On the concept of self-learning.
Look: you are using a large language model. On the one hand it is not a genius, and on the other hand it is. It is smart, but its knowledge may be a little out of date.
That is, the model is very smart, it has a lot of knowledge, it can reason, handle facts and draw conclusions. But that knowledge is not always fresh.
For example, if you ask: "Which sites can I buy such and such a monitor on, one that is actually in stock?" If the model does not know that, it has to go and look the information up on the internet.
What I have worked out: a model needs a division of duties. It can be very strong at reasoning, but for current information it needs to be able to search.
When I talk about a model self-learning, I do not mean fine-tuning, changing the model's weights or the facts it operates on.
I mean something else: you have a brilliant assistant. In my case that is the ideal engineer. But he does not always know everything. He needs to rely on fresh knowledge.
He needs to be able to educate himself.
I taught my LLM to run deep research on its own. Loosely speaking, my Claude goes straight to Gemini, Grok and ChatGPT and orders deep research there, at any depth and on any topic.
Which is to say, my model does not try to guess or pull data out of memory. I treat it as a brilliant employee.
But even a brilliant employee, like an ordinary person, has to educate himself and not be shy about googling.
There are two kinds of research.
The first option is simply to go round the internet yourself and look for information.
The second option is to order deep research in several different models. You will get different results and a better quality trail of reasoning.
Plenty of people are afraid of getting a one-sided assessment. But when the model goes out 3 or 4 times and orders deep research in different places, it gets a many-sided opinion from different sources.
And by the way, all of this deep research is ordered within an ordinary subscription. It does not burn a giant number of tokens.
When the result is assembled out of several deep research runs, the model gets a more reliable and more rounded picture.
How should you treat your LLM?
You should treat it as a brilliant employee who works for you. But you cannot lean entirely on its memory alone. It has to verify information every time.
It does the same thing you would do: it goes and looks. It can look in different places, it can go to another model, it can turn to deep research.
What is deep research?
Deep research is different from an ordinary internet search. It is research across dozens or hundreds of pages. If we run deep research in the large LLMs, the result is a thorough and reliable report, one you can actually rely on.
My message is this: do you want to teach your LLM to learn by itself?
I have skills that make your LLM always rely on the freshest and most correct information.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261026.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261026.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)