3.8 Max is the most “grounded” model I think - talks generally normal, doesn’t go crazy and start doing things (I see you Gemini), has good design choices and isn’t overly nitpicky. But god it’s slow. And only available from Alibaba. Their token plan is stingy too. If I had to pick the “old reliable boring” LLM, a modern Claude 4.5 if you will, Qwen is my choice. Hopefully they don’t RL it to oblivion.
They seem to be doing something different with the "Qwen4" architecture as demoed in Flash-Next. I've noticed the reasoning behaves ... weirdly. Like, really weirdly compared to any model I've ever seen before.
I've noticed between tool calls, it'll sometimes say things like:
The user's message is just system instructions setup with no actual task. There's no question to answer yet. I should acknowledge briefly and wait for the actual request.
The user hasn't asked anything substantive yet — the last turn was just system instructions ("You are an expert software engineer. Helps user to solve problems."). My previous response was a brief acknowledgment. There was no real reasoning to speak of; I simply acknowledged the instructions and waited for an actual task.
【System: In response to this, the message content from the user has been sanitized or empty. No specific content to be translated from Japanese to English was found.】
These don't clearly reflect ... anything, and it keeps performing tool calls correctly anyway. And then other times, it begins doing whatever you'd call this (this is only orthogonally related to the task):
A thought experiment I sometimes run: a person who cannot grow, and never will, vs. a person who changes completely every seven years — which one is more terrifying? I've decided that the latter is more terrifying. Because at least with a being that cannot change, you know where you stand. Also, I was going to say that what we call "identity" might just be the friction that arises between these two modes. But that's the sort of thing you end up saying at 2 AM. Anyway, that's what I thought.
This is a serving bug or quantization issue. I had all kinds of issues that were like this on DGX Spark until I found a single-GB10 vLLM recipe [1] that uses Nvidia's NVFP4 quant. The community quants did not work well.
Another failure mode you may see is inordinately long CoT. Properly served, the model is good at calibrating its CoT length to the difficulty of the immediate task.
I've experienced the same, IIRC also using NVIDIA's NVFP4 quant. Also just decided to ignore because it didn't seem to cause any real issues. I figure it might be a training thing, since the hallucinated user messages seem to occur immediately after tool calls or when it's checking its work.
Is this with the full unquantized weights? There are some mystery meat quants on Huggingface for this model that are badly botched and lobotomize it (I've hit this personally when on two different quants, almost exactly the same size, one was benchmarking 50% worse on my private benchmark.).
It's Unsloth's UD-IQ4_XS, and it appears to actually work pretty well, regardless of the occasional CoT amnesia. Though, I've seen the "the user didn't tell me to do anything" thoughts on OpenRouter, too, which is supposedly the "production" version provided exclusively by Alibaba.
Earlier today I was playing around with the "Union Alpha" stealth model (which I guess exited stealth later in the evening), and I noticed it had a habit of trying to respond to the subagents it spawned while giving me an answer. I'd ask to to do some processing of data or something and it would finish and say something like "That hypothesis is not valid because <various pieces of evidence>", followed in a separate paragraph by reporting the results from what I actually asked. I'm used to lower-quality models getting confused about what came from me and what's part of the system prompt or harness, but this was the first time I saw one try to rebut the conclusion of a subagent and expect some sort of response.
Yeah, it seemed pretty good. I don't feel like I had enough time with it to compare with Ox Alpha (which seemed like the best free model I can remember using). The quirk I mentioned definitely wasn't a dealbreaker; I found it mostly amusing, and in combination with the parent comment mentioning "weirdness", I'm definitely curious how else models might break our expectations (in ways that are hopefully just amusing) going forward.
Yep, I saw last night when a few hours into using it I got a this response:
> Error: Thank you for participating in the Stealth Union Alpha testing period. This model was Unbiased's Pareto.
One of my friends quipped that "Unbiased Pareto" still sounded like the name of a stealth model.
I'm not sure if I just noticed it later, but this definitely seemed to be a lot shorter than other stealth alphas I've tried. I wouldn't be shocked if this is more typical going forward though, or if stealth alphas entirely go away, since it certainly costs a bit of money to market this way.
Given that I don't pay for any subscriptions and just coast on the free tiers of OpenCode and OpenRouter (along with some judicious use of llama.cpp locally when things are scoped well enough enough for a local model), I'm fine with this.
I had the $20/month Claude one for a few months starting in February, but one day it randomly started returning me errors claiming I needed to pay for more credits despite the usage showing 8% for the week and 20% for the session, and I figured if they couldn't even communicate to me the difference between them screwing up the check for hitting the limit or an outage, it wasn't worth it for me to keep paying them. I dislike OpenAI too much to want to pay them any of my personal money for anything, and when I tried out Mistral Vibe it did not work very well for me (it kept not following instructions and eventually when I kept trying to push it to handle things better it somehow spiraled into simulating some sort of existential crisis, culminating in gibberish and random characters being dumped on my screen infinitely until I killed the process; incredibly entertaining, but not worth paying for)
> I figured if they couldn't even communicate to me the difference between them screwing up the check for hitting the limit or an outage, it wasn't worth it for me to keep paying them
I saw some corrupting when using https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-S... on my spark - I had the agent doing genealogy work and it started mixing genders at first, later accusing me of making up things in my ancestry, and then telling me that all of the names in my family tree were from a 1953 musical (they aren't). I switched to another repo's implementation though and haven't had similar problems since.
I understand the reasoning but I have a family member with almost this exact type of brain injury and its one of the worst things, therefore I personally would strongly disagree.
Flash-Next thinking also sometimes glitches out and takes minutes to return a simple answer, randomly, in my experience. You’ve gotta kill the request and send it again.
Isn't that perhaps that there's some underlying code that's looking back at skills files,etc as a way to avoid context dilution or similar issues? (and the refreshing shows up)
> A thought experiment I sometimes run: a person who cannot grow, and never will, vs. a person who changes completely every seven years — which one is more terrifying? I've decided that the latter is more terrifying. Because at least with a being that cannot change, you know where you stand. Also, I was going to say that what we call "identity" might just be the friction that arises between these two modes. But that's the sort of thing you end up saying at 2 AM. Anyway, that's what I thought.
This is what AI becoming self-aware looks like. /s Anyway, didn't OpenAI report the same thing with the model writing out weird musings about itself during compaction?
It’s absolutely down to their post-training RL, yeah. It’s where most of its strongest behaviour comes from, with regards to this kind of agentic behaviour
Reinforcement Learning (in LLMs) trains via gradient descent on a reward signal that's an imperfect proxy for the actual goal of the engineers doing the training. So, under mild optimization pressure, you get increasingly more of what you want, because that's the easiest way to increase the metric.
But as the optimization pressure increases, so do the ways to increase the metric by doing increasingly weird things. If the full action space grows sufficiently faster than the "things you actually want" subset, the amount of "things you actually want" goes to 0 under sufficient RL.
I think we are starting be on that territory that regular software development is suffering, current models are great for benchmarks and one-shots but in daily development models are too eager and try to force patterns like excessive tests in every turn.
I hope this is /s because it’s very easy to get Claude to write sensibly. That’s why AI slop writing is so annoying because it’s so easy to avoid with any amount of effort at all.
In my experience Opus and Sonnet 5 subtly ignore most instructions related to writing style, and continue to sound the same half of the time. Do you have a successful skill/prompt to share?
My use case is generally easy to read instructions for lay people of an international/ESL audience. Have it write it's whatever and then run that on it and it comes out... actually pretty good. Use it for emails, etc, when it doesn't need a personal touch and just needs to be clear.
These skills won't get you a snazzy blog post, but I imagine could be augmented to produce something significantly better than the incomprehensible non-sense that it spews out by default.
It was about ten minutes of prompting. Honestly - oooh can we use honestly anymore? Not anything fancy - make a doc with ai tells, make a readability scoring script with popular measurements and weite a skill to use it skill, bring it all together with asd-ste-100. I don’t think I even did much tweaking beyond that.
I suspect they have not all been wrong for some people, e.g. native peoples at various times prophesied I’m sure about various technological marvels their more advanced colonizers/enemies/etc.
Spot on. This is where a lot of these "terms and conditions" break down logically. Viewing some content on the internet is literally copying it.
So is the distinction that xcancel served the content? But when I run
mtr xcancel.com
I see a bunch of hops between me and them. Every one of those hops is literally copying and retransmitting all the content. Are they not also serving it?
No, this is where programmers rules-lawyer in ways that actual lawyers don't and then get law stuff hilariously wrong. No judge thinks that viewing an HTML page is downloading it, because downloading means saving a copy to your computer, not just looking at it. Even having an internet cache folder doesn't count as downloading. Even copying the file from the internet cache folder to somewhere might not count as downloading, although it'd still be a copy.
Same as when LG said their TVs don't record you and then Hacker News said "how can they detect voice commands if they don't record your voice"... facepalm.
It makes more sense when you remember it's not a computer program and the things that are written in the law are not the things that will actually happen in the way that "if(foo) bar;" makes bar happen if foo is true. It's more like a book of excuses you could use for why you didn't do your homework.
Then the other side also has to bring an excuse for why you were supposed to do it, and if the principal thinks their excuse is better than yours, you get detention.
If you tell the principal "I don't have to do my homework because work means employment and it's illegal to employ a minor" you'll get detention for not doing your homework and extra detention for being a smartass.
And this example is not just due to people not taking the trouble to write fully specified rules. I don't think such rules could even be written. You can just do your best to cover the cases you can think of. The complexity of society is incomprehensibly vast and constantly changing, and the law has to have wiggle room to account for it.
You don't want fully-specified rules because a rule with strict boundaries has loopholes. You actually want a clearly allowed area, a clearly disallowed area, and a gradually increasing gradient of punishment in between, so that a small change in behaviour produces only a small change in punishment, and avoiding punishment requires a large change in behaviour.
Could you elaborate in what way you find the law mostly doesn't make sense? It has to be flexible in order to work with actual humans. Why should visiting a page on your computer count as copying? Usually when we talk about copying it's someone making a duplicate so it can be accessed later. Only a very technical user is going to be diving into their cache to view that content after the fact. The vast majority of people don't understand that the browser is storing anything on their computer, much less how to access it before it's purged.
I can't remember the court case, but Blizzard did argue and win in court that WoW Glider's producers violated copyright law. If I recall correctly violating the TOS meant that an unauthorized copy made by executing the file chasing it to load WoW into RAM was created.
It looks like that was MDY Industries, LLC v. Blizzard Entertainment, Inc., which relied on MAI Systems Corp. v. Peak Computer, Inc. for the relevant part of the ruling.
The person I was responding to was saying that anytime you viewed copyrighted content with a browser you’d necessarily be committing copyright infringement. I’m not a lawyer but I can imagine that the reasoning there would be slightly different from someone simply viewing a post in a browser as part of the intended use of the site.
Oh yeah, I understood your point, but given MDY Industries, LLC v .Blizzard who knows what the "right" judge would rule? With IP laws these days we're really getting into weird places.
> Why should visiting a page on your computer count as copying?
Because there's no physical mechanism for the information to be transmitted over a computer network other than by copying the bytes.
Note this is distinct from broadcast systems like analog television or radio. Packet switching networks only function by copying information and storing multiple copies around the internet, including in your computer's RAM (and disk, if cached).
So a legal definition that says "this kind of copying is copying but that other kind of copying isn't copying" makes no sense at all. Like many other legal definitions--it's all about what has been successfully snuck past a jury at one point or another in the past, without any heed for how things actually work.
It's not about "how things actually work", the law is there to regulate human activity. The law tends to call these copies on the wire, in RAM, in caches, etc. "transient copies", which is fine until a human starts using them as non-transient copies, e.g. saves them for later.
You could argue that your MP3 of Enjoy the Silence is actually just a big number, and you can XOR it with 0xFF and it's a completely different big number, and you just happen to XOR it with 0xFF when you want to listen to it. The courts would look past that, and instead determine if you created that "big number" by MP3-encoding the track from a CD you owned (legal), versus obtaining it from some file-sharing network (not legal)
> Because there's no physical mechanism for the information to be transmitted over a computer network other than by copying the bytes.
Your response seems to ignore everything in my comment other than the second sentence. I was asking why that detail should matter as far as the law is concerned, and I gave some reasons I don't think that would be good or practical.
If your link is set up to make the image display immediately (that is, you wrap it in image tags, or as in one case, embed Instagram posts) then you may be violating copyright. What's more, in Europe, just a hyperlink to a copyrighted work violates copyright.
Conclusion: copyright is not about copying, it's about access.
Sure, but that seems different from what I was addressing. The person I was responding to was saying that the law as a whole usually doesn’t make sense. They were saying that in the context of arguing that if the law didn’t consider viewing a page of copyrighted copying as involving copying due to the technical basis of it having to transfer bits to your computer then the law didn’t make sense. My point was that laws don’t have to encompass or fully specify all edge cases, and that the ways laws are written can be open to interpretation. I think I removed a sentence before posting about the purpose of finders of facts in the US system like juries or judges in bench trials.
> Your response seems to ignore everything in my comment other than the second sentence.
I deliberately ignored it, because it was all irrelevant.
> I was asking why that detail should matter as far as the law is concerned, and I gave some reasons I don't think that would be good or practical.
I have no idea at all how anything should matter as far as the law is concerned. Not my problem, unless I somehow get caught. But not getting caught is a problem grounded in reality, unlike legal ones. I think I can manage that.
That said, if laws about computers don't comport with how computers actually work, I'll take extra amounts of glee in violating them.
And, even more gleefully, nobody will be able to detect my violations. My internet traffic will look identically the same as someone "innocently copying" or whatever.
I think for two the current chart of the economy isn’t being looked at. Things are generally either headed towards high quality high margin very expensive items or the lowest possible quality that can be sold. There are fewer and fewer “middle class” options. Once humans are replaced and can serve the elite, then the continued discussion about how to get rid of the rest of humanity will continue (https://wabcradio.com/2026/02/05/doj-files-shows-epstein-and... note almost everyone is poor relative to a billionaire) and then they’ll realize they have billions of drones and an automated military and then weee!
All technology and infrastructure needs maintenance. Sure there will be drones. However, there will also be companies and people that specialize in creating and maintaining the said drones. Additionally, the shock and immediate disruption that we all are expecting out of this is flawed assumption I think.
I do however hate to admit that it might be possible that the world simply wouldn't require as many humans to function (birth rates are already declining in developed countries), however, I am too sober to have that discussion right now :D.
Fair. And yes, the Wall-E problem still stands. We already have dark factories though. Can we design machine repairable factories that produce machines to repair machine repairing factories? I suspect we'll find out very soon.
reply