Example 2B: Medium Format — Experiment
--
Context 1 — GPT-2 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset. GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text. Try it here: Link to generator.
Context 2— Researchers from Harvard University and the MIT-IBM Watson AI Lab have developed a new tool for spotting text that has been generated using AI...it exploits the fact that AI text generators rely on statistical patterns in text, as opposed to the actual meaning of words and sentences.
“By GPT-3”
Adversarial paragraphs written to imitate GPT-2 in order to spoof AI-generated text detectors. It actually ended up being quite fun writing these.
This article is based on true accounts of stories well documented as well as stories heard under the grapevine, all of which have been fictionalised for prose. This article for the most part tries to veer away from any third rails. It softly segues from one tale to the other hoping to elicit a form of elliptical prose. This is nothing new, these issues have been perpetuated across the ages, in problem discovery you are better of betting that its medieval or renaissian. History will yet again be our magistra vitae in dealing with new decision-making tools. It is tough to make predictions, especially about the future. There is not incentive for it, the world doesn’t have great solutions — it is a world of partial solutions. As is quoted by the Adams family what is normal for a spider is chaos to a fly. If it’s not true to the letter it’s true to the spirit. This is the start of a real intellectual journey, like being in the forest by yourself.
As these models allow you to offload some cognitive complexity. The future is not what it used to be. Expect the rise of duck food data scalping, consultants provide you with a conundrum just to repropose your data for future projects. Firstly, I see the whole thing from afar and then all the text has to be licked into shape. You would have to think that once the decisions are triaged through management, solutions would open up like the red sea. A bout of bias can take hold of your thinking and let on every strain of evidence that completely refutes your model. Machine learning has turned into a caricature of itself. Internet it is not what everyone is doing but what everyone thinks everyone is doing. The greatest healing is when you realise you were never sick; you just had a bit of dirt on your monitor. Outrage is now a sort of forking mechanism that slices new communities into being. I ensure that I am proportionately vengeful within the bounds of moral law, sometimes I unfortunately tit for half-tat. The tools have surpassed your biological competence. The power we wield have surpassed your brains.
Machine learning industry application is a cesspool of bike sheds. It is easier for a committee to approve nuclear plants than a bicycle shed. No one on this committee knows much about nuclear power plants. Opposition and dissent are lacking. Management unstudied in quantitative sciences might be barking up the wrong dog. Journalist could be monkeys on typewriters, but no one wants to bad mouth monkeys. If you don’t move, you won’t realise you’re in chains. I am not here to be a master of any future, I am here to absorb the past, the future is an apathetic void of no interest to anyone. Often when we read or learn from something, we see in our adversary an opponent we reactively devalue. Like all impartial people I am partly impartial. The worst part of online shopping is having to get up and get your card.
Probing the large shifts towards automation someone should take on the role of the fifth column. Machine learning did not put itself on the top shelf. Able to harness data to discard false epiphanies and intuition. Machine learning has not yet been criminalised, but the war on machine learning certainly begun, and while it is legalised the over-prescription of machine learning could be a problem. Here the point is to take a strong opinion, the pendulum is still too far on the side of technocrats, the point here is to grab the pendulum and shake it, not to kick it over. Think about the sheer number of hidden mistakes, obfuscation and secrets, by multiplying your own proclivity into the exponential growth of a complex society.
We are to label and dissect what is happening around us. Great you are able to predict consumer preferences, but what about the voting paradox. Would you not have been better off with a ranked choice system for more equitable results. Explainable is a way not to be as firmly divided among partisan lines. You can get people to listen just by adding the word boo, baa or AI. The adoption of AI causes some changes in physiological processes, some managers risk the complete annihilation of their frontal lobe. This speculation comes from experiments involving real humans in corporate captivity. If your desk is aligned from east to west, the magnetic alignment of the earth predicts you enjoy afternoon sun. Some scientist argues that the positioning is not willy-nilly on corporate pastures. Employing a data scientist could be a fool’s errand how more gingerly can one be.
The return of the vanquished disease of overanalyses, overprediction, and over quantification reflects historical amnesia. It is true that they have arrived on the block, the question is once they leave for the next thing, have I been adequately pushed towards the utility axis to ensure that the velocity of churn doesn’t cause a nosebleed to adoption. This trope fundamentally pulls on your inner Hobbesian. You write so clearly that I cannot understand what you say. The developing character normally just leaves you in tatters forcing you to readjust so that you can once more fend off people with bad but commendable private-public splits.
It is often the case that models can’t be both explanatory and accurate, like housing which cannot both be affordable and a good investment. There is a level of mutual exclusivity. Because knowledge have become more specialised, we rely on the people in the field to tell us how much they are progressing. Only an intellectual could believe in progress. It is starting to feel like I am creating some sort of mental list here but there is an extension of the charitable interpretation and it involves a few moving parts. Wealthy people with smiles and ostensible humility, while in their off time they ponder on the purpose of life having seemingly reached the pyramid-top constructed for us by Maslow. Stanford students who insist they don’t study hard, just too keep appearances. Executive forced out of the interest of shareholders to play games as the figurehead of a corporation, while most are craving for purpose and have developed an excessive interest in Robbins and Carnegie on the way to work.
Like Kant said, machine learning as an animal spirit that will vie for attention in a zero-sum world, the question is will positive sum benefits transpire. You needn’t kowtow to the machine learning monster; some traditional alternatives still bode well in accuracy and certainly in explanatory brilliance. This three is a bad tirade of traits. It’s also morally salutary to understand the effects it would have on your employees. I hear fine sentences every time from the man on the street. English is both a Latin and a Germanic language. For any idea you take you have two words, two registers, those words do not mean exactly the same. Fraternal is not the same as brotherly, regal is not the same as kingly, dark is not the same as obscure, ghost not spirit. King is to man as queen is to wife. Some articles can be translated from Latin to Germanic English and get away with it. Without lying, every long-running soap opera would be over in 5 minutes. Lying is a hydra, it has one consequence that you do expect but three or four consequences that you don’t expect. The externalities of calling someone out is their reputation in favour of your prestige.
Something is locked into a sequence. It is good to write something actionable every so often. I was casting for a different phrase but would settle on this one for now. In the time of burgeoning capital experiments are abound, slicing and dicing at the status quo. Reminiscent of the social worlds of Versailles and the Tales of Genji managers are seeking methods to toil in more time with less resources. Algorithms are designed to emphasise popularity, impressiveness and likability as efficiency is just a bit ho hum. Companies are playing a new social game as the corporate world becomes safer with less stringent rules. Art, design, and accomplishments are taking precedence over engineering.
There is a disjointed nature between discovery and observation, Galileo discovered the observation of a potential star in 1612 but put it down to observational error. The one who discovered Neptune was not the first to see it, but the first to realise what they were looking at. Schmidhuber noticed and documented a star in 1650 in an obscure journal only read by me and long finger. If this insanity is hereditary you receive it from surrounding text in previous iterations. It is true that it is sometimes simpler to compute things outright as opposed to relying on approximations. The law of small number and all other grossly underestimated biases magically disappear once it enters my black-box. Convenient but not performant. False discovery rates are the largest contributor to startup success. The alchemy of transforming people to money is the most intricate form of alchemy.
In some scenarios non-sensical features slide their way into the top list of predictors. Canadian hockey players have a disproportionate number of top players born in the first quarter of the year. The result is confounded. No astrological magic is at play. Bert was built in Jan, not a nice man. Clustering is anathema. A data scientist risks digging themselves into mushy inarticulate holes at team visitations. A data-scientist will scratch their head, while a domain expert will know that the eligibility cut-off in the hockey league is January the first. And that small difference of age has large effects at the extremes.
Too many functionally similar packages to screen with different abbreviations. The modernist cult of originality has spurred us on to swap mnemonical order as a stamp of uniqueness and the everyday user is suffering to make sense of it all. Too much is reinvented from first and zeroth principles. It’s even worse when they are reinventing a square wheel. The main inclusive and exclusive difference between machine learning and statistics is p-values and cross-validation. There is not much purchase to that way of thinking.
You first have to understand how things are traditionally understood, to understand the traditional weaknesses. I would argue statistical significance should only be used for comparative reasons across multiple pairs and selections; the problem is that once statistical significance is used for comparison you might as well settle for effect size. The problem with using statistical significance blindly for a single pair is that by virtue of testing two different things you would expect significance with a large enough sample. In reality null hypothesis are always false.
You have to test your results in the real world before you discard it. Some results might show bad performance in a sandbox or in the lab, supposedly some of the first military radars did better in the field than in the lab. Multiple effects could be at play. You don’t always feel like saying xylophone and zebra after a long day, but we don’t complain, we incorporate even the rarest of examples. We tend to underrate our simple works, lemmas, and appendages. William Tell’s Overture of four minutes has cemented him in the minds of many, not his four-hour opera. It’s Schwarz, Poincare, Fatou and Zorn’s lemma that stands the test of time.
Different people are affected differently by anything that affects anyone. The whole point of a raw machine learning model is to statistically discriminate. Regulators step in to optimises for the reduction of bias and capitalism for the optimisation of bias. What sort of questions are we asking, I quote John Tukey, “what is happening, where am I”? What is the point in predicting which employees are more likely to be absent when one can predict which employee will perform the best? Absenteeism need not be related to job performance. In good times average policy can look celestial and good policy can look diabolical.
Eric Raymond in the Cathedral and the Bazaar compares open source software to Bazaars, as much as this is true, you can approach it in a cathedratic manner. Although bazaars emerge without coordination, a business has to learn to apply structured thinking to their use of open source software and research. You have to pick well and become fast switching computational minimalists. The best performers of a larger group are typically better than the best performers of a smaller group. Irrationality is not uniquely an Australian preserve. We intuit the same relationship with the size of the data.
The uncanny valleys in business are filled with corporate relationship hubris. Nothing looks uncanny with managerial smiles plastered across overpriced website. The innovation vacuum is being filled with stock with efficiency grinding machines. It is not like we are experiencing a Cambrian explosion of sentience. Although machine learning cannot throw fists yet, some rules might be necessary, speech can’t throw punches still have a physiological impact. We need compassionate policies from dispassionate analysis. Machine learning should be put on a maladaptive adaptive continuum.
Hypothesis swapping, the salmon story. A speciated generation who wants no ambiguity. Most data scientist anatomically adapted to stare at a screen waiting for model to converge. Best way to fight science, get another predictor. Use salmon instead of humans, see if the same effect prevails. Predict student’s average height by teacher, if it is correlated the average test scores per teacher does not mean much. With too many interactive entities there is no general closed-end solution of a predictive algorithm that will suffice. Numerical methods are needed.
Many companies do not need to nor have the ability to predict a single task to inform its business processes, yet they still employ data scientists. Data scientists are not terribly beneficial without data. Your PhD’s will grift over percentage point of performance, making you poorer by the EC2 minute. Maybe you are better off waiting until the Chinese solar power subsidies kick in or even better for creative finance to foot the bill to replace the Zen garden on your roof with the latest GPU servers. The Ptolemaic earth-centric model fell afoul when man saw sunspots and Jupiter’s moons, like so the human-centred model will end. Since Copernicus our human mediocrity have become all to evident. With each additional shift in paradigm scientifically, technologically or otherwise, we become ever more aware of our mediocrity and an ever-increasing possibility space to which the initial reaction is awe and our delayed emotion, despair. The best way to deal with this is to substitute out this naturalistic cosmic layer for something that provides a satisfactory answer be it myth, purpose or religion.
Overfitting does not necessarily mean that the solution is not generalisable. A solution can overfit the near future for short term gain with negative long-term repercussions and generalisation. If machine learning can be approached as a system of concordance where models are run in parallel, until one is satisfied with the results. Malthus’ prediction was historically correct but wrong after the industrial revolution where oil out of earth pushed us forward. Machines don’t have trailing data of something that is ultimately unknowable. There can also just generally be pollution in the information ecology. Externalities in the way datasets are captured.
Oh, homo economicus, fear girts and gashes the strings cording you to movement, halted by the play dead of life, pulling furiously on the frictionless grip you thought you had. You escape what can be escaped, just to be caught again by yourself. It’s easy friend, have you not seen that its easy friend. It’s just thought you say, it’s just thoughts. To think is to throw your mind in the cloud praying for knowledge to rain back on you. Some people have an uncanny ability to envisions their future emotional state. I see traces of this in prescient authors like Frankl and Hesse. They seem to understand the unwavering viscidity between their current and future state of mind. People generally struggle to predict their emotions due to a hesitance in picturing modest change.
Large machine learning failures risks sucking the legitimacy out of AI. The showstopping blowouts have not happened yet, although we have started to see the smoke. Public intellectuals are splintering in real time, fracturing over issues of AI safety and implementation. The issues being debated are far removed from reality; An article like this could allow us to square off on issues that exists right in front of us. From the dark recesses of this mind something quite operative emerged. Today is nothing but the morning after the night before. A lot of intelligence is about ignoring irrelevance. We live in a post-parody society where it’s impossible to invent anything more ridiculous than reality.
Public spending is feckless if it cannot be translated into solving practical problems. One should not consider scientific disagreement as an indication of dubiousness, it’s part of the process of falsification in the public square. Machine learning has given managers and CEOs a new bastion, the great generals can’t do much between dry spells of innovation. While writing this I am sure I would someday find myself in a hypocritical loop. Everybody for a long time looked at the universities for advice on how to address complex problems, because the universities built up a great reputation with the hard sciences. So, the other departments at the universities can live on the fumes of deserved reputation. It is better to think back from the fifth dimension towards the first dimension, makes it easier to see that everything is preordained.
We have a great sense of the logarithmic, while using these models, we are slowly made aware of small spikes in turbulence making us afraid of a fractal spike. We come across breadcrumbs of inferior performance to traditional models. We become weary of putting our business on autopilot. It is not easy to hark back technological breakthroughs. The answer to the game theoretic meaning of life leads to infinite regress, constant why’s, or infinite loops, why because of because, or axioms this because of God. Meaning of life because of life.
You are immediately priced out if you don’t pick a side for yay or nay as it is quite hard to hold a sensible position between the two. The skill level is too high. And even if you knew how to hold one, it’s not necessarily fun, like Carl Smith remarks, humans are never going to agree as that is part of the human condition. The position held by management were absolutely anathema to the data scientists. This article is just writing itself to provide some fodder to the easily entertained. Before UBI first pay the salary for fake jobs to retain meaning, bar-sitter, city-walker, dog-whisperer to keep faces smiling.
Unfortunately, the existence of AI and the speed of adoption could force us into a two-tier system where it becomes expensive not to utilise ai tools. Most companies are Jacksonian and don’t care about what is happening outside their own. Even worse, they don’t seem in favour of contributing to open source research. I think it best captures the qualia of being a machine learning engineer. “I can’t be hypnotized.” Says the person who spends the majority of the day wearing earbuds and staring at a screen. Maybe the world would be better off with management hiding behind piles of paper as opposed to machine learning models that so easily get the blame. From this position it is all a corporate flatland, no noose-restrained suit dweller can see the large data lake for what it is, they can’t gauge the steps of descent, they don’t understand the uncanniness of the valley. Meanwhile the data scientist has a front row seat to the lake, and it is their intention to purify it; they would do this without knowing that the purification of the pond leads to the water lily’s death.
Apropos of nothing, these models are applied left and right, believe you me the inanity and hyperbole get worse with every step up the corporate ladder, and let’s not start talking about the consult. There is a serious need for remedial work if you step from the business world into the world of quanta or the other way around. I kid you’re not we are almost at the point where it is better once more to repackage yourself as a statistician, quick go reretrofit those LinkedIn profiles. All the data scientists applaud for their small win at the level of simulation. You can make a pretty good living talking to people who already agree with you. To test a good model don’t let it play with acolytes’ package and send you model to enemy.
Don’t upset your data scientists, they might quit and like a researcher write textbooks for themselves. Next year we might need to change the description of this banner to podcast. Most researchers are Pangloss about their work. For every scientific solution we have to adjust for bias based on researcher degree of freedom. All the other consultancies are hot on your trail, all it takes is a head fake to shake them off. Open up a job for a quantum computing researcher, that would keep them busy. Smart individuals are tactical and will reach a much deeper more striking point if you allow them to flourish forward from their prelude. That is unless they start some hogwash about the anthropic principle. Production is all a bit round about nowadays, we harvest crops with machines that were manufactured by other people who used machines that were manufactured by people.
Kahn’s studies show that in terms of human well-being technological nature is better than no nature, but not as good as actual nature. We should develop and use technological nature as a bonus on life, not as its substitute, and re-envision what is beautiful and fulfilling and often wild in essence in our relationship with the natural world. Are we made to exist in this constant position of uncertainty between machine and human? The best way to stop romanticising about the future is too remember that humans are coming along for the ride. Humans who can’t fight their way out of a plastic bag.
We should be more willing to substitute the factual for that what make us feel better for daily living, while inconspicuously acknowledging the factual truth whenever such acknowledgement benefits the collective in some utilitarian sense. Substitution is an obvious tool used to hammer down the ignorant, our brains insistently find easier questions to answer as proxies for hard questions we can’t conceptualise. This concept should, however, not be put down and should be investigated for its remedial properties. On the topic of questions, many a wise man has said that the questions someone asks can be quite telling of their disposition. Question are very powerful and should be phrased with extreme caution especially if you want to pre-emptly architect the respondents answer.
Reinforcement learning exists for reasons of plasticity, they remain embryonic for very long waiting to learn from action and feedback. It is the way we learn during the day, whereas supervised learning is the way our brain processes information at night. As of now it is a step function increase, the open source turnover rate is quite fast. The question as to whether AI would make us more productive is still up for grabs. In the same way that the quality of pottery deteriorated after the fall of the roman empire, it might be hard to argue that it was a simple change in culture. Something as revolutionary as AI should surely up in the productivity figures. But Rome wasn’t burnt in a day.
Almost akin to evolutionary psychologist that would not have been needed had evolutionary biologist not been too afraid to study what happens from the neck upwards, statisticians would have been all we needed if they just adopted non-linear models faster and started emphasising validation sets. Like the classics, you might be better off believing in cycles as opposed to progress. Maybe in the future corporate strategy will turn into gobbledygook only understandable and executable by machines. The grass is always greener with the shoe on the other foot and the grass is probably greener because you are not over there stomping your feet.
The Renaissance and the reformation were both backward looking movements. This is critical, what has been true could be true once more. The Renaissance repeated that the classics are supreme, the scientific Revolution replied that the Greeks are wrong, the enlightenment professed that religion is supposition for promoting reason. Maybe humans can be the hormeses to the models. My friends, through who evolution bread reality, are sophisticates, I met them on the internet. I find it depraved that some companies are focusing on machine learning without the business foundation and strategy to carry the technology. In New York City people are disconnected from their local governments and don’t insist on better metro systems, they are looking down on their phones; they don’t care that they are moving about in 100-year-old infrastructure.
An accountant after reading how-to-automate-the-boring-stuff decided that he would predict next year’s budget variance using the last decades financial statements. He decided on a random forest model after recalling it from a Coursera sound bite. Yes, your solution is great to the extent that it is the only solution available. After reporting back the results to data scientist, the accountant was appalled to learn that he was trying to knock a fly off a wall with a wrecking ball. The data scientists soon stepped in to minimise the damage. Hope you enjoyed the harrowing ride to the terminus of this elliptical prose. With an arrow buried in our chest we argue over the scent of the wood the shaft was hewn from, and the song of the bird of bird whose feathers make the fletching.
There is a need for a concerted effort to release the models from bias. In these domains, the perception of unbiased models is as important as unbiased models, you have to show the world before insiders do the job for you. You can’t possibly believe in machine learning essentialism. The importance of intuition, when the ocean receded if means that the ocean monster was hungry this native tribe in Thailand knew this and their lives were spared. No professors alive can resists writing non sequitur, prima facie, and ceteris paribus as the sight of a page. They are squashed into intellectual submission to spew out Latin. Perfectly good sentences get ruined by perfectly awful words. Machine learning comes in two flavours, one that wants to support you and the other that wants your job. For now, let me sleep let her sleep for when I wake the world would move.
