If I understand correctly, threat model here seems to be to protect against accidental issues that would impact performance, but doesn't cover malicious actor.
For example, Sketchy Provider tells you they are running the latest and greatest, but actually is knowingly running some cheaper (and worse) model and pocketing the difference. These tests wouldn't help since Sketchy Provider could detect when they're being tested and do the right thing (like the Volkswagen emissions scandal). Right?
Providers like OpenRouter default to the cheapest provider. They are often cheap because they are rediculously quantized and tuned for throughput, not quality.
This is probably kimi trying to protect their brand from bargain basement providers that dont properly represent what the models are capable of.
I'd take it at face value. Since they release open weights they would appear to genuinely want other providers to serve this as well as themselves, but the benefit of this depends on it being served accurately.
Kimi, GLM, and Minimax are the "Big Three" of open source Chinese AI startups. There's also Qwen and DeepSeek but they are all subsidized by other lines of business.
The Chinese AI models are generally 5-6 months behind high end SOTA western models (and as of the time of this comment it's Opus 4.7 and ChatGPT 5.4 Thinking, it's rumored however that the Mythos and Spud codename models are even better).
To gain market share, the Chinese startup use open source as a distribution strategy and essentially made mid-high end AI a commodity. The best models are still Western but for any application that doesn't require the highest performance in the market or if there's a need for extensive customization or alignment (imagine if you are an oil rich petro state and you don't want your national AI strategy to be tied to liberal international order ideology).
It creates a lot of pricing pressure on the low and mid end, and it's also why Anthropic is desperately trying to go full B2B instead.
However if the third parties hosting the Chinese models at near cost doesn't perform good quality control, it ruins the strategy because customers are not inclined to use chinese models anymore (and first party hosting on chinese infrastructure is out of the question because of geopolitical reasons, so everybody hides behind the polite fiction of using resellers like OpenRouter, Fal.ai, Wavespeed, fireworks AI etc.).
I've been burned on openrouter getting routed through terrible quants with equally terrible quality. While paying maybe 15% less.
Nearly a year ago it was impossible to avoid it due to silly openrouter routing algorithm and the api. You had to set multiple things just right to make it work.
Similar to their other api quirks. You want valid json format response? sure, set response_format to "json" just like our documentation suggests. Oh, it only works some of the time? How silly, why would you expect it to work all of the time? If you want it to work more often, set require_params to true. We may still use other providers that don't offer it, but you want that, right? You don't? Well, then set our "very_require_params" to "very_true". And then switch a few toggles in the frontend. Oh and also add these 7 lines just so your other config options don't break. Oh wait they will break, how silly of us Is there any way to make it work as advertised? Of course no!
Sorry for the semi-offtopic rant. I still use them every day though, but not for open models anymore.
Catching accidental drift is still worth a lot. It's basically the same idea as performance regression tests in CI, nobody writes those because they expect sabotage. It's for the boring stuff, like "oops, we bumped a dep and throughput dropped 15%".
If someone actually goes out of their way to bypass the check, that's a pretty different situation legally compared to just quietly shipping a cheaper quant anyway.
Also it's not just about running an obviously worse quant.
Running different GPU kernels / inference engines also matters. It's easy to write an implementation that is faster and thus cheaper but numerically much noisier / less accurate.
For a truly malicious actor, you're right. But it shifts it from "well we aren't obviously committing fraud by quantizing this model and not telling people" to "we're deliberately committing fraud by verifying our deployment with one model and then serving customer requests with another".
I suspect there's a lot of semi-malicious actors who are only happy to do the former.
I love how many interviews Larry Tesler did (he passed away in 2020), he was so influential and it's interesting to see what that looks like from the inside.
And it's not mentioned in this ACM interview but rather this one with the Computer History Museum https://archive.computerhistory.org/resources/access/text/20... that implementing a modeless editor was easier too, since you could use a simple case-switch instead of having a bunch of explicit modules for each mode.
It's interesting how many people I know who jump instantly from hobby to thinking about hustling, Etsy, Patreon, fame, etc. and the thought that they'll never be good enough to go pro is a real barrier. You don't need to monetize your joy.
I’ll use an LLM, often Claude, to tighten up my writing. I have a tendency to use too many words.
I brain dump my candid reaction / thinking, and then I’ll get something to tighten it up. No LLM used for this follow up.
I apologize if my use of Claude to tidy up my thoughts was offensive — here was my unfiltered, original comment:
> There's a new type of product and service that's now possible with LLMs improving each month. The new value prop is shifting from time savings to stress relief.
Tools need to be built around human psychology like the self-checkout example. It's not faster, but it provides relief. Some tools, while powerful, can add anxiety to one's day, especially if it's built promising efficiency, but the user feels like they're not getting more done, getting things done faster, or both.
This article is a great example of "strong + weak = weak".
I only made it to the interesting stuff because of Carreyou's name, otherwise I would have stopped.
The email timing and lack of email metadata were also strong, in my opinion. But all of this nonsense like "Wow, these guys both talk about PGP??" distracts from it.
> The process of generating this data is labor intensive, because it requires sound ID experts to listen to each audio file carefully.
Oh man. This is THE ONLY REASON why AI at scale works...and it's entirely powered by extremely repetitive classification done by people in third-world countries (for now; there are similar jobs in US and Canada for harder domains like math and law). It's definitely the biggest reason why autonomous driving works.
(Cornell, who maintains Merlin, probably has students do it, though I know there is data crowdsourcing in the app too.)
As far as I understand it, classification data is basically the Brent crude of the AI industry (well that and the datasets used for training LLMs).
It paid well for the area until the company that spun up these services decided to move operations to SEA to save on cost. I'll try and link to it if I can find it.
Thanks for sharing this. I love Merlin but never knew how they got it to be so good. Blood, sweat, and tears - of course - as everything actually valuable and useful requires.
Deepseek will regularly spit out Chinese (汉字）during English sessions. They generally seem to be syntactically related but it makes me think that there's some overhead of using English with an engine that's primarily trained in Chinese.
I'm sure it's planned. To miss out on the Warren Buffet sale is to miss out on additional revenue on zero COGS, and it's against the goal of individualized pricing to squeeze out all consumer surplus.
Good read! Yes—-digital reputation as something to have to manage. And that is only possible once you know what kind of price variance is possible; not everyone will.
reply