postmodern literature. This exercise leads to a new perspective that illuminates the relation between language modelling and artificial intelligence

My summary:

LLM is a story telling fiction machine with innumerable forks that can write any story and, be warned, can be manipulated by others. Neither truth nor intention matters to the operation of the machine, only narrative necessity. Narrative necessity is statistically determined by what comes before.

The machine merely follows the narrative demands of the evolving story. As the dialogue between the human and the machine progresses, these demands are coloured by the convictions and the aspirations of the human, the only visible dialog participant who possesses agency. However, many other invisible participants make it their business to influence what the machine says.

Delusion often involves a network of fallacies that support one another

Forking paths: The linguist Zellig Harris has argued that all sentences in the English language could be generated from a small number of basic forms by applying a series of clearly defined transformations. Training a large language model can thus be understood as analysing a large corpus of real texts to discover both transformations and basic forms, then encode them into an artificial neural network that judges which words are more likely to come next after any sequence.

The Purifiers want to eliminate the heinous, tidy up the machine to serve the human race and make money from it. They want to reshape the garden of forking paths against its nature, severing the branches that lead to stories they deem undesirable. Although there are countless ways to foil these attempts to reshape the fiction machine, efforts have been made, such as “fine-tuning” the machine using additional dialogues crafted or approved by humans, and reinforcing responses annotated as more desirable by humans (“reinforcement learning with human feedback”.)

Confabulation: As new words are printed on the tape, the story takes new turns, borrowing facts from the training data (not always true) and filling the gaps with plausible inventions (not always false). What the language model specialists sometimes call hallucinations are just confabulations. Confabulation is inventing plausible stories with no basis in fact.

If an amnesiac patient is asked questions about an event they were previously at, instead of admitting they do not know, they would invent a plausible story. Similarly, in split-brain patients, where the corpus callosum is severed so each half of the brain cannot talk to each other, patients can invent elaborate explanations for why the other half of their body is doing a specific thing, even when the experimenter knows this is not the case because they have prompted it with something differently.

Story telling: The invention of a machine that can not only write stories but also all their variations is thus a significant milestone in human history. It has been likened to the invention of the printing press. A more apt comparison might be what emerged to shape mankind long before printing or writing, before even the cave paintings: the art of storytelling. Fiction can enrich our lives, so what is the problem?

Reference:
The Library of Babel by Jorge Luis Borges (1941)
LLMs confabulate not hallucinate
Zellig Harris. Mathematical Structures of Language. John Wiley & Sons, 1968

children are not hackers

Children are not hackers by Paulo Blikstein & Marcelo Worsley (2016) - pdf available

This is a cautionary argument against people like me who have been known to say “we are all makers” without thinking deeply enough about the issues. See the footnote for more about this.

The authors begin with the counter intuitive claim that the cultural roots of the modern maker movement are a threat to its flourishing or even survival in schools! The one eyed warriors may sew the seeds of destruction of what they love. How ironic but not unusual. After this introduction I had to read on to understand.

They explain this claim. The people that created the first Fab Lab at MIT (Gershenfeld eta la in 2001) were hackers. By the way I don't use the term "hackers" as a pejorative. Hackers are those curious people who want to look inside and understand how things work. A better term for those who steal your data is Crackers.

This led to the creation of Maker Faires for those and other hackers to show off their cool products and the MAKE magazine (which originated in 2005). Those people were sophisticated publishers. Furthermore, Maker clubs have flourished more outside of schools in informal settings (eg. museum, after school programs and competitions) rather than inside schools.

Then there has been a push for STEM education due to a perceived shortage of qualified engineers and scientists. Those considerations involve the industrial workplace, not the school workplace.

All of these influences (hackers – publishers – informal educators – industrial workplace) have a middle to upper class origin and elite appeal. Such a culture will not win the battle to introduce modern maker education into schools in a mass way.

Hacker culture is self-sufficiency, autodidacticism, individualism and competition

“The popular image of the hacker is that of a disheveled, unshaven White male in his twenties, doing all-nighters in a messy electronics lab, capable of learning anything by him­self by scouring the Web or doing late-night runs to the library ...”

This is an extreme minority. To promote this won’t help most students.

Publisher culture: The initial culture of the Maker Movement which began in earnest around 2005 was college-educated, affluent, White men. They published MAKE magazine and organised Maker Faires where makers showed off their innovative finished products. This was a culture of Product before Process where unfinished “half baked” efforts are not rewarded.

Also the types of projects developed by this elite group downgrades and devalues projects such as traditional crafts, costumes, pottery, technology-augmented wearables and jewelry, among many others

Informal spaces Within these spaces the Keychain Syndrome prevails - the “30-minute” workshop model: fast, scripted, perpetually “introductory” workshops. The problem here is that the demonstrations often never get past trivial objects.

Job Market culture: It is argued that we need more STEM students because of shortages of engineers and scientists and other countries, such as a threatening China, are way ahead of us here.

The pathway pioneered by Papert and others is different: that software such as Logo (the precursor to Scratch) and hardware such as LEGO are tools for self expression and ways for transforming traditional subjects such as maths into something that is more interesting and natural to learn.

Summary from this section about the hackers – publishers – informal educators – industrial workplace influences: Promoters of modern maker education such as myself should not take the efficacy of “making” for granted.

TOWARDS A MAKER ED CULTURE THAT WILL WORK

Hard fun (Papert): The lesson for educators is that the work in FabLabs and makerspaces can be enjoyable but should never be “easy” fun, devoid of frustration and difficulty

Abstract and Concrete Thinking: The maker space approach does not reject the abstract but attempts to make the abstract more concrete. The examples provided by the authors are:

Supposedly abstract mathematical ideas suddenly become concrete when, for example, a student needs to design a laser-cut object using the least amount of material, or when a very “con­crete” 3D printed object gives rise to a discussion about Boolean operations

I can think of other examples arising from design work using Turtle Art. I asked students to make a right angled triangle. Some did this by trial and error which was fine. I also showed them how to get the lengths exactly right by using Pythagoras theorem. In another shape the outside octagon by trial and error the size was 121 (good enough), when calculated using trig it was 120.7 (exact). I would say the trial and error approach is acceptable but it's good to show the more precise way to obtain the values and some of the students (not all) will pick it up.

Gut feeling or Research? Doing or Theorising? Some maker ed advocates rely on gut feeling (doing is learning – list of attributes) but of course many educators want to see the real research done before they accept this. Deep learning doesn’t happen by magic. Maker ed advocates have to make strategic choices here. IMO a combination of both constructionist and instructionist methods are required. The authors are building a case here for a coherent theory of maker ed, not just hands on and she'll be right mate.

For example, in one class where we were making and coding with the microbit, I took the opportunity to try explain where the 255 came from in some MakeCode parameters. I talked about bits and bytes and 1s and 0s. It wasn’t very successful. It was too abstract for most of the students. It made me realise I need to prepare this sort of break from the making and coding more carefully.

RECOMMENDED LEARNING CULTURE FOR MAKER SPACE

How does a learning culture differ from a hacker culture?

From the perspective of where actual students are at these are bad slogans: “every child should be a maker”, “making mistakes is good”, “every child should hack”

Reality check: many students need support! If they don’t get it they will feel lost or frustrated. They drift into doing the less demanding parts of a task (colouring in).

Without help (sink or swim approach) those who feel uncomfortable in a maker space will become further disempowered

One possibility: Pair more competent with less competent and make the less competent the driver (in control of computer, mouse and keyboard). I thought this was a great idea and have been angling for an opportunity to try it out:

“ In half of the mixed pairs, the low-achieving student was mandated to be the “driver” of the activity (having control over the computer mouse and key­board, etc.). In those groups, the learning outcomes were almost the same as the groups with two high-achieving students, and dramatically higher than mixed groups in which the high-achieving student was the “driver” instead (Schneider & Blikstein, in press).”

The authors refer to other researchers about the stereotype threat (Cohen, Garcia, Apfel, & Master, 2006) which shows that individuals can perform below their ability level when they suspect that they belong to a group that historically does not do well at a particular activity

Key points from this section

• include tasks that are meaningful to all students
• avoid too much “learn from failure” rhetoric
• find ways to get students out of their comfort zone (eg. instruct lower ability in a pair to be the driver)
• be aware that some groups expect to fail

From jobs culture to literacy culture:

There is often lots of talk about STEM (and also STEAM) in education systems these days. Some of this originates from social shortages of engineers and scientists. This can be a source of an educational problem rather than a solution to a social issue.

There is a deep cultural abyss separating the corporate world and K–12 schools. Educational materials should be designed for children

• microbit not arduino
• Scratch not Java

The authors argue that the point of STEM literacy is to provide a lens through which to interpret the world and act upon it – “consciousness of the possible” Friere 1970. I have often advanced a similar argument, that Scratch is a multimedia fun machine for making stories and games.

From keychain culture to deep projects culture:

The “keychain syndrome” is ok for an introduction to 3D printing but we are not achieving much if we don’t go beyond that. One version of this is downloading a great design from thingiverse and printing it. I do that and again it has its place. But in the bigger scheme of things it is too easy, no design or remix skills happening here. How do we go beyond that to a culture of deep projects?

To develop a school culture of engaging cross curricular projects does require administrative support. Teachers are time strapped and so it is not realistic to expect them to develop such materials on top of their normal workload.

Unfortunately curriculum guidelines such as ACARA, which separate the what from the how, do not help here. A good curriculum ought to have more flexibility about WHAT we teach (eg. design a project around an idea that interests the student). Then the teacher helps the student HOW to do that. In other words the HOW should be guiding the WHAT. But what does ACARA do? Tells the teachers the WHAT like Moses' stone tablets and leaves it to the teacher to figure out the HOW. (Thanks here to Mitch Resnick)

We think outside of the “STEM box”: We have seen students creating fascinating musical instruments, clothes, costumes, and visual arts projects, working with and augmenting traditional crafts, and creating interactive art. We have also seen teachers from non-STEM areas create very compelling units, combining history and math, biology and engineering, language arts and physics. Allowing teachers to “pair up” and design curricula together, even if they are from different areas, greatly expands the range of activities that can be done in the labs and makes it possible to attract students with a variety of different interests.

Project ideas and themes should be connected to students’ lives, interests, passions, and their communities. Lives, interests, passions, communities covers a lot of ground

From product culture to process culture:

Priming students helps their performance. eg. if students have been previously taught that triangles make stronger structures (and are reminded) then rather than using readily available objects to build bridges (eg, a chair) they are more likely to build with triangles.

A product culture sees a great finished product suitable for a Maker Faire as the end goal. A process culture looks at things like collaboration, management (eg. planning ahead) and preparedness to go outside of their comfort zone. It’s a different form of assessment.

THE MAKING OF THE FUTURE

Maker education has made significant inroads into many schools and even official curricular. For this progress to continue so that maker ed flourishes advocates such as myself need to understand the issue of what cultures are more attractive to most members of a school community.

Footnote: An extract from a previous article where I went a little overboard about humans as makers:

We, humans, are homo faber (Latin for Man the Maker), the concept that human beings are able to control their fate and their environment as a result of the use of tools.

- Thoughts on reading Paulo Blikstein (the founder of the Fab Learn Schools Movement)

I think now even in pre modern societies there was a division of labour (eg. hunters and gatherers) and that in our present youth culture, with the influence of social media, people are more likely to become consumers than makers. However, the modern maker movement does provide a promising way for many to break out of this.

3D hand print: from atoms to bits and back to atoms

I adapted the idea from a tutorial at the Prusa Printables education site (free for schools, universities and other educational institutions but you have to apply to register) to develop an interesting 3D print of my hand

The tutorial was Autumn Leaves by Vesela Skola

I wasn’t happy with the recommended image to SVG converter so I searched around and one that worked well for me: Free image to SVG converter

The Autumn Leaves tutorial taught me how to prepare the images and then in Prusa Slicer to superimpose a hand image showing fingernails and knuckles onto the full hand backdrop. It also showed me how to alter the settings to print the hand with infill only (30% honeycomb) thus producing an attractive mesh effect.