In short
Neuron podcast episode with Alexandre Pasquiou (Neuronk) arguing that “tabular foundation models” (for spreadsheets/databases) will replace many custom predictive ML models across industries. He claims language models are “structurally failed” for production prediction on tabular data, though they work as UI/orchestrators. He describes how Selden projects table cells/rows into latent space without flattening structure, enabling better statistical inference and dataset-specific adaptation (meta-learning).
Key claims
one general model can handle churn, fraud, demand, pricing, risk, etc.; tabular models scale beyond language-model context limits (up to ~20M rows, 600+ columns); enterprises will use tabular foundation models for most productive workloads by 2030, with agents orchestrating workflows.
Notable examples
predicting product categories across different retailer taxonomies (Amazon/Walmart/Carrefour), fraud thresholds varying by country/context, and “extravagant” uses like predicting flight delays and World Cup winners (they predicted Spain).
Guests
Alexandre Pasquiou, CEO/co-founder of Neuronk; background includes PhD using fMRI/video-game experiments and transformer-based architectures to study brain regions tied to syntax/semantics.
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Chapters
Tap a time to open that second in VOThe Limitations of Language Models in Prediction
0:00 to 1:12
Explore the inefficiencies of language models in making predictive tasks.
“Do you think we're wasting billions and billions of dollars by trying to scale language models to do a task like what you're doing with tabular models?”
Alexandre's PhD Research and Insights
1:55 to 3:18
Learn about Alexandre's research involving fMRI data and language models.
“So I guess to start off, I heard a story about your PhD days where you were working on studying brains when people played video games.”
Differentiating AI Models: Data-Rich vs. Data-Poor
3:18 to 5:51
Understand the two lanes of AI models and their data requirements.
“And I also did a bit of decoding, which is reading the brain.”
Challenges of Language Models with Tabular Data
5:51 to 12:43
Discover the structural issues language models face when handling tabular data.
“So if today's language models become 10 times larger and 10 times cheaper, let's say, which business problem would they still be fundamentally bad at solving given the problem that you just described?”
Understanding Structure in Tabular Models
12:43 to 14:00
Delve into how tabular models differ structurally from language models.
“They do meta-learning, so they learn to learn during the pre-training so they adapt to each new dataset and they can make a prediction that are adapted to a client to a dataset.”
Understanding Tabular Foundation Models
14:00 to 14:58
Learn about the structure and capabilities of tabular foundation models.
“and others will associate a given row of your data set to a token.”
Time Encoding in Predictive Models
14:58 to 18:51
Explore how different models encode time and its implications for predictions.
“If you mean like you want to transform it to tell you the time or if you want...”
The Need for Modern Predictive Models
18:51 to 20:52
Discover the challenges businesses face with legacy predictive tools and the need for innovation.
“or they go to external consultants to build a model that would be specific to one question and one data set.”
The Disruption of Predictive Technology
20:52 to 22:38
Understand the market disruption caused by tabular models over traditional ones.
“while all the classic machine learning algorithms, they just cap to the information that is available in the data set of the client.”
Cost of Misaligned Models
22:38 to 23:14
Learn why using language models for tabular data is not cost-effective.
“Language models are not adapted to making production and tabular data.”
Show all 23 chapters
Future of Predictive Models
23:14 to 24:24
Envision a future where multiple models work together for enhanced predictions.
“And then you could dispatch a tabular model from that orchestrator, like you said, to go run data on your business if you're working with an agent, for example.”
Interacting with Tabular Models
24:24 to 26:48
Learn how to interact with Seldon and Neurolc for prediction tasks.
“Obviously, we've built MCPs and skills to allow any agent to make prediction using our predictive model.”
Unexpected Use Cases for Predictions
26:48 to 28:04
Discover some surprising applications of tabular foundation models.
“of course people will start with churn, demand, risk, and all of the operational problems.”
Operational Predictive Models in Finance
28:04 to 29:04
Learn how predictive models can be integrated into finance workflows for better data analysis.
“So getting back to the operational side of things.”
Understanding the Context Window Problem
29:04 to 30:38
Discover the challenges posed by context windows in data analysis and how new models address them.
“Previously, without me having the full context on how language models aren't fully fit for tabular data, to me, the most difficult part about using a language model to try to analyze data is the context window.”
Replacing Traditional Models with Foundation Models
30:38 to 32:09
Explore how foundation models can transform prediction processes in businesses.
“We can scale to 20 million rows and more than 600 columns.”
Impact of AI on Business Decisions
32:09 to 34:24
Examine the potential economic impact of AI-driven decision-making in companies.
“This is a huge transformation of the economy that we are pushing.”
Challenges in Model Adoption
34:24 to 36:29
Understand the hurdles faced in adopting new predictive models compared to traditional ones.
“and deliver the best precision yeah well pretty pretty huge impact yeah totally so what's the What's the biggest barrier to you?”
User Interaction with Predictive Models
36:29 to 38:44
Learn about the evolving role of users in data interpretation alongside advanced models.
“I would love to use it even just analyzing our YouTube data and putting that in a spreadsheet and saying, hey, predict this out six months from now and help us improve.”
Benchmarking AI Models for Predictions
38:44 to 41:25
Discover how benchmarking helps compare AI models and their effectiveness in various industries.
“So you can see all the different models performed for finance, healthcare, retail, and so on.”
Predictions for Enterprise AI by 2030
41:25 to 42:05
Hear predictions on the future of enterprise AI and its implications for various industries.
“All right, before we wrap up, give me three predictions for enterprise AI in 2030.”
The Future of Automated Workflows
42:05 to 44:44
Explore how automation will disrupt workflows and industries.
“like energy or telecommunication where you have critical data, but like most of the workflows, 90 % will be automated with orchestrators.”
Promoting Innovative Technologies
44:45 to 45:24
Learn about resources available for understanding advanced models.
“Well, Alexander, thank you so much for joining us.”
Transcript
Automatic transcript. May contain errors.0:00The Neuron: AI Explained Host:Do you think we're wasting billions and billions of dollars by trying to scale language models to do a task like what you're doing with tabular models?
0:07Alexandre Pasquiou:Language models are not adapted to making prediction and tabular data. They are great orchestrators, they are great UI. They help accelerate productivity, increase productivity of people. They structurally fail on this predictive task. If you look at any industry, any company, they will have these predictive problems. And these people that need to make predictions, they are the ones that just can't make it. And think about all the money that has been lost because the company was not able to predict the customers that will churn in this time. And think about the gain that could be generated if they were able to better predict the churn.
0:45Alexandre Pasquiou:If I have to make a prediction in two or three years, this tabular foundation model, they will be everywhere. Every company that will need to make a prediction will be using this technology. You'll be able to better predict your demand, better predict what you should offer, at which price, in which region. It will be like having the brain of your company that is super efficient, autonomous, that can act at scale and deliver the best precision.
1:11The Neuron: AI Explained Host:Welcome humans to the latest episode of the Neuron podcast. I'm Grant Harvey, writer of the Neuron Daily AI newsletter. And today I'm joined by Alexander Pascu, CEO and co-founder of Neuronk. Neuronk is building Selden, a foundation model for tabular data, meaning the structured information businesses store in spreadsheets, databases, rows, and columns. The company's larger bet is that businesses should not need to train and maintain a different predictive model for different things like churn or fraud, demand, pricing, risk, whatever, and every other use case you can think of that you'd want to predict.
1:47The Neuron: AI Explained Host:Instead, there should be one general model that could adapt to many prediction problems. Alexander, welcome to the Neurot. Great to meet you, Grant. And thanks for having me. Of course. So I guess to start off, I heard a story about your PhD days where you were working on studying brains when people played video games. Is that correct?
2:11Alexandre Pasquiou:Yeah, I spent a lot of time during my PhD playing with fMRI data and transfer-based architecture. So I was trying to better understand the human brain by using language models especially. So human brains are black box and there is so little we know about them. And one idea that was underlying my PhD is maybe we can do some mapping, some correlation between the neural network that are in your brain and the neural network that compose AI models like language models. But the two are black boxes. And so to advance, I had to make one of the black boxes a bit more transparent. The only one that could be more understood was the language models.
2:55Alexandre Pasquiou:And so I spent most of my PhD training language models, so making them learn specific things. And so the goal was to control for the information that was learned by the language models. I had to control the architecture, control the pre-training data that I knew that I was training models that learned about semantics, about syntax. And this allowed me to identify brain regions that are involved in the processing of syntax, so the structure of sentences, semantics, so the understanding of words, meanings. And I also did a bit of decoding, which is reading the brain. So it's kind of futuristic experiments where we had a human playing a video game in an MRI machine.
3:40Alexandre Pasquiou:And from the fMRI brain activation, I was reconstructing the video gameplay the person had in front of the eyes. So pretty cool stuff, yeah. So is it accurate to say you can read people's minds? For that, I would need fMRI data. I would need a world model rebuilding what you're looking at. and a lot of requirements before that.
4:05The Neuron: AI Explained Host:Fair, fair, fair, fair. Well, I'm really excited to talk to you because I'm very intrigued by this concept of tabular language models. How is a tabular language model different from a large language model? Or I guess I should say a tabular model different from a large language model.
4:22Alexandre Pasquiou:Maybe there's first a broader differentiation that needs to be made between these two models. In fact, there are, I would say, two lanes of AI. They are the data-rich people. So the model train on vast amount of publicly available data. Think about language models, audio model, image model. And so these are models that you just, these are categories for which you just need to scrap the internet. You will get enough data to pre-trained foundation models. And there, the biggest mode is money, because more money means more GPU, means you train bigger models. And this is exactly the fight that OpenAI and Anthropik are fighting for.
5:08Alexandre Pasquiou:The other lane in AI are the data-poor people. So the ones that don't have any data, you cannot scrap the web to find data, and you would find autonomous driving, robotics, AI for science, and predictive AI. us. And these are categories where you need to generate data, synthetic data, that you can use to pre-train a foundation model because there is not enough freely available data to train at scale these foundation models. And here the main mode is about data as you can guess. And the one that will have access to the data and to qualitative industry like data will be the one that wins the category, building the best
5:53The Neuron: AI Explained Host:models. So if today's language models become 10 times larger and 10 times cheaper, let's say, which business problem would they still be fundamentally bad at solving given the problem that you just described?
6:09Alexandre Pasquiou:So yeah, given this differentiation between data rich and data poor category, you also have to understand that each foundation model is specific to a modality. language model we're trained on text vision model we're trained on images, multimodal vision language model we're trained on both a combination of text and images, but if you want to use a foundation model to process a modality on which it hasn't been trained then it will fail try to use a language model to generate an image without training it on image data, it's impossible the reverse operation is the same and tabular data which is at the core of every company in fact you look at any company in the world they run on tabular data all their operations churn, risk, fraud, demand they all are based on tabular data on classic machine learning models and this is what they use to predict and to make the best decision to run the strategic decision and to run the business.
7:24Alexandre Pasquiou:And so this table of data, they have, of course, information in each cell that could be numbers, categories, text, images, whatever, but there is also implicit information that is encoded in the structure of the dataset. The co-occurrences of values on a given rule or on a given column makes sense. And this information is quite is super useful when you try to make a prediction. And this information is captured by current machine learning algorithm, but it's not captured by language models. Language models, in fact, will completely flatten the structure and convert it into one single sentence, one string.
8:08Alexandre Pasquiou:So you will retrieve the core semantic components of each cell, but first you will destroy the semantic of the structure, and you will destroy the semantic of numbers because language models are quite bad at numbers. In fact, because they have a fixed vocabulary, a vocabulary of tokens in which they operate to generate the next token given a sequence, they have a fixed vocabulary of tokens, so they cannot represent any numbers that they will face. They have to split it into a group of numbers and map it to their token of the vocabulary.
8:49The Neuron: AI Explained Host:Is that because tokens themselves are numbers or is that just a structure of how you train it?
8:53Alexandre Pasquiou:So tokens are in fact an ID that relates a series of characters to a vector. In fact, any model, any foundation model works in the latent space, which is a high dimensional space composed of numbers. and this is the only way you can do operations you cannot do operation on characters you do operation on numbers and so you have to find a mapping between a pixel to numbers or a token a series of characters to a numbers or a category to numbers this is how any foundation model will operate and so language model to work to process text or images in the case of multimodal model they need to transform the input into this latent space of vectors.
9:43Alexandre Pasquiou:So they work by having this vocabulary of tokens, which is not adapted to tabular data. The way a tabular foundation model will process information is by projecting every cell directly into the latent space without choosing vocabulary.
10:02The Neuron: AI Explained Host:So what happens when you take, let's say, I'm copying data from a table in my spreadsheet into, let's say, ChatGPT or something, and it still retains all of the information, or it seems to, because it's able to work with it. But what happens when you flatten that structure like you're talking about? What's the negative consequence?
10:26Alexandre Pasquiou:So with language models, you can interact with tabular data. You can do text to SQL, for example, that will allow you to aggregate past information, to do some visualization. This is what we call descriptive analytics. But one of the core weakness in today's AI, in this language model, is the inability to understand the raw distributions of a given data set. I mean, they cannot do statistical inference. They cannot adapt to a given data set from a given client. and generate prediction at scale, they just memorize the trending distribution of tokens. Language model generates the next token given the sequence of previous tokens, but they won't be able to adapt to a new distribution.
11:16Alexandre Pasquiou:And when you predict on tabular data, each table is really specific, really different from one another. A concrete example would be, let's say you want to do product categorization. So you have tables where each row represents a product. You have the title, description, a few attributes, and you want to predict the category. And this will be useful for the retailer to know where to put it on their website or in the store. If you take three retailers, like Amazon, Walmart, and Carrefour, they will all have a different taxonomy of categories, meaning that for the same product, they will label them, categorize them differently.
11:55Alexandre Pasquiou:and so you see that when you try to do this predictive problem you will have to adapt to each retailer's taxonomy it's the same when you do fraud detection for example a given transaction in a given country made in a given context might be quite different in another country with another context if you you send me one million dollar that might be some fraudulent activity on an account while if just Bezos make a 1 million transaction that could be considered as non-fraudulent. Him getting coffee. So yeah, you need to adapt to each dataset, to each distribution and this is a specificity of Tabular Foundation model compared to large language model.
12:45Alexandre Pasquiou:They do meta-learning, so they learn to learn during the pre-training so they adapt to each new dataset and they can make a prediction that are adapted to a client to a dataset.
12:58The Neuron: AI Explained Host:So is that sort of equivalent to, I believe the term, you can correct me, test time compute or learning and inference time where it's actually, let's say you're training it on the dataset for Amazon in this case, you do some sort of training run to fine-tune it for that data? Or is that inherent in the structure of the tabular model? And maybe to answer that, it would be helpful to differentiate how structurally the tabular model is different from the language model.
13:29Alexandre Pasquiou:Yeah, so they're both based on transformer-based architecture and transformers. So a stack of layers that will compute attention mechanisms across different tokens and with multi-layer perception for the classic stack. The difference is that the tokens in transformer-based architecture, and there are different architectures in the ecosystem, some of the architecture will tokenize your table by associating each cell to a token, and others will associate a given row of your data set to a token. You can have a mix, of course. So everything is possible. The difference is in this structure, meaning that language model, you have a very specific set of tokens which are part of the vocabulary of the language model.
14:23Alexandre Pasquiou:Tabula foundation models, the tokens are not fixed in a vocabulary. They are point in a high dimensional space. And so you have a wider variety of tokens. And they relate to the structure of the table itself. They represent cells or rows.
14:41The Neuron: AI Explained Host:Does this also solve the problem of traditional language models not really having any concept of time? Like in theory, could a tabular model have a stronger concept of time because it can track time through the rows and the columns?
14:57Alexandre Pasquiou:Depends on what you mean by having concept of time. If you mean like you want to transform it to tell you the time or if you want...
15:06The Neuron: AI Explained Host:No, more so for predicting into the future and knowing how things change and that sort of thing.
15:13Alexandre Pasquiou:So it really depends. It depends on the architecture. Some tabular foundation models that are made for classification or regression, meaning predicting discrete or continuous value from observation, most of them, they do not encode time information. So they won't have a positional encoding, while time series foundation model will natively encode this time information. Like I said, the answer is more complex, meaning that even for regression model, you could include time as a feature. But in fact, it all depends on the architecture. Language models, they have a position encoding. They can encode the position of each token.
15:54Alexandre Pasquiou:So you know that this token follows this one, follows this one, and so on. In tabular foundation model, you want to avoid that because you want to have the same prediction whatever the orders of the columns or the orders of the rows. So you want it to be positioned invariant, whatever the shuffling of the rows or columns. So you want to remove this encoding. For time series, you want to have the encoding because you want to encode the information that this point follows this point and so on. So these are all architectural choices depending on the application that you want to build.
16:29The Neuron: AI Explained Host:So how did you go from your PhD where you're looking at fMRI data and trying to predict the human brain to decide, you know what, we can't do this with language models, we need to do this with tabular models? How did you come across this as the path that you wanted to pursue and you thought we needed? Yeah,
16:53Alexandre Pasquiou:during my PhD, I applied transform-based architecture to so many modalities, text, images, I believe in world model to reconstruct these video game plays. And so I dissected in all directions this transformers architecture, all the attention mechanisms, all the positional encoding, the layers that I torture in every way that you could think of. And I was so surprised to see that there were absolutely no application of this transformers architecture to structure data in particular to tabular. All the applications were for unstructured modalities. So think about text, images, audio. And before my PhD, I did one year of data science.
17:37Alexandre Pasquiou:So I knew that one of the biggest problems today in AI, the fact that companies run on prediction and their prediction has been made by the same tools for the last 15 years. And why do they make prediction? In fact, like all the business owners, so the people that have business knowledge, budget data, that makes predictions that drive their rent, the economy. And we're talking about the entire economy. So that's a huge market. They need some insight about the revenue that is generated, the demand, how much stores they should open to put their product. So all the operational information.
18:21The Neuron: AI Explained Host:Or think about the federal bank, like trying to make predictions about inflation. and whether or not they should change interest rates. They need to be able to make sure that the data they have and the prediction models they use are incredibly... Exactly.
18:34Alexandre Pasquiou:If you look at any industry, any company, they will have these predictive problems. And these people that need to make predictions, they are the ones that just can't make it. So what do they do? They ask for the data science team to do it for themselves or they go to external consultants to build a model that would be specific to one question and one data set. And this is a super laborious problem because it usually takes six months to one year or even more just to go from the idea of I want to build a churn model to I have this churn model in production. And think about all the money that has been lost because the company was not able to predict the customers that will churn in this time.
19:24Alexandre Pasquiou:And think about the gain that could be generated if they were able to better predict their churn. And so you have to look at this impact of one use case, which is usually tens of millions up to billions for the company per year, and multiply it by the number of use case and the number of companies that are in the world. And so we're talking about a market that is huge. and data sense is still today some consulting it's not there is no predictive infrastructure like you have for text or for images there is not a this unified model that this unified infrastructure that will allow anyone to interact with this tabular data and this is what we are building we are building this predictive infrastructure to allow any business owners to make prediction at scale so that they win months, year of time before to act and to take decisions.
20:23Alexandre Pasquiou:And we also allow us to make better prediction at scale with our models because these tabular foundation models, they are not limited to the information that is available in the dataset. As we train them on billions of datasets, their knowledge compounds and they are able to transfer this knowledge to the data set of the clients, meaning that over time, our model keeps increasing, getting better and better, while all the classic machine learning algorithms, they just cap to the information that is available in the data set of the client.
21:02The Neuron: AI Explained Host:So why wouldn't a company switch over to this? They just don't know it exists? Why are so many people still stuck in the past with these previous models? Just because they've worked for them? Because it's a recent disruptive technology.
21:17Alexandre Pasquiou:Fair. So the companies building this tech only existed for like two years. And two years is the creation of the company. So let's say that the tech has been existing for the last year. And you know that everything, especially when it's usually disruptive as an inertia, And so it takes time to educate the market, to build the product around the core technology. And yeah, but this is in a good way. There is a lot of traction. We clearly see the impact of what we're building, whether because it accelerates the decision making or because the ROI that generates from better prediction is clearly a no-brainer for the company.
Read the full transcript
22:06Alexandre Pasquiou:No, no, this will completely disrupt the market. And if I have to make a prediction in two or three years, this tabular foundation model, there will be everywhere. Every company that will need to make a prediction will be using this technology.
22:25The Neuron: AI Explained Host:So I'm going to put you on the spot for a hot take. Do you think we're wasting billions and billions of dollars by trying to scale language models to do a task like what you're doing with tabular models? Definitely. Yeah.
22:38Alexandre Pasquiou:Language models are not adapted to making production and tabular data. They are great orchestrators, they are great UI. They help accelerate productivity, increase productivity of people, but they're not, they're structurally failed on this predictive task. you can do whatever you want to to adapt your language model to tabular data in the end they're not made for it so it will be super costly way less efficient than even classic machine learning models and you would just lose money time and uh and ROI yeah totally so do you
23:22The Neuron: AI Explained Host:think then we're going for a future with uh let's say multi-model future where you have you know Let's say like the interface of a language or a multimodal model that you talk to. And then you could dispatch a tabular model from that orchestrator, like you said, to go run data on your business if you're working with an agent, for example. How do you see this all playing out?
23:46Alexandre Pasquiou:No, definitely. The economy of tomorrow will be agentic. And so we will see different foundation models for different modalities. But I guess the best way to interact with them will be through natural language, like we do when we communicate between each other. So a language model will be the interface, the orchestrator of these different foundation models. But for the different application, you will find a vision model if you need to interact with images, if you need to generate images, modify them. Tabular foundation model if you need to interact with structured tabular data and other foundation models for the different modalities.
24:21The Neuron: AI Explained Host:So can the two interact? Can a large language model, like can a tabular model, specifically Seldon, can it interact with text or chat?
24:32Alexandre Pasquiou:Obviously, we've built MCPs and skills to allow any agent to make prediction using our predictive model.
24:42The Neuron: AI Explained Host:That is so cool. Yeah, tell me a little bit more about Seldon and how it works. So if I go to neuralc.com, I see that there's a box that lets me chat. Am I chatting with Selden when I enter, help me predict the 2026 midterm election results?
24:59Alexandre Pasquiou:We have this core predictive platform that hosts our different foundation model for classification, regression, and forecasting. And you can interact with this model in different ways. so we have SDK where you can just code in Python and interact directly with our models we have Excel plugins so if you want to use our model directly in your Excel sheet to make predictions it's quite simple to use we have MCPs and skills so you can just ask Claude or Chargpt predict this and this from this data set and they will be able to do it in a few seconds it will have taken like a few hours, days or weeks for a data scientist and we have a platform also that we sell to a client that allows them to interact to predict in a few clicks connect the data from the problem and get prediction.
26:00Alexandre Pasquiou:So we have different ways of using it but in all the case for now you need to have the data on which you want to make a prediction. So if you want to predict the outcome of the next election, either you have the data or you need an agent to scrap the data for you to make the prediction.
26:19The Neuron: AI Explained Host:Right. You'd have to do a lot of surveys to get informative information on that.
26:26Alexandre Pasquiou:That's a really good point that you're making because with coding agents, I think no one was really anticipating all the application. People were thinking it will accelerate developers, but in the end, even non-technical people can build their own application. That's going like way beyond what was anticipating by OpenAI or even Anthropic. For tabular foundation models, of course people will start with churn, demand, risk, and all of the operational problems. But in the end, they might also use them for a use case that we did not anticipate. For example, using an agent to scrap web data, build your own data set, and then build your own custom predictive problem where you will use our tablet foundation model to make a prediction.
27:19The Neuron: AI Explained Host:Have you seen any interesting unexpected use cases from people who've already been using Selden and Neuralc right now? Have you seen anything that totally surprised you on your platform?
27:29Alexandre Pasquiou:For now, most of the use cases that we've seen are mostly related to operational predictive problem inside companies. We've been using it for, let's say, more extravagant use cases, like predicting flight delay by scrapping airport data. We use it to predict the World Cup winners. Oh, amazing. And I think we'll use it for other predictive problems like that. It's more about marketing in this case. but yeah the core of the value so far is mainly based on operational predictive problems
28:10The Neuron: AI Explained Host:right which makes perfect sense on the world camp front did was it able to predict spain or did it think uh france was going to win and by the way my apologies to france they were my pick
28:21Alexandre Pasquiou:i thought france was going to win yeah a lot of people were predicting uh france we we predicted Spain. But yeah, that was a close match.
28:30The Neuron: AI Explained Host:Okay. So getting back to the operational side of things. So let's say I'm a finance, you know, someone working in finance and I'm already using Claude's financial plugins to do a lot of my work. I could right now add the Selden or Neurok MCP, and I would be able to actually take my spreadsheet, take all of my data, maybe I have an external data source I've connected, and send off a prediction right now.
29:04Alexandre Pasquiou:Yeah, exactly. It's super simple to use. That's amazing. You can either use it directly into your Excel sheet and you select by end the data that you want to use to contextualize the model and to make a prediction on, or if you use MCP, you can just plug your cloud on it and your cloud for Excel will use it directly in the backend and you will have even less steps to do by yourself.
29:31The Neuron: AI Explained Host:Previously, without me having the full context on how language models aren't fully fit for tabular data, to me, the most difficult part about using a language model to try to analyze data is the context window. So the context window, to me, was the number one biggest barrier, and specifically the lost in the middle problem. So just because you have, let's say, like 1 million tokens of context doesn't mean that when you run Claude to analyze a spreadsheet with that much data inside of it, it's actually going to analyze all 1 million of those tokens. So how does Selden, or in particular the models that you've built, how do they address the context window problem?
30:19The Neuron: AI Explained Host:And what is the limit with how much context you can put into these things?
30:23Alexandre Pasquiou:And this is exactly one of the biggest reasons why you can't apply long-winded model to prediction at scale. We have an architecture that is way more scalable. In fact, Selden is the most scalable model that you could find and the fastest one. We can scale to 20 million rows and more than 600 columns. So you have to multiply 20 million per 500 and you will get the order of magnitude of number of tokens that we can process, which is way beyond any language models. And this tells you the discrepancy in terms of tech that is between OpenAI, Entropic, and what we're building.
31:04The Neuron: AI Explained Host:So I guess let's talk very quickly about the general nature of this, right? So you mentioned earlier that lots of companies have traditional statistical machine learning models that they've used to do a lot of this analysis for a long time. And they've also perhaps hired consultants or they have internal teams to create their own custom models. So if one foundation model can replace hundreds of custom predictive models, how does that change the nature of the business of prediction?
31:37Alexandre Pasquiou:Did coding agents replace developers? No. No. So this is exactly the same. You still need this expertise and these people to source more data, frame the problems, push the adoption of this tech internally to business owners. So you don't just replace people. You change the ecosystem. In fact, we are changing the paradigm of how companies are interacting with their data and how they are making protection. This is a huge transformation of the economy that we are pushing.
32:13The Neuron: AI Explained Host:Do you see it kind of similar to how the formal verification models in math are kind of bringing a lot of new people into math, where now all of a sudden anyone can spin up some of these models and try to solve unsolved problems or verify proofs on the math side of things? Obviously, some people are very naive when they're attempting to do this, but do you see it having a similar effect there?
32:37Alexandre Pasquiou:I think these are two really different problems. First, on the scale of the market that is at rest, I think what we're building is way bigger. And you won't put a formal solver into the end of everyone, but everyone is making decisions inside a company and needs to make predictions to decide. So this is why what we're building is core to the economy and will be quite disruptive for all the companies you can think of.
33:09The Neuron: AI Explained Host:what happens if every company is using your model do we reach some sort of like economic equilibrium
33:18Alexandre Pasquiou:that's a good point i didn't think of that let's let's uh do a follow-up meeting in two years to see what happens yeah i think that's a good idea yeah or maybe you just make custom variants
33:32The Neuron: AI Explained Host:you know and then you start to switch things up just to add some variety and spice a life
33:37Alexandre Pasquiou:But imagine that a company accelerates every decision by months or even, let's say, six months. That's a huge impact. First, you can predict your financial flows, which is pretty useful, which will avoid any bad ending for a lot of companies. companies uh you'll be able to better predict your demand better predict uh what you should offer at which price in which region if you need to open a store recruit people so it will be like having the brain of your company uh that is super efficient uh autonomous that can act at scale and deliver the best precision yeah well pretty pretty huge impact yeah totally so what's the
34:33The Neuron: AI Explained Host:What's the biggest barrier to you? We talked about data. As you're training new versions of your models, what is your biggest hurdle that you're trying to overcome?
34:45Alexandre Pasquiou:I think unlike language models, where you had absolutely no baseline, before language model, you couldn't interact with text. You couldn't interact with images. So even the first iteration, you know, the first version of, I don't remember the name of the models, like the first VIT, Vision Translators, that were generating images that were absolutely horribles. But yet it was impressive. And people were like, wow, I can generate an image of Donald Trump even if he has four eyes. That's amazing.
35:19The Neuron: AI Explained Host:Right.
35:20Alexandre Pasquiou:But when you're doing prediction, you're addressing the core problem of every company and you need to make it right. And the baseline are way stronger because you're facing the internal models that have been optimized for years by data scientists. And so even if you look at language model, the real business enterprise application, they're quite recent. In fact, the end of 2025, and before that, there was mostly POCs that did not convert to anything in production. so whether you're talking about foundation models for text or foundation models for tables there are high barriers to the entry you need a sufficient level of quality before being able to deliver to integrate and go in production but this is this is the way it will go and we our model keep improving, they will be better and better in six months, in one year.
36:27Alexandre Pasquiou:And at some point, it will be a no-brainer for anyone to adopt it, even the bigger companies.
36:33The Neuron: AI Explained Host:Yeah, it makes sense. I would love to use it even just analyzing our YouTube data and putting that in a spreadsheet and saying, hey, predict this out six months from now and help us improve. I can see so many different applications of it.
36:47Alexandre Pasquiou:It's freely available. You can just go to a website. There is a documentation you can implement. You can integrate the Excel plugin or the MCP servers and play with it. So feel free to use. Thank you.
37:00The Neuron: AI Explained Host:Okay, a couple of lightning round questions before we wrap up here. So let's say I'm using CELTON and I am using it to form predictions for me or analyze my data. where do you believe the role is of me looking at the data and interpreting it? Where does that role begin and end when I have these production models?
37:25Alexandre Pasquiou:So what is the role of the model and what is the role of the users? Yes. And where do they diverge? In fact, the model takes as an input the data and output prediction. And so today, foundation model, they still need the data scientist to prepare the data. For the next version of our model, they won't. So we'll be able to plug directly the data, the model onto the raw data and just output the prediction, which is pretty powerful because it's like 80 % of the time of the data scientist today is spent on data preparation and feature engineering. And this is the most laborious part where you need this field expertise of what really matters for the prediction.
38:13Alexandre Pasquiou:And we will do it natively with the first layers of the model that will be able to natively build implicit features. Beyond that, there is not much involvement from the users. Just send the data and receive prediction, especially with all the layers of all the wrappers and the way to consume our models that we've built. it's super easy to use and this is all the feedback that we get which is that it's just three clicks from getting your protection once you install the plugin or the MCP Wow
38:47The Neuron: AI Explained Host:have you been following, I know there's at least one benchmark that tracks all of the models and how good they are at making predictions, have you been following these benchmarks at all and if so, how does what you've built stack up?
39:02Alexandre Pasquiou:so like for every modality every model you you need indeed to build benchmark to compare to one another we have our own benchmark which is one of the most extensive one tab bench everything is reproducible we have public dial set all the foundation model are included in it and we have like almost 200 dial set for classification which is quite extensive. We split it by industry. So you can see all the different models performed for finance, healthcare, retail, and so on. And we did it because we wanted to be as clear as possible for any data scientist that needs to select a model. Whether they're in finance, healthcare, telecommunication, we wanted them to get an understanding of what model they should use.
39:58Alexandre Pasquiou:If you look at the different benchmarks for language models, they're quite complex, and I think no one will select their language model based on this benchmark.
40:09The Neuron: AI Explained Host:I think of trying to compare benchmarks with models to the difficulty of trying to compare multiple products and people sort of default to, you know what, I would rather just get word of mouth. Like, what does my friend in my industry use? And I'm going to use that. Like, just to a certain point, it's like it's uninterpretable by the average person. And you'd rather just someone, yeah.
40:35Alexandre Pasquiou:It's super empirical. Like, everyone will decide based on their usage. They will test one, two, three models, and then they will keep the best. And the most, let's say, adventurous one will do this comparison every week to always have the best. And I think like 90 % of the people will just stay with the one where they subscribed a few months ago. And this will be insufficient.
41:02The Neuron: AI Explained Host:But what I like about what you've built is that it's integratable with whichever one you're using, whether it's ChatGPT or Claude as your, you know, call it daily driver AI agent.
41:12Alexandre Pasquiou:I think this is a key component for every foundation. You need to be pluggable everywhere. otherwise it would be easy just to change a tool, change a model. So we need to be everywhere to be used everywhere.
41:28The Neuron: AI Explained Host:All right, before we wrap up, give me three predictions for enterprise AI in 2030. What do you think is almost certain? Maybe one that most AI researchers might disagree with, so out of distribution take, and one that you're betting your company on.
41:43Alexandre Pasquiou:First, easiest prediction, which is in 2030, all productive workloads will be based on tabular foundation models. The second one will be that all, like 90%, I wouldn't pronounce myself for the critical industries like energy or telecommunication where you have critical data, but like most of the workflows, 90 % will be automated with orchestrators. I'm talking about agents. Agents will run like most of the workflows. And so the last consequence of this second point is that we will see a massive disruption of the economy because we will completely shift the scale of progress. And it could go really wrong.
42:50Alexandre Pasquiou:It could be a good thing. And for that, I don't have a prediction. It's fair. It's fair. Because it's at least one of the best.
42:58The Neuron: AI Explained Host:What do you think is the barrier there? Like when will a tabular foundation model or something you've built be reliable enough and trusted enough by critical industries to be able to use them? What is the true barrier to those really, really mission-critical industries using a tool like this?
43:22Alexandre Pasquiou:I think the first barrier, a barrier that we already destroyed, is that this company needs to be able to host the models themselves on their servers. And this, we can already offer it. You can deploy Selden anywhere on any cloud provider, so even on-premise. that's the key yeah and that's quite different from language model if you want the best language model you wouldn't be able to afford to deploy it yourself and the second one regarding the quality threshold there is no general answer to that and it will depend on the use case, it will depend on the data of the company, it will depend on the industry so unfortunately there is no answer but this means that for each of these critical application, there would be a need to build partnerships, to iterate with them on the data, to improve the model, be sure that we have the right benchmark to assess the quality of the prediction, and that in the end, it's good enough to be deployed.
44:25The Neuron: AI Explained Host:Yeah, that makes sense. It's like a testing, you know, you're iteratively testing it. You're not going to rush into that. You're going to make sure, hey, do we know that this is going to work? Because if it doesn't work, it's a lot of trouble.
44:38Alexandre Pasquiou:Exactly, yeah. but we're pretty optimistic about the future.
44:43The Neuron: AI Explained Host:That's awesome. Well, Alexander, thank you so much for joining us. For everyone listening, where can they find more about what you're building?
44:53Alexandre Pasquiou:On our website. We have lots of documentation about the different ways to use our models, but we have also blogs where we explain what are these stable defundation models, why they will have a huge impact on the economy, And we even enter details. So we regularly post technological blog posts to explain the tech behind what we're building, the application, and the impact it has on the business. Awesome.
45:24The Neuron: AI Explained Host:Thank you very much. For everyone listening, you can find more practical AI news and explanations at theneuron.ai and our daily newsletter, which you can subscribe to with the link below. Follow the podcast wherever you listen. Give it a like if you're on YouTube. And we'll see you next time. Farewell, humans.
From the publisher
AI can write code, summarize documents, and hold a conversation. But it still struggles with the structured data businesses rely on to predict churn, fraud, demand, pricing, and risk.
Alexandre Pasquiou explains why language models flatten the relationships inside spreadsheets, how tabular foundation models learn from rows and columns, and why Neuralk believes one general model could replace hundreds of custom predictive systems.
The conversation also covers Seldon, AI agents, enterprise adoption, and Alexandre’s prediction that tabular foundation models will power every predictive workload by 2030.
Learn more about Neuralk: https://www.neuralk.ai/
Subscribe to The Neuron newsletter: https://theneuron.ai