Why AI Progress Often Comes in Jumps
AI progress often looks strangely uneven. Months may pass without an obvious change, then a new release suddenly appears much more capable.
The improvement may be real, but it probably did not happen overnight. Public releases reveal only selected moments in a much longer development process.
A five-part series about what may shape future AI systems and why confident forecasts often fail.
From the outside, AI progress can resemble a staircase.
A model is released. For several months, improvements seem small. Then another model arrives with better reasoning, stronger image understanding or a new ability to use tools.
It can feel as though the breakthrough happened suddenly.
Usually, the public is seeing the release date rather than the full development process.
AI progress can look sudden because training, testing and integration happen out of public view. Several gradual improvements may then be released together.
A public release is only a snapshot
People usually encounter a model when a company announces it or adds it to a product.
That moment is easy to see. The months of preparation before it are not.
Before a new system reaches users, teams may have worked through:
- data collection and filtering
- training experiments
- architecture changes
- post-training and feedback
- safety evaluations
- speed and cost improvements
- product interface changes
- tool and retrieval integration
Some experiments fail. Others produce small improvements. A few changes may work well enough to become part of the final release.
Users see the combined result, not every step that led to it.
Training begins long before the model starts learning
Large training runs do not begin with someone simply pressing a button.
Data must be assembled, cleaned, filtered and organized. Hardware must be prepared. Training settings must be selected and smaller experiments may be used to test whether the plan is likely to work.
These preparations can affect the final model as much as the visible training run.
For example, poor-quality or heavily duplicated data can waste computing resources. A better data mixture can help the model encounter a wider range of patterns without simply increasing the total amount of text.
Much of this work creates no immediate public change. It still contributes to what eventually appears as a jump.
Scaling can improve several abilities together
Increasing the size of a model, the amount of training data or the computation used during training can improve performance across many tasks.
This relationship is often discussed through scaling laws: measured patterns showing how performance tends to change as certain resources increase.
Scaling laws are useful, but they are not promises that every ability will improve smoothly or forever.
Some skills may improve gradually. Others may remain unreliable. A larger model can also cost more to train and operate.
What matters to users is that several small improvements can combine into a much more noticeable experience.
A model may become slightly better at reading instructions, keeping track of context, choosing relevant information and correcting mistakes. Together, those gains can make it feel substantially more capable.
Post-training can change visible behaviour
Initial training teaches a model broad statistical patterns from large datasets.
After that, additional training can shape how the model responds to instructions, handles difficult questions or avoids unwanted behaviour.
This stage may include examples of preferred answers, human or automated feedback and practice on particular types of problem.
A base model and a post-trained version can therefore feel quite different even when they share much of the same underlying training.
The change users notice may come from better instruction following rather than a completely new store of knowledge.
Tools can make an existing model appear much stronger
Not every visible improvement comes from the model itself.
A product can add retrieval, calculators, code execution, memory or access to external software.
These tools can compensate for weaknesses in the underlying model.
For example:
- retrieval can provide current documents
- a calculator can handle exact arithmetic
- code execution can test a proposed solution
- memory can preserve selected information between sessions
- tool access can turn a written suggestion into an action
To the user, the entire product may appear more intelligent. Mechanically, some of the improvement comes from connecting the model to better external resources.
Construction may continue for months while most of the building remains hidden.
When the scaffolding is removed, the change appears sudden. The visible reveal is sudden, but the work was not.
Several changes may arrive in one release
Companies rarely release every small internal improvement separately.
A new product version may combine:
- a stronger model
- better system instructions
- a larger context window
- faster infrastructure
- improved retrieval
- a redesigned interface
- new tool permissions
Because these changes arrive together, it becomes difficult to tell which one produced the largest difference.
A model may seem better at research because retrieval improved. It may seem better at long documents because the context system changed. It may feel more responsive because latency was reduced.
The public experiences one large step even though the product contains many smaller steps.
Some abilities may become useful only after a threshold
A gradual numerical improvement does not always create a gradual improvement in usefulness.
Imagine a task that requires ten dependent steps. If the model succeeds at each step only some of the time, the full task may fail frequently.
A modest improvement at each step can produce a much larger improvement in the probability of completing the whole task.
This can create the appearance of a sudden new capability.
However, it is important not to assume that every plateau hides an inevitable breakthrough. Some methods reach practical limits. Some experiments do not work and some capabilities remain difficult even after additional scaling.
Benchmarks and everyday use may move differently
A model can improve on formal evaluations without producing an equally large improvement in ordinary use.
Benchmarks usually test defined tasks under repeatable conditions. Everyday work contains unclear instructions, unusual files, missing context and changing goals.
The reverse can also happen. A model may improve only slightly on a benchmark while becoming much easier to use because it follows instructions more consistently or works better with tools.
This is another reason AI progress can be difficult to measure with one number.
Why one jump does not predict the next
After a dramatic improvement, it is tempting to extend the trend forward.
If one generation improved quickly, people may assume the next will improve by the same amount.
But the next step may face different constraints:
- high-quality data may be harder to obtain
- larger training runs may cost much more
- remaining errors may be unusually difficult
- new capabilities may create new reliability problems
- product deployment may move more slowly than model research
A staircase does not tell us how tall the next step will be or when it will appear.
AI progress often looks sudden because users see selected releases rather than the full process. Training, data work, post-training, tools and product design may develop gradually before appearing together.
Better performance still leaves a deeper question: what would it take for an AI system to understand the world more reliably?
Comments
Post a Comment