Does this mean Claude no longer outputs the full raw reasoning, only summaries? At one point, exposing the LLM's full CoT was considered a core safety tenet.
Anthropic was chirping about Chinese model companies distilling Claude with the thinking traces, and then the thinking traces started to disappear. Looks like the output product and our understanding has been negatively affected but that pales in comparison with protecting the IP of the model I guess.
When Gemini Pro came out, I found the thinking traces to be extremely valuable. Ironically, I found them much more readable than the final output. They were a structured, logical breakdown of the problem. The final output was a big blob of prose. They removed the traces a few weeks later.
I don't think it ever has. For a very long time now, the reasoning of Claude has been summarized by Haiku. You can tell because a lot of the times it fails, saying, "I don't see any thought needing to be summarised."
It also gets confused if the entire prompt is in a text file attachment.
And the summarizer shows the safety classifier's thinking for a second before the model thinking, so every question starts off with "thinking about the ethics of this request".
I'd get confused if I was a LLM and you put my entire prompt in a text file attachment. I'd be like, "is this the user or is this a prompt injection??"
If you paste a long enough prompt into either GPT or Claude they turn it into an attachment, so it can happen. I think it's invisible to the model, but somehow not to the summarizer.
They are trying to optimize the circus trick that 'reasoning' is. The economics still do not favor a viable business at these valuations or levels of cost subsidization. The amount of compute required to make 'reasoning' work or to have these incremental improvements is increasingly obfuscated in light of the IPO.
Genuine question, why have you chosen to phrase this scraping and distillation as an attack? I'm imagining you're doing it because that's how Anthropic prefers to frame it, but isn't scraping and distillation, with some minor shuffling of semantics, exactly what Anthropic and co did to obtain their own position? And would it be valid to interpret that as an attack as well?
I don't think that learning from textbooks to take an exam and learning from the answers of another student taking the exam are the same.
Joking aside, I also don't believe that maximum access to raw Internet data and its quantity is why some models are doing better than Google. It seems that these SoTA models gain more power from synthetic data and how they discard garbage.
Firehosing Anthropic to exfiltrate their model seems materially different than Anthropic downloading all of the Internet to create the model in the first place to me. But maybe that's just me?
I don't see the material difference in firehosing anthropic vs anthropic firehosing random sites on the internet. As someone who runs a few of those random sites, I've had to take actions that increase my costs (and burn my time) to mitigate a new host of scrapers constantly firing at every available endpoint, even ones specifically marked as off limits.
Yes, what the LLM providers did was worse and impacted people financially a whole lot more in lost compensation for works as well as operational costs that would never reach the heights they did solely because of scrapers on behalf of model providers.
Very cool that these companies can scrape basically all extant human knowledge, utterly disregard IP/copyright/etc, and they cry foul when the tables turn.
We should treat LLM somewhat like patents or drugs. After 5 years or so, the models should become open source. Or at very least the weights. To compensate for the distilling of human knowledge.
All extant human knowledge SO FAR. Remember, by the nature of the beast, the companies will always be operating in hindsight with outdated human knowledge.