Artificial intelligence cannot run without code, because every AI system is software: a program that loads a set of learned numbers and applies them to new input. What can be skipped is the writing of code by the person using it. Browser-based tools such as Google's Teachable Machine let people train a small image, sound, or pose model with no programming, while the tool itself rests on code written by engineers.
Key takeaways
- A trained model is two things: code that defines the calculation, and numbers (weights) learned from data.
- GPT-3, described in a paper submitted on May 28, 2020, has 175 billion parameters, none of them typed by hand.
- No-code tools hide the programming from the user but do not remove it from the system.
- Skipping code usually means giving up control over data handling, model design, and testing.
- Before relying on a no-code model, check what it was trained on and how it was tested.
What "operating without coding" can mean
The question mixes up three separate ideas. The first is whether an AI system can exist without any program behind it. It cannot: a model is a mathematical function, and a computer needs software to evaluate it. The second is whether a person must write code to get a working AI model. Often not. The third is whether the intelligence comes from the rules a developer typed. In modern machine learning it largely does not, and this is the source of most of the confusion.
Code and learned numbers
In older software, a developer states the rules: if the temperature is above a threshold, send an alert. In machine learning, the developer writes a training procedure and a model structure, and the data fills in the details. The details are the weights, plain numbers that decide how strongly each input influences the output.
An illustrative example: suppose the training data are three pairs of numbers, (1, 2), (2, 4), and (3, 6), and the model is the formula output = w times input. The training code tries values of w and keeps adjusting until the predictions match the data. It settles on w = 2. Nobody typed the rule "double the input"; the data produced it. The code that did the adjusting, though, was written by a person.
At scale the same idea applies. The GPT-3 paper by Tom Brown and 30 co-authors describes an autoregressive language model with 175 billion parameters. No team could type 175 billion values, so they are the output of training, but the architecture and the training loop are code. The 2017 paper "Attention Is All You Need" introduced the Transformer design that many current language models use, and that design is a specification engineers turn into software.
Layers of an AI system and who writes each one
| Layer | What it is | Written by hand? |
|---|---|---|
| Model structure | The shape of the calculation, such as a Transformer | Yes, as code |
| Training procedure | The loop that adjusts weights to reduce errors | Yes, as code |
| Weights | The learned numbers | No, produced by training on data |
| Serving software | The program that loads weights and answers requests | Yes, as code |
| Interface | The web page, app, or chat window | Yes, or generated by a no-code builder |
Only one row is free of hand-written code, and it cannot work without the others.
What no-code tools actually do
Google describes Teachable Machine as a web tool that makes it fast and easy to create machine learning models, with no coding required. The user supplies example images, sounds, or body poses, the tool trains a model, and the result can be exported for use on websites and in apps. A classroom can, for example, teach it to tell two hand gestures apart by recording a few dozen samples of each.
The programming has been moved, not removed. Engineers wrote the interface, the training routine, and the export function. The user contributes data and judgment about whether the output looks right.
What is lost without code
Convenience has a price. A simple point-and-click tool offers a fixed menu of project types, and the user cannot usually change the model structure, inspect how data are split for testing, or add checks for rare cases. A model trained on 30 photos taken in one room may fail in another room with different lighting, and nothing in the interface forces that test. Skilled practitioners write code partly to control these steps. For a classroom demo the trade is reasonable. For decisions that affect people, such as screening applications or flagging medical images, the missing controls matter, and the regulatory and privacy duties of the organization stay in place regardless of how the model was built.
A short checklist before using a no-code model
- What data trained it, and does that data resemble the situations where it will be used?
- Was it tested on examples it never saw during training?
- What does it do with input unlike anything in its training data?
- Who is responsible for monitoring it after launch?
- Can its decisions be reviewed by a person when the stakes are high?
FAQ
Can AI write its own code and run without developers?
AI assistants can generate code, but they run inside software that people built, and their output still needs review. Generated code is still code.
Is a neural network just a pile of numbers?
The weights are numbers, but they only produce answers when a program multiplies, adds, and combines them according to the model structure. The numbers and the program belong together.
Is programming needed to work with AI?
Not to try simple models or to use finished tools. Programming becomes useful when a project needs custom data handling, careful testing, or integration with other systems.
Is a no-code AI tool less accurate than a coded one?
Not by definition. Accuracy depends on the data and testing. The difference is that coded workflows give more ways to inspect and improve both.
Related reading
- Whether artificial intelligence requires coding
- Automation without programming
- What AI cannot do
- Machine learning with little or no data
More in AI Basics.
Last verified: October 2026
Sources
- Brown T. et al., Language Models are Few-Shot Learners, arXiv:2005.14165, submitted May 28, 2020
- Vaswani A. et al., Attention Is All You Need, arXiv:1706.03762, submitted June 12, 2017
- Teachable Machine, Google Experiments, 2026