Such a negative tone they put on this. Distillation is amazing, because it means anthropic and openai fail to keep a monopoly. Who even are they who claim it's unethical? If it is truly unethical, then so is the mass data scraping they do on my personal website on a regular basis (without my consent), and all the unauthorized use of content produced by authors, blog writers, wikipedia contributors, and creators everywhere. If it is truly unethical, then anthropic, openai, meta, google... all these companies should have deleted their LLMs long ago. This wording disgusts me.
Heck, it would be amazing if we had more models without guardrails - some of the models that are produced via heretic[1] are actually quite nice to use - in particular, I've enjoyed investigating Chinese censorship by interacting with an abliterated model of Qwen3.8-27b. If security is really a concern, then secure your systems - don't attempt to dumb-down the tools we use. If someone breaks your window, then they are responsible, not the hammer they use to do so.
IMO the biggest problem with distillation is that not enough people are openly doing it. I would love to see more small, competitive US labs instead of having the eggs in 2~4 baskets (depending on how you count).
An even smaller fraction of the cost if they do it by buying AI access at as much of a discount as they can find, including black market resellers, and then reselling that access to paying users again with a proxy. As is common.
This gives ruthless "fast followers" an economic edge over the innovator that's putting in the real work.
The dynamics are very much alike to what patents and copyright law are supposed to prevent. Same type of "we took the products of your work and used them to undercut you". Except there are no laws against distillation - so most of the enforcement happens on model provider level.
Is there actually that much capability transfer from non-logit-matched distillation, or is Anthropic just another unwilling source of data?
Even the early papers on distillation techniques found that surprisingly small distillation datasets can improve task performance noticeably on some specific task types - and that valuable adaptations like SFT/RLHF instruction following can be distilled from one-hot non-logit traces.
A big part of what distillation really gets you is: paving over the mismatch between pre-training and final performance. A base model is trained to spit out fitting text, but not to instruction follow, reason autoregressively, self-check or use tool calls - like an AI has to. There is transfer straight from the "text prediction" pre-training objective, and pre-training sets the foundation for all that follows - but the capabilities you get "out of the box" with it are often unrefined and fragile. Which makes some sense - internet text doesn't often include raw chain-of-thought autoregressive reasoning. It's not the kind of thing humans tend to write.
Reasoning traces? They let an AI learn proven techniques and adaptations directly, from an AI that was already taught "how to be an AI" in other ways.
It's why this kind of distillation typically plugs into mid-training and post-training, not pre-training.
Now, I'm not saying that all Chinese companies do is eat tokens, distill and lie. That just isn't the case. They developed or refined numerous training techniques and architectural adaptations - like deep fusion for high performance visual input, RLVR with GRPO, trunked MoE, storage-efficient and bandwidth-efficient attention formulations, or residual routing techniques like AttnRes. Some of those are used widely now, and some are still on the uptake but show good promise.
But Chinese labs are enjoying massive efficiency gains from being able to distill from the frontier instead of doing things the hard way. It's a leg up. It lets them put their supply of R&D effort and RL compute elsewhere. They wouldn't be nearly as advanced if they couldn't do it.
That's the moat. Mistral has the capability but not the legal protections.
Instead of instilling confidence, it was overwhelming. Not sure if I'm the only one.
It would be great to know if this was Opus 5.5 or a lesser incremental improvement, as otherwise it's difficult to judge whether Opus 5.5 is expected to be a big improvement.
It's frustrating that there isn't more transparency here.
They write that at the top, but then on benchmarks, it beats literally every other model, including Fable and Astra?
Will be interesting to see how people's opinions of it line up IRL, but so far I've loved Fable so hopefully will love this one too
Considering fable gives me a refusal at least once a day on my very mundane reasonable requests (in a funny example - one of the subagents suggested bypassing the rate limit for running a report inside my own cluster and that caused a refusal) and my only solution is to switch to opus - seems like my next step will be switching to Astra or K3/GLM
Sounds like they noticed the complaints. I'm curious to see what LLM-isms this one may have.
> The Vercel target is hard-coded. That's common and not wrong, but it's opaque; nobody reading this later will know which Vercel project it belongs to, and if the project is recreated the target changes silently. A comment or a named variable would help.
> Pointing a DNS name at Vercel is only half the job. The domain also has to be added to the project in Vercel's dashboard, otherwise requests will arrive and Vercel will reject them. That step lives outside this code, so it's easy to forget.
> Finally, [CENSORED] existing only in production is slightly odd on the face of it. It may be perfectly deliberate (perhaps a single shared testing tool that only needs one public address), but if you're reviewing this rather than just reading it, that's worth confirming.
It has the same annoying cadence and writing style with slightly less prominent claudisms.
* Consider leaving a comment about the hard-coded Vercel target. It's not clear where does it come from.
* [This is just a bullshit point, because the domain is not "added to" Vercel, it's provided by Vercel]
* Are you sure that [CENSORED] is prod-only? The name suggests otherwise. [also, what "if you're reviewing this rather than just reading it" even means?]
It means "I'm treating you as lay-person punter, not a developer working on this project." Opus 5 feels like it's constantly trying to reward-hack me into treating it as intellectually honest and epistemically humble, while in the same breath it talks down to me and tries to smuggle its own bullshit assumptions and assertions into the conversation unchallenged. No progress on this front apparently. Glad I cancelled.
Claude is just comically bad nowadays.
Seems like it based on my first session. It still does the whole “bury the important thing in a pile of words” coupled with the “it might actually be important” thing… so basically you never really know what it’s talking about.
Honestly I trust opus so little that the entire “opus” brand is completely tarnished. Its writing style is so god awful that it needs more than just a point release. Either dump the name and ship a different model entirely or at minimum call it “opus 6”. Calling it 5.5 makes it sound like it’s basically a continuation of the same garbage output that 5.1 had but with some minor adjustments. And based on my single first test, that is what it appears like to me.
I don't mean to pick on this comment in particular. The majority of my work day is now spent reading AI generated text, and I look at HN (too much!) because I want to read human commentary. Humans pretending to be obnoxious AI on repeat is net negative to say the least.
I wouldn’t be surprised if Opus 5 was trained on content written by other LLMs
and it's not about the verboseness (even though it obviously contributes to the fatigue and loss of focus), I swear the vocabulary of the llms change working on the same task on the same codebase significantly.
I wonder if there are studies around this.
https://openai.com/index/where-the-goblins-came-from/
Small quirks can quickly add up in posttraining if not caught. Although TBH with how obvious Claude language is, I do feel like this is something Anthropic probably noticed and just assumed people would not care about. Now that people have obviously cared, they're probably actively looking to alleviate it
"Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5"
and
"We’ve made major improvements to the way Opus 5.5 writes and communicates, one of the most common areas of feedback we heard about Opus 5."
and
"In our own use, this has made Opus 5.5’s work easier to follow and check—which is a safety benefit as well as a practical one."
I realize it is corporate communications but "most common areas of feedback" and is a bit sterile. If the company wants authenticity and trust its easy to say that they found it hard to follow. And that it did not meet a quality bar they generally expect from their releases.
If this is not true, that it Opus 5 output was generally acceptable and we might see something like that again, that is an important consideration for potential customers or investors.
Wdym Opus 5.5 scores 14.7% higher than GPT Astra for Terminal Bench 4.0?
How would this alleged difference (most likely bs) actually show up in reality?
GPT Astra was literally the best model in the world by a margin until 1 hour ago or so.
>> On our benchmarks, Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work. That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.
> how would this alleged difference (most likely bs) actually show up in reality?
Furthermore: so they admit it's bs but still placate it like its the next biggest thing ever ... alright
All I'm saying is I refuse to buy into it anymore – yet many on here still do, including ... you?