ignoramous

> spyware to be installed as a device administrator, granted accessibility privileges, and then given every possible Android permission via AppOps

Correct, and apps using these setups (even for genuine reasons) have been under the cosh.

> At that point, ADB no longer matters.

"on-device adb" wasn't meant for scenarios it is being used for (via Shizuku, for example). It is a pointless attack surface.

> anyone with physical access to the phone would still be able to install and configure their spyware via a USB cable.

This same case can be made in support of removing on-device adb. Anyone with physical access can install & configure for their use cases. If you claim that's more hassle than worth, then you have your answer (that is, added hassle for stalkerware sellers, too).

Disallowing on-device adb is restrictive (I co-develop a security app that can absolutely make more from elevated permissions, via Shizuku or Device Admin; let alone root), but I don't think Google will do so because they want to close Android (the platform) more than they need to. To me, given the very human costs of stalkerware & financial fraud, it is an understandable, even if disappointing, security decision. Otoh, I do get that the road to hell is paved with good intentions...

No.

The article overlooks security implications from spyware, which is a huge problem not only for financial applications, but personal safety, too [0].

Per FTC, a stalkerware will: geo locate, read call list & record calls, read notifications, texts, & possibly emails, access gallery, camera, & files, and monitor network activity. [1]

You could do all of those with "on-device adb" (in some cases, with just the appropriate permissions), without root access. Stalkerware & financial fraud enabled merely due to the scale & reach of Android (half of humanity uses it!) and lack of basic security literacy warrants such protective measures, as (Thaler & Sunstein would like to remind us) defaults matter.

With conspiracies abound, we must not lose sight of tech safety and related issues, which almost exclusively affect the most vulnerable & the most disadvantaged.

[0] https://www.techsafety.org/spyware-and-stalkerware-phone-sur...

[1] https://consumer.ftc.gov/articles/stalkerware-what-know

You need update your blog posts.

  Our whole stack is radically open source — frontend and backend alike, Apache-2.0 licensed — and so is everything behind this benchmark. That is how a benchmark number earns trust: verifiability, not hype.

The repos have since been moved to BUSL-1.1: https://github.com/Lore-Hex/quill-router/commit/8155ac666ae0...

Do-able in many low-income countries but it'd be borderline miserable, by modern consumption standards.

Mixture of Models, perhaps? tbf to OP, the setup they're proposing has also been recently evangelized by other "ai gateway" products (like OpenRouter, JusCode, Fireworks etc), so there's likely something useful here.

I see eventuality here as job cuts or salary cuts.

Don't think that day is far when "software people" are paid as if they were taxi drivers.

> I can't use it for theoretical physics because I can't evaluate the responses.

A tool, even if it is a chisel, in the hands of a master sculptor would obviously result in a wildly different outcome.

The free OpenCode Zen models (I've only had the need for DeepSeek v4 Flash & MiMo v2.5 for one-time / throw-away tasks) did work, for me, with GitHub Copilot & Reasonix.

Sounds like 3 more iterations on K3/Fable, and we'll able to ask it to break computational hardness assumptions like DLP?

> I'm bad at math ... he includes the GPT5 prompts for his conversation, which are easier to follow

You were kidding, right?

Simplest way is to signup to OpenRouter and filter out all non ZDR (zero data retention) providers. Paying "API rates" however can prove expensive compared to coding plans (for instance, MiniMax $20/mo coding plan allows 1.7b tokens; depending on input/output/cache ratio, it is worth $200 to $500+), except for Hy3, DeepSeek v4 Pro, and MiMo v2.5 Pro (whose API rates are cheap and/or discounted already).

If you're looking to not have to deal with Chinese providers, AtlasCode ($20/mo), OpenCode Go ($10/mo), and Cline Pass ($10/mo) provide 2x to 6x usage for some of the popular open weights (depending on the model).

Personally, I subscribe to Z.ai ($17/mo), and pay API rates for MiMo v2.5, Hy3, Qwen 3.7 Plus, & DeepSeek v4 to the original providers (Xiaomi, Tencent, Alibaba, & DeepSeek).

> focusing on what they can get wins in like speed instead

Speed as a differentiator has always been Google's thing. They (used to?) show the microseconds it took to query & rank web-scale search results. Chrome, notoriously, focused on speed at the expense of resource use. The very many efforts to efficiently speed up Android & its runtime since its inception, and so on...

> their big model underperforms chatgpt 5.6

Possible but TFA claims:

  We have started our most ambitious pre-training run yet, for Gemini 4 ...

> "The highest tier Chinese models are not more economical than US frontier models. Try GLM 5.2 and see how much it costs to do real work. I did, and it was more expensive than GPT 5.6." This is a flatly false statement.

It may not be false but may be a "category error" [0]. Reserved GPU pricing & bulk inference pricing is 3x to 6x cheaper than "API rates", but renting your own GPU cluster (in this crunch) to run a 600b+ open weights is going to be "more expensive than GPT 5.6".

Even then, it remains to be seen if Huawei will pull their weight (and match up to Nvidia) as spectacularly as their fellow Chinese AI Labs have. If so, the WAICO alliance is ready to go all-in.

[0] Ben, and probably other "influencers" in this space, may be prone (knowingly or unknowingly) to favour points that make their conclusion for them (https://en.wikipedia.org/wiki/Motivated_reasoning).

Don't think it is down to Wang or MSL but Meta's focus on "personal AI" led them to whatever strategy (OAI missed the developer market too, which Ant then captured; leading to several high profile departures at OAI, coincidentally hired from Meta). The original Llama team themselves started Mistral which hasn't gone anywhere. The simple fact of the matter here is, Chinese firms have the money and the talent to rival the US ones in this field, should they as much miss a beat.

Don't think HN is anti-AI but experts of course may have a different perspective. A comment in this discussion points out that LLMs have been doing math that's been long overlooked in favour of perhaps more impactful work: https://news.ycombinator.com/item?id=48974274

  There isn't a lot of money in academic math, and the ones that love it don't look for low value findings. Proofs like these are ... usually the domain of hobbyists ...

This makes me wonder, what if anyone uses Fable-class LLM and passes of its novel results as their own work?

There's no shortage of folks doing that in software, right now.

Erstwhile seafaring powers went to war over trade routes for these. Surprising it isn't that well known.

Curious: On which plan? In my experience, their $20 plan is super inadequate as a daily driver (ditto for Qwen's $30/mo), though given K3, I'm tempted to trial $49/mo or $99/mo.

> It's cheap enough to where I don't think about cost, made even cheaper by DeepSeek having the most effective and cheap caching in the industry

I use DeepSeek v4 Flash & MiMo v2.5 Pro. Prefer the latter over DeepSeek v4 Pro because it costs the same while being equally good & less chattier for coding workloads. Although, I've begun experimenting with Hy3 (as an in-between Flash & Pro) & GLM 5.2 (for long-horizon tasks).

> tried Kimi K3 on a task I've done with every other model I use regularly and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan

ArtificialAnalysis puts Kimi K3 just below DeepSeek v4 & GLM 5.2 in token use per task, which is about 2x to 3x more tokens than Grok 4.5: https://x.com/ArtificialAnlys/status/2077832879187620192 / https://archive.vn/zBbFi 2 other open weights MiMo v2.5 & MiniMax M3 are comparatively thrifty.

> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare

I always put my coding subscriptions (that allow it) through "AI gateways" (Cloudflare & OpenRouter are free) which help track token use.

In my experience, Kimi & Qwen Cloud have opaque & restrictive limits, their "credits" drain faster. I now make it a point of subscribing (directly [0]) with providers that are transparent like MiniMax, DeepSeek, Xiaomi, & Z.ai.

[0] OpenCode Go, Cline, and AtlasCloud have generous limits for open weights, otherwise.

Open weights are indeed a commercial risk because there's no shortage of companies that'll make a business out of running MaaS (model as a service), close & resell a fine-tuned or distilled variant (DeepSeeking of Llama / Cursorification of Kimi, for example).

> Android was initially developed for phones with a keyboard (similar to Blackberry).

It was codenamed "Astro Boy". Btw, the team Andy Rubin assembled to build "android" first built OS for Digital Cameras viz. FotoFrame.

> I'll make a prediction

You just hope that BigLabs & BigTech doesn't gut out the talent from Chinese labs. They certainly have the money & impetus.

Astonishing. Considering none of the BigTech except Google (Microsoft, Apple, Meta, Amazon, Nvidia, SpaceX) have managed to challenge OpenAI & Anthropic frontier models, such achievements are scarcely believable.

Re: GLM-5.2: For a ~750b model, it holds up pretty good against models 3x its size (and ~10x the cost). Same goes for Tencent Hy3 and MiniMax M3, which almost match Opus 4.6 levels with ~295b params.

  The full [Kimi K3] model weights will be released by July 27, 2026. Further details on the architecture, training, and evaluations will be released alongside the Kimi K3 technical report.

https://archive.vn/KBzXr

OP's point is, as a consequence of providing "useable security", the secret keys remain "exportable" (in this case, unwrapped in a URL fragment), which is an undesirable trait in some threat models. Let alone web-based cryptography, which in on itself is perilous: https://news.ycombinator.com/item?id=39436238

> continuing to refuse to use chinese models due to security and IP concerns

One can run open weights in an exclusive TEE'd GPU too, which still comes out cheaper than closed weight LLMs. Ex: https://chutes.ai/pricing