[00:00.520 --> 00:02.320]  Right, thank you all for coming today.
[00:02.600 --> 00:06.400]  Thank the organizers and thank the people who are working through this entire conference.
[00:06.600 --> 00:07.400]  It's much appreciated.
[00:07.720 --> 00:08.940]  And speaking here is an Alan.
[00:09.420 --> 00:10.320]  So let's begin.
[00:11.600 --> 00:15.140]  The proponents of AI claim that a new world lies before us.
[00:15.660 --> 00:17.740]  LLMs and generative AI are the future.
[00:18.880 --> 00:21.760]  This new world is usually presented in one of two ways.
[00:22.000 --> 00:29.880]  In the first version, machines emerge that are smarter than humans, that can teach themselves and will take your job.
[00:31.280 --> 00:33.220]  And ultimately take over the world.
[00:33.700 --> 00:40.540]  In the second vision, we hear tales of a utopia, a life of leisure, without the need for any of us to work.
[00:41.080 --> 00:45.700]  But as I see it, there are ulterior motives for presenting both these futures.
[00:47.120 --> 00:48.920]  Fear makes people pay attention.
[00:49.740 --> 00:54.780]  That fear makes politicians pay attention and they write laws in response.
[00:54.780 --> 01:05.020]  The advisory committees to those politicians are stacked with corporate visionaries, who make suggestions that favour the same corporations that are provoking our fears.
[01:05.980 --> 01:11.580]  Having the utopian narrative available provides a handy counterbalance to this fear.
[01:11.580 --> 01:21.940]  So in this way, the AI will need to be regulated away, nor unregulated, but regulated just enough to benefit those driving the agenda.
[01:23.000 --> 01:29.880]  So as impressive as AI is, and specifically generative AI, we're a long way from intelligence.
[01:30.320 --> 01:35.280]  You might all recall that we've been told that we'll have self-driving cars since 2016.
[01:35.760 --> 01:39.180]  So perhaps 2025 will be the year we finally get it.
[01:41.140 --> 01:45.940]  In this talk, I'm going to present a third vision for our AI future, or a possible future.
[01:46.560 --> 01:51.940]  Not the Hollywood version, there'll be no Skynet, no Terminators, nor a utopia.
[01:52.740 --> 01:56.440]  I think the most probable future is a grey, dull, and dingy one.
[01:57.320 --> 02:02.320]  Where we follow our corporate guides to greater shareholder value in the coming quarter.
[02:02.940 --> 02:08.940]  We approach mediocrity, a loss of creativity, and a cultural and technological dark age.
[02:10.360 --> 02:12.620]  Although I paint a drab picture, there is hope.
[02:13.220 --> 02:18.820]  The hacker culture and its ethic form a direction with which we can fill our sails and propel us to the new world.
[02:19.620 --> 02:27.140]  In the rest of this discussion, I'll walk you through my reasoning, provide some key terms, and suggestions on how people can make a difference.
[02:28.000 --> 02:29.960]  Let's avoid being caught in the doll rooms.
[02:29.960 --> 02:36.400]  So I'll begin the journey with the building blocks of AI and modeling.
[02:37.420 --> 02:43.620]  Without them, it's all too easy for people to evangelize and obscure many of the real outcomes.
[02:47.730 --> 02:50.090]  The process of modeling begins with data.
[02:51.050 --> 02:53.890]  Data that represents an input and an output.
[02:54.430 --> 02:59.770]  The job of a model is to take some input data and generate the desired output.
[03:00.810 --> 03:05.870]  So the model is just a machine or a function containing one or more variable.
[03:06.690 --> 03:11.070]  The utility of the machine is measured on how frequently it predicts correctly.
[03:12.230 --> 03:20.550]  In order for it to achieve this utility, we just have to adjust the model variables until the model gets closer to what we desire.
[03:21.590 --> 03:26.610]  So during training, we twiddled the knobs of the machine until we've minimized the error.
[03:27.970 --> 03:29.190]  So how do we twiddle?
[03:29.450 --> 03:33.450]  Well, the common approach is gradient descent, which is why it's part of the name of this talk.
[03:33.690 --> 03:41.470]  But before we go there, let's address some fundamental components of data, the input, the output, the error, and our accuracy measures.
[03:43.810 --> 03:48.810]  So if we ask 10 people to define what data is, we'll almost certainly get 11 different answers.
[03:49.710 --> 03:53.230]  I'll choose a broader description than the one given here by Merriam-Webster.
[03:53.790 --> 03:55.290]  And soon you'll see why.
[03:56.050 --> 03:59.330]  I'll describe data as anything innumerable and accessible.
[04:01.010 --> 04:07.450]  We normally think of data as something that's factual and concrete, such as when two people measure a distance.
[04:08.190 --> 04:11.250]  The people may disagree on the measure being used.
[04:11.490 --> 04:13.150]  We use feet, inches, millimeters.
[04:14.010 --> 04:19.310]  And they may disagree on the degree of accuracy that they measure it with.
[04:20.170 --> 04:25.470]  But the values are in some distribution around a truth, a true value for that.
[04:27.070 --> 04:31.870]  That distribution around a true value is ultimately the motivation for statistics.
[04:33.130 --> 04:36.490]  There's some physical, external concept of a truth.
[04:38.730 --> 04:42.370]  But when we enumerate human abstractions, things get really messy.
[04:43.190 --> 04:48.130]  Two physicians can look at the same patient vitals and come to a different diagnosis.
[04:48.730 --> 04:59.550]  Two people can look at the same piece of written text, and one might say that person is happy, and the other may say that they're very sarcastic and very unhappy.
[05:00.910 --> 05:03.850]  And in these cases, there's no inherent truth.
[05:04.470 --> 05:07.330]  There are distributions on distributions.
[05:08.130 --> 05:11.650]  There is an inherent bias in our system.
[05:12.930 --> 05:18.950]  Language in the written word is an abstraction on the collective human experience as we interact with the physical world.
[05:20.750 --> 05:27.190]  Where we take our data from the Internet, it represents a dominant worldview with the most prolific generators of language.
[05:27.810 --> 05:32.310]  Out of convenience, we're just using the data that's accessible.
[05:33.750 --> 05:38.510]  But the most prolific content isn't necessarily what you want to train your model on.
[05:39.310 --> 05:44.430]  Yet corporations building these models consider the training data to be proprietary.
[05:45.110 --> 05:50.530]  So we simply don't know what's actually in there and what's used to train the LLAs.
[05:52.550 --> 06:11.670]  If we did know there is some data that we actually thought was a problem and we decided to remove it, let's say related to child protection laws and compliance with good ethical guidelines, how would we effectively filter petabytes of information?
[06:13.590 --> 06:18.410]  When you have lots of data, I go further and say, how do we know it's correct?
[06:19.430 --> 06:21.170]  What is it that we want to model?
[06:22.930 --> 06:26.410]  And then how do we measure error in that data?
[06:28.990 --> 06:31.210]  So we can ask ourselves, what is an error really?
[06:32.410 --> 06:40.770]  We can begin by noting that there are systematic and random errors, commonly described for our concrete truth-related data.
[06:40.990 --> 06:43.790]  But that doesn't cover everything that we actually see.
[06:45.030 --> 06:47.430]  Who decides what an error is?
[06:47.750 --> 06:50.090]  Because somebody actually decides that.
[06:51.790 --> 06:53.570]  There are different kinds of errors.
[06:53.950 --> 06:56.650]  Logically, we have to see different kinds of errors.
[06:57.450 --> 06:59.990]  And some errors we care about more than others.
[07:00.370 --> 07:04.070]  Some errors have greater negative outcomes than others.
[07:05.350 --> 07:11.510]  And I see this is a key point that is often missed, certainly by Merriam-Webster's definition.
[07:12.870 --> 07:17.290]  We're reducing the problem of optimizing the machine down to simple error values.
[07:18.270 --> 07:21.930]  In the case of our target, this is the distance from the center.
[07:22.370 --> 07:25.310]  And if we get really fancy, the distance from each other.
[07:26.090 --> 07:29.430]  But we also have to think that someone designed that target.
[07:29.950 --> 07:32.210]  Somebody set the distance of that target.
[07:32.210 --> 07:33.930]  All of which make a difference.
[07:35.310 --> 07:38.190]  We're completely ignoring the cost of all those errors.
[07:39.210 --> 07:41.750]  And who gets to decide what the error is?
[07:42.550 --> 07:47.470]  So I'd ask, what if the cost of an error is a human life?
[07:48.430 --> 07:51.190]  What if the error itself causes an injustice?
[07:52.910 --> 08:01.350]  In our target example, are the points on the third ring more harmful to people than those on the second ring?
[08:01.870 --> 08:06.930]  Does the left versus the right side affect different groups more significantly?
[08:08.390 --> 08:13.250]  And I'll introduce one more error here that's not given sufficient attention.
[08:13.530 --> 08:14.210]  That of time.
[08:14.650 --> 08:20.170]  Our existence, the data we generate, and the error labels that we assign are temporal.
[08:21.250 --> 08:25.870]  What is considered by some as an error today might not be tomorrow.
[08:26.570 --> 08:36.150]  For example, if we built a model of recruiting to recruit shipyard workers, trained today, it would predict that advertisements should target men.
[08:36.810 --> 08:42.350]  But if this was during World War II, that same advertisement model would target women.
[08:45.660 --> 08:49.700]  And of course, we're looking at 20 years of the Internet as our data.
[08:50.420 --> 08:51.800]  And that's a long time.
[08:53.620 --> 09:03.820]  So now that we have the concept of data and error, the inputs and the outputs from the model, and how we measure differences from real values, let's consider how to train a model.
[09:05.940 --> 09:14.120]  Where we know the mathematical description of our model, we can determine how quickly each parameter will move us towards the desired output.
[09:15.280 --> 09:22.740]  The rate of change of the output model caused by each parameter in our model is its gradient.
[09:23.460 --> 09:30.460]  By rearranging our model formula, we can reduce the error by descending along its gradient until our model fits.
[09:32.760 --> 09:43.540]  All we need to know is the equation of our model, its rate of change by each parameter in that model, and the error between the model output and our desired output.
[09:44.520 --> 09:51.000]  Where we have two outputs, we can compare them, square the result in the least squares method.
[09:51.560 --> 09:58.720]  So simply, when we compare the machine output with our real output, how far off is it in absolute terms?
[09:58.740 --> 10:00.380]  This is the error that you see.
[10:01.400 --> 10:08.420]  If the error points match perfectly in all cases, then our least squares error value is zero.
[10:09.720 --> 10:19.060]  Logically, we are descending downhill along the gradient until the error value is zero, which means we get closer to our desired result.
[10:22.710 --> 10:34.370]  So, let's examine what optimization can do if we take accuracy as our only measure of error, and we don't attempt to weight specific outcomes of that.
[10:34.370 --> 10:40.030]  Consider the case of airport security, as handled by the Transportation Safety Administration.
[10:40.710 --> 10:46.350]  In order to prevent issues, all passengers are screened.
[10:47.790 --> 10:51.230]  But how often do passengers actually have an ill intent?
[10:52.230 --> 10:55.330]  The answer, of course, is astonishingly infrequently.
[10:56.150 --> 10:59.350]  Almost never, maybe one in millions upon millions.
[11:00.990 --> 11:05.450]  So, if we build a model based on accuracy, what would be the optimal one?
[11:08.100 --> 11:17.530]  As an example, I'm going to take a less extreme case, and we'll say issues occur 2% of the time, such that I can visualize it here on the grid.
[11:19.430 --> 11:29.230]  Given the grid of truth versus the model's predictions for passengers with issues and those who are normal, we have four possible logical outcomes.
[11:29.690 --> 11:34.310]  The horizontal represents the true cases and the vertical the model's predictions.
[11:35.590 --> 11:39.190]  Here the diagonal represents the true positives and true negatives.
[11:39.650 --> 11:43.090]  The off-diagonal the false positives and false negatives.
[11:43.690 --> 11:49.110]  The accuracy, then, is the ratio of the true cases along the diagonal to the total.
[11:50.270 --> 11:56.730]  If our model were to only use accuracy, we could simply label everything as normal.
[11:57.830 --> 12:02.010]  And because of the data distribution, we'd be 98% accurate.
[12:02.770 --> 12:04.350]  So, let me emphasize that.
[12:05.250 --> 12:13.490]  We would need exactly zero parameters in our model to make a 98% accurate model.
[12:15.150 --> 12:16.990]  It's like we unlocked a cheat code.
[12:18.450 --> 12:25.510]  And in the cases where the data is unbalanced like this one, gradient descent will unlock that very cheat code.
[12:26.430 --> 12:31.070]  When you build a model, you have to be careful that it doesn't do that.
[12:33.230 --> 12:42.730]  So, at this point, we'll move on to consider how LLMs function and how the logical complexity of error and accuracy can affect them.
[12:46.230 --> 12:55.050]  LLMs simply predict what word or token, word part, is coming next, given the context of a prior sequence of tokens.
[12:55.630 --> 13:03.670]  The generative part is when the prediction token is added to the prior sequence that we used to predict it.
[13:04.050 --> 13:09.370]  The model applies this approach over and over again until it reaches some defined stopping point.
[13:10.130 --> 13:17.670]  From the animation, you can see that the LLM process begins with the transformation of those tokens into a numerical representation.
[13:18.070 --> 13:19.830]  That step is called the embedding.
[13:21.630 --> 13:31.210]  Then, using the transformer self-attention mechanism, it combines the information with the complete set of input words and their relative position.
[13:32.730 --> 13:38.330]  It then has a representation of the word and its context relative to the other words in its set.
[13:39.070 --> 13:47.650]  There are multiple self-attention mechanisms that are generated in parallel, which means that more than one aspect of the priors can be attended to.
[13:49.110 --> 13:58.590]  Note that the transformer model is able to represent the input word or token and its relative position, but not the time between those events.
[13:59.590 --> 14:11.550]  Therefore, this approach can represent a written word, but not the spoken word, nor something like aircraft flight telemetry or any other system where time is significant.
[14:12.430 --> 14:18.250]  An additional limitation comes from the attention mechanism itself, which is constrained by its size.
[14:19.450 --> 14:23.690]  Building a model to predict the next token is done with as much data as available.
[14:24.230 --> 14:29.250]  So basically, the last 25 years of the Internet is trained on our past.
[14:31.070 --> 14:32.910]  4chan, watch and all.
[14:34.130 --> 14:36.810]  At this stage, the model isn't particularly useful.
[14:37.050 --> 14:41.270]  It's a little feisty and reflects a past that we're all embarrassed by.
[14:43.250 --> 14:48.210]  So this culture of the past, not the current environment, is what we use.
[14:48.630 --> 14:51.750]  And it's certainly not a culture that we aspire to be.
[14:54.350 --> 14:59.270]  So, now we're going to take our petulant child of a model to a finishing school.
[15:00.630 --> 15:04.290]  So what we get is a model that acts more like an indulge.
[15:07.000 --> 15:11.060]  The finishing school is called Reinforcement Learning from Human Feedback.
[15:11.060 --> 15:16.460]  Here, humans provide the model an education by grading its responses.
[15:17.100 --> 15:23.800]  This feedback is used to retrain the model to output responses that will better generate rewards from the humans.
[15:24.600 --> 15:27.260]  The petulant child is trained to be nice.
[15:28.680 --> 15:33.580]  And not that here I'm using the archaic, ungendered, meaning of nice.
[15:33.800 --> 15:37.920]  That is, foolish, ignorant, frivolous and senseless.
[15:39.820 --> 15:45.660]  But we go one step further, and we ask whose idea of nice are we talking about?
[15:48.860 --> 15:56.060]  The finishing school is staffed by a group of people, each with their own individual worldviews, experiences and education.
[15:56.820 --> 16:01.520]  They are given directions on how to grade the model responses by their employers.
[16:02.000 --> 16:06.520]  The model is then retrained to generate responses that are more appealing to their staff.
[16:07.900 --> 16:10.300]  In my mind, this begs several questions.
[16:10.800 --> 16:12.680]  Who selects the model's teachers?
[16:13.780 --> 16:16.480]  Who decided and approved the grading system?
[16:17.220 --> 16:20.700]  The responses are appealing to whom exactly?
[16:21.680 --> 16:24.560]  And what does this training actually result in?
[16:26.180 --> 16:27.860]  We can all guess the answers.
[16:28.720 --> 16:30.860]  Nothing that will make the brand look bad.
[16:31.180 --> 16:33.680]  Nothing that will increase company liability.
[16:34.200 --> 16:36.500]  Nothing that could be interpreted as dangerous.
[16:37.080 --> 16:40.040]  And nothing that could be seen as culturally unacceptable.
[16:40.840 --> 16:42.960]  So we are going to get rid of most of the 4chan data.
[16:44.500 --> 16:50.660]  So the model is being trained not only to be nice, but also to co-toe the corporate lie.
[16:53.320 --> 16:56.800]  So in this regard, the word nice doesn't quite cover the scope.
[16:57.860 --> 17:02.260]  I am personally fond of the word glib, because I think it reflects the corporate slick.
[17:03.400 --> 17:06.800]  But in recent academic work, it's been described as bullshit.
[17:07.620 --> 17:11.460]  So perhaps a better description would be, it's glib bullshit.
[17:14.140 --> 17:19.360]  So now that we've reviewed error optimization of our LLM, let's consider the model itself.
[17:19.600 --> 17:22.580]  And how it differs from the world that we're trying to understand.
[17:23.440 --> 17:25.400]  Remember, the model isn't human.
[17:27.980 --> 17:32.720]  So we've all heard the term hallucinations as it relates to LLMs.
[17:33.360 --> 17:36.380]  Where they produce convincing but false responses.
[17:38.180 --> 17:40.300]  But why does an LLM do this?
[17:40.420 --> 17:42.560]  And why can't we just fix all of the problems?
[17:44.540 --> 17:51.840]  What is presented by the benign term hallucination is actually something intrinsic to the model and the data that it was trained on.
[17:53.000 --> 17:58.520]  In order to understand this phenomenon, let's turn to the study of adversarial examples.
[17:59.500 --> 18:05.060]  An adversarial example is one where a model produces an erroneous result.
[18:06.920 --> 18:08.820]  In quotes, erroneous.
[18:09.280 --> 18:14.060]  Given an input only fractionally different from another with the correct result.
[18:15.880 --> 18:19.080]  An example of an adversarial attack is shown by the figures.
[18:20.320 --> 18:23.320]  You can see the original image of the bus on the left.
[18:24.400 --> 18:26.620]  In the center, there's a noise-like pattern.
[18:27.220 --> 18:31.800]  And on the right is the mixture between the noise and the original bus image.
[18:32.900 --> 18:35.460]  We can all clearly see that it's still a bus.
[18:36.160 --> 18:40.580]  But in this case, the model, the AI, labeled this as an ostrich.
[18:42.180 --> 18:57.800]  In order to describe this phenomenon, all you need to do is recall what I just described was the model will mathematically always optimize itself to maximize accuracy above all else.
[18:59.580 --> 19:07.260]  As a result, the model will exploit the most effective predictive features from the data, whatever they may be.
[19:08.460 --> 19:14.960]  The data may contain features which, when viewed from a human-selected notion of similarity, are fragile.
[19:15.860 --> 19:17.400]  It sees things that we don't.
[19:18.380 --> 19:22.740]  For a classification task, this might mean that the bus is labeled as an ostrich.
[19:22.740 --> 19:26.900]  From the human perspective, it's comical, right?
[19:27.580 --> 19:31.640]  Yet, it demonstrates that the image contains features that we simply aren't aware of.
[19:34.520 --> 19:37.620]  I've got a reference here to Lias et al's paper.
[19:37.960 --> 19:47.020]  The implications of which are quite subtle, but have tremendous ramifications for the model-building process and everything you see in AI.
[19:48.160 --> 19:50.360]  And I'm going to quote the research paper here.
[19:51.400 --> 19:55.840]  Adversarial examples can be directly attributed to the presence of non-robust features.
[19:56.540 --> 20:04.100]  Features derived from patterns in the data distribution that are highly predictive, yet brittle and incomprehensible to humans.
[20:05.000 --> 20:12.500]  After capturing these features within a theoretical framework, we establish their widespread existence in standard datasets.
[20:14.120 --> 20:15.760]  Let me restate that point.
[20:15.760 --> 20:21.620]  The brittle, incomprehensible features are present in all datasets.
[20:23.000 --> 20:27.980]  So hallucinations are features that cannot be removed from an LLN.
[20:31.520 --> 20:35.840]  So sadly, what comes next is in shitification overdrive.
[20:37.260 --> 20:43.100]  Consider the media portrays the authoritative information as it comes from the big platforms.
[20:43.460 --> 20:50.020]  Voices from Google, Facebook, Microsoft, Twitter, OpenAI dominate the news cycle.
[20:50.360 --> 20:53.660]  And they maneuver themselves into government advisory committees.
[20:54.840 --> 20:59.480]  In this period of peak media hype, academics are a quiet voice.
[20:59.840 --> 21:04.500]  And in some important cases, those working for said corporations are silenced.
[21:05.700 --> 21:09.640]  Business leadership hears again and again that they need AI.
[21:10.280 --> 21:12.460]  Without it, they're going to fall behind their competition.
[21:13.560 --> 21:20.900]  We are all hearing that software developers need to learn AI prompting and quick before all our jobs are gone.
[21:21.680 --> 21:22.820]  I don't believe it.
[21:23.960 --> 21:27.600]  The models are expensive and biased and not fit for many purposes.
[21:28.640 --> 21:30.960]  But what they are is convincing at first glance.
[21:31.800 --> 21:37.860]  They amaze all that try them and they pass the sniff test for use in many applications.
[21:38.560 --> 21:40.880]  So share prices go up along with the hype.
[21:41.360 --> 21:42.960]  It's simply a gold rush.
[21:44.080 --> 21:51.120]  So the AI vendors pitch their products to executives and managers who are just as impressed by the magic tricks as we are.
[21:52.180 --> 21:58.300]  When they hear about the projected savings and increases in productivity, they think that they are viable.
[21:58.560 --> 21:59.680]  It looks viable.
[22:00.580 --> 22:03.840]  But once you scratch the surface, the grift shows itself.
[22:04.620 --> 22:08.160]  This is in shittification overdrive and the shit is bullshit.
[22:09.220 --> 22:20.020]  In a market where the resources required for entry are so high and the returns so incorporeal, platforms are abusing both users and business customers.
[22:21.160 --> 22:27.580]  We are all being sold on very expensive services that have ill-defined or non-existent value.
[22:28.040 --> 22:32.100]  We are buying a promise based on a few well-rehearsed demonstrations.
[22:33.040 --> 22:34.920]  So where does this cost stakers?
[22:35.420 --> 22:47.140]  If we follow it into the doldrums, a place with little money available for investment, little appetite for research, stagnated individual personal growth, and a zombified Internet.
[22:53.740 --> 22:58.000]  Year 2023 marked the widespread adoption of large language models.
[22:59.280 --> 23:01.820]  The LLMs are a synthetic voice.
[23:02.380 --> 23:04.440]  They are now prolific on the Internet.
[23:05.360 --> 23:09.420]  And that voice is not easily delimited from real human data.
[23:12.150 --> 23:19.690]  The scammers simply don't have to work so hard to create content, to build sock puppets, or bypass filters.
[23:20.710 --> 23:24.410]  Students don't have to struggle to understand the unending wit of Shakespeare.
[23:25.030 --> 23:28.090]  And meeting minutes don't have to be decomposed and considered.
[23:29.450 --> 23:34.630]  But this inhuman voice is a saccharine sweetness on the palate of the true human condition.
[23:35.610 --> 23:43.410]  And my allegory here is still made after the detonation of the first atomic bomb is contaminated with nuclear fallout.
[23:43.410 --> 23:50.450]  It's because of this contamination that Geiger counters and other sensitive scientific instruments can't be made of it.
[23:51.110 --> 23:58.230]  The Internet prior to 2023 is the last generally available uncontaminated research data for language.
[23:59.350 --> 24:06.790]  The contamination affects future models and not only those models in English.
[24:11.310 --> 24:18.230]  Training a model on data that contains a significant amount of synthetic data is subject to model collapse.
[24:18.650 --> 24:23.130]  This is where subsequent generations of models are trained on the results of the prior model.
[24:24.070 --> 24:27.590]  In every generation, the next model is more likely to collapse.
[24:28.610 --> 24:35.830]  Even in the cases where real data farmed the majority of the data present in the system, the models don't perform as well.
[24:36.810 --> 24:40.570]  As subsequent models are trained, the defects become more common.
[24:41.770 --> 24:49.590]  So, LLMs based on the current technology are eventually going to get worse, as is the Internet in general.
[24:53.500 --> 24:58.400]  So, I think the generative AI is just about past peak hype.
[24:58.900 --> 25:03.540]  There is an increasing body of research and evidence in the practical applications of LLMs.
[25:04.200 --> 25:07.960]  Of the many applications, few are actually financially viable.
[25:09.860 --> 25:13.760]  Economists now are predicting that negligible improvements in productivity...
[25:13.760 --> 25:19.300]  are predicting negligible improvements in productivity, along with negative social outcomes.
[25:20.020 --> 25:24.960]  And of course, we all know that the automated scams online are getting way, way better.
[25:26.600 --> 25:31.120]  So, where are the gains and what is the killer app that we see for LLMs right now?
[25:32.640 --> 25:36.240]  So, let's keep in mind that generative AI is a probabilistic system.
[25:36.640 --> 25:38.840]  And it's expensive to run.
[25:39.520 --> 25:43.640]  And then we can consider what industries it's most palatable to.
[25:46.930 --> 25:51.310]  So, on the vertical axis, we have risk.
[25:51.770 --> 25:53.970]  Higher meaning that we have a greater impact.
[25:54.650 --> 25:57.390]  High on that scale, we can place aviation and medicine.
[25:58.190 --> 26:00.090]  Here, if human lives are on the line.
[26:00.690 --> 26:04.030]  On the horizontal axis is value to our society.
[26:04.670 --> 26:09.170]  A greater economic impact places an industry further to the right.
[26:09.710 --> 26:12.630]  So, let's put aviation and medicine on the top right corner.
[26:14.050 --> 26:16.430]  Law on that scale will be entertaining.
[26:16.910 --> 26:18.670]  No one dies if the movie is bad.
[26:19.710 --> 26:22.830]  So, there's no question that it's a valuable industry.
[26:23.690 --> 26:26.190]  But low on risk, so we'll put it in the lower right.
[26:27.090 --> 26:31.890]  So, it makes perfect sense that OpenAI's app store is filled with AI girlfriends.
[26:33.390 --> 26:36.190]  It's their sweet spot of low risk and high value.
[26:36.950 --> 26:39.410]  Even though it's against their own terms and conditions.
[26:41.170 --> 26:46.730]  And we've got one application with low identifiable risk that we've had for some time.
[26:47.330 --> 26:49.190]  And it's familiar to many here.
[26:49.450 --> 26:50.810]  And that's code generation.
[26:52.710 --> 26:57.130]  Software development is a good example of a field where the results are bespoke.
[26:57.530 --> 26:58.390]  It's custom.
[26:58.550 --> 26:59.250]  It's made to measure.
[26:59.630 --> 27:02.770]  And the work requires creativity in application.
[27:03.870 --> 27:06.610]  The same is true in many fields, but this is the first one.
[27:07.810 --> 27:16.110]  In this domain, what separates juniors from seniors is a deeper meta-analysis of the problems and solutions.
[27:16.910 --> 27:21.430]  Fixing errors and architecting elegant solutions requires experience.
[27:24.080 --> 27:30.700]  But what happens when you provide neophytes with a boundless set of just slightly broken tool?
[27:32.720 --> 27:34.020]  They can build a car.
[27:34.620 --> 27:38.700]  Hit the gas and go from zero to 60 in two seconds flat.
[27:39.460 --> 27:44.420]  Only to realize that they forgot to build the brakes or any kind of safety features.
[27:46.180 --> 27:48.680]  So they're being compromised in the long term.
[27:51.040 --> 27:54.960]  Where we measure human endeavors by tasks and balance sheets.
[27:55.620 --> 27:58.200]  It all looks great for a while.
[28:00.500 --> 28:03.980]  Clearly then, the inshitification doesn't just affect economics.
[28:04.340 --> 28:05.720]  It affects culture.
[28:06.000 --> 28:08.260]  And not just the culture that created it.
[28:11.930 --> 28:15.610]  Generative AI predicts the next word that is most likely in the training data.
[28:16.350 --> 28:21.670]  But that data, both in content and construction, reflects the culture in which it was captured from.
[28:23.130 --> 28:32.230]  With the Internet being a U.S. military project, and the bulk of the data coming from Western nations, there are implicit assumptions and biases.
[28:33.630 --> 28:36.910]  These are potentially damaging to other cultures and populations.
[28:37.310 --> 28:43.210]  In recognition of this, the United Nations hosted a workshop resulted in the formation of the Indigenous Data Alliance.
[28:44.410 --> 28:48.510]  This organization promotes some simple principles for digital sovereignty.
[28:49.290 --> 28:55.530]  They state that data ecosystems should be designed to function in ways that enable people to benefit from the data.
[28:56.770 --> 29:00.250]  People's rights and interests in their data should be recognized.
[29:00.770 --> 29:03.350]  And they should have the authority to control that data.
[29:04.570 --> 29:09.530]  Those working with data should inform people on how that data will be used to support them.
[29:10.070 --> 29:16.330]  And people's rights should be the primary concern at all stages of the data's life cycle across the ecosystem.
[29:17.390 --> 29:23.430]  Their surprisingly simple and sensible rule, yet widely ignored.
[29:25.610 --> 29:29.010]  So, in summary, that brings us neatly to ethics.
[29:31.810 --> 29:36.050]  Human attractions are not a great fit for machine optimization.
[29:37.370 --> 29:39.270]  Why doesn't everybody see that?
[29:40.430 --> 29:47.730]  A viable explanation is that in a Western world, at least, we see life through the lens of a Protestant work ethic.
[29:48.330 --> 29:52.410]  In this view, life is focused around work and work around tasks.
[29:52.410 --> 30:01.590]  You'll hear phrases from economists like, We assume that human work can be decomposed into atomistic tasks that differ in their complexity.
[30:02.630 --> 30:15.970]  So, it's not surprising then that those in business and people with MBAs see AI as an opportunity to replace tasks and replace people, because fundamentally that's how we see life or they see life.
[30:17.770 --> 30:22.330]  But unlike during the Industrial Revolution, this approach can't work.
[30:22.730 --> 30:29.570]  Because our future progress relies on human understanding and abstractions, not on the mechanization of a process.
[30:32.030 --> 30:35.290]  But our predominant economic culture obscures this fact.
[30:36.470 --> 30:42.810]  Our corporate business approach isn't what's needed to advance humanity beyond industrialization.
[30:45.050 --> 30:48.090]  LLMs don't shift or enhance our worldview.
[30:49.370 --> 30:54.270]  Remember, they're optimized to give you the answers you want.
[30:55.630 --> 30:59.150]  They're optimized to keep you in your comfortable little bubble.
[31:01.330 --> 31:03.490]  So then how do we avoid the doldrums?
[31:05.810 --> 31:10.630]  If I've achieved much so far, you'll be asking yourself questions that cut through the hype.
[31:11.670 --> 31:14.710]  The hype is being driven by those in a position to benefit from it.
[31:15.450 --> 31:19.090]  We are centered on a culture of the Protestant work ethic.
[31:19.630 --> 31:23.490]  And corporations are trying to maintain their position and leverage it.
[31:23.890 --> 31:25.110]  They're trying to build them up.
[31:26.170 --> 31:29.270]  But there is no more, nor can there be in this case.
[31:30.350 --> 31:34.530]  The creative spark from LLMs came from Google's own research teams.
[31:35.510 --> 31:40.890]  The transformer model that underpins LLMs was published to an academic journal.
[31:42.070 --> 31:45.290]  And that was counter to Google's corporate interest.
[31:46.090 --> 31:51.590]  Why will Google or any other corporation allow such a valuable idea to become public?
[31:55.450 --> 31:56.910]  In one word, culture.
[31:57.630 --> 32:01.690]  If you want original research, you have to engage creative people.
[32:02.210 --> 32:05.730]  And accept they all have a deep desire to share their passion.
[32:06.030 --> 32:10.570]  They go to conferences to tell everybody who will listen what they're excited about.
[32:13.500 --> 32:17.640]  And that they all want to share their victories so that all others may benefit.
[32:17.920 --> 32:20.500]  They think it's cool to forward human knowledge.
[32:21.080 --> 32:23.520]  They like to have multiple ideas that they're working on.
[32:23.900 --> 32:26.580]  Knowing full well that they'll never finish some of them.
[32:27.000 --> 32:31.840]  They like to measure themselves by the books that they haven't quite gotten around to reading just yet.
[32:33.040 --> 32:37.700]  And none of this works if we assume that humans can be decomposed into atomistic tasks.
[32:37.700 --> 32:40.620]  None of this works with tickets and deadlines.
[32:41.600 --> 32:45.220]  So we could imagine a conversation between a project manager and a researcher.
[32:45.640 --> 32:51.560]  If our project manager said, Well, when will you be finished with this groundbreaking approach to text prediction?
[32:52.580 --> 32:57.740]  Our researcher might say, Well, I'll have a right good think about it this sprint.
[32:59.220 --> 33:06.400]  And after a couple of weeks, our project manager comes back and says, Do you have that text prediction approach completed yet?
[33:07.200 --> 33:10.500]  And our researcher would be, No, not even close.
[33:10.960 --> 33:13.540]  But I did explore some really interesting stuff.
[33:15.580 --> 33:18.340]  And I want you to take careful note of what I say.
[33:19.240 --> 33:21.100]  Explore some really interesting stuff.
[33:22.260 --> 33:28.420]  Exploration requires open access and an assurance that published materials could be of interest.
[33:28.880 --> 33:34.920]  The word interest means that it has some substance, which requires a public and professional scrutiny.
[33:35.320 --> 33:40.160]  It's that openness of the academic process that allows anybody to be a knower.
[33:41.680 --> 33:44.120]  Anyone can publish to an academic journal.
[33:45.080 --> 33:49.240]  They have a blind review process where the viewers don't know who the authors are.
[33:50.740 --> 33:54.120]  And it's at this point we get into some ethics and epistemology.
[33:54.280 --> 33:55.720]  So let's be specific for a second.
[33:57.340 --> 34:02.640]  I'm making the claim that there's little difference between an academic researcher and a hacker.
[34:02.940 --> 34:04.340]  Except perhaps getting paid.
[34:06.760 --> 34:11.720]  The dominant commercial culture that leads us towards the doldrums can be characterized by a focus on work.
[34:12.460 --> 34:13.080]  Money.
[34:14.000 --> 34:14.620]  Optimization.
[34:14.800 --> 34:15.340]  Agility.
[34:15.540 --> 34:15.980]  Stability.
[34:16.440 --> 34:17.060]  Determinancy.
[34:17.180 --> 34:17.780]  And culpability.
[34:20.480 --> 34:25.820]  You'll see the same sentiment in all get-ahead-in-business books.
[34:25.980 --> 34:28.400]  It's the cultural approach that brought us to the glib.
[34:28.540 --> 34:28.920]  Bullshit.
[34:29.280 --> 34:34.580]  The post-industrial world of AI is in need for some creative maladjustment.
[34:35.300 --> 34:40.100]  And if we want something other than a shittification, we need to embrace a hacker culture.
[34:40.780 --> 34:42.380]  So what is the hacker ethic?
[34:42.860 --> 34:46.660]  I'm taking the description from Pekka Himinen's book from 2001.
[34:47.180 --> 34:51.420]  In this book, Himinen takes us on a journey starting from the Protestant work ethic.
[34:52.140 --> 34:52.700]  Calvinism.
[34:52.860 --> 34:53.580]  The hegemony.
[34:53.720 --> 34:55.340]  And the emergence of the hacker culture.
[34:56.120 --> 35:00.800]  So let's enumerate Himinen's characteristics of the hacker ethic.
[35:02.260 --> 35:02.820]  Passion.
[35:03.100 --> 35:05.820]  This is where we take joy in the things that we do.
[35:05.980 --> 35:07.340]  We do things because it's fun.
[35:08.600 --> 35:09.160]  Freedom.
[35:09.480 --> 35:11.480]  The creative work melds with our passion.
[35:11.940 --> 35:14.020]  And it happens in its own time.
[35:14.300 --> 35:16.680]  One does not simply create on demand.
[35:18.140 --> 35:19.260]  Social worth.
[35:19.620 --> 35:22.920]  Money isn't the primary motivation in and of itself.
[35:23.640 --> 35:29.680]  Sharing ideas with the community and others is creating value that makes our primary motivation.
[35:30.180 --> 35:31.100]  That's why I'm here.
[35:32.100 --> 35:32.600]  Openness.
[35:32.740 --> 35:33.980]  Anyone may be a knower.
[35:34.420 --> 35:37.360]  Anyone may gain the respect and adulation of the community.
[35:43.520 --> 35:44.080]  Activity.
[35:44.420 --> 35:52.880]  Our expression in action meaning that people are free to express themselves in their actions and have a privacy necessary to protect an active culture.
[35:55.020 --> 35:55.580]  Caring.
[35:55.900 --> 35:58.200]  Is a focus on the long-term benefits to society.
[35:58.220 --> 36:00.660]  With particular attention to those who are marginalized.
[36:01.660 --> 36:02.640]  And finally, creativity.
[36:03.020 --> 36:05.600]  Is to surprise yourself with your own creations.
[36:07.840 --> 36:12.660]  And if we apply this hacker worldview to the things we discussed, things don't quite sit right.
[36:16.010 --> 36:20.070]  AI is being deployed widely in our daily lives.
[36:20.590 --> 36:24.910]  Few of those outside academia have the language to describe their experience.
[36:25.570 --> 36:29.790]  We're being given benign terms like hallucinations.
[36:30.070 --> 36:33.130]  When the truth is that this is being deliberately...
[36:33.130 --> 36:37.430]  We're being deliberately denied the knowledge with which to describe our experience.
[36:39.310 --> 36:42.630]  And those who do speak to inform face the wrath of employers.
[36:43.050 --> 36:45.290]  Or they're just muted out of the conversation.
[36:46.710 --> 36:51.610]  As we've seen, generative AI provides an additional barrier to becoming a knower.
[36:51.810 --> 36:54.930]  And it has the potential to create further injustice.
[37:01.160 --> 37:04.120]  With generative AI, people can build things that they don't know.
[37:04.120 --> 37:07.840]  How to fix our worst cannot be fixed.
[37:09.400 --> 37:11.760]  But failure is an essential part of learning.
[37:12.000 --> 37:14.360]  Especially where we're pushing in the boundaries of human knowledge.
[37:15.180 --> 37:18.020]  I would argue that creativity requires that we fail.
[37:18.400 --> 37:21.820]  Where AI reduces the little failures to build our understanding.
[37:22.120 --> 37:24.180]  We're failing in the uncool way.
[37:27.350 --> 37:29.050]  But is there a better course?
[37:30.230 --> 37:32.370]  And this is really my last side.
[37:32.550 --> 37:33.290]  So this should be okay.
[37:34.350 --> 37:35.270]  Remember that...
[37:36.030 --> 37:36.590]  Economist...
[37:36.590 --> 37:40.050]  Assume that worm and work can be broken down into atomistic tasks.
[37:40.410 --> 37:41.690]  This really doesn't work.
[37:43.710 --> 37:47.910]  So what we need to do is create a framework by which we can...
[37:47.910 --> 37:50.250]  Build our understanding moving forward.
[37:50.490 --> 37:55.410]  And leverage AI as a tool to inform and change our worldview.
[38:00.060 --> 38:02.920]  And that's why I end a little word.
[38:03.220 --> 38:03.580]  Thank you.
[38:05.880 --> 38:06.520]  Bye.