Andrew Ng: Unbiggen AI sounds perfectly plausible as a 2026 take. Small problem: the IEEE Spectrum interview was published in April 2022.1

That timing makes it more useful, not less. Four years have passed since Ng argued that some AI work should stop treating size as the automatic answer and start engineering the data more deliberately. We can check what happened rather than merely repeating the slogan.

The answer is messier than “Andrew Ng was right” or “Andrew Ng was wrong.”

In 2026, the most ambitious models still consume more compute. At the same time, the smallest model needed to reach a given level of performance has sometimes collapsed in size. Both trends are happening together.45

That is where unbiggen becomes useful. It separates two questions that are often folded together: how far can the frontier be pushed? and how large a model do I actually need for this job?

Fifty images

Ng was not saying in 2022 that large models had reached a dead end.

In the IEEE interview he explicitly says scaling still has “steam,” particularly for language foundation models and potentially video. He then separates that race from another class of problems: those without hundreds of millions of examples or a use case broad enough to justify a gigantic general system.1

His favorite example was industrial visual inspection. A factory needs to recognize a specific defect on a specific part. It may have only a few dozen good examples because the defect is rare, the line recently changed, or there simply is no public corpus of scratches on this component under this lighting.

Ng gives a deliberately uncomfortable number: 50 thoughtfully engineered examples can sometimes be enough to build something useful for that sort of task.1 It is not a universal rule, and certainly not a promise that any vision problem can be solved with fifty images. His example is a narrow industrial problem with a controlled domain.

The more interesting move comes next. If a dataset contains 10,000 images but the 30 examples of a rare class are labeled inconsistently, he would rather direct the team toward those 30 cases and make the labels consistent than immediately swap in a new model.1

LandingAI defines data-centric AI in similarly plain terms: systematically engineering the data used to build an AI system.2 That includes label consistency, example selection, augmentation, finding weak subsets and giving domain experts a practical way to correct the corpus.

It is less photogenic than adding another zero to a parameter count. It is also much closer to normal ML product work: inspect the failures, understand their source and change the thing feeding the error.

Chinchilla, already

The funny part is that in 2022, almost at the same time, DeepMind published a result pointing in a related direction without talking about “small data.”

Gopher had 280 billion parameters. Chinchilla had 70 billion. They used roughly the same training compute budget, but Chinchilla saw far more data: 1.3 trillion tokens, about four times as much as Gopher in the comparison.3

The model with four times fewer parameters beat Gopher and several larger models on nearly every benchmark DeepMind reported.3

That does not prove Ng’s industrial small-data thesis. Chinchilla is still enormous and its corpus is anything but small. It demonstrates a different point: even at the frontier, parameters and data share a budget. Making a model bigger while undertraining it can be a poor way to spend compute.

It feels obvious after the fact. A bigger machine is not automatically a better-fed machine. In 2022, DeepMind still had to demonstrate it by training hundreds of models across different parameter counts and token budgets.3

142 times smaller

Two years later, the Stanford AI Index gives us another thermometer.

In 2022, the smallest model in the report scoring above 60% on MMLU was PaLM at 540 billion parameters. In 2024, Phi-3-mini crossed the same threshold with 3.8 billion.4

Stanford calls that a 142-fold reduction in the size of the smallest model clearing the threshold in two years.4

The number deserves less celebration than it first invites. MMLU is a benchmark, not a universal definition of intelligence. The models do not share training data, recipes or product goals. “Same MMLU threshold” does not mean “same product” or “same capabilities everywhere.”

As a measure of efficiency, though, the direction is hard to miss. Something that required hundreds of billions of parameters to reach a particular test score in 2022 fit into a few billion parameters two years later.

That kind of reduction changes very ordinary decisions: memory, latency, local execution, inference cost, and how many requests a fixed budget can serve. AI progress is not visible only when the frontier moves up. It is also visible when a capability moves downmarket.

Meanwhile, bigger

It would still be a mistake to look at that drop and conclude that the industry followed Ng into a general slimming program.

The 2026 AI Index says almost the opposite at the frontier. Publicly reported parameter counts have hovered around the trillion range for several years, with disclosure declining at some labs, while estimated training compute has continued to rise.5

The same report estimates global AI compute capacity grew about 3.3× per year since 2022, reaching 17.1 million H100-equivalents.5

So no, AI did not “unbiggen” at the infrastructure level. We are building substantially more machinery for training and inference.

At the same time Stanford gives an unusually neat example of movement in the other direction. OLMo 3.1 Think 32B, with nearly 90 times fewer parameters than Grok 4 according to the report, reaches comparable results on several benchmarks through techniques including pruning, deduplication and data curation.5

Again, “several benchmarks” is not “equivalent at everything.” The number remains useful precisely if we do not ask it to say more than it does.

The word small

The trouble with “large model versus small model” is that it often chooses the size before choosing the work.

For a frontier lab trying to discover how far a general system can go across science, code, video, agents and multimodality, scale remains a lever. Nothing in the 2026 evidence says every frontier model can be replaced with a carefully labeled 3B model and everyone can go home early.

For a company that needs to classify five defects on a part, extract three fields from a document, detect an anomaly in a signal, or answer questions inside a closed domain, the question is different. It should start with the required quality, the real error distribution, latency, memory, cost, privacy and available data.

Then choose the model.

That may be the part of Ng’s proposal that aged best. “Data-centric” does not mean the model becomes irrelevant. It means the dataset stops being a file downloaded at the start of the project and treated as a constant afterward.

The corpus is part of the architecture. Its errors have effects. Maintaining it costs time. Its quality may allow a smaller model to work, or reveal that a more powerful model is still necessary.

A shifted budget

Four years after the IEEE interview, Unbiggen AI therefore reads as both right and misleading.

Misleading if it is heard as a prediction about the size of the industry. Data centers, compute and frontier systems have kept growing.5

Useful if it is read as a manufacturing rule: do not pay for size before you understand where the error lives.

Chinchilla showed that a model could be too large for its own data budget.3 Stanford’s tracking shows a performance level once requiring hundreds of billions of parameters moving rapidly into much smaller models.4 Recent systems add curation, deduplication and post-training to that equation.5

The work did not disappear. It moved.

That is inconvenient for presentations that want one number to represent progress. It is rather useful for people who actually need a system to work: there are more levers available than “buy the biggest model you can access.”