Pattern Sense

A spreadsheet of sales figures does not look like a conversation. There are dates in one column, numbers of items sold in another and perhaps temperatures in a third. We can add the numbers up and draw a curve. Predicting next month’s sales begins a different kind of work: Which factors matter, how are they connected, and which method should we use?

That work can now begin somewhere else. With an AI that has already learned something about how sequences can develop over time, before it has seen our figures.

Google’s TimesFM-3 is trained on more than a trillion time points from real and artificially generated time series. It can take a new sequence of numbers and produce a forecast without first being trained specifically on it. It can also combine several series and incorporate information about upcoming events. In Google’s example, planned promotions are included in a forecast of ice cream sales.

There is no conversation in that example. No friendly voice explaining itself. Even so, there is something familiar about the situation: we give a machine a context it has not encountered before and receive an informed estimate of what comes next.

That was also how many of us first experienced language models.

Language was not the boundary

A language model can continue a text that has never existed before. It has learned regularities through its training and applies them to the context we give it. The result can be an explanation, a poem or a piece of code. How well it succeeds varies, but by now we have grown used to the ability itself.

I call it pattern sense. The phrase captures something easily lost when we speak only of recognition: that a learned capability can be used in a new situation. The model need not have seen this particular text or this particular sales curve to offer a useful estimate.

It is a metaphor. There is no being sitting there with a hunch that Tuesday will be a good day to sell ice cream. Technically, we are still talking about computation and learned statistical relationships. But from the user’s perspective, the result can resemble getting help from someone with a feel for that kind of sequence.

The connection to language models runs right into the architecture. Both TimesFM and Google’s tabular model TabFM are built on transformers, the family of neural networks behind most modern language models. One central mechanism is called attention: the model calculates which parts of the input are relevant to one another. In text, these can be connections between words. In TimesFM-3, this happens both along time and between different series of numbers.

The numbers are not simply typed into a chatbot. The systems are built for their data types. TimesFM, for example, groups time points into small blocks that the transformer can work with. The shared idea takes a different form when the material changes.

We have devoted a great deal of attention to machines learning to write. Now we can see related abilities emerging in places where there are no sentences at all.

A sense practised on something invented

TabFM works with tables. It can take rows with known outcomes and use them to predict the outcomes for new rows. This can be a number or a category, such as whether a customer is likely to leave a company.

The known examples are supplied as context. The model adapts its answer to them without changing the parameters it learned during its original training. It therefore still receives information about the specific task. The difference is that we do not have to build its ability to work with tables from scratch each time.

The most surprising detail is where that ability comes from. According to Google, TabFM is trained exclusively on hundreds of millions of artificially generated datasets. The researchers constructed different relationships between variables, used them to generate tables and gave the model tasks involving predictions of unknown outcomes.

It has practised on something invented and can subsequently apply some of what it learned to real tables.

That recalls the possibility I explored in The Childhood of Machines: that machines can build capabilities through artificially generated experience. Then, the subject was robots practising in simulated worlds. Here, the practice world can consist of rows and columns. It need not have floors, walls or gravity to teach the machine something that becomes useful outside it.

There are, of course, limits to what the constructed tasks prepare the model for. A table can be misleading, and the relationships in training may be poorly suited to the task we set. But the possibility is remarkable: we can train an ability to infer relationships without having access to all the real situations in which that ability will later be used.

It is happening in several places

Google is not alone in this development. Amazon’s Chronos also produces forecasts for new time series. In the original version, numerical values were converted into tokens, small units like those in a language model, and a transformer learned to predict continuations. The researchers called their work Learning the Language of Time Series.

In tabular data, TabPFN is an important predecessor. It too learns through synthetic tasks and uses known examples as context when making new predictions. Google itself points to TabPFN and TabICL as inspirations for TabFM.

Other researchers give transformers an entirely different material to practise on. InstaDeep and its collaborators developed Nucleotide Transformer, which learns representations of DNA sequences. After further adaptation, the learned capability can be used for different biological prediction tasks. Microsoft’s Aurora is likewise broadly pretrained on Earth system data and then adapted to tasks such as weather forecasting, air pollution and ocean waves.

These models cannot simply swap places. A sense for DNA does not automatically become a sense for customers’ purchases. Some can be applied directly to new tasks; others require further training. Even so, it is hard to look at them together without becoming absorbed in the development: pattern sense is spreading into more domains.

Nor is this happening only with transformers. Google’s 2024 research on flood forecasting described a different type of neural network, trained on data from thousands of river basins. It could use weather data and information about the landscape to predict streamflow in rivers without local measurement series for training.

This is a different and more limited form of transfer. But its consequence makes the idea tangible. What the machine has learned elsewhere can become a warning here. For people living along the river, that may mean time to move before the water arrives.

What do we notice now?

Machine learning has always been about learning patterns and applying them to new observations. Transfer learning and models trained on many time series also existed before these systems arrived. We are not at a sharp boundary between machines that could only remember and machines that can suddenly generalise.

But it makes a difference how much learned capability can be reused and how much work it takes to make it function somewhere new. With a traditional tabular model, training typically takes place on the particular dataset. With a pretrained tabular model, some of the capability is already there before work on the new table begins.

Imagine a small association with years of attendance records that has never been able to afford an analytics project. Or a workshop that stores measurements from its machines without getting much out of them. Better access to pretrained models could make those numbers worth investigating. The examples are possibilities, not documented results of the systems mentioned. The data still have to be usable, and predictions must be tested against something the model has not been given the answers to.

Even so, the threshold may shift. Where the question used to be whether one had the resources to build a solution, one can begin by investigating whether an already learned ability can help. More small questions then become worth asking.

In AI Finds the Connections We Overlook, I wrote about the possibility of AI opening paths through knowledge that people had not previously seen together. Tabular and time-series models show another path to a broader view. You do not have to be at the research frontier. There may be something to investigate in a spreadsheet that has been opened every Monday for five years.

A forecast does not necessarily tell us why something happens. If ice cream sales follow the temperature, that does not give us a full explanation of customer behaviour either. A machine can overlook a break in how the world works, just as an experienced person can. Pattern sense is therefore an estimate to work with, not an answer to obey.

But a good estimate can change what we try. The workshop may inspect a machine earlier. The association may plan an event differently. A researcher may choose an experiment that would otherwise never have been considered. The value also lies in the actions and questions a prediction makes possible.

We have grown used to AI answering when we write to it. That has made language our most visible image of the technology. Meanwhile, a family of methods has learned to work with far more of what the world leaves behind: measurements, tables and biological sequences.

The spreadsheet of sales figures is still a spreadsheet. No one has added a column containing certain knowledge of the future. But if we can get help from a pattern sense built through far more sequences than we could ever examine ourselves, there is suddenly more we can do with the numbers.

And there are many spreadsheets. Many measurements. Whole areas of the world that we have recorded more of than we have so far been able to use.