Mirendil and Recursive release RSI results, NVIDIA researcher discusses survival rules for researchers

We have just witnessed history!

Recently, there have been two major releases in the field of "recursive self-improvement"!

Over the past decade, pretraining a base large model has been a "arms race" for tech giants.

In the open-source community, this requires dozens of top experts.

In Silicon Valley AI labs, this requires hundreds of people: some specialize in architecture and scaling laws, some clean trillions of tokens of data, and others watch tens of thousands of GPUs day and night to prevent cluster crashes.

But just today, this seemingly insurmountable iron curtain has been torn open!

Two startups almost simultaneously dropped bombshells, smashing "recursive self-improvement RSI" from a sci-fi concept directly into reality.

1 person takes on a hundred-person army! Two god-tier models emerge

Today's first nuclear bomb comes from the startup Mirendil.

They announced a crazy experiment that pushes the "human efficiency leverage" to the extreme: with just 1 researcher, they hand-built from scratch a coding base model with a parameter count between 100 billion and 500 billion (100B–500B)—Mirendil-Base!

The entire process cost less than $10 million in compute (including all preliminary exploratory experiments and Agent token consumption) and took 3.5 months.

The most scalp-tingling part is the details:

The project started from an empty code repository, bought no commercial licensed data, used no model distillation or synthetic data, and was purely based on PyTorch, JAX, and public GitHub code and web data.

So who did the training infrastructure, model architecture improvements, data mixture tuning, and low-level kernel optimization that used to require a big-tech team to divide and collaborate on?

The answer is: AI Agents wrote over 99% of the code, ran all the experiments themselves, and even diagnosed training failures themselves!

And the one and only human researcher did only two things in those 3.5 months: decide which research directions to try, and judge which experimental results were worth trusting.

You think a model trained by a single person is just a "toy"? One look at the evaluation results and everyone fell silent.

In the code repository prediction task, the Mirendil team directly brought out the five strongest monster-tier coding models currently on the market to compete:

Kimi-K3 (total parameters 2.8T / activated 104B)

DeepSeek-V4-Pro-Base (total parameters 1.6T / activated 49B)

Inkling (total parameters 975B / activated 41B)

GLM-5.3 (total parameters 753B / activated 40B)

Nemotron-550B-Base (total parameters 550B / activated 55B)

Facing these MoE (mixture-of-experts) giants with trillions of parameters and hundreds of top-tier team members behind them, the 100B–500B-scale Mirendil-Base not only traded blows but even scored astonishingly high on multiple newly released external code repository tests after training!

Measured by bits per byte cross-entropy for fair comparison across tokenizers, lower is better

Even more terrifying is that this is just the beginning. Mirendil revealed that the larger Mirendil-Base-Large (500B–2T parameters) has already been 80% trained, while the flagship Mirendil-Base-XL (2T–10T parameters, a true ten-trillion-level behemoth) has just reached 19% training progress, and its performance is already approaching the Large version's current score!

Moreover, their next-generation model will be directly built with Mirendil-Base's own participation.

Another big shot, Tian Yuandong's startup Recursive, officially released an open-weight 40 billion parameter (40B) coding model—Flint.

In agentic coding benchmarks, among mixture-of-experts (MoE) models with total parameter counts below 50 billion, Flint took the top spot.

Its performance is comparable to the strongest dense models of similar total parameter scale, yet it runs much faster.

Even more surprising is that Flint itself is just a "bonus"; the real star is the Eureka Machine behind it—a system that automatically completes the full research loop of "propose ideas → implement → test → decide next step."

Using Flint, an actually delivered model, proves that automated research can indeed build things, not just run toy experiments.

Recursive's goal is to build superintelligence capable of recursive self-improvement, to automate knowledge discovery.

Mirendil is "humans command AI, compressing manpower and time to the extreme"; Recursive is "using the Eureka Machine to make the research process itself highly automated."

Both companies independently target RSI:

The loop is closing before our eyes!

UC Berkeley professor and Databricks co-founder Ion Stoica and other big names reposted in amazement.

As industry insiders say: "Soon, everything will depend on just two things: how much compute you have, and how good your 'discovery loop' is."

Junior researchers break down, NVIDIA big shot issues urgent call

The wheel of history rolls forward, but those crushed under it feel a biting chill.

After seeing the news about Mirendil and Recursive, a newly entered junior researcher completely broke down. Late at night, he sent a private message to NVIDIA senior researcher Ali Hatamizadeh, every word filled with frustration and fear:

The entire pretraining process was just one person, no R&D team, nothing...

If one person plus a computing cluster can do the work that used to take an entire team, then hasn't the expertise I spent years painstakingly accumulating become obsolete before it even had time to warm up?

Faced with the panic of younger colleagues, NVIDIA researcher Ali didn't offer platitudes; instead, he wrote an extremely hardcore, even somewhat stinging long essay, issuing an urgent call to all technologists: don't be scared out of your wits by marketing headlines! You are not obsolete, but your "rice bowl" has indeed changed!

Ali tore open the underlying logic behind "a single person training a hundred-billion-parameter model" and pointed out four survival rules for ordinary researchers in the era of RSI (recursive self-improvement):

First, face reality and stop coasting by "repeating old recipes"

Ali states bluntly that RSI (recursive self-improvement) is real.

In September of this year, OpenAI disclosed that the agent workload recorded by its research department had reached 3.1 agent-workdays per human workday, and it plans to achieve "automated AI researchers" by March 2028. If your five-year plan still assumes AI won't take over pretraining, change the plan now.

But why can one person at Mirendil train a model? Because the "recipe" for conventional pretraining has become thoroughly mature and public!

The major labs have long since fed open-source weights and detailed technical reports right to your mouth.

"If your main job is just to tweak last year's training scheme slightly, swap in a new batch of data, and run it again, then the room for that kind of work is shrinking fast."

Since AI writes 99% of the code, why are humans still needed? Because automated scientific research has simply not gotten through the most critical steps!

For example, on the EinsteinArena platform, AI agents did set records on 11 unsolved math problems (raising the best known lower bound for the kissing number in 11-dimensional space from 593 to 604).

But the premise is: humans defined the problems in advance and wrote the automatic scoring code. And the recently hyped "fully AI-generated paper passes peer review"—look at the details and it only passed the initial review of a workshop with a 70% acceptance rate, not the main conference, and it was later retracted.

Inside OpenAI, among tasks that humans need 4–8 hours to complete and that agents ultimately complete successfully, more than half still require at least one human intervention.

Why? Because agents give answers extremely fast and always in a confident tone. It will excitedly come running to tell you: "Some metric jumped 3 points!" and thoughtfully attach an ablation table and a self-consistent explanation.

But in reality? Those 3 points may be entirely illusory—they come from data leakage in the evaluation, a low-level bug in the data loader, a random seed that happened to get lucky, or a tokenizer mismatch!

In an era when the cost of execution has dropped to zero, the judgment of "which experiment to run, which result to believe, which direction to abandon" is your most ironclad new rice bowl!

Combine the speed of agents with your own ideas and judgment, and even if you are a junior researcher just entering the field, you can instantly become a "ten-times-efficiency researcher."

Ali gives a thunderous warning: this is the paradox of this era—the more work you hand over to agents, the more important your fundamentals become. Because you can never supervise something you don't understand!

Being able to see at a glance that the results an AI produces are fishy depends on a bone-deep understanding of the whole system: why does the loss curve take this shape? How large an improvement is consistent with physical laws? What kind of result is "too good to be true"?

How do you develop such a sharp eye? Ali gives the most practical advice:

Build the wheels by hand: implement the attention mechanism from scratch by hand at least once, then implement a linear variant, such as Gated DeltaNet.

Run sabotage experiments: train a small model yourself from start to finish, then deliberately break it and observe with your own eyes what happens when it collapses.

Read papers with skepticism: don't skip even the appendix; before you see the authors' experimental numbers, build an expectation in your head of what a reasonable result would be.

With this hardcore knowledge, you can truly command the machine, rather than believing whatever it says.

Fourth, large models are far from fully understood, and golden research questions are everywhere

Finally, don't think there's nothing left to research about large models.

One person being able to train a base model is only because the old road has already been paved flat. Ali directly lists 7 golden topics that no one has yet perfectly solved; breaking through any one of them would be enough to sustain you for a lifetime:

Few-shot efficient learning: how can we learn more from less data?

Exploring new architectures: can linear attention like Gated DeltaNet catch up with full attention on retrieval and long-context tasks? How should the two be proportioned in hybrid architectures?

Continual learning: how can a model keep learning after training ends without forgetting what it has already mastered?

Long-horizon reasoning reliability: why does reasoning become less reliable as tasks get longer? How should this be solved?

Eliminating hallucinations: why do models fabricate content? Can they be taught to honestly say "I don't know"?

New evaluation systems: when benchmark scores saturate, or test content gets mixed into training data, how do we measure whether models have truly improved?

Interpretability and control: what actually happens inside a model? Faced with increasingly autonomous models, how do we ensure they act according to human intent?

If models really do train models in the future, someone must decide: what should they optimize? How should we measure it? And how do we prevent this self-evolving loop from drifting further and further off course?

The machine is getting better and better at racing along the highway we have already paved; but where the next road should lead—the steering wheel is still firmly in your hands!

References:

This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: ASI Revelation, published by 36Kr with authorization.