Move the storage around at home, tidy it again a few years later, and the word for it may shift from putting things away to organizing, and then to decluttering. The thing is still the same thing; only the name has changed. The same happens in technology, only faster, fast enough that outsiders assume each round is a brand-new invention, when in fact one thread is simply being renamed again and again.
Many names bring a problem with them. People in a discussion each use their own, so it sounds like three different things are being discussed; ask how far things have actually got and it turns out nobody has drawn a clear boundary. Sorting out the names along one thread looks like antiquarian work, but it is really laying the ground for the judgments that follow.
According to a media report, Fei-Fei Li and Jeff Dean once had a public conversation on exactly this: how many names the idea of machines taking part in improving machines has gone by over the decades. This article treats it only as a discussion that can be checked, names no platform or date, and supplies no scene for the conversation.
The Names One Idea Has Had
According to a media report, Fei-Fei Li ran through the line in that conversation: as a graduate student it was called incremental learning, then continual learning, meta-learning, lifelong learning, and later architecture search; today it is called self-improvement, and recursive self-improvement, abbreviated RSI in English. The names kept changing while the question stayed much the same.
Each of these words has its own origin. Incremental learning watches how new data is added bit by bit; continual learning worries about forgetting the old while learning the new; meta-learning asks whether a system can learn how to learn; architecture search treats the structure of a model itself as something to be searched. Different entry points naturally produce different labels.
She put the purpose of this sorting lightly. According to a media report, it was to separate fiction from fact. That is worth pausing on: the louder a word gets, the easier it is to assume the capability behind it has already arrived. Lay the names out first, and only then ask how far things have actually got today.
The Code From Ten Years Ago
According to a media report, Dean pulled the thread back about ten years: at the time, he and several colleagues worked on neural architecture search, letting one model find good structures for a class of tasks and feeding the results into the next round. That can be seen as a very early version of recursive self-improvement, though nobody called it that then.
In plain terms, it is a bit like letting a machine try building with blocks itself: put one arrangement together, measure how good it is, keep a record, then try another. A person stands by and fixes which blocks may be chosen and by what rules, leaving the repeated trials to the machine. Run enough of them and it can indeed land on arrangements a person might not have thought of first.
The difference lies in how the request is handed to the machine. According to a media report, back then one had to write a pile of Python code to specify the optional search space, line by line; today the task can simply be stated in natural language. The same job, moving from hard-coded rules to saying what is wanted, lowered the barrier by more than a little.
A Sentence Handed to a Model
According to a media report, he gave an example closer to the present: attach an agent system to a very capable model, ask it to explore model architectures that might help read magnetic resonance images, and it will go and do it. This article does not turn that into a line in quotation marks; it keeps only the meaning.
The agent system here can be understood as a helper that breaks a job down itself, tries to carry it out, and adjusts the next step from the result. In the past that breakdown had to be written as code by a researcher; now a single instruction may be enough to start. What changed is not that the machine suddenly grew smarter, but that the step of handing over a request got shorter.
The second half matters more. According to a media report, he added a qualifier: if such things can be made to work reliably, a great deal would be unlocked, especially for science and engineering. The word reliable matters more than unlocked; getting it right once proves little, getting it right every time and recovering when it goes wrong is what earns the right to be given real work.
Can the Loop Run Without End
There is a point in that discussion that is rarely raised but very concrete. According to a media report, Dean does not deny that such loops can speed up research, but cautions that they cannot run infinitely fast: computation has to run, results have to come back, data has to be stored, and then another model has to look at it. Every step takes time and takes machines.
An experiment finishing at two in the morning, with a person back at a desk at eight to see the result, and a system that sees the result and sets the next step on its own, are two different rhythms. Automation saves the idle waiting in between, but an experiment still takes as long as it takes, and a computer does not skip a calculation just because a loop has been wired up.
Put differently, the bottleneck is not only in ideas but in hardware and process. However automatic it gets, the computer still has to finish the work. Stating that physical limit plainly is more grounded than arguing over what to call the concept: it is a reminder that a system able to improve itself still stands on real things like server rooms, power and time.
What the Name Debate Is About
Many names are not necessarily a bad thing. Each generation of researchers enters from its own patch: some watch how data is added bit by bit, some how model structures are searched, some whether experience can accumulate. Different cuts give different names. Once models could understand natural language, these scattered threads were gathered under one word.
What is really contested is not the name but the boundary. Say self-improvement and one has to answer whether what changes is the solution to a single problem or the system itself; say recursive, and one more layer has to be answered, whether the ability to change the system is changed along with it. Without a boundary, a fashionable word easily wraps in things not yet done.
Self-improvement and recursive self-improvement are not the same thing either. According to a media report, the conversation used cooking as an analogy: doing the same dish better after a few more tries is the former; having the very knack of cooking, even the way one reads a recipe, rewritten, is closer to the latter. The analogy pulls an abstract term back to the stove.
What Was Left Out
A few places are deliberately left blank. The specific names of several colleagues could not be matched one by one to an original source in public information, so this article writes only of several colleagues; departures and causes not stated outright in the conversation are not filled in either. Every wording whose original could not be checked is turned into indirect speech without quotation marks and prefixed with according to a media report.
Precisely because it accepts only indirect speech, this article does not say which platform the conversation took place on, on what day, or who was in the room. A reprinted transcript can serve as a lead, but confirming a verbatim line requires going back to the original release; until that step is done, it is better to sound clumsy than to let a reader take paraphrase for a word-for-word record.
Some lively notes are left out as well: this article does not claim that some end point is near, does not play up the imaginative space the concept carries, and does not judge whether any company is doing the right thing. With one idea and six names, what is worth writing is how the thread got here, not a forecast of where it goes next.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.