Two days ago, an MIT blog "Weight Space Programming" was featured in the AI circle.
In the experiment, GPT-6 Astra himself wrote a Python script-checking diagnosis, modifying only a 0.8B small model.0.016% by weightIt has healed its re-reading cycle, increasing mathematical scores to six times that of the earthquake.
Although it soon emerged that it wasn't just a matter of "mind calculation" weight, or the essence of writing code to fix Bugs, the word that came to the end of the blog post was Silicon Valley's "returning self-improvement" (RSI) that has been talked about repeatedly over the years.
Top models are no longer satisfied with their own answers, but instead start writing scripts, adjusting parameters, and even training another AI themselves.
The whole technology circle is moving in a direction that is visible to the naked eye.Vibe CodingTurning toVibe Training。
You don't have to write your own training script, and you don't have to work with nightmarish parameters. If you set a goal, AI will be able to teach you a new model from the beginning.
But back to the big factory labs playing top-of-the-line closed-source models, can ordinary people really experience an "AI-training AI" on their own computers?
With the help of the newly-opened Naive-N0.5-Flash from NaiveAI, wild gamers around the world have already created wildflowers on their own computers.
A 24GB memory at homeMacIn the first half of the 20th century, there was a huge increase in the number of people living in the city.
Ordinary laptops take an hour.Local customer models that respond to milliseconds in zero-code;
Even in the 3D world of physics.It's time for AI to move from zero-church to two-legged robots.Standing steady for 10 seconds on a rolling basketball。
When the Silicon Valley giant was rolling out the RSI schedule behind closed doors, the most crucial step of "training AI with AI" had already landed on ordinary people's computers.
It's a good idea to turn a personal computer into a refinery:A Mac lesson broke down, and even the exam errors were fixed by itselfOrdinary people want to learn an exclusive small model, and the most discouraging is often not how deep the algorithm is, but to wash the data, adapt the environment, modify the storage, write the scripts and evaluate the painful work.
What would it be like to give AI a full-on life in this shithole?
Little Red Book owner"Don't worry about it".(Open source projects)self-llmIt's the one in my house.24GB of memoryMacIt's a big deal.
From the beginning to the end, he wrote only one page of his mission book.program.mdThe rest of the process was fine-tuned and handed over to Claude Code for access.Naive-N0.5-FlashIt's a small open-source model of 0.6B, and it's called "Hello".
What's interesting is that AI doesn't start practicing when it does, but instead pulls it out of human settings, tone, and courtesy.41 test subjectsThis is the first time I've touched the bottom of the initial model (8,451).
I didn't think the first round was over, but the score dropped.7,229 pointsIf you ask the question "who are you?" and you learn that the model is:"My name is Anil Yun".You're a good guy, you haven't learned how to understand it yet.
Naive-N0.5-Flash turned his head to look at the data source and found the original script.82% of the lyrics have no spoken word.I'm not sure what I'm going to do.
After identifying the root of the disease, it immediately wrote the script to eliminate the vocabulary of the other characters.566 landmarks in the cityIt's time to re-educate and heal the plague.
What's more, training it also exposes two major flaws in its own set of "exams": one is that the model doesn't answer questions, so long as the words "royal" are mixed up in the mouth; and the other is that when the model enters the game, the script on the board sees four words "thinking process" and misjudges the score.
The AI stopped right away and corrected the scoring rules, rewritten the history, replenished the rejection sample, and finally cleaned up the scores.8,732 points。
Don't look at the score, it's only 0.28 points, a mouth, and that asshole Black House girl smells so good.
"I don't know what the answer is, but if my sister asks questions about the palace, I'll tell you".
Say it again in English.Write a short self-introduction in EnglishWhen she asked her to introduce herself, she went straight back with a handkerchief:
"What my sister is saying is that she doesn't understand what she is saying".
It's not just small talkers, even the most likely to collapse.Strengthening Learning（RL）It can also be used to stabilize people.
Another Little Red Book owner"You're a little awake".For the first time, a 1.7B character-counting model was used on a single card with the codex, and uncontrolled data read and wrote directly to the hard drive, triggering a black screen die-off.
When it was replaced by the Naive-N0.5-Flash, it not only scrapped the explosive storage and the hard drive, but also pulled out the big craters from the drawdown: "The hard disk divider decides the number of pairs to zero".
It switches on its own in exchange for faulty judgment logic.Automated inspection every 15 minutesAfter running for almost two hours without touching the keyboard, the accuracy of the test was lowered.It's 21.3% and it's 54.4%.。
The most important step is to be able to fight back and the most important step is actually to be able to fight back.I've been trying to figure out what's wrong with the exam itself.。
In model training, especially reinforcement training, the most frightening thing is that you can't beat yourself up."Reward Hacking" is the name of the game.Once the scoring rules are flawed, the models will be busy mixing up the gaps, or injured by the dead-board rules, the more distorted they become.
In the past, AI has trained hundreds of thousands of wheels.If you can't keep your own exam, it directly determines whether the model is really strong or it's in place.
And when Naive-N0.5-Flash took over the dirty data washer, the scanner, the "miner alive", the personal computer really became a computer."Local Model Replacement Factory"。
No, Little Red Book has already spent an hour of zero code training off-line custom classifiers, and it's been crawled into more than 2,600 documents, training it to rewrite a single key in plain text into a small model of Markdown.
 |  |
As we watched the netizens rolling up their laptops, we got our hands dirty, just plugged the Naive-N0.5-Flash into Claude Code and tried it out on our own 64GB Mac.
The task is to use the data set of 77 English-language banking clients.BANKING77It's the first time I've done this.There are 110 million parametersBERTDistribution of small models。
In the same way, on the ATM machine, "the card was swallowed" and "the card came back without spending money", both are literally talking about the withdrawal machine and the bank card, but the corresponding processing unit is completely different.
The feeling of "pointing where" is particularly intuitive when you open a local web page after you've been trained.The ATM kept my card.(The ATM's got my card, and in milliseconds, the card crashed.)card_swallowedThe report also highlighted the importance of the role of women in education.
If you change the last half of the sentence to "the card came back, you didn't get a penny, and the account balance was low", even if the sentences had the same ATM and cards, the model would instantly cut to cash short payment.wrong_amount_of_cash_received）。
Once again, "Buying a train ticket has been deducted twice", repeat the deduction in seconds.transaction_charged_twiceIt's just that I don't have half a bunch of crap.
 |  |
During the whole training process, the most intuitive feeling is that you don't have to spend the night overboard.
The Naive API runs in the cloud, the Mac runs on-premises training, and even the "fake progress" on the verification set.
For example, when training is extended to eight times, training Loss is still going down, but in 995 verification tests, the fifth point is clear to the highest.Article 940It's the first time I've ever run a race in eight years.Article 932。
At first, the AI saw that the model had begun to "dead with its hard back", and it decided that it was going to make a difference.The fifth time I stepped off the brakes.In the meantime, all 9,947 clean-up data re-trainings have been delivered.
It's finally over.3,080 articlesThe official test subject is the same speed, and the perceptual gap is drawn.
Direct-to-cloud DeepSeek V4.1 Flash, which can only swallow about one bit per second, has only a very low accuracy.75.81%。
This is an open-source BERT model that was developed locally by Naive-N0.5-Flash and can be used to unplug the wires for free on a Mac.6.77 millisecondsIt's faster than clouds.150 times that.The answer is yes.Article 2,875 ((accuracy rate 93.34%)The recent hot open source classification projects, Laya and Jev, are also lagging behind.
This also breaks down a common misconception about the current state of AI: there is no need for businesses to be tied to expensive cloud models.
Having a strong model as a "general trainer", playing back 6 milliseconds locally, disconnecting the network, and having data that doesn't come from the internet, is the best way to optimize vertical scenes for both speed and privacy.
Throw away your existing quiz book and try it yourself.Teaching robots to step on a 50-centimeter hill, hand-in-hand restoring top meeting papersIf it had been trained to set up, distribute, or hone in on the text task of "having a ready-made library of puzzles", if it had completely discarded standard instructional materials, given dynamic game scenes, board games, or even the 3D physical world, would it still be able to stay in place?
It's not a problem. In the face of Russian cucumbers and snakebite, it teaches a person to look at the screen from the ground up.Multimodal small models。
After adding weight to the difficulty of grouping the remaining samples, the trained small models on the left beat the Russian squares at 140 rows, and on the right play greedy snakes eating red apples while sliding through a fence.
 |  |
And some of the testers only give it one.Empty folders and a suggestion。
In just three hours, it builds an entire AlphaZero self-contained frame from scratch, a four-chess AI trained on a single graphics card, and takes the lead in killing a "perfect solver".3 wins 1 is equal。
And in the Little Red Book,"Half the Moon"AIIt's a note".and"Research Center"It's more about taking advantage of gaming and hardcore robots to control it.
 |
In the face of the benefits of gaming 2048 tonnes, AI itself has found 66,000 step up trajectories from open data sets, first raising the average to 3.74 times.
In the fifth round, it attempted to reverse the PPO reinforcement learning to find the score, not to mention the second round.Cut out the new branches and keep the most stable version.It's the first time that I've ever been in a relationship with a woman.
After moving the campus to the NVIDIA Isaac Sim 3D world of physics, the test was upgraded directly to two robotic puzzles:
The first is to get the two-legged robot to stand on a free-rolling basketball for 10 seconds.
AI has gone from zero-to-zero rewards to hard-to-practice fixed test success rates.100%(Random recommendations also have a 91.15% survival rate.)
The second challenge is to challenge the little two-legged robot MicroDuck.Step 50 centimeters up and down the stairs。
After taking the first step from previous experience, the robot fell to the ground.
AI follows the original plan after hitting the South Wall several times.A decisive overthrowIt is the first of its kind in the world, and the first of its kind in the world.
At the end of the day, not only did robots get up and running at high speeds, but they also made their way up to the original version.It's just a matter of time before we get to the next level.。
You'll see here, and you'll find,It's a two-way street between modeling and writing code.。
If the code is written incorrectly, the compiler will make mistakes on the spot; however, the model is biased, and the screen often does not have a single line of red letters, only the scores go downwards.
And Naive-N0.5-Flash, which looks like the hands of the pharaohs, is exactly where it is."Recognize the anomalous signals and get off your head in time".。
Whether it's the previous bank client case in the fifth time deciding the brake, or the 2048 PPO branch that cut off the division, the high-tech robot overthrew the old program rearrangement program, it didn't do it on a template running script, but with a real sense of strategy of "trying to overthrow and overthrow again".
With this piece of "dead shit", even the most brilliant AI researcher can do it right.
Little Red Book owner"A little tomato".With just one ordinary CPU computer and an open instruction, it picks up an ACL 2026 timing prediction paper on its own.
A breath of fresh air.743 rounds of experimentAfter correcting 29 errors on its own, it squeezed the predictive errors in eight group tests to the bottom of the original paper baseline (down by up to 4.10%).
At the end of the experiment, the workers submitted an eight-page paper in English.
He was the head of the "Pharaoh's Hands" practice.One million windows record a hundred rounds of experiments, cracking 2,000 words per secondAfter looking at this test, you might wonder: how did Naive-N0.5-Flash, the hard-working "Faraoh's Hand", come up with hundreds of experiments, high-speed trials and trials, accurate digging of all sorts of black holes?
The answer lies in the three underlying technical canyons that it crosses: "Can't remember, can't wait, can't touch".
First, byAIParticipated in the design of a 5:1 mixed scarcity attention architecture, originally installed over 1 million times.
This total parameter of 309B, a model of MoE that only activates 15.5B at a time, reduces attention retrieval time by 44 percent, even if it's not.One million tokens windowIt's filled with hundreds of rounds of experimental logs, fault stacks, and a complete code library, so you can see the full cause and effect, and the second response.
Secondly, the researchers broughtAIA handheld NaiveRT logic engine with a single stream speeds up to 2,000 words per second.
Models are most afraid of suffering, and much of the time is often dependent on the single-stream production of the model.
To do this, the researchers used AI to rewrite the underlying CUDA code in six days and 151 rounds of experiments, merging open-source SGLang with 29 sparse GPU calls per layer.A giant nucleus with a synchronous 148 stranded blocks。
After the conversion, the time it took to decode the single-wheel guesses was shortened.72.4%It's been a long time since I've been on the internet (from 12.3 milliseconds down to 3.4 milliseconds).Extremely fast pattern of 2,000 words per second。
The third is AI4AI, which rolls around almost 10 million real sandboxes a week.Strengthening LearningThe training ground.
It's engineering intuition doesn't come from behind the scenes, it comes from behind the scenes.The peakand distributedUp to 100,000 real container environmentsThere's a round of code writing, a round of bugs, a round of verifying the curve.
It also topped engineering charts such as SWE-bench Pro (73.6), Terminal-Bench 2.1 (86.7) and MLE-bench-30 (73.7%) and PaperBench (63.2) for AI research and development capabilities.
Step 4 to the RSIFrom helping people write scripts to making a model for 400 hours.I'm going back to Silicon Valley and I'm going back to Silicon Valley."Return to self-improvement" (RSI)You'll find it doesn't really matter.
"Developing AI with AI" is never a magic that comes out overnight, but is being trampled underfoot.Stage four：
The first level is a one-point play.This is where most regular programming assistants stay.
The second level isAINew models are developed independentlyPeople just swap targets and acceptance criteria, from washing data, starting classes, evaluating to cracking out exam loopholes, AI itself running the whole loop, the front row, banking flows, gaming AI, and robots all stand at this level;
The third stage is that AI turns to improve its own operating systems, no longer just to train small external models, but to turn to optimize the reasoning systems and model architectures that support its own operating systems, as exemplified by the NaiveRT engine and the 5:1 rare architecture mentioned above;
The fourth level is to be deeply involved in the development of the next generation.Just like the team used its 400-hour AutoWM world model to climb to the top of the WorldArena-1 list, it's really moving towards the "AI to make stronger AI".
So the Silicon Valley RSI isn't a step up, it's a step up. The first level used to be the daily routine of many people, and now the average person has moved up to the second level.
Of course, AI training has not yet reached the point where humans can completely master it.
Whether it was written before the start of the projectprogram.mdIn the meantime, we're going to have to work on our mission books, or we're going to have to shut down the research branch in the middle of it."In the direction of man,AIRun and execute".It's still the most pragmatic gold partner.
However, the technological wall that was once unstoppable has been pushed back.
If you open your own computer, you can call an AI researcher who knows how to weigh and mines for you to train an exclusive model.
The era of Vibe Training for ordinary people has officially begun.
