82 minutes
SDS 833: The 10 Reasons AI Projects Fail, with Dr. Martin Goodson
Subscribe on Website, Apple Podcasts, Spotify, Stitcher Radio or TuneIn
Martin Goodson speaks to Jon Krohn about what he would add to his viral article “Ten Ways Your Data Project is Going to Fail”, why practitioners always need to be present at AI policy discussions, and Evolution AI’s breakthroughs in computer vision and NLP.
About Martin Goodson
Martin Goodson has worked in data science and AI for over 20 years. He is the Chief Scientist and CEO of Evolution AI, a multiple award-winning data extraction firm. Artificial intelligence-powered products he developed are in use at companies such as Time Inc, Staples, John Lewis, Top Shop, Conde Nast, New York Times, Buzzfeed, etc. He is also the former Chair of the Royal Statistical Society Data Science Section, the professional body for data science in the UK. He also organises the largest community of AI and machine learning practitioners in Europe: the London Machine Learning Meetup.
Overview
Martin Goodson speaks to Jon Krohn about what he would add to his viral article “Ten Ways Your Data Project is Going to Fail”, why practitioners always need to be present at AI policy discussions, and Evolution AI’s breakthroughs in computer vision and NLP.
In 2016, Martin Goodson wrote “Ten Ways Your Data Project is Going to Fail” to hit back at his boss at the time, who took erratic and illogical approaches to devising, developing and delivering their data projects. The article struck a chord with many data scientists around the globe who felt that they were in similar situations, and Martin’s emphasis on thinking about data science as a highly technical discipline that needs specialists, not buzzwords, is key to ensuring a project’s success. Now, eight years later, Martin believes that the points in his article still stand, alluding to the recent hype surrounding AI and its applications. He believes that the visibility of AI and LLMs entering everyday use have led to “over-inflated expectations of what’s really possible” [51:23]. He calls out LLM use in particular, which, in his view, is “really good at fooling people” who have little background knowledge of how these models actually function.
At Evolution AI, Martin and his team work on improving automation for the finance and banking sectors. Over the years, they have helped companies from Deutsche Bank and NatWest to agile FinTechs. In the past, much financial documentation was procured and managed manually. Since 2015, Evolution AI has been automating these processes. Accuracy is of utmost importance to the company, and making sure that machine learning algorithms understand the variation between documents, such as invoices, so they can be read and processed without causing delays or errors. For Martin, understanding spatial reasoning is key to developing an accurate algorithm. He says that when we read documents, we are “using the characters in context” [11:28], and context is essential to understanding the document.
Martin notes the limits to maintaining accuracy with automation alone, saying that ultimately, it will be necessary to employ humans to proofread the results, a standard operating procedure when extracting financial data that must, above all, be accurate. While Evolution AI is developing well-calibrated confidence scores for their data to manage large infrastructures where it would not be possible for a human to proof every single data point, he acknowledges that “sometimes, we need to use human beings” [13:39]. He also says that hallucinations remain a problem and that Evolution AI is implementing clear strategies that help mitigate those risks.
Listen to the episode to hear how Evolution AI mitigates AI hallucinations, where Martin stands on AI company leadership styles and his views on the UK’s global ranking in science and technology.
In this episode you will learn:
- What Evolution AI does [04:25]
- How to maintain accuracy in large infrastructures [11:41]
- How to cultivate innovation and creativity while meeting market demands [21:22]
- Potential knowledge gaps for machine learning practitioners [24:27]
- Martin’s viral article, “Ten Ways Your Data Project is Going to Fail” [30:57]
- Strategies for the UK to become a key player in AI [59:54]
Items mentioned in this podcast:
- This episode is brought to you by Keith McCormick (follow #SDSKeith)
- Evolution AI
- London Machine Learning Meetup
- NeurIPS
- “Sparse Instrumental Variables (SPIV) for Genome-Wide Studies” poster by Felix V Agakov, Paul McKeigue, Jon Krohn, Amos Storkey
- “Ten Ways Your Data Project is Going to Fail” by Martin Goodson
- An AI for Humanity, Martin’s talk to the European Commission in Brussels in 2023
- Fireside Chat with Andrew Ng
- Google Gemini
- AlphaGo (film)
- Jupyter Notebook
- The Human Genome Project
- What’s Next
- SDS 547: How Genes Influence Behavior — with Prof. Jonathan Flint
- SDS 781: Ensuring Successful Enterprise AI Deployments, with Sol Rashidi
- Vision and Brain: How we perceive the world, by James Stone
- The Evolution of Cognition
- ScaleUp:AI
- Web Summit
- Jon Krohn’s Generative AI with Large Language Models, hands-on training
- The Super Data Science Podcast Team
Follow Martin:
Podcast Transcript
Jon Krohn: 00:00:00
This is episode number 833 with Dr. Martin Goodson, CEO of Evolution AI. Today's episode is brought to you by epic LinkedIn Learning instructor Keith McCormick.
00:00:17
Welcome to the Super Data Science podcast, the most listened to podcast in the data science industry. Each week we bring you inspiring people and ideas to help you build a successful career in data science. I'm your host, Jon Krohn. Thanks for joining me today, and now let's make the complex simple.
00:00:36
Welcome back to the Super Data Science podcast. I think you're really going to enjoy the conversation today with my guest, Dr. Martin Goodson. Martin is CEO and chief scientist at Evolution AI, a firm that uses generative AI to extract information from millions of documents a day for their clients. He's also the founder and organizer of the London Machine Learning Meetup, which with over 15,000 members is the largest community of AI ML experts in Europe. He previously led data science at startups that apply ML to billions of data points daily, and he was a statistical geneticist at the University of Oxford where we shared a small office together. Today's episode will be of interest to anyone even vaguely interested in data science, machine learning or AI. In today's episode, Martin details the 10 reasons why data science projects fail and how to avoid these common pitfalls. He provides his insights on building AI startups that serve large enterprises and the importance of open source AI development. All right, you ready for this fun episode? Let's go.
00:01:48
Martin Goodson, welcome to the Super Data Science podcast. Where are you calling in from?
Martin Goodson: 00:01:57
I'm calling from Oxfordshire.
Jon Krohn: 00:02:00
Very nice. That is where you and I know each other from. We used to share a little office in Oxford.
Martin Goodson: 00:02:06
Very small, very small office.
Jon Krohn: 00:02:07
It was, but it was nice. And if I remember correctly, a lot of the time we were in there we had four desks but only three of us in there, so that felt pretty luxurious.
Martin Goodson: 00:02:15
We were very lucky.
Jon Krohn: 00:02:16
Do you ever see Jérôme Nicot the third occupant of that office?
Martin Goodson: 00:02:21
I don't, unfortunately. I wish I did.
Jon Krohn: 00:02:22
Maybe this will spurn a reach out.
Martin Goodson: 00:02:25
[inaudible 00:02:27]
Jon Krohn: 00:02:27
I'm pretty sure. He could be. He could be.
Martin Goodson: 00:02:27
Hey Jerome.
Jon Krohn: 00:02:31
All of the geneticists that cannot resist listening to this Super Data Science podcast. I mean he does sit at the intersection of data analysis I guess, so it's great to have you on the show. Only Martin would know this, but years ago I asked Martin to be on the show when I first took over as host of this podcast and you say that you remember the conversation. Well, the way I remember it was, "No. Not going to do it."
Martin Goodson: 00:02:57
I remember very well. I said, "I don't have anything to say." I still don't think that-
Jon Krohn: 00:03:07
Yeah, that's it. And now our listeners are going to get to experience that for an hour. That was the plan on the show. Yeah, no, it's great to be connected with you. We had a lot of laughs before we even started recording, which is really nice. You are one of those people that I really miss being around you all the time. You're brilliant. You do always have a lot of thoughtful things to say. I also really miss the way you would sit in meetings. You would often sit on a desk chair cross-legged and with a very upright posture you looked kind of like an all-knowing Buddha in the corner of the room.
Martin Goodson: 00:03:45
I didn't know that.
Jon Krohn: 00:03:49
It adds to your gravitas.
Martin Goodson: 00:03:52
Thanks just telling me that.
Jon Krohn: 00:03:53
It's probably things like that that led to you becoming CEO and Chief Data Scientist at Evolution AI, which is according to our research, it's a generative AI powered data extraction platform specializing in financial documents. That's the kind of thing that wasn't obvious to me when I was just looking at your LinkedIn profile, but according to our researcher Masis, you specialize particularly in financial documents and you surpass traditional OCR, optical character recognition capabilities. Do you want to tell us about the advancements that you and your team have made in computer vision and NLP that allow your AI models to achieve human-like accuracy in data extraction?
Martin Goodson: 00:04:36
Yeah, yeah, sure. So what we do, the data extraction thing is customers, they have some kind of documentation. So typically it might be a commercial lender like a bank, someone who lends money to businesses, normally SMEs. They've got loads of documentation that they need to read, transfer some data from, and then run the data through their risk models, credit risk models, credit decision engines, whatever, to decide whether they want to lend money to an organization, to another business normally. So that kind of documentation might be a financial statement, it might be a bank statement, it might be an invoice, maybe they're doing some kind of asset financing, they need to look at the invoice. Traditionally all of that stuff has been done manually. People have literally copied and pasted from documents or just typed in stuff into Excel, which is obviously a huge waste of time. So we've been automating this process for lots of banks and other financial institutions since about 2015.
00:05:35
So yeah, in terms of what the difference is between traditional OCR, OCR has been around since the '70s. Optical character recognition, it just extracts characters, optical character recognition. It is really, really good at extracting characters. If you've got something that looks like a check where you know exactly where the check number is on every check because all checks look the same, it works great. You don't need to bother with anything else, but most documents are not really like a check. If you look at an invoice, there's loads of variation. The invoice number could be anywhere, you need to really just look at it. A human would just read the document, figure out where the invoice number was, and then use that information. So we've worked hard to try and design machine learning algorithms that can go through that kind of process, a similar kind of process.
Jon Krohn: 00:06:18
Now, I know that from other research that we've done that like many people working in technology, you prefer not to use patents because patents then actually kind of disclose your secrets on how you're doing things, and because it would be so difficult to know if somebody has read your patent and is implementing your technology behind the scene. So I totally understand that, and given that there's probably limitations on what you want to tell us on air, but if you've been doing this for nearly a decade, I can't help but imagine that the approaches that you're using have presumably evolved a lot at Evolution AI. Is there anything that you're able to tell us in terms of behind the scenes capabilities, technologies you're using and how that's changed over time?
Martin Goodson: 00:07:01
Yeah, yeah, I can speak pretty freely about that actually. So I think there've been three main phases. There's kind of traditional OCR, traditional machine learning phase for this stuff where people were using things like hidden Markov models to understand text and maybe some neural networks as well, like classical neural networks. Then there's a deep learning world, and that's kind of where we entered into the field. I went to what was called NIPS now called NeurIPS conference in about 2011 maybe while we were still working together actually, and I went to a deep learning workshop. There's like 30 people in this workshop, but I didn't know what this thing was about, but I learned all about convolutional neural networks and stuff. Now obviously NeurIPS is massive and 10 times bigger than it was in those days.
Jon Krohn: 00:07:58
I had a NeurIPS paper in 2010, I don't know if you remember this. I had a collaboration at the University of Edinburgh and there was a guy, Felix Agakov who was, he was the first author on the paper and-
Martin Goodson: 00:08:11
I remember, yeah.
Jon Krohn: 00:08:11
Yeah, he went to NeurIPS in 2010 and it ended up being only the top 10% of papers ended up being selected for the proceedings. We managed to clear that threshold, and I still have not been, but today I bought my ticket to go this year for the first time finally.
Martin Goodson: 00:08:26
Oh, wow.
Jon Krohn: 00:08:26
It's one of those huge gaps where I'm like, "What the hell have I been thinking?" It's exactly because of the things that you're describing. In 2011, you had somebody talking about convolutional neural networks and you'd never heard of them before and that kind of information. I'm like, "Why didn't I go in 2010? What was I thinking?" Instead, I went to the complex Trade Consortium in Chicago.
Martin Goodson: 00:08:49
So useful. Yeah, really useful knowledge. But no, you're going to have a great time. I'm sure you're going to have a great time. Obviously it's a very different thing now to what it was then, but it's amazing. It's more amazing now than it was, I think. Anyway, so I did this thing and I came back and these kind of ideas kind of percolated and I'd worked in a different startup where we tried to extract information from tax documents. We were doing an automated tax thing, but that startup completely failed because we just couldn't extract the data accurately using traditional OCR. And then I realized that convolutional neural networks would be a good technology for this thing, so that's why we started the company. I started it with a friend. So originally your question is about how have we changed the technology? So originally it was all about convolutional neural networks and deep learning, Haskell deep learning, LSDMs and stuff like that. And then now gradually everything's moved over to transformer-based ways of working. I think basically everything is a generative model now.
Jon Krohn: 00:09:48
That is not surprising at all then, so it sounds like there is a lot more happening than... Because of maybe the OCR starting point and also the way that you've mentioned online that you're looking for the way a document is structured and which parts of the document are more important. It sounds like you're doing a lot more than with a typical kind of transformer approach today. You might use something like the PDF to text utility in Unix to convert a PDF into a text file, and then you could just pass that text file into a transformer architecture. You could use an off-the-shelf, OpenAI API or Cohere or whatever to be doing any kind of document processing in that way, and so it sounds like you are doing something more sophisticated where maybe the entry point is pixels as opposed to characters.
Martin Goodson: 00:10:46
Yeah, I mean the approach that you just said, it actually works really well. It's going to give you some really good results, but probably not good enough to use in some automated process where you're trying to lend money to tens of millions or hundreds of millions of dollars or pounds to other companies. You're like, "You need something that's not going to have so many errors."
Jon Krohn: 00:11:05
You don't want the page number in the bottom right corner of the page being inserted and adding a zero into the-
Martin Goodson: 00:11:12
Exactly. That would [inaudible 00:11:14]. So you need something with a bit higher accuracy. So it's really about spatial reasoning. When we read a document, when humans read a document, you're not just transcribing all of the characters and then figuring out what it means. You are using the characters in context and you use spatial relationships to try and understand the document.
Jon Krohn: 00:11:33
Nice. Very cool. So that gives us a sense in terms of a single read of a document, but you guys process over a million pages of documents a day. What's that like? Can you tell us a bit about the infrastructure of the challenges of scaling to that high volume while we're trying to retain accuracy? And maybe if you can tell us about some of the specific technologies that you use in order to be able to do that.
Martin Goodson: 00:11:59
Yeah, I mean I don't think we do anything particularly clever in terms of the engineering, but in terms of scaling up, the most important thing is that typically in this industry, if you want really accurate data, you're going to use some automated method and then you're going to employ some human beings to check that data, and that really is still the kind of standard operating procedure when you're trying to extract data that needs to be of high quality. The problem is when you've got... Some of our projects are huge, like you said, there might be 200,000 pages a day for one project. You can't really employ people to check that data, it is impossible. So either you use no humans at all, in which case you need a model which is really, really accurate, or you've got some really, really very, very good well calibrated confidence score that you can use to only allocate your meager human resources to those cases where you think it's a dubious prediction, it needs to be checked by a human. Obviously you need the vast majority of the documents, the pages to go through automatically.
Jon Krohn: 00:13:05
Okay. Yeah, those are definitely two options. Are you going to tell us which one you do?
Martin Goodson: 00:13:11
Well, we do both.
Jon Krohn: 00:13:13
Oh, you do both.
Martin Goodson: 00:13:15
We do both. I mean, we've got loads of different projects going, so in some projects we just have really very accurate models. We've invested a lot in training data and in algorithm development, and we don't really have any human component in them at all, but others, we just don't have algorithms that are that accurate, frankly. It's hard and sometimes we need to use human beings, so we do employ humans to check data.
Jon Krohn: 00:13:45
Now that you're in a transformer world and using generative AI models, something that people gripe about a lot, although a lot less than they did a year or two years ago, is hallucinations. And according to our research we pulled out that you at Evolution AI, you implement specific strategies for mitigating and managing the risks of hallucinations in your AI outputs.
Martin Goodson: 00:14:06
Yeah, I think it's definitely a huge problem, and one thing we're seeing is when you're trying to extract information from let's say a financial statement, like a profit and loss statement, the problem is that you're going to extract some information and the LLM is just going to make up a line item. It's going to completely make up this line item, look at appellate sheets, it's going to say current assets is X millions and that number is just going to be completely made up. The problem is that that number is going to make sense in context, typically. It is going to be some other numbers added together, it's going to make sense, but it really matters that it hasn't actually been reported. These are kind of official documents. You can't just make stuff up if it hasn't been reported you shouldn't really be saying that you've extracted this information.
00:14:55
So it is a real problem and we can talk, we could maybe get into that, why the limitations of LLMs in this world, but also why they're so powerful in this world for what we do. But I think the main thing is that you really do need some domain-specific knowledge at the moment to sort this out. Obviously, nobody really has solved this issue of hallucinations, and we haven't either, but I think we have some stuff specifically to do with, for example, financial statements, some tests that we can do. And some other technology that we've built on top of LLMs to, for instance, give you a good sense of confidence, whether the result was accurate or whether it's just been hallucinated.
Jon Krohn: 00:15:46
Keith McCormick, the data scientist, LinkedIn learning author, and many-time guest on this podcast, most recently in Episode #828, Keith will be sharing his “Executive Guide to Human-in-the-Loop Machine Learning and Data Annotation” course this week. In this course, Keith presents a high-level intro to what human-in-the-loop ML is, which will be intriguing even for consumers of AI products. He also introduces why data professionals and need to understand this topic for their own AI projects, even if they delegate data annotation to external companies.You can access the new course by following the hashtag #SDSKeith on LinkedIn. Keith McCormick will share a link today, on this episode's release, to allow you to watch the full new course for free. Thank you, Keith!
00:16:36
Yeah, it sounds like that's a key part to your proprietary tech stack are these models that are assigning confidence and allowing you depending on the project to bring in, as you say, some meager human resources to double-check things, and so getting that calibrated well is critical. Obviously you can't, you're like, "Okay, we promised the client that we would get through 200,000 documents today and our model found that just 2000 documents have something suspicious about them." It's only at a 10th of 1%, but obviously it ends up being a huge deal.
Martin Goodson: 00:17:13
Yeah, I think it's kind of interesting to think about what's the role of the startup here now? When ChatGPT came out, it was a kind of panic. There's no role for startups anymore. AI startups, they don't do anything. You just need to use a LLM to do everything, and I think we're starting to see that actually there is a lot of domain-specific knowledge that's really valuable. That's where the startups can use some open source LLM or some commercial LLM and add a lot of value.
Jon Krohn: 00:17:41
For sure. I'm a big proponent of there being a lot of opportunity in startups at the application layer, like you're saying with some domain-specific expertise, building a verticalized solution, the biggest players are going to be occupied for some time with just building out the core capabilities, the LLMs that are underneath the hood that you can be leveraging. And it's great for startups that companies like Meta are open sourcing such huge powerful models that we can use, fine-tune very cost-effectively with approaches like LoRA that allow us to (low-rank adaptation for our listeners who aren't familiar with that term), that allow us to add a very small number, like single-digit percentage model weights into a huge LLM and fine-tune it to our specific use case for your specific application, your client-specific needs. Yeah, it's a really exciting time to be building application-layer startups. Your company evolution AI is used by major companies across a range of sectors, Time Incorporated, Staples, Condé Nast, The New York Times. Is there anything that you're able to tell us around how you tailor your solutions to meet demands across diverse industries?
Martin Goodson: 00:19:00
So I think that list of names maybe your research have found out, that's basically a list of different companies that have used various technologies that I've built in various other companies, so not just our company. For us specifically, we are mainly in financial services, so we are a lot like Deutsche Bank, NatWest, some smaller banks as well, mainly banks. And so the question is how do we work across such diverse... I mean really, I don't think we do work across the [inaudible 00:19:38]
Jon Krohn: 00:19:37
Right. And so I guess then the better question then is how do you tailor... I mean, so even if you're talking to NatWest or Deutsche Bank, they have different needs depending on different projects, different scales that they're working at. Yeah, I don't know if there's anything interesting to say.
Martin Goodson: 00:19:54
There is something interesting to say, which is that we just don't really work so much with those larger banks anymore. We like to work with smaller banks and fintechs because it's such a hugely painful process. You're absolutely right. How can you work with the smaller companies and the really big companies? I don't think you can really do that, both. At least we can't. They work at a different timescale, but everything's on a different scale, and so we gradually have moved from the bigger banks to smaller fintechs. Actually, we love working with fintechs.
Jon Krohn: 00:20:29
I really appreciate you saying that on air. I think that that might be the kind of thing that... That feels like a really honest insight that will be really valuable to a lot of our listeners. I'm sure we'll be making that into a YouTube short.
Martin Goodson: 00:20:39
I'm probably just losing loads of customers by saying that.
Jon Krohn: 00:20:44
Yeah, our Deutsche Bank listeners are like, "I'm out of here." Nice. Speaking of people being out of here, you've previously touched on the balance between being a dictator and a democrat in leadership. So you've said that on air before in high stakes environments like an AI startup where one big Deutsche Bank client walks out the door and all of a sudden it's all hands on deck trying to figure out how we're going to make revenue come together, not avoid some kind of fiscal cliff that's coming up. It can be really intense. So how do you cultivate a culture that encourages innovation and creativity while still maintaining decisive leadership to meet market demands?
Martin Goodson: 00:21:30
Yeah, well, I definitely haven't solved that. They did say, in fact, I asked Andrew Ng that question when I interviewed him, who I'm sure your listeners know about, and he gave me a really great answer that I wasn't expecting, which is that there's no real answer to this question about whether you should be a dictator or a democrat. You really need to think and be introspective about what your experience allows you to be confident about. If you have really spent many years thinking about something, then you really should be confident and you really do need to say, I think we should go this way and please come with me.
00:22:10
I think he used the word, he said something like, "Just give me a chance on this," or something like this. "Take a gamble with me," I think he said. And I've had this myself. There's lots of stuff that I just don't know anything about really, and it's really important, so just be really, really self-aware that there's so much stuff and the field's moving so quickly that there's so much stuff. You basically don't know much about anything. There's a few things that you do know about. I think I'm quite strong on certain things. I think I'm quite strong on core scientific method, data analysis, experimentation, design of experiments and stuff like that. Quite classic stuff.
Jon Krohn: 00:22:51
I remember when we were working in the same office, you were a researcher, so you were past postdoc, you were doing research, but you undertook, you audited or maybe it must not have even been audited. I mean you would've remember better than me, but you did a graduate level mathematics course at Oxford while you were working full-time as a researcher and if I remember correctly, you got the top grade.
Martin Goodson: 00:23:16
I got the second top grade.
Jon Krohn: 00:23:19
The second top. That actually, I had a few neurons fired that were second top, but I didn't want to guess that and be wrong.
Martin Goodson: 00:23:25
Yeah, rather be wrong than the other way. Yeah, yeah. Okay. Second top. Yeah. Yeah, so I did the statistics course. It was like the applied statistics to Oxford, but I didn't do it. I just sat the exams and stuff, but I would not officially... Yeah, that was fun. So I learned a lot doing that. I really enjoyed that actually. So I feel like I'm pretty strong with that kind of basic stuff. I'm not really super strong on anything else, but stuff like that I am. But then I do think that a lot of machine learning research does come down to understanding data, understanding statistics, just understanding the way that data can screw you over and confuse you. I've been confused so much by data probably with you many times in the past, and it's just really stayed with me to my core.
Jon Krohn: 00:24:09
I remember when you're the first person who ever told me that I should learn Python. That was a good tip.
Martin Goodson: 00:24:14
That was a good tip, right.
Jon Krohn: 00:24:18
I was doing everything in R at the time and you're like, "What are you doing? Everything's going to Python." And yeah, you're the first person that I was like, "Man, I guess I should learn Python." In terms of statistics, do you think that there's a big gap amongst a lot of machine learning practitioners that don't study statistics? I think that there's a lot of value there. I mean, I have a lot of confirmation bias on this because I was a statistician before getting into machine learning. And so I like to always think that there's a lot of value in that, but I think in terms of understanding your data and especially cleaning up your data, it can be hugely valuable.
Martin Goodson: 00:24:50
I mean, there's a lot of stuff that is in statistics and that I've learned in statistics that is just useless now and it's a waste of time. It's quite a traditional field and it hasn't moved very quickly and people are still being taught stuff. Well, they definitely were then being taught stuff that is not really that useful. There should be much more emphasis on computational methods, I think. But having said that, I mean there's a lot of stuff that you can just safely ignore in statistics.
Jon Krohn: 00:25:20
P value.
Martin Goodson: 00:25:21
Yeah, P values, the delta method, they've quite a lot of mathematically intensive stuff. You just don't really need any of that, but you do need the approach and the kind of attitude and the skepticism about data and really understanding bias, really at your core understanding bias, selection bias and all of that stuff, that statistics has got really important lessons to teach about that. That's really, really critical. Most of the other stuff, you don't really need any of the techniques. Like the techniques, do you really need any of that stuff? You need to know linear regression, you could do a lot with linear regression. I don't think you really need much more. I don't know, that's a bit of a blanket statement maybe.
Jon Krohn: 00:26:05
I think something that for me has come in handy is actually interesting for things like evaluating, not statistical bias, but unwanted bias, so your model's behaving in a way that, so for example, at our company at Nebula, we're doing a lot of things in human resources and so for example, we would be ranking people for a specific job description. We have database of 180 million people in the US and somebody puts in a job description. We rank everyone in the US for that role. We want to be sure that it's not ranking men better than women, for example. And so we have a test data set that we can test our algorithm on and obviously there's going to be some variation. The mean score on men is going to be somehow different on women. You're never going to get exactly the same number, and so statistics has been useful for me there to be able to say, "Okay, but there's no statistically significant difference between these two groups." I don't know.
Martin Goodson: 00:27:08
Yeah, no, absolutely. I think the concept of statistical noise is just a really deep concept and it needs to be really, really understood, and I don't think it really... Some of these basic concepts in statistics, they are just not that widely understood by people that come really from a computerized background. And also stuff to do with data visualization. Like today we're showing someone how to make a normalized histogram, and it's quite typical that I'm teaching someone how to use a scatterplot. People who are really, really good, they got great background, they come from computer science background, but they don't really know how to do relatively straightforward things in data analysis, like make a good scatterplot or make a good histogram or whatever. It's kind of boring this stuff, but it's also really, really important.
Jon Krohn: 00:27:57
Yeah, it is also the kind of thing that we're getting really good at having your LLM, Google Gemini built into your Colab notebook can do a lot of that and make a scatterplot. But you've got to be able to, especially on big important decisions, like you're talking about situations where tens of millions of dollars or hundreds of millions of dollars are involved in a transaction, you want to be sure. You don't just want to press the Gemini magic button and see some results and believe it. You want to be able to dig into it and note it yourself, sign off on it.
Martin Goodson: 00:28:30
Well, I was speaking to someone, a friend of mine, he works for a big consultancy, like one of the top kind of strategy consultancies, and I won't say who it is, but they're using LLMs to analyze financial statements, but they are basically doing what you just said. They're just pressing the button, which analyzes the statement, gives them results, summarizes it all, makes all of these financial decisions. It is a magic button. It just does one thing in one big step. I think you need to split these things up into multistep processes and have a look at the results at each step.
Jon Krohn: 00:29:04
I bet they're charging a lot more than you are too.
Martin Goodson: 00:29:07
Yeah.
Jon Krohn: 00:29:10
Yeah, those consultancies. So in addition to your work at Evolution AI, you're also widely known for being the founder and organizer of what is now a nearly thirteen-year-old meetup in London, the London Machine Learning Meetup, and I don't know exactly how you prove something like this. You in the same way that I guess I can say Super Data Science is the most listened to podcast in data science because I'm aware of the other data science podcasts, and I know how many downloads they have and it's smaller than ours. So in the same way, I guess you're well aware of all of the AI and ML communities in Europe, and so you can confidently say that with 15,000 members, the London Machine Learning Meetup is the largest community of a AI ML experts in Europe, so that's cool.
Martin Goodson: 00:29:55
Well, I don't know if we can confidently say it, but we can definitely say it. We do say it. We definitely do say it. We definitely do say it.
Jon Krohn: 00:30:01
Nobody's raised their hand and said, "You're wrong."
Martin Goodson: 00:30:03
Nobody's disagreeing with us. I should say that I didn't actually found the meetup. It was founded by somebody else. He gave it over to me, someone called Jurgen who he handed it over to me, but I think I've been running it maybe for 10 years or something, quite a long time.
Jon Krohn: 00:30:18
Well, we'll take that experience. That's enough for us. I'll continue on with the same kind of line of inquiry.
Martin Goodson: 00:30:23
Yeah. Yeah.
Jon Krohn: 00:30:26
So you've been working in data science for longer than it's been called data science. You've witnessed the evolution of AI from ad-hoc experimentation to strategic, continuously integrated workflows from niche to mainstream and from exciting novelty to at times over hyped technology. And one of the things that I was really excited about to have you on air on the show even four years ago when you thought you had nothing to say was that you'd written an article that reached number one on Hacker News called Why Data Projects Fail. And we're obviously going to have a link to that in the show notes.
00:31:06
One of my favorite things about it is that it's funny, it's short, it takes a few minutes to read and it's funny and it's spot on. If you're looking at getting started on a new data project, I frequently, I probably don't remember all 10 of them verbatim, but there are items from it from reading when you wrote this eight years ago that come back to me when I'm thinking about getting started in a machine learning project. Very simple things like, how are the data structured now? Has anybody been doing any kind of modeling with these data yet? And if they haven't, you need to add months onto the timeline for this project in order to be able to confidently say to the client that you're going to be able to do something.
Martin Goodson: 00:31:50
Yeah, I mean, those were some hard learned lessons in that article, mainly just based on me just screwing up loads of times previously and trying to learn the lessons from it.
Jon Krohn: 00:32:00
I'm going to quickly go through them on air. They're so simple and they're so useful. So number one is that the data aren't ready, so that's kind of the first one that I just said. You've got to look at the data before committing to a project, and I think that's eight years later, as AI has become so powerful through transformer architectures, there's this kind of expectation that more and more magic is possible. And so executives in a company think, "Oh, my competitor is doing this." But maybe the competitor has been tracking data, logging data that's useful to this task, and the executive at the company that you're considering doing this consulting work for hasn't been logging those data.
00:32:44
And so how are you going to... you can't magically create an AI capability without some kind of underlying data. Your number two is that somebody heard data is the new oil, which is something that people today, it's more like people are saying data is the new electricity or things like that. It's such a weird thing to me because these resources like oil, electricity, they're finite, whereas data, I mean data are finite, but data you can copy easily. And so there's some ways, and it's kind of like the inverse of an asset like oil, where part of what makes data so valuable is that you can copy them.
Martin Goodson: 00:33:24
The thing is if you've got loads of oil, you can sell it. And oil has a defined price, you know that it's valuable and you can sell it, but you try to sell data and it's just not like that. There is some guy, there's always some guy or girl whose job it is to sort the data out and maybe their life is going to be a misery because the data is horrendous and it's got so many problems that it has literally negative value. It is not worth anything to anyone. But it's easy at the top of the shop in the executive team to think that we've got all of this data, and we can say that idea was particularly prevalent at that time when I read the article. Maybe it's not so prevalent now.
Jon Krohn: 00:34:09
Yeah, well said there. Number three was that your data scientists are about to quit, and this was because of issues like access, where it's wild to me how many big organizations don't allow their data scientists or software developers to have root access to their machine to be able to install libraries, Python libraries. It's wild. And so I think that was kind of the main point.
Martin Goodson: 00:34:36
Yeah, I don't know if that's still an issue. That was a huge issue at the time. Is that still an issue? It can't be, can it? Really? Oh my God, that's terribly if it is.
Jon Krohn: 00:34:44
I regret to report, I've seen it firsthand recently.
Martin Goodson: 00:34:48
Wow.
Jon Krohn: 00:34:49
Yeah, it definitely happens. Actually, something that ties in perfectly to what we were just talking about with respect to statistics is you're number four, which is you don't have a data scientist leader on the project. So you mentioned specific things like selection bias, measurement bias, Simpson's paradox, statistical significance at the time you thought was important eight years ago, and if somebody is going to be building... The term data scientist is used to describe so many different kinds of roles. You really have to dig into a job description or a project description to understand what's really required in a role.
00:35:32
But for me, something I use as a like litmus test of whether this is really a data science role or not is whether there's going to be predictive models built. So not just analytics, not just analyzing things, but building a model that could go into production that will be making predictions on data that the model hasn't even seen before. That to me is for sure a data scientist, and it doesn't need to be a machine learning model. It could be a regression model, a statistical regression model, but you can end up, it would be easy if you were creating a data science project where you want to be building a predictive model and you have somebody who doesn't have experience creating predictive models and putting those into production, and you could end up with another wildly popular blog post probably around that same time called the no-limit credit card of software debt.
Martin Goodson: 00:36:33
Yeah, no, I remember the article, I don't remember the specifics, but specifically when I wrote the thing that you are talking about, at that point, I was really talking about leadership. And I feel like that is still a huge problem that you get these teams where the leader of the data science team has never seen any data and they never really spent a lot of time with data. They have some other background, some great background in something else, but they just don't really get it. They just don't really get data.
Jon Krohn: 00:37:01
They come from one of the big three strategy consulting firms. That happens a lot.
Martin Goodson: 00:37:06
Happens a lot, yeah.
Jon Krohn: 00:37:09
Very highly paid, very well-educated, but they're used to pressing the magic button.
Martin Goodson: 00:37:13
Yeah. Yeah, exactly. Yeah, and you get lots of money wasted and your projects all fail. Nobody wants to admit that they failed because it's embarrassing and it's just a complete waste of time.
Jon Krohn: 00:37:26
Your number five is the inverse of four, which is that you shouldn't have hired data scientists at all. So again, it's kind of a misunderstanding of the project where you hired data scientists, but in fact you need data engineers or you need BI analysts. And yeah, I guess it's easy when you're an executive, time is short, you know have this exciting project. You think there's an opportunity and you're like, "Oh, this is AI, we need a data scientist." But very often you don't, and I think that's something that I've talked about on the air before is that while, yes, there is a lot of demand for data scientists out there in terms of that being a job that people aspire to have.
00:38:03
I'm sure there's probably quite a few listeners out there listening to the Super Data Science podcast who are interested in getting a job as a data scientist. And yes, there is a future for you in this space, but a career like being a data engineer, a software developer, I mean, there's orders of magnitude more vacancies per interested applicant in those kinds of areas. When we have people on the show and we say, "Are you doing any hiring?" Invariably, people are hiring software engineers, data engineers, they're not always hiring data scientists.
Martin Goodson: 00:38:33
Yeah, I think that's true. I think early days of the industry, everyone thought they needed data scientists, and then they gradually started to realize, actually, you don't need as many as you thought. You need way more engineers.
Jon Krohn: 00:38:46
Number six is actually basically what I just said, which is that your boss read a blog post about machine learning.
Martin Goodson: 00:38:53
Yeah. I've never said this before, but actually I partly inspired to write this piece as a revenge thing, because my boss at the time was a complete, I can't say, he used to come in, he used to come work and just come up with some rubbish about ensemble models or something, and I thought, "How I'm I going to get revenge? I know, I'll get something on Hacker News, specifically calls out this behavior, which is very dysfunctional." And that's what happened, and he was quite embarrassed in the company at the time, I think. He left shortly afterwards.
Jon Krohn: 00:39:30
Oh, wow. Well done.
Martin Goodson: 00:39:31
That's how petty I am.
Jon Krohn: 00:39:36
No, I mean this has happened to me many times where people... Yeah, there's a kind of personality in business where they think somehow they can get the gist of an idea. Oh, one of my favorite ones is supervised in unsupervised machine learning where somebody who's... it sounds like there's an intuitive concept there where a supervised machine learning model needs somebody, needs a human in the loop. And then unsupervised, you don't. That's kind of what it sounds like. And so I've had that conversation a number of times where an executive, they've read about, somehow they came across a blog post where the term unsupervised learning was in there and they're like, "Oh, if that's what we need. We need no more human in the loop." I mean, just to be clear on this for our listeners, what unsupervised learning actually is where you have an algorithm, where you're training machine learning model and you don't have labels.
00:40:37
So supervised learning is a paradigm of machine learning where you have say a whole bunch of images. The classic example is to say that half of them are dogs, half of them are cats, and they're labeled as such, and your supervised machine learning model learns how to classify dogs from cats. In the unsupervised learning paradigm, you don't have those labels. You don't know whether the images are of dogs and cats, but there's still machine learning algorithms out there that can recognize patterns in the data and kind of categorize, figure out how to sort things into buckets vaguely. And so you're not necessarily going to end up with dogs and cats. You might end up with being able to distinguish dark images from light images or kind of... I'm getting into a very vague example.
Martin Goodson: 00:41:23
I think the critical point is... What I'm trying to say in this article and in many other articles that I've written, is in data science, it's a really technical discipline and it is a discipline and you need to be a specialist to really perform well. And just reading a blog post and kind of spouting off about unsupervised, supervised whatever, just because you kind of heard the word before, it's just not going to get you anywhere. In any other technical discipline, people don't do that. Well, why is this different? It's because it's new. It's a young discipline. People don't treat it like a real discipline, like a real scientific discipline.
Jon Krohn: 00:41:55
I just remembered another one where with reinforcement learning, when reinforcement learning started to become a buzzword around the time of AlphaGo, I was having lunch with someone that I really respect actually, who's a startup founder. Brilliant guy, has been very successful, but he'd done the same thing that we just described with unsupervised. He'd heard this reinforcement learning and he, I guess, read a blog post or watched the AlphaGo movie or something. And so he was talking about how within his platform, he wants to get reinforcement learning involved so that when humans use his platform and add data, that will reinforce for the machine learning algorithm what it needs to learn.
Martin Goodson: 00:42:42
And there's some data scientist whose job it is to execute his vision. He's not having a fun time.
Jon Krohn: 00:42:48
Exactly. That's exactly right. Number seven, I can't find a way to thematically tie it very nicely to number six. Actually, no, we can. So number seven is that your models are too complex, which is kind of this classic example of, so if number six was your boss read a blog post about machine learning. Number seven is that your models are too complex, and that's a situation where it could happen. It's like your boss says, "We're going to do reinforcement learning, or we're going to use a large language model." When you don't need that at all, when you just need a simple statistical logistic regression, and that'll get you all the way.
Martin Goodson: 00:43:21
Yeah, I'm quite a simple person. I like to do things in a really simple way if you can. I definitely like to start projects off with very, very simple methods that everyone really understands really well, just to figure out what's going on, what's going on with the data, just to have a really simple idea of what's happening that you can really understand in a deep way. I've heard of many, many projects where somebody's come in and they just want to use some really advanced technique that they've just heard about or just learnt. It's natural to have some excitement, but you use this complex thing, you don't really understand it. It's really difficult to interpret the results. And I've known so many stories where six months down the line, "Oh, we used the wrong column in the input data set because we didn't really understand the data, and it was really hard to interpret the results." Seriously, six months of work just completely wasted because people use some complex method that they couldn't really understand. I'm such an advocate for simplicity.
Jon Krohn: 00:44:22
Where possible, and it also allows you to be prototyping a lot more quickly, potentially save a lot of resources, lots of reasons to start simple, for sure. Reason number eight that data projects fail from your blog post is that your results are not reproducible. And so you specifically cite the kinds of tools like Git, code review, automated testing, data pipeline orchestration, which yeah, I mean, luckily today we have more and more tools that make that easy.
Martin Goodson: 00:44:48
We do. We also have more and more tools that make it hard. Like Notebooks, right?
Jon Krohn: 00:44:53
Yeah. Yeah. Yeah. Jupyter Notebooks are a pain. I do love them, but yeah, they're such a-
Martin Goodson: 00:44:56
So do I.
Jon Krohn: 00:44:57
Yeah.
Martin Goodson: 00:44:58
Yeah. I love them. They are my guilty pleasure.
Jon Krohn: 00:45:01
Yeah, exactly. Eager to learn about large language models and generative AI but don't know where to start? Check out my comprehensive two-hour training, which is available in its entirety on YouTube. Yep. That means not only is it totally free, but it's ad-free as well. It's a pure educational resource. In the training, we introduced deep learning transformer architectures and how these enable the extraordinary capabilities of state-of-the-art LLMs. And it isn't just theory; my hands-on code demos, which feature the hugging face and PyTorch Lightning Python libraries, guide you through the entire life cycle of LLM development, from training to real-world deployment. Check out my generative AI with large language models, hands-on training today on YouTube. We've got a link for you in the show notes.
00:45:47
Pre-pandemic, we got to a point where my whole data science team was off of Notebooks, and I was the only one as the chief data scientist using that, and it was such a pain for everyone else in the organization. I am that boss.
Martin Goodson: 00:46:05
But it is still a problem. I don't think the industry has really, or the field has really, really figured this out. I definitely don't have a perfect solution to this, but definitely on a daily basis, I'm still encouraging people in my team, we're a bit confused with these results. We can't really look at the code. We can't really reproduce these results. We haven't really been rigorous enough at this research. We need to just simplify things, get everything down in Git or whatever so that we could start again really. So it's still an ongoing thing. I don't think we've figured it out. I don't think that's a good solution.
Jon Krohn: 00:46:38
Yeah, there's no perfect answer. It's one of those things where having a leader, at least an engineering leader, is helpful in figuring this out. There's I think a Goldilocks sweet spot on a lot of these things where if you go too far, you have too many processes, you can kill the project with overhead. So finding that sweet spot depends on expertise. Number nine is that an R&D lab is alien to your company culture. This is so huge. I remember actually after reading this, I implemented things like an internal talks club. So R&D is a high-risk activity. If you don't have things like you say in here, lab meetings, talks, publishing papers, it's difficult to have that R&D culture. If your business is looking for tangible near-certain return on investment on every data science project, it's going to be hard to do anything interesting.
Martin Goodson: 00:47:46
Yeah, I think companies really need to take a hard look at themselves and really think about whether they want to do research or not, because lots of companies say they want to do it, but when the hard reality comes up, confronts them, I think they realize we don't really want to be doing this. We want certainty. We don't want someone saying to me, "I don't know how long it's going to take to do this, and I don't even know whether it's going to work ever."
Jon Krohn: 00:48:10
Something that I have been doing over the years that has I think been helpful to management is that I'll typically, I'll kind of break up the amount of time that my team is spending on different levels of risk and say, "We're going to spend 30% of our time on these relatively high-risk projects where if it works out, it's going to be a game changer for us and our positioning relative to our competitors." And then another third on medium-risk projects, and then another third on projects that I'm like, "We'll have something at the end of the quarter to show management even if these go through."
Martin Goodson: 00:48:51
Yeah, really good idea.
Jon Krohn: 00:48:54
Final one here, number 10 from you is people designing data products without seeing live data. You describe this as doing taxidermy without looking at live animals, and this is huge. And this is one that I think data scientists themselves are... A lot of these ones in the list, the data scientists are all left thinking, "Ha, ha. But with number 10, that's the one that data scientists often get wrong. They've collected data or they've scraped data and they anticipate what the user or what the production use case is going to be like, but they don't know for sure. And it ends up being that due to drift or due to real world use cases being quite different from what you trained your model on, you get vastly different results in production.
Martin Goodson: 00:49:41
Yeah, I mean, I personally made this mistake really badly. I did a project, so it must've been about six months, and yeah, I just completely screwed this up, but I had this vision for how things were going to look in this data product, and I spent a lot of time with customers really trying to understand their use case and trying to understand their workflow. But I made loads of wireframes and stuff, and I tried to be really good about the whole product design thing, but I just didn't really use enough real data. And when I used real data, I realized everything that I thought was just a load of rubbish and it was not going to work, and the project was a complete failure. So that was a really heartfelt, that was me screwing up massively.
Jon Krohn: 00:50:25
Yeah, I've done it too. That's an easy one to make. I asked online, knowing that you were going to be coming on the episode, I posted that you would be coming on. I posted a link to this article on LinkedIn and on Twitter, and I asked my audience if they had any thoughts on why their own AI projects fail. We got a good one here from Peter Anderson in Denmark and so Peter, he gives us a hark back at episode number 781 with Sol Rashidi, where Sol Rashidi talks, actually, that's another episode about basically why AI projects fail.
00:51:08
She's written a whole book about it, and her main point, and also Peter Anderson's in recalling it, is that projects fail more often, he's seen projects fail for one thing that they have in common, which is a lack of proper business reason. So it's kind of this where you have a hammer, and so you go around looking for nails as opposed to having some problem that you know is actually a business problem, that there is a chance of there being an ROI if you succeed at that. I think we've seen this a lot in recent years with generative AI, haven't we, Martin, where every application is trying to build generative AI into it, and I bet most of those projects are a complete waste of time and resources.
Martin Goodson: 00:51:53
Yeah, I can't say, I have to be careful about what I'm saying, but let's say that I know of many projects, real projects where people have made the decision to use AI. First of all, they made the decision, what kind of method to use as a second step, they started working on this method as a third step, and then as a fourth step, somebody says, "Do we actually need this thing?"
Jon Krohn: 00:52:22
Yeah, exactly.
Martin Goodson: 00:52:23
Very, very common. And it's a problem. It's a huge problem. It's always been a problem. I'm not seeing it change. It's been a problem for 10 years at least.
Jon Krohn: 00:52:32
Yeah, it is one of the big ones, for sure. Yeah, is there anything else that you think you would add, now almost a decade later, since you originally published this wildly popular post? Is there anything else that you think you'd add in or something that's changed a lot over the past decade?
Martin Goodson: 00:52:49
Well, I think it's just got worse. The problem has got worse because of the hype. The hype has increased, so the problem has got worse, and people want to use AI for everything, and they have overinflated expectations of what's really possible. I'm seeing this a lot, actually. The thing is that the LLMs are really good at creating plausible output, right? That's what they're really good at. That's what that's designed to do. They're really good at fooling people. I spoke to many people, they try some stuff out with some commercial LLM and everything looks great, the output looks great, but they don't really do any evaluation and they don't think they need to because it just looks perfect. Everything's fine.
00:53:33
If a human came to you and you asked them some accountancy question and they gave you this answer, which was perfectly formed English for five minute exposition on this accountancy concept. You'd be like, "That person's really intelligent, and they're obviously an expert in accountancy." But that world has just changed where somebody can tell you something, which is very fluent and very long-winded, very eloquent actually. But it's wrong. And you need to do all of that boring stuff like quantitative analysis and evaluation to figure out whether it's actually useful and it's going to do the job for you. So these problems are even more prevalent now, I think, than they used to be. So I don't think there are any new things that I've seen, but it's still definitely, most of these are things are real issues.
Jon Krohn: 00:54:21
Very nicely said. Well, yeah, thanks again for that blog post, which yeah, as I said in my own mind, it's something that I've referred back to so many times when thinking about getting going on a project. Switching gears here to something that you've done a lot more recently. Last year you gave a talk to the European Commission in which you advocated for a publicly funded open source in which you advocated for publicly funded open source AI for humanity writ large, and you drew lessons in that talk from the Human Genome Project. And you want to ensure that the benefits of AI are spread widely across the global population and not just concentrated in the hands of big tech. Do you want to talk about that talk? I mean, how that even come about? How did the European, do they send a letter? How does the European Commission invite you to speak?
Martin Goodson: 00:55:09
Yeah, I think the European Commission asked Nature Publishing group to organize this thing and convene some people to give some talk. So there was a panel of three of us, and they selected me. I don't know why they selected me, actually. Good question. But I thought, "Okay, you know what? I'm going to advocate for this because I think it's really important. I really think that we need to get..." Because at that time, it really wasn't clear what was going on. It really seemed like we were going to have a handful of commercial players with LLMs who were going to take over everything. And there was nothing going to compete with them, and that hasn't really panned out.
00:55:43
I mean, there've been obviously huge and really successful open source projects, which are catching up if not exceeded some of the commercial offerings, which has been amazing. But at that time, it was kind of a huge worry, like what's going to happen to the workforce, unemployment, employment considerations. All of this power, there's so much power concentrated in the tech players that the public publicly funded initiatives need to get in there. And the reason why I spoke about the Human Genome Project is that I had a lot of people saying, "Well, what can public funding achieve? You can't do anything in the public sector. It's never going to be able to compete with the commercial world." And that's just so historical because the Human Genome Project did compete with the commercial world and beat the commercial world.
00:56:33
This is well documented, the selects Aerogenomics and Venter trying to patent the human genome and the lots of public-spirited scientists basically said, "No, we're not going to allow that to happen." And they worked together on this huge project. So they got public funding, partly public funding, it wasn't all public funding, and they won. And they managed to release the human genome and put it in the public domain, and because they put it in the public domain, Craig Venter could not patent those genes. And they beat him, and they stopped that happening. And who knows what kind of world we'd be living in if he had managed to patent those genes. So I really wanted to inspire people to say a public sector can achieve stuff.
00:57:15
Scientists and researchers and academics can achieve stuff together. Let's stop thinking, "Oh, everything's so doom and gloom." Like, "Oh, let's just give up our research and everything's going to be owned by OpenAI and Google." I don't believe that. And I think that's been born out because you've had loads of really cool open source projects that have been built. They've trained a model using eight A100s or whatever, some academic lab has done, like Lava, I think it was trained in eight A100s. It's open source model and is really, really interesting. And that despondency was totally unwarranted, I think. But as any impact that I had with the European Commission, I think I had zero impact. I got the words out there.
Jon Krohn: 00:58:01
You might've helped Mistral get some funding.
Martin Goodson: 00:58:04
Maybe it's possible. It's possible, yeah. But I think there would've been a lot... Okay, so the UAE did that. The UAE released Falcon, they funded United Emirates, they did fund Falcon. So this is the kind of stuff did happen. It's just that the UAE didn't do it, and the UK didn't do it. Mistral accepted, I guess, but definitely the UK didn't do it. And it wasn't just about building an open source model. It was about building capabilities and skills. I really thought that in particular, if the UK did this, then we would be building really good skills. Instead of that we've kind of spent money on stuff that I didn't think was very useful and there's a whole topic in itself, which is for UK government, and it really didn't take advice from the right people, I think. It didn't take advice from technical people, so it ended up wasting a lot of money. I think that money would've been a lot better spent if we invested in open source, open source LLMs or other things. It would've been great for building network skills and those other stuff.
Jon Krohn: 00:59:13
As you mentioned, the UK there, and so you have in previous podcasts and articles, you've lamented how the UK used to lead science and technology and now on AI, at least they do have a backseat, so it seems like Mistral, the UAE, the US, China, there's lots of LLMs in those countries that are pushing the state of the art. We don't typically see the UK or British universities with LLMs on the top of AI leader boards. What kinds of strategic moves do you think the UK should consider to regain leadership in AI, which has historically been a strength for the country?
Martin Goodson: 01:00:02
Well, the big thing that we did that we shouldn't have done is completely ignore language models and neural networks. We completely ignored it. In the UK, we've got this big flagship research institute called the Alan Turing Institute, and it completely ignored this research. All of this, it just didn't do anything on language models, even BERT's era language models. It just didn't do anything on this. I've written about this before. They were publishing blog posts on fungible, non-fungible tokens or whatever, and just stuff that was just completely unrelated to the mainstream of AI research.
Jon Krohn: 01:00:41
Cryptocurrency.
Martin Goodson: 01:00:43
Yeah, but using AI to predict something about cryptocurrency or something. That was their biggest blog post in 2022, I think was the thing. They came out with all this stuff. They had this group called the AI Council that the government convened, and it was just all of these things. The UK is just great at having lots of people with letters after their name, getting together, writing these long reports, but they just invited the wrong people. They didn't invite any startup CTOs. The data scientists, the actual practitioners who were doing the work who understood this. They didn't do any of that. They didn't invite any of those people to any of these bodies. They invited, like I said, people with letters after their name. They had the ingredients of a really lovely garden party, but they didn't have the ingredients of a hardcore engineering expert group.
01:01:42
That's the problem that the UK has, I think. But that's to be negative, but the positive thing I think is really just listen and talk to the practitioners. I shouldn't need to say this, talk to the people who are actually doing AI, who understand the field. There are many people in the UK with great talent. There's some amazing talent in the UK. There's a great community as well. It's a great technical community. The government needs to get, just to talk to them and they're just going to tell them what to do. At that time, when I read that article that you mentioned, I really think the right thing to do was to work at open source LLMs. I'm not sure, and now whether that's right anymore because lots of other groups have come up and have done that. The timing is maybe not so good anymore, but I said I don't think I have good answers to your question.
Jon Krohn: 01:02:33
Something related to your leadership in the UK is that you were elected chair, and I think you no longer do this role, but for several years until recently, you were the elected chair of the data science and AI section of the Royal Statistical Society. That sounds pretty fancy. You were just talking about garden parties there, the Royal Statistical Society, is that like, do you have little sandwiches with the Queen and talk statistics with her?
Martin Goodson: 01:03:02
No, but it's great in the RSS, it's really old. I think Florence Nightingale was the first president of the RSS. Is that right? Wow, I just made that up. I think that's true. It's great, and they really wanted me to help them have a presence in the AI world, data science world and they quite rightly felt like their statistics was being sidelined in this world and they wanted to come to the fore, and I think I was part of that. Yeah, I was the chair of this organization. I'm no longer the chair. It's Janet Bastiman, is the great chair right now. She's doing a great job, so I've taken a step away from that. But what we wanted to be was the voice of the practitioner because no one was the voice of the practitioner.
01:03:56
And still that's the case, that no one is really putting the views forward of data scientists, the actual people doing the work. In any other industry, you're going to have some industry body who are going to represent the fishermen. If the government want to take some policy decision that affects fishing communities, they're going to go and talk to some groups who represent the fishermen or they're going to go to the communities. That never happens in data science. They just don't go and talk to the communities who are involved. So we wanted to be that voice and represent their survey membership, make sure that we're representing their views accurately and stuff like that. That work is still going on. I'm no longer doing it, but it's still happening.
Jon Krohn: 01:04:39
Nice. Martin, I did a fact check here. I believe that this episode is actually, it looks like it's going to be released on the day of the US presidential election. One of those candidates has been refusing interviews with fact checking, so luckily you did agree to come into one of these fact-checking interviews, and we've got one here. It doesn't look like Florence Nightingale was the first president of the Royal Statistical Society. I thought that that-
Martin Goodson: 01:05:04
Was she a president?
Jon Krohn: 01:05:06
It doesn't look like it. No.
Martin Goodson: 01:05:08
She was something.
Jon Krohn: 01:05:11
So the first president of the Royal Statistical Society was Henry Petty-Fitzmaurice, the 3rd Marquess of Lansdowne.
Martin Goodson: 01:05:19
Yeah, that's true because it's actually really old. It's a lot older though, because Florence Nightingale is not actually going back in time that far, is she? Okay, I'm not going to Google this now. I'm going to spoil the flow and my computer's too slow. But she is something.
Jon Krohn: 01:05:35
We'll find out. We will take a break in filming.
Martin Goodson: 01:05:38
Oh, she was the first female fellow of the RSS in 1858.
Jon Krohn: 01:05:42
There you go. We got it. And our listeners only had to hear a small amount of you typing while speaking. Nice. Well, very cool. Nice to get that kind of historical context and yeah, must have been, I mean, I guess was that a cool experience being in the Royal Statistical Society?
Martin Goodson: 01:06:01
It was great.
Jon Krohn: 01:06:02
Nice.
Martin Goodson: 01:06:02
It was great. Yeah, no, it's wonderful. I met some really interesting people and it did give me a voice, and it did mean that... Really, why did I get invited to the European Commission? It's probably because it's something to do with being part of that learned society. It gives you some credibility. So yeah, it was a great platform.
Jon Krohn: 01:06:18
So you talked a lot in your last answer about people, about actual practitioners providing guidance on AI. My last question for you is related to the public's perception of AI, which seems to be influenced a lot by high profile tech personalities. So I, for example, at the time of recording, I had been watching Bill Gates Netflix special called What's Next, which is at least the first episode is all about AI, and it became... I have been laughing to myself because I mean with apologies to Bill Gates, who is a very impressive individual and quite learned, but it became quite obvious, at least at the time of this show being filmed, which looks like it was about a year ago. It looks like it was 2023 based on the ChatGPT related things that they're talking about. And it's pretty clear that Bill Gates does not have an understanding of AI that I expect the vast majority of the listeners to this podcast have.
01:07:18
And so that was a really interesting experience for me because I would think that he is the kind of person that would understand these things well, but in the first episode, the funniest part for me so far is Bill Gates has this yellow notepad and he has the words train in a box and then reinforce written off of the box or something. And the people who are filming it, as he's explaining a bit about these same kinds of things that we were talking about earlier, supervised learning, reinforcement learning, they made the directorial decision to use footage zoomed in of his notebook on this thing that is like, it's like some important, "Oh, Bill Gates notebook, look at this great schematic that he drew." And I'm just like, "What? The word train in a box."
01:08:10
And so yeah, my expectations of at least 2023 Bill Gates, my expectation of Bill Gates 2023 knowledge of AI was much less than I would've anticipated from him. And so you have a quote from another podcast that you did where you said, "It's true. I don't really know anyone in the field of AI who thinks of Elon Musk as an expert on AI." And that is the kind of person that I would expect to be more off the mark than Bill Gates, I guess, on their understanding of AI. But yeah, we have this problem where the public's perception is being influenced by these kinds of high profile tech personalities. It doesn't seem like people like Fei-Fei Li or Jeff Hinton who really know what's going on, share the same kind of reach as these other kinds of people who the public seems to think, "Oh, Elon, Musk, Bill Gates, these are AI experts."
Martin Goodson: 01:09:07
Yeah. What's the question?
Jon Krohn: 01:09:10
Right. I didn't really ask a question, did I? Well, guess I just evoked your point.
Martin Goodson: 01:09:18
Who asked the questions? You asked the questions, right?
Jon Krohn: 01:09:20
I don't think I can do [inaudible 01:09:23]
Martin Goodson: 01:09:23
I can ask a question. I can ask a question, what can we do about this?
Jon Krohn: 01:09:27
Yes, yes, yes. That's it.
Martin Goodson: 01:09:28
What can we do about this? You mentioned before that I've got this machine learning meetup that I run, part of the organizing committee, let's say, and we have lots of academics who come on and give talks, some of them give really great talks. It's really amazing actually. I really love it and the talks are absolutely amazing, but I have to say that we quite often, we sometimes, let's say sometimes we get academics on who give talks and they're really over hyping stuff, and it's very easy if you are outside the field, like some of the people that you're talking about to read some papers, but you could become an expert, like a self-proclaimed expert quite easily by reading stuff on archives. You could sort of read loads of papers and stuff and get to become a self-proclaimed expert quite easily. The problem is, one of the problems is, there are many problems, but I'm just going to highlight one.
01:10:28
One is that the academics are publishing stuff and they're over claiming, the titles of the papers are just wildly, they don't have the evidence to claim what they're claiming. And I won't mention names because it's kind of unfair, but we do have people who come to the meetup and they give a talk and they just make up the stuff that's just very over hyped claims, and you really put them under scrutiny. Once you put them under scrutiny, it just falls apart. They don't have the evidence. Both you and I, we met in a world-class research institute in genetics. It was world-class, so we learnt at first hand what it means to be really rigorous and what the scientific method is at the highest level. I'm not saying that I was the highest level, working at the highest level, but we definitely worked around people who were working at the highest level, and we took on board a lot of lessons then. And I actually sometimes get quite annoyed with some of our speakers, I have to say.
01:11:30
We have someone who recently who came, they gave a talk and they said, "Oh, I'm not going to talk about any of the technical stuff here because I don't think you're going to be... We don't have time to talk about the technical stuff." And you're in a technical meetup. You should be under scrutiny. And I think we all need to do better in terms of raising the bar of the scientific culture within machine learning, and I think that if we did that, we would do much better. This goes towards some way to solving the problem that you're talking about. Back in our day, working in genetics, people used to write papers in your university and then you'd have a PR department who would make up these massively over claimed headlines that would go into the newspapers, but now just people skip the PR team and they just do it themselves. The academics do it directly. They just cut the PR people out of the job, and I just don't think that's a positive. So I guess, what should we do about it? We should stop doing that.
Jon Krohn: 01:12:32
Nicely said. You talking about a high level of rigor and science reminds me of something that happened. You probably don't know this, it might've actually happened, I might've still been doing a PhD and you'd moved on, I can't remember what the timeline was exactly, but Jonathan Flint who ran the lab that both of us were in, and he's in episode number 547 of the show if people want to hear more on genetics and the intersection with machine learning, it's a great episode. He is an outstanding leader in the space. He's probably the world's foremost psychiatric geneticist, and he was a big believer in following the data, following the scientific method to the extent that there's something that he made me do, which was extremely embarrassing.
01:13:23
And he must have known that something like this would happen, but he made me go through it anyway, was you know what? This was actually, I now remember, it was before I'd even started the PhD, I did a master's project with him, and this was a result that came out of my master's project. I had looked at the relationship between... So at the time that I was doing my master's research, there had been a paper that had made a lot of headlines that had shown that magnetic fields, that looking at satellite imagery, animals in the fields like cows tended to be aligned somehow magnetically. There was statistically more likely to be facing north or something, magnetic north. I don't remember the details. But so Jonathan had me look into behaviors related to fluctuations of the moon. It was something really bizarre, and there were some statistically significant effects.
01:14:31
There was some biochemical things in mice, turned out to be related to moon cycle, and that couldn't be related to light because lab rats, lab mice, they don't see daylight. They're in the dark, they're always in some interior room. There's no windows in the rooms that they're in, so it would have something to do with magnetism, something like that. And so I can't remember what exactly the chemical was. Let's just say it was calcium. Calcium turned out to be statistically significantly related to fluctuations in the moon, to the moon cycle, and so Jonathan made me go to the Radcliffe Hospital, which is this huge hospital, and meet with researchers that had similar kind of biochemical data on humans and say to them, "I found this relationship between calcium and the moon. I would like to have some human data, please." Hey, Jonathan made me do that, and they took one meeting with me and then refused to answer any of my subsequent emails.
Martin Goodson: 01:15:51
Brilliant, but that's a great experience for you to learn that.
Jon Krohn: 01:15:53
I guess so, I don't know what he was teaching me there. Something, I guess it's about following the data, as unlikely as it seems, probably it was just as spurious correlation. But if you could show it in a completely unrelated dataset, I don't know, something interesting and publishable, PR team would've loved it.
Martin Goodson: 01:16:15
Yeah, I guess you're right. You have to follow the data and you need to be dispassionate and even something that you believe to be completely crazy. If the data says it, then it's worth investigating it. I mean, you don't need to be wedded to it, but it might be worth investigating.
Jon Krohn: 01:16:29
Yeah, I guess that's the fear is that this person in this position, they don't really know me, so they're like, "This guy must be a whack job." But really I'm in there like "I don't believe this. I don't want to believe this. Please provide me some more data to prove that this isn't true." Well, Martin, it's been awesome having you on the show. I've really enjoyed reconnecting here on air. Before I let you go, I always ask my guests for a book recommendation. Do you have anything for us?
Martin Goodson: 01:16:57
Well, I do have, I have two recommendations actually. So these are books that I really love and not really about AI, although you might be... they're kind of related. So one is Vision And Brain. Whenever anyone asks me this question, I always mention this book, which I absolutely adore. It's called Vision And Brain, it's by James Stone, and it's just about human vision system, and it's just such a lovely book about how our brain processes visual information. But it goes into technical detail, but it explains it in such a way that you don't need to be a specialist to understand it.
01:17:40
For me, it just really opened my eyes to so many different things. I really adored that book. And the second book, it's called The Evolution of Cognition, it's by David Bowles, which also opened my mind to so many different things, and it really is about how, really about humans, how mammals, no, how animal... Actually, that's just not true. It talks about human cognition, but it starts from single cellular animals, so it really is talking about biological cognition, and so I'm interested in AI, but I'm also interested in animal cognition, and it's just a really amazing book. I really think it's beautiful.
Jon Krohn: 01:18:19
That is super cool. I also am fascinated by this. I mean, technically my PhD was in neuroscience, this kind of how our brains allow us to perceive the world to think all the thoughts we think and do the actions we do is pretty wild. And yeah, so that's a great recommendation. I may have to check that one out myself because I hadn't heard of it. Martin, what's the best place for people to follow you after this episode? If my memory's correct, you aren't a prolific social media user these days.
Martin Goodson: 01:18:56
I'm on LinkedIn. I'm on LinkedIn. Yeah, I do post quite regularly on LinkedIn. You probably jut-
Jon Krohn: 01:19:03
Oh, my apologies.
Martin Goodson: 01:19:04
... ignored my posts.
Jon Krohn: 01:19:06
Oh, yeah, you do have, I don't know what I was thinking.
Martin Goodson: 01:19:10
Just like every day, at least twice a day.
Jon Krohn: 01:19:15
I must have looked into it years ago, maybe around the time that I asked you to be a guest four years ago.
Martin Goodson: 01:19:20
Yeah, I didn't used to use LinkedIn very much, but now I think LinkedIn's got a lot better, and you can actually have some really interesting conversations about AI on LinkedIn now. It didn't used to be, that wasn't the case years back.
Jon Krohn: 01:19:32
Yeah, Elon Musk has done a great job of making it the de facto platform.
Martin Goodson: 01:19:38
Yeah. Yeah. That's what happened with me. Yeah, that's exactly what happened.
Jon Krohn: 01:19:41
Nice. All right. Thanks Martin, for taking the time. Really appreciate it and yeah, exciting all the things you have going on over at Evolution AI. And yeah, even just seeing in the background for people who are watching the YouTube version, they get to see Martin's in this beautiful, what's called the garden room of his home, and so you can see garden on opposite sides of the room, which is a really beautiful thing. And it makes me miss Oxfordshire, which when it's sunny, which is rare, but you are experiencing right now, it seems like the most wonderful place on earth.
Martin Goodson: 01:20:13
It is. Well, thanks so much, Jon. I've really, really enjoyed it. It's been really good fun. Thanks again. Thanks for having me.
Jon Krohn: 01:20:25
Nice. To recap, in today's episode, Dr. Martin Goodson filled us in on 10 reasons why data science projects fail, including due to issues like lack of data readiness, overly complex models and poor reproducibility. He talked about how there's a need for more publicly funded open source AI development to ensure benefits are widely distributed. How practitioners and technical experts need more of a voice in AI policy discussions that rigorous scientific methods and healthy skepticism are crucial as AI capabilities advance rapidly. And he provided his insights on how interdisciplinary knowledge spanning fields like stats, computer science and biology are invaluable for AI development.
01:21:03
All right, as always, you can get all the show notes including the transcript for this episode, the video recording and materials mentioned on the show, the URLs for Martin's social media profiles, as well as my own at superdatascience.com/833. And next week if you'd like to connect in real life, I will be in New York on November 12th conducting interviews at the ScaleUp:AI conference, which is run by the iconic VC firm Insight Partners. This is a slickly run conference for anyone keen to learn and network on the topic of scaling up AI startups. One of the people I'll be interviewing will be none other than Andrew Ng, one of the most widely known data science leaders. I'm very much looking forward to that. Cool that Martin also interviewed Andrew in the past. I've got a link to that interview in the show notes for you. And right after that conference in New York, I will be flying overnight to Portugal to give a keynote and host a half day of talks at Web Summit, so that runs from November 11th to 14th in Lisbon, Portugal, with over 70,000 people in attendance. It's one of the biggest tech conferences in the world. It'd be cool to see you there.
01:22:13
All right, thanks of course to everyone on the Super Data Science podcast team, our podcast manager, Ivana Zibert, media editor Mario Pombo, operations manager Natalie Ziajski, researcher Serg Masis, writers Dr. Zara Karschay and Silvia Ogweng. And yes, of course, as always, last but not least, our founder, Kirill Eremenko. Thanks to all of them for producing another fun episode for us today, for enabling that super team to create this super podcast for you. We're so grateful to our super sponsors. You can support the show by checking out our sponsor's links, which are in the show notes, and if you would like to sponsor an episode, you can get the details on how by heading to jonkrohn.com/podcast.
01:22:55
Otherwise, you can help us out by sharing this episode with people who would like it, reviewing it on your favorite podcasting app or on YouTube. Subscribe of course, if you're not a subscriber, that goes without saying, come on. But yeah, most importantly, I really just hope you'll keep on listening. I'm so grateful to have you listening, and I hope I can continue to make episodes you love for years and years to come. Till next time, keep on rocking out there, and I'm looking forward to enjoying another round of the Super Data Science podcast with you very soon.
Show all
arrow_downward