Route A is good for rapid, disposable prototyping. If you ever have a stray thought, “I wonder how this would work if the whole paradigm were turned sideways”, you now have a chance to preview a “working” version of your idea. If you like it, discard the code and reimplement it correctly. In this way, I think it can be a good adjunct to sketches and other lofi prototyping techniques. Just don’t outsource the creative ideation to the LLM, because all you’ll get are the same solutions as everyone else.

Have you seen many of the replies in this thread and others? They’re not doing that. They’re putting it into production code. Their company’s thought leaders are trying to say that review itself is a waste of time, etc.

That’s not “disposable prototyping” that’s whole versions of the codebase written with barely a human in sight.

> The bitter lesson is about hand tuned AI vs computational general methods. However in truth today’s AI uses both. We have general compute heavy models which require narrow expert instructions

The models are not even really trained bitter lesson-style anymore. That concept peaked during the era of pre-train scaling, back when it was thought that making a bigger and bigger GPT-3 would automatically solve all problems through prompting. In 2026, the most important part of training is post-training, which uses vast quantities of niche, hand-curated data to fit the models for specific tasks in domains like tax law.

Does anyone have an idea how this might perform on a DGX Spark at longer contexts? I've been trying to investigate their performance with these medium-sized MoE models, but I'm seeing a lot of incomplete and conflicting information. The 273 GB/s bandwidth looks awfully bad on paper...

A single Spark alone is not worth the price. You are paying $1000 just for networking equipment you aren't using. At 2x it starts to maybe become worth it if you don't want to deal with Apple. Outside of the newest Macs, I can't think of anything else you can get 256 GB ~550 GB/s memory bandwith for $8200. Even at 3-4 Sparks it scales relatively well.

With 2x Sparks, I am getting 40 t/s. I'd guess that without MTP you'd get 12-15 on 1 Spark, maybe 20 with MTP?

Given the new architecture, speeds are harder to estimate. Max would be ~40 tok/s.

I can run Deepseek flash 0731 version (ds4, esl3) on single DGX spark. getting around ~20 tok/s. It's great. Quantized version of this model would probably run on the DGX spark. I am excited to wait for quantized models that fits in single DGX spark.

How about proof that black-box distillation can deliver these results without a very sophisticated RL pipeline doing the heavy lifting?

"Black-Box On-Policy Distillation of Large Language Models", Microsoft Research, https://aka.ms/GAD-project

> 'GAD consistently surpasses standard sequence-level distillation, delivering superior generalization and achieving performance that rivals the proprietary teacher. These results validate GAD as an effective and robust solution for black-box LLM distillation.'

No RL, although I'm a little bit surprised to see MS Research publishing a paper on distilling GPT5?

That's an interesting paper, but there is virtually no discussion of reasoning behaviors or optimization for long-horizon tasks (i.e., all of the recent advances in LLMs that people care about). The evaluation methodology also is pretty dated:

> We reserve 500 samples of LMSYS-Chat-1M-Clean as the primary test set. We also include test datasets consisting of a 500-sample subset split from Dolly [6], the 252-sample SelfInst dataset [37], and the 80-question Vicuna benchmark [3] to evaluate out-of-distribution generalization. We report the GPT-4o evaluation scores [45, 10], where GPT-4o first generates reference answers and then scores the output of the student model against them. We also conduct human evaluations on the LMSYS-Chat-1M-Clean test set for qualitative assessment.

It's fun to imagine that it could be GLM 5.3-Flash. Between GLM 4 and 5, the flagship's total parameters doubled and the active parameters went up 25%. GLM 4.7-Flash was 30B / 3B active. If this model were 60B / 4B active, that sure would hit a sweet, currently empty spot in the lineup of open models.

But I don't know. There are vagueposts on X about this model running on two DGX Sparks. If they did the same scale-up with Air instead of Flash, that would probably be about right.

This is such a tough problem. Anthropic would need access to some kind of technology that could, like, intelligently handle unforeseen circumstances and nuances. Yeah, that’s definitely not something we should expect of them.

It’s interesting how much Anthropic itself is a demonstration that its own hype is false.

Just use any product going all in on AI hype. In 5 minutes you will see annoying bugs, server is down, non-sensical press releases, confusing UI. This is GitHub, this is Cursor, this is Anthropic, this is Google, this is all of them.

Including most of vide coded apps one sees, even from people who they'd trust before.

Anecdotal example, I downloaded a new alerting app recently from an indie dev who had a small following back in the day in iOS space. It asked for a subscription, like $20/year.

I thought, let me try this the (final version, from Mac App Store) app first. Well, it's a barely-there vibecoded shit. There's a bare-bones list, everything looks like my nephew designed it, the macOS "app" is a iPhone-size view of the iOS one, it has a bug that if you click on it it opens multiple duplicates of the same list view for no reason that you have to manually close, and in general it barely works.

Yay for vibe coding.

Not to mention if their tools were so clearly useful they wouldn’t spend so much time making UI updates designed to force, trick, or confuse me into using their tool when it wasn’t my intention. SaaS companies with assistant integrations are the worst about this (looking at you, HubSpot)

Agreed. I think LLMs are best used as pair programmers or typists for users who already know what they’re doing. Or as tutors for users who want to learn.

Vibe coding is mostly garbage. But it can be useful for creating instant, disposable prototypes to investigate an idea or design direction.

The issue is that, if there is a “good enough” point approximately here, it is only a matter of time before models become small and efficient enough not to need all those data centers. Though, it should be good for companies that sell computers (like Apple) rather than putting a toll booth in front of a pile of numbers.

GPT-OSS 20B didn’t really merit the fanfare even when it was released; it’s definitely not competitive now. Even the 120B version has been well eclipsed by smaller LLMs at this point. The last version of Qwen 27B/35B was better, and now the new one is even better than that!

> We've been seeing that headline for a few weeks now and I really don't understand the problem.

It’s powerful symbolism. It reminds me of that tone-deaf iPad ad that sparked outrage in 2024. The one where all the cultural artifacts were crushed in an industrial press to make a soulless slab of glass. And that was Apple, who is generally well-liked by the public.

But it’s the most useless canary ever. We already know that certain topics are taboo in China.

As for all the other uses the models have, it seems pretty clear they’re not doing anything weird. If they were, people would be posting examples of that and not of Tiananmen Square.

It's a fine canary for the question "is this model Chinese?" Which is pretty much where this thread started.

What it really helps with is "is this _provider_ likely Chinese?", as we've seen, many of the big Chinese models have no issue on their own discussing the taboo subjects. It's the higher level provider that is filtering output.

If you run the same query a bunch of times you'll see filtered responses mixed with model output where it says random stuff varying from 'nothing to see here' through 'The government of China cares deeply about its people...'