Beyond the Version Number: How Do You Really Improve Artificial Intelligence?

In recent weeks, intrigued by the frantic speed at which new AI models are released, I asked myself a question as simple as it is complex and ambitious: how do you concretely improve an AI model that already seems capable of doing everything? What makes it able to do that one extra thing? What makes it that .1 better than before? What exactly has been done in those few weeks between one release and another?

I did some research, even with the help of AI itself, among technical reports, papers, and announcements from OpenAI, Anthropic, Google, and DeepSeek. What emerged was a picture different from what I imagined, and much more interesting.

To understand the answer to this question, we must stop looking at AI as if it were a new smartphone with a faster processor. Improving a language model is a fascinating mix of algorithmic discoveries, extreme optimization, and business strategies.

1. The Data Diet

In the beginning (4 years ago), the rule was: "the more data we input, the smarter the model is" (Chinchilla, DeepMind, 2022). Half the internet was downloaded and fed to the servers. But later research (FineWeb, 2024) showed that quality beats quantity. Today, to release a better model, companies spend months "cleaning" the library. They eliminate duplicate texts and select very high-quality data (perhaps scientific papers or perfectly written code). The result? The new version doesn't have a "bigger brain," but it has studied better texts, drastically reducing hallucinations and egregious errors. This is generally the most important, heaviest, and most expensive phase to carry out.

2. Post-training, teaching how to interact

If pre-training gives the model knowledge of the world, post-training teaches it how to use it. And this is where I found the most curious story of all.

For years, this phase was based on human trainers who rewarded responses perceived as most useful. The problem lies entirely in that "perceived": it ended up rewarding what sounded good. The result was sycophantic machines, capable of writing long, well-formatted, and persuasive texts, which, however, concealed logical errors and fabricated facts. A 2023 Anthropic study indeed measured the phenomenon on five different assistants and found that in certain cases, the systems used to evaluate responses preferred a compliant answer over a correct one.

Today, the industry has changed course by adopting techniques based on verifiable rewards. Instead of asking a human if an answer "sounds good," AI is made to solve programming problems or mathematical theorems, where the result is objectively true or false. This rigorous approach forces the model to develop true deductive reasoning, eradicating the habit of elegantly guessing.

It is the method behind models like DeepSeek-R1, whose operation was published in Nature in 2025.

However, this has not replaced human feedback; it has been added. Tone, confidence, and format still pass through there.

3. The Architecture

Thirdly, intervention is made on the architecture, which is the way the model is internally constructed: how many layers it has, how information circulates within it, how it relates one word to another ten pages earlier, how much text it can hold in mind all at once. There are dozens of design choices, and each can be re-evaluated with every version.

Many improvements we now take for granted come from there. The fact that today you can paste an eighty-page contract and receive a sensible answer doesn't depend so much on the data the model has studied, but on how the part that manages its working memory has been redesigned.

Among all these choices, the one that has changed the game most in the last two years has a name: Mixture of Experts.

The principle is intuitive: instead of firing up the entire brain for every question, the model activates only the useful portion. In simple terms, instead of gathering the whole team, you only call the specialist.

DeepSeek-V3 had the merit of putting numbers on the table, and they are impressive figures: 671 billion total parameters, but only 37 billion active for each piece of text processed. About 5%.

This is also one of the reasons why the cost of a single request has plummeted in the last two years, making it a commercial lever even before a technical one.

4. The Reflection Time

Until recently, we expected AI to respond instantly. Reasoning-oriented models, from OpenAI's “o” series in September 2024 onwards, have been programmed to do the opposite: stop and think before speaking.

But what does it concretely mean to think? This feature is divided into 4 phases.

The first: the model writes out its reasoning. Before giving you the answer, it produces a long intermediate text in which it sets up the problem and breaks it down into pieces. Some models don't show it at all, others show a summary, but it's there, and it's the longest part of the work.

The second: it tries multiple paths. Instead of following the first path that comes to it, it generates several and compares them.

The third: it re-reads itself. If it finds a contradiction halfway through, it goes back and changes its approach, instead of pushing straight to the end as previous models did.

The fourth, the most interesting: it decides how much to think. Not all questions deserve the same effort, and a good part of the work in recent years has been teaching the model to dose it.

Improving reasoning therefore means tweaking the parameters of these phases and finding the right balance that brings economy, speed, and precision. It seems like a leap in intelligence, but in reality, it's a leap in method: not a bigger brain, but a better processing path, activated only when truly needed.

However, this lever does not apply to all models, but only to the more powerful ones like Opus 5, GPT-6 Astra, Gemini 3.1 Pro, etc. Alongside those that reason, there is an entire family of fast models, those you find called flash, mini, or lite, calibrated for the exact opposite: to respond immediately, without intermediate steps. They are not substandard versions. They are used for tasks where the answer is almost obvious and what matters is that it arrives in half a second and costs very little. It's the same lever, turned to minimum.

5. Distillation, how flash and mini versions are born

The four levers analyzed require months of work and millions of dollars. But often only three weeks pass between one release and another. How is this possible?

Because a fifth lever is used, namely distillation: a large model is taken, made to produce thousands of responses, and a smaller model is trained to imitate them. The result is a lightweight model that performs like a heavy one on certain tasks, at a fraction of the cost.

It is the least spectacular lever and probably the most used. Almost all "mini," "flash," or "lite" versions that are released a few days after the large model come from there.

6. The Commercial Part, the non-technical lever

Behind a new version number are these five things. And then there is a sixth, which you won't find in any technical report: the commercial calendar.

Companies don't release a new version just because a team has discovered something. They release it also because there is an ongoing race. If a competitor introduces a faster and cheaper model, the others must respond within weeks, or they lose the developers who built their products on their APIs, and a developer who has migrated usually doesn't come back. There is also a second audience to convince, more demanding than the first: investors who have put tens of billions into these companies and want to see that the growth curve does not flatten.

This does not detract from the improvements, which are real and measurable. However, it adds a key to understanding: the moment a model is announced tells as much about the market as it does about the technology.

To Conclude

This analysis describes a precise window, the development of generative AI from 2022 to today. The five levers are not laws of physics; they are the answers that the sector has found to the questions of these years. Will they still be the same in two years? No one can guarantee it.

In fact, there is already a signal: Epoch AI estimates that the stock of public text produced by humans will be completely consumed between 2026 and 2032. Lever number one has an expiration date, and at that point, something else will have to be invented.

If I think back to what a model could do twelve months ago, the distance to today is almost embarrassing. Perhaps the curve will flatten, and we will settle nearby. Maybe a lever will emerge that no one thinks about today, as happened with reflection time, which was not on anyone's list two years ago.

One thing, however, I believe will remain true. The leaps have never come from doing the same things on a larger scale, but every time someone has changed the question: not how much data, but what kind. Not how big it is, but how much of it I turn on. It is the essence of human history's progress.