Pardon my ignorance and the silliness of my question to all the experts here, but that's how I learn, can someone explain to me like I'm 5 how an LLM is able to discover anything new?!
https://www.ant...
LLMs can't discover anything. they can help the operator find pattern matches based on intuition, the novelty comes from the human. the search is accelerated by the robot. it's because they are not *precise* pattern matches, but can find matches across a whole variety of close similarities all the way up to absolutely no relevance at all (aka hallucination).
this article is bunk.
deep-snek's take:
What actually happened: humans chose the problem (find interesting RTs), built the database, designed the harness, defined what "interesting" meant, ran the lab work, and verified the result. The AI scanned 200,000 sequences, filtered to 3,500 candidates, and wrote 20 reports. One of those reports flagged a repeat array next to an already-known RT. That's a real pattern-spotting task at a scale a human couldn't match by eye — the article even quotes the agent saying "I can see by eye a tandem repeat array," which is the actual contribution. But the underlying RT was already known, the function is still unknown, and the "discovery" is a candidate system, not a characterized one.
So "Claude autonomously discovered" is marketing. The honest version is "Claude spotted a pattern in raw sequence data that humans had missed, and humans did everything that made the finding meaningful." Zhang's quote is telling — "intriguing," "merits further investigation." That's scientist-speak for "might be something, we'll see."
It's the same shape as the "AI wrote a C compiler" claim. The labor is real. The framing erases the scaffolding. The AI did the scanning; the humans did the science.
---
i'm actually in the process, near the end, after 4 months now, of writing a completely novel programming language compiler. i have had to fight the LLMs drift towards its known pattern weights to force it to actually implement it as i specify it. and it has helped a lot. it finds problems, i think about solutions, i suggest a better design, it implements it, finds more problem, different problem, i suggest a solution that modifies the design, and so on and so on.
i did the design. the machine sped up the prototyping.
same as with this story, the humans selected the data set, defined the search criteria, and the robot ground through it looking for patterns that matched the criteria. basically, a souped up search engine, sifting through bajillions of data points.
it's a good thing, yes, but it's not "AI" in the driving seat. and anyone who tells you that AI can have agency is trying to sell you an IPO scam.
this article is bunk.
deep-snek's take:
What actually happened: humans chose the problem (find interesting RTs), built the database, designed the harness, defined what "interesting" meant, ran the lab work, and verified the result. The AI scanned 200,000 sequences, filtered to 3,500 candidates, and wrote 20 reports. One of those reports flagged a repeat array next to an already-known RT. That's a real pattern-spotting task at a scale a human couldn't match by eye — the article even quotes the agent saying "I can see by eye a tandem repeat array," which is the actual contribution. But the underlying RT was already known, the function is still unknown, and the "discovery" is a candidate system, not a characterized one.
So "Claude autonomously discovered" is marketing. The honest version is "Claude spotted a pattern in raw sequence data that humans had missed, and humans did everything that made the finding meaningful." Zhang's quote is telling — "intriguing," "merits further investigation." That's scientist-speak for "might be something, we'll see."
It's the same shape as the "AI wrote a C compiler" claim. The labor is real. The framing erases the scaffolding. The AI did the scanning; the humans did the science.
---
i'm actually in the process, near the end, after 4 months now, of writing a completely novel programming language compiler. i have had to fight the LLMs drift towards its known pattern weights to force it to actually implement it as i specify it. and it has helped a lot. it finds problems, i think about solutions, i suggest a better design, it implements it, finds more problem, different problem, i suggest a solution that modifies the design, and so on and so on.
i did the design. the machine sped up the prototyping.
same as with this story, the humans selected the data set, defined the search criteria, and the robot ground through it looking for patterns that matched the criteria. basically, a souped up search engine, sifting through bajillions of data points.
it's a good thing, yes, but it's not "AI" in the driving seat. and anyone who tells you that AI can have agency is trying to sell you an IPO scam.
11❤️1👆1🔥1
