Are they “stealing” training data or are these other companies inputting specialized problems to Claude with answers and solutions that can be used to back out Claude’s internal logic/processes? It sounds more like the latter to me.
Do you know how LLMs work? There's no 'logic', it's all statistical modeling. What DeepSeek et. al. are doing is taking inputs from specific benchmarks, throwing them into Claude, then training their (smaller, cheaper, faster) models to match the output, bypassing the need to scrape the broader internet (or pay) for training data. Which is (IMO) why Anthropic is actually salty; they paid for publicly available data (to avoid being sued) and someone outside the legal system bypassed that process.
I guess you could call that 'backing out' or 'reverse engineering', except none of these companies actually understands the inner workings of their models (they have billions of parameters and are just too complicated), just the processes used to create them. It's a black box full of linear algebra.
There’s got to be some pattern they can acquire though? Some way Claude works a problem through a gazillion matrices that makes it superior and worth stealing from?
They are using Claude to create "training" data. Essentially they are asking Claude random questions and then training their model to output Claude's answer.
If I were Anthropic, and therefore morally bankrupt, whenever I detect these other companies querying Claude start outputting bullshit answers to make their models worse.
Are they “stealing” training data or are these other companies inputting specialized problems to Claude with answers and solutions that can be used to back out Claude’s internal logic/processes? It sounds more like the latter to me.
It's basically them crying "waah wahh, you can't look in the mystery box"
Do you know how LLMs work? There's no 'logic', it's all statistical modeling. What DeepSeek et. al. are doing is taking inputs from specific benchmarks, throwing them into Claude, then training their (smaller, cheaper, faster) models to match the output, bypassing the need to scrape the broader internet (or pay) for training data. Which is (IMO) why Anthropic is actually salty; they paid for publicly available data (to avoid being sued) and someone outside the legal system bypassed that process.
I guess you could call that 'backing out' or 'reverse engineering', except none of these companies actually understands the inner workings of their models (they have billions of parameters and are just too complicated), just the processes used to create them. It's a black box full of linear algebra.
There’s got to be some pattern they can acquire though? Some way Claude works a problem through a gazillion matrices that makes it superior and worth stealing from?
Tl:dr; we've already reached the point that LLMs are cannibalising each other.
They are using Claude to create "training" data. Essentially they are asking Claude random questions and then training their model to output Claude's answer.
If I were Anthropic, and therefore morally bankrupt, whenever I detect these other companies querying Claude start outputting bullshit answers to make their models worse.