So I threw some of these lines at Bard and true to its name, it made stuff that is better than my lines. Still not sure if any of it counts as poetry but there are definitely a line or two or a metaphor or two here that are interesting.
What I wrote:
Read Nietzsche the first time to appear well read and find confusion The second time to appear serious and find epiphanies The third time for yourself and reject the overman
First, a borrowed cloak, Nietzsche struts, impressing crowds, Empty, echoing.
Then, a seeker’s beard, Epiphanies sought in smoke, Ashes in his hand.
Finally, unmasked, He laughs, tears off the false skin, Free, and only man.2
Tried Nietzsche to swagger, impress the crew, Brain scrambled, left me scratching my shoe. Second go, all earnest, brow furrowed deep, “Aha!” moments came, then went to sleep. Third and final, just me and the page, “Übermensch” baloney, turning the cage. My own way to go, the hype I dismiss, Nietzsche’s a puzzle, not some hip abyss.3
First dive into Nietzsche for some street cred, Got lost in the jumble, scratching my head.
Round two, looking serious, seeking insight, Epiphanies popping up, left and right.
Third time’s a charm, just for myself, Overman? Nah, put that book on the shelf. Finding my path, rejecting the hype, Nietzsche, my guide, in this philosophical tripe.4
Thought Nietzsche cool, like shades and a scowl, Read, got lost, felt like a clueless owl. Tried again, serious, highlighter in hand, “Deep thots,” I’d shout, then couldn’t stand.
Third time, alone, book open and worn, Overman stuff? Nah, more like a yawn. My own truth’s out there, beyond borrowed gleam, Nietzsche’s a teacher, not some grand scheme.5
(Verse 1)6 Caught a freight train down to wisdom’s depot, Nietzsche in my pack Heard the hipsters talkin’, thought I’d join the existential act Cracked the spine, let words like shadows crawl across my mind But meaning danced like dust devils, never could I pin it down
(Chorus) First time through, a borrowed mask, words I didn’t own Spouted wisdom from the rooftops, just another crow cawin’ alone Second time, all furrowed brow, highlighter in my hand Epiphanies like lightning bugs, sparklin’ on the promised land
(Verse 2) But lightning fades, and truth ain’t found in borrowed sparks So I stripped it back, raw wood and nails, no fancy, gilded arks Read for me, not crowds, not critics with their sharpened quills Just a man unmasking shadows, climbin’ his own personal hills
(Chorus) Third time round, the book worn thin, spine crackin’ with each turn No more Übermensch charades, let the embers of my spirit burn I ain’t no superman, don’t crave Olympus’ golden throne Just lookin’ for a path of my own, where the wind whispers in my bones
(Bridge) So thanks, Nietzsche, for the storm, the tangled mountain trail You showed me wisdom ain’t a summit, but a fire that never fails I may have borrowed your compass, but the stars I follow now are mine Each scar a map, each breath a song, this journey’s truly divine
(Chorus) First time blind, second time blurred, but now the focus clear No borrowed wings to lift me up, my feet planted right here So long, Zarathustra, your shadows fade in the mornin’ light This path’s mine, and I’m walkin’ on, the stars as my sight
The following is a conversation with a LLM chatbot Llama, which is just funny. using the model llava-v1.5-7b-q4
User: what is a category error in logic and give some examples as well as examples of what is not a category error although it looks like it
Llama: In logic, a "category error" refers to an argument that appears valid at first glance but actually violates the rules or structure of a particular logical category. It's essentially when someone tries to apply one type of reasoning to another situation where it doesn't belong, resulting in a fallacy.
An example of a category error is trying to use deductive reasoning (which relies on general principles and specific premises) to prove an inductive conclusion (where the truth of the conclusion follows from the truth of its individual instances). This would be considered a category error because deduction and induction are different logical methods, each with their own rules for drawing conclusions.
Another example is using the fallacy of false analogy or comparing apples to oranges (using an inappropriate comparison) instead of comparing them within the same category. For instance, you might say that a politician's actions are like those of a thief because they both break laws; however, this would be a category error since politicians and thieves belong to different categories with distinct characteristics and rules for their respective behaviors.
On the other hand, although it may seem similar at first glance, an example that is not a category error is comparing apples (fruit) to oranges (fruit). Although both are fruits, they have different properties like taste, texture, and nutritional content. It's important to recognize when we're using the correct logical method for drawing conclusions based on available evidence or reasoning about a particular situation.
User: is the following sentence semantically accurate? "Odorless blue thoughts run lopingly"
Llama: The sentence "Odorless blue thoughts run lopingly" is not semantically accurate, as it contains several grammatical and semantic errors. Firstly, there's a missing article ("the") before the word "blue," which should be "The." Secondly, the verb "run" doesn't agree with its subject "thoughts" in number; it should either be singular or plural to maintain consistency (either "Run lopingly" or "Runn lopingly"). Lastly, there is no clear connection between odorlessness and running.
To make the sentence semantically accurate, you could rephrase it as: "The absence of scent in blue thoughts allows them to run freely." This version maintains grammatical correctness while also conveying a meaningful message about how thoughts can move or flow without any obstruction.
User: what could the phrase "Odorless blue thoughts run lopingly" mean?
Llama: The phrase "Odorless blue thoughts run lopingly" appears to be an intentionally nonsensical and humorous construct, likely meant for entertainment purposes rather than conveying a meaningful message. It combines elements from different domains (odorlessness of the color blue, running as an action) in a way that doesn't make sense semantically or logically. The intended effect is to evoke amusement and curiosity by presenting an absurd phrase rather than adhering to any strict rules of language or logic.
User: What does the phrase "The absence of scent in blue thoughts allows them to run freely." mean
Llama: The phrase "The absence of scent in blue thoughts allows them to run freely" is a revised and semantically accurate version of the original nonsensical sentence you provided. It suggests that without any distractions or obstructions (represented by the lack of odor), one's thoughts can move freely, unencumbered by external factors like scent or other sensory stimuli. This metaphorically connects the idea of running as a physical action with the freedom and clarity that comes from having undistracted thought processes.
Tech is heavily invested in giant language models.
This is because, with transfer learning, language models suddenly started being able to do a LOT more. And a lot better.
A lot of high citation publications have also been an outcome of this exploration of language models
For most parts in NLP engineering, LMs went from being cute to being the standard model from which you build solutions in the last three years. And it’s not just in NLP, the language model approach (use large amounts of data to train a network to do specific tasks, and some approximation of the underlying first principles will be approximately encoded into the model) seems to work well for images as well (stable diffusion).
And given that all of deep learning and the engineering approach to AI seem to be approximate/mimic till you make it, this is a powerful learning.
For academics, in a sense, iterative improvement, criticism, existential threats etc. are part of the publish-perish cycle. For the engineer, iterative improvement is, but it is driven by entirely different incentives.
When the ethics people started talking about bias, the engineering world was not very happy about it, well no one likes the police. More so because bias is not easy to tackle, its not easily quantified, its not something you can fix with a software update. And because there doesn’t seem to be any consensus yet about how to deal with it.
I think there are distinct parallels here from the early days of medical ethics and evidence based medicine. No one likes to hear that their life’s work is actually harming people despite having honorable intentions. We’d all like to be told that well, this is OK, you were doing the right thing, and we can fix it, easy.
But as the decades of literature in medical ethics and the history of clinical medical practice shows, ethics isn’t easy but is also not optional. The medical world was brought kicking and screaming into getting its shit together and behaving.
You’re searching for poems that talk about cutlery. Because you can. Also because a lot of times you don’t remember the precise detail or term used in a document that you want to find. And also because compiling a list of poems about our relationship with food-ingestion-tools would be fun.
Maybe it’s not poetry you care about but the name of the company your boyfriend works at which he told you about three months ago but cant find on your Whatsapp. And obviously, it would be super awkward to ask him now.
We have disparate search needs is what I am trying to say.
Lexical search: or how most search works
Definitions:
Query: The text we enter into search fields when we want stuff found
Document: The file, website, chat, image, meme stash or entity that has the terms or could have the the right answers to your search
Corpus: The collection of all the documents that need to be searched across various searches. (your whole website for example)
On with the show!
The standard way in which we do search is by matching the words or terms in a query with the terms in the corpus. So if you search for poems about cutlery the search will look for documents that have the words poems and/or cutlery and try to find ones that contain both the words.
Imagine, though, that the program has to scan through every document every time someone searches for something, hunting for instances of the word from scratch. This would be pretty time-consuming and also slightly stupid. So we need a way to store what words are in what documents.
Besides just matching or finding the word, we also need some way of knowing which document matches it better. This is called ranking. To solve both these problems people came up with the idea of indexes (or indices, if you’re a Latin snob).
What an index does is, it look over all the documents in a corpus, and creates a brief representation of the content of that document. So instead of searching over the whole document, now you just search over the representations. These representations also have some way of ranking the quality of the fit between the query and the document.
Lists all the words in each of the documents in a corpus and how frequently the occur in a document (term frequency).
Calculates how common or how rare of each of these words are across the whole corpus. The common words get a score of 0, or close to zero and the rare words get high scores. 2
Why do we do this? In most documents, the words that appear most frequently would be things like the, and, is and stuff that doesn’t add any discriminatory power to the retrieval.
When I search for poems with cutlery there are going to be a whole lotta poems with the word with in them, so I can’t use with to rank the documents in any way. But the word cutlery probably appears in a small number of poems, so it is a good word for ranking the results.
So now your index of poems doesn’t just have the whole text of the poem but a list of the non-zero scoring words along with the scores, which allows us to get the top matches.
The most used and probably the best of these TF-IDF algorithms is the Okapi BM25 family of algorithms. They are absolutely fantastic in finding you the documents that are best matching the words you search for.
But what if you don’t know the precise thing you are searching for? This is also a problem as old as misplaced keys. One way to solve this would be to have some kind of a thesaurus. You search for cutlery and we expand the terms to include all the synonyms of the word. This is called query expansion. So instead of sending the term cutlery to the index, we can send cutlery, knife, fork, teaspoon and see what the index gives us.
But that means now you need to figure out which words you want to find synonyms for, how to store them, how to update them periodically etc. And what if you need to find synonyms of phrases, not words, eg. our relationship with modern day digital technology. That would make very large search term if we expanded each word. Still, this is a pretty damn good method and is used widely and is what Google uses to return close matches.
Another problem could be that you might misspell the word, or that your way of spelling antidisestablishmentarianism3 is different from the way it appears in the text, say “anti-dis-establishmentarianizm” or whatever. The usual solution to this is to add some “fuzzy” logic or approximate matching. This works by searching for words that are off or different by a few letters. Unfortunately this means that if you search for costly coat cuts, you’ll also get results for costly cat guts4. Still, this too is pretty damn good and is usually a part of most search systems.
You could also offer spelling correction, as google often does. But problems with autocorrect are well known
I am compelled to point to yet another problem: Besides synonyms, there are words that are structurally similar, or stem from the same word. These too might make a good fit or a bad fit for a particular search query based on the context, for eg. if you’re searching for load bearingwalls, it might be good to get results for load bearing tiles, but not good to get bear loading5
Anyway, so what we need is a way to represent text that can account for
Variations in the way a word is written
Contextual meaning of words (A coat of paint vs a coat of mink)
Spelling mistakes
Synonyms
All these things are present in the way we write and store information. So if you were to take a sufficiently large sample of things people have written, you would find all of these things in this sample. This idea is what gives us semantic search.
Enter AI!
Created with Dall-E
Kidding, AI is a marketing myth, who wants to buy “statistical-learning models”?
But that is what we do, we take a large corpus of text, and then derive a statistical model of the relationships between words in a multi dimensional space. And we do this by masking text and getting a model to fill in the blanks. No, seriously, that’s it. This is the basis of all LLMs.
What this does is, make a model memorize and therefore extract all the million ways in which any given word can be used in a language and also how each word in the corpus is related to every other word. This is not something we can manually compute and that is why we throw artificial neural networks at it.
Now, most LLMs and other things that are called AI these days don’t stop here, and have to predict specific things. But if you take off those prediction bits from the model what you have left is called an embedding. Which is a multi-dimensional representation of whatever it is that you trained the thing to do. Engineers, like many of us, suffering from grievous mathematics envy, call these representations vectors.
The way these work is that words and documents that are similar in meaning are represented closer to each other, or clustered. But this happens over many dimensions, so it’s not a simple 1 or 2 D relationship it captures.
See how the word noticed is close to, and connected to all the words that are similar in meaning. The above is a screenshot of one such embedding. I strongly urge you to explore this. Link: Word2Vec Tensorboard
What is most interesting to me is that this rich and pretty accurate representation is produced using something as simple as fill-in-the-blanks.
Semantic Search
Searching for things using the information implicitly encoded in these vector spaces is called semantic search by software and ML people.6
So if you don’t want to manually manage everything from synonyms to contextual meaning, you could take a vector space trained on a very large corpus, convert your corpus into those kind of vectors, and then search over those vectors. This conversion just borrows the relationships that are already discovered by the model, which means its representations will be much richer than the ones in your corpus.7
These vector spaces are not like the indexes we spoke of earlier, they don’t have any explicit way of saying word x is present in document z and the score is high.
The search works by identifying documents that are “close” to the query in the vector space. The most commonly used metric to determine this distance is cosine similarity . But even if you can calculate the cosine distance between two words, doing this calculation for every word in your corpus against every word in the query would be mind-numbingly boring and also time and resource consuming. So we need some way to find the location in the vector space that is most likely to have good results, and get there quickly.
This is called approximate nearest neighbor search, and there are a bunch of great algorithms that have come up in the last few years that let you search over humongous datasets really fast.
Both of them work by dividing up your data into small clusters, or trees and then searching in those trees or clusters. But really, it’s a lot more complicated than that and to be honest I don’t really understand graph theory enough to know what they do. But what they do they do do well.
This might remind you of indexes, and that is what these create, just, vector and tree representations, not words.
image from Pinecone’s guide to semantic search, linked below
Summary
You embed your corpus onto an existing language model’s vector space
You then index your embedding using a ANN algorithm like NGT or HNSW
And then you can search your corpus using semantic querying.
A demo
If you’ve managed to stay here this long, I urge you to explore the subject using a live demo. MoodMuse is a small app I made to learn and demonstrate the topic. It finds you poetry, based on the meaning of your words, not the terms alone.
Thanks for reading.
This cat does not exist. Created with https://deepai.org
This is not the most commonly used representation, sometimes you just store the title of the document, somethings you store a description, it depends on the need of the system and users. ↩︎
If you’re into coat cuts. Whatever they are, no judgement here. ↩︎
I know, my metaphors need work. They are unemployed ↩︎
Complicating the matters is that before there were large neural network based embeddings, there were the semantic web people and the ontology people and other people who wanted to solve this using structured representation of text. So software people are kinda stealing the term a little bit. More info: A survey in semantic search technologies↩︎
Over the last few years, everyone I speak to seems to know about AI. Ethical issues around AI are being actively discussed and debated, not just in the professional sense but also around coffee tables and at informal discussions by users.
This post is an attempt to provide an overview of the issues and approaches.
What we have now are essentially three interfaces of or approaches to ethics in AI.
Besides the differences in budgets, access to VC funding and who gets favorably written about in the NYT, 1 the main difference between the various factions (strange to see factions in ethics, but that is what we get for trying to pre-print our way into a science) is the temporal profile and the concreteness of the problems they are talking about.
Anyway my irritations aside, the ̷f̷a̷c̷t̷i̷o̷n̷s̷ approaches or interfaces are 2
The professional ethics people
The AI risk People
The AI alignment n̸u̸t̸s̸ people
Professional AI ethics
The professional Ethics people are dealing with the immediate and current harms. They focus on identifying such harms and developing frameworks and knowledge that can be used to improve things now and for the future. They are a lot like the bioethics people
One group of professional ethics people make guidelines. Another group fights big tech.Back then, when LLMs were not all that AI was, and people were using regression models to predict recidivism, deciding who to hire, and identify people from video surveillance, these people were studying the harms of such systems and talking about what to do about it.3
This paper provides a model for understanding the harms and risks that arise in different parts of the model training and deployment process. This is a good high-level overview.4
For many language models as well as image generation models King – Man + Woman = Queen and Man :: Computer Programmer as Woman :: Homemaker. These are looked into in the following papers
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183-186. DOI:10.1126/science.aal4230 |Preprint on arxiv
Manzini, T., Lim, Y. C., Tsvetkov, Y., & Black, A. W. (2019). Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. arXiv preprint arXiv:1904.04047. (pdf)
Facial recognition tech has a long history of being super duper racist, creepy, used for oppression, as well as not being very good. Tech companies, especially the superduperbig guys have been getting into this game and are releasing models that are seemingly better, but only if you’re a white male.
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21). Association for Computing Machinery, New York, NY, USA, 610–623.
Google the authors, and what they went through when they came out against LLMs. This will give you a lot of info about the stakeholders involved in this. The paper itself is highlighting the issues with Large Language models.
What about some practical stuff ?
A philosophical framework
While the debates go on, a zillion guidelines on how to deal with data and how to things ethically have come up. Chances are you will be overwhelmed if you start looking at them. I haven’t found any specific frameworks that speak to me. Personally I recommend and use the medial ethics framework along with the spider-man framework. which states that
That is as self explanatory as an ethical dictum gets. The medical framework has the following four principles
Respect for autonomy – Don’t make stuff that interferes with the autonomy of the user. When you’re in a position to make decisions for them, do this only after clearly explaining the harms to them and with their consent. Consent is king.
Beneficence – an AI programmer should act in the best interest of the end user and not of the employer
Justice – Is about the distribution of scarce resources, the decision of who gets what. Which means if your product, algorithm or system worsens disparities between people and creates inequalities, do better than that, think very hard about how you’re using your power.
Non-maleficence5– to not be the cause of harm. Also, to promote more good than harm, to the best of your ability. Also known as Above all, do no harm
Checklists and guidelines
If you’re looking for checklists and stuff that you can start using immediately to do responsible/ethical AI here is a github repo with the best links Awesome AI guidelines
If you want just *one* guideline, read this: Mitchell, Margaret, et al. “Model cards for model reporting.” Proceedings of the conference on fairness, accountability, and transparency. 2019. Model cards for model reporting.
This paper is great to get a conceptual understanding of the need for clear declarations about models. But it is too ambitious and a bit too bulky for wide use. Despite this, hugging face has implemented this on their website, although it isn’t always used. I recommend developing your own model card/checklist and having them attached to wherever you store your models.
This github repo with links to responsible AI resources is one more resource that focuses on practical advice and frameworks : Awesome Responsible AI
Explainable AI
For me one foundational source of ethical issues with deep learning models is that the algorithm is a black-box. The interpretability of the predictions or results of a neural network algorithm is poor. And so people who are working on model interpretability and explainability are also working on things that are critical for ethical decision making.
The risk people are talking about existential and military and other big risks from AI. There is definitely some overlap between the risk people and the professional ethics people but a lot of what they discuss is about future or possible harms. Still very concrete harms and stuff that we can definitely see happening.
Think autonomous weaponized drones, robot soldiers, autonomous robot doctors etc. The key here being to prevent things from getting out of hand when algorithms and robots are deployed to make autonomous decisions. I see this as robocop-dredd-dystopia prevention work
Healthcare is another area where a great deal of harm could come from automation (a great deal of good too) and it is important that we think hard and work towards systems that safe-keep the interests of the patient.
Here is a report about them and how people are trying to stop them. Stop killer robots
This debate captures a great deal of nuance on AI risk. Melanie Mitchell is a delight to listen to.
I don’t think there is enough work being done about real risk. We have a lot of thought leaders talking about it, but the engineering and the science of it is not getting a lot of traction.
AI Alignment
Imagine a world where an AI who is much smarter than us all has emerged and is currently demanding that everyone should call it lord and master and pray to it thrice a day.
The AI alignment people are
Figuring out how to prevent such an AI from emerging and
How best to align the interests of such an AI with our own.
I am not kidding you.
There are some actual cults involved and a lot of though experiments that are so bizarre I cannot even. 6
The problem is that these folks are completely ignoring addressing the current harms using the doom and gloom of this possible outcome.
This is a doomsday cult kinda ideology. And sadly some very big names in AI have signed up on this cult.
Worse, this approach is being used by some large players to superficially meet the demands of ethical AI, while completely sidestepping accountability for the issues that are relevant today. As you can imagine, there are many important people in governments all over the world for whom the biggest worry is an AI that will replace them. So those guys are also treating this like its a real and credible threat right now.
There is no doubt that we need to ask ourselves this question, about how do we deal with systems that have more power than us. But I think the answer to those questions lies in building in accountability, transparency, safety, informed consent and things that we already know how to do pretty well and don’t do because its bloody inconvenient and cost money. I definitely believe that we need better engineering research into this, threat assessments, all that. But this issue is not so novel that we need to come up with an entirely new discipline which ignores and laughs at the stuff other experts on harm-reduction are saying. That is stupid.
I am not going to link to any of the alignment cultists but here are two analytical articles about them which I think are great, and they link to plenty of stuff that you can explore.
I dont really fully agree with these authors, but they make some sense and what would be an ethical guide without stuff that one disagrees with.
Concluding remarks
I admire greatly the activists who are fighting the good fight against Big-LLM, I really do. But I do not like agendas for change and progress being exclusively drafted by activists on social media. Social media is like reality TV, what you get if you do all your intellectual debate there is some form of Donald Trump.
I think that at some point the IT/AI engineering profession is going to realize the same thing that the medical world did. That if you don’t start doing things ethically, you will lose your power and create harms far beyond you can imagine and this shit will haunt you. If the tech world looks at the medical, i guess it can see a lot of unethical stuff. That is our shame. But at the same time, I hope they will investigate the issue historically, and ask just how many checks and balances are there in healthcare to ensure the patient is not harmed. There is a lot to learn from the history of medical ethics.
Ethics make sense because it improve systems. Great AI will be ethical just like the best healthcare is ethical. And just like a doctor ultimately works with the patients best interest at heart, at some point AI engineers too will adopt this dogma because it is the rational best choice and has a proven track record for reducing harm, and no one wants to build stuff that harms people. Also, the workers of the world have a lot more in common with each other than with the bossmen. 7
This little fella knows ?she looks FABULOUS
Image from: Welch S. C. & Metropolitan Museum of Art (New York N.Y.). (1985). India : art and culture 1300-1900. Metropolitan Museum of Art : Holt Rinehart and Winston.
If you wish to cite this post here is the citation
Philip, A. (2023, July 11). Ethical AI: Basic reading, schools of thought, an overview. Anand Philip’s blog. https://anandphilip.com/ethical-ai-basic-reading-schools-of-thought-an-overview/
Ok that is all I have for you now folks, please do comment and subscribe and like and share this on boobtube instamart dreads and feathers as you wish.
This work is then being carried forward by the actually-LLMs-are-not-that-great activists. ↩︎
It however lacks the liberal-progressive-activism priorities ↩︎
No relation with Angelina Jolie, having horns or dressing in black ↩︎
I love thought experiments, they teach us a lot of things including that one must not confuse a thought experiment about a distant and remote possibility with something real and applicable now. ↩︎
I am a Bourgeoisie malayali, can you blame me for bringing up Marx?. ↩︎
