Exactly what I am saying for months now. And it's exactly the reason why I am shifting to open weight models now. Just bought myself a 2x DGX Spark Cluster. Will run Qwen3.8 Flash Next on it, maybe Qwen4 when it comes out.

Not only do I have full control over quantization and inference, but also will I experience a constant level of quality. It won't be frontier. But it will be stable, and that's enough reason for me to switch. Also I will likely save some money on subscriptions.

Perhaps reverse the question: Are your build/construct requirements just really counter-productive to how LLMs need to understand things?

I've found constructing the code, writing the tests, adding the docs; then running through them gets most of the way there.

I've also found that making a simple obvious edit is a useless endevour when the LLM is primed for the long context tasks.

So, again, the question is reversed: are you over reliant on the LLM to do even stupid simple likes like editting a css variable?

If I ask something like: "I'm building a simple X as a learning exercise, I'm writing the code so please only answer the question I'm asking and don't try to solve the problem directly. How does ..." There's a 30% chance it starts reading and writing code immediately and a 20% chance it argues with a "design decision" that will bite me in the non-existent future of my learning exercise. If I ask a follow up question, naively assuming that the context from my original question still stands without repeating, it will almost assuredly start making modifications to my code.

Co-Authored By: Haiku 4.5

But here’s the standard question: At what speeds/other limiting factors?

But also, everyone says "it's the harness" and almost nobody ever gives good examples, it gets a bit tiring to read everywhere, as if everyone wants to sell a harness to us.

Yeah, expecting the world when all you have is a 8GB graphics card? You're going to be disappointed.

16GB is table stakes (IQ3_XSS). 32 GB is better.

The problem is they can't fit any frontier level open models.

I'm asking, not arguing, because I'd like to understand. Is Astra so much more expensive for those tasks, and are they frequent?

Say I'm down to about 5% of my 5 hour limit, I'll stop what I'm doing to call the handoff. Then I logout and login to the other account and call the resume skill. It adds about 5 minutes of friction to do all this and resume work.

Where'd you get 30 years from? Show your work.

Yet… even Altman called out Anthropic for serving dumbed down models.

Shits weird man

Even Altman called out Anthropic? Isn't Anthropic the biggest competitor Sam Altman has?

Not by a mile. They're even (probably illegally and there's apparently a class action lawsuit oncoming: at least something to that extent was posted on HN today) teaming up, as a duopoly, to push for the same bullshit regulations / "we need to slow down AI research".

The reason they're teaming up is the real competition is, as in many other domains, China.

Unlikely. The $200 Claude subscription allows for billions of tokens/month, and that kind of hardware will take years to amortize.

How do the numbers pan out? Ack that it isn’t always juts about cost, so even if it is pricier to self/host it might still be better for you for other reasons.

I'm getting really sick of the constant rot and "magic breakthrough" cycle, so im going full local, at expense on paper but being able to trust something which I need to understand the reliability of is priceless.

I like predictable. I'll take slightly less capable over unreliably capable, since with reliable i can calibrate my expectations and learn what aspects of my workflows to entrust and trust it will work. You simply cannot do that with models you don't control and in my experience they will all errode after the initual marketing wave passes, likely you eventually get fed heavily quantized versions and are expected to accept degraded service when what convinced you to pay was a far superior product. No such issues with local.

This wouldn't explain progress on benchmarks (including closed sets), or the fact that newer models are providing solutions to major math problems that older models cannot.

Is there any training variable here? For example, can a model released in October perform better on the same benchmarks vs its predecessor released in July just by virtue of training on newer data that was made available on those 3 months?

Sorry if it’s a dumb question, I don’t really know much about the topic.

Opus 5 came out with better benchmark results than Fable, but it really did not feel better to use at all.

* Get everyone to talk about you as the first model to ever score 100 on the 100benchmark.

* Tune it down over time so that you end up only scoring 75 on the benchmark and people get used to it, gaslight them into thinking it never changed or that it's just a harness problem, they can't run the old version locally anyways to verify. This also cuts your costs in half. Your gross margin on API calls goes from 70% to 150%.

* Release new model that scores 120 on the benchmark and advertise it as 50% better than the current model, while it's only in practice a minor increment. Everyone praises it as the second coming of Jesus Christ.

* Get everyone to talk about you as the first model to ever score 120 on the 100benchmark.

Bis repetitae.

Where?

I see so many accusations of this happening and it's so easy to check, but nobody ever proves it.

I imagine some people have their own personal in depth benchmarks they could do this for.

Your benchmark didn't get 100 ? It's normal, it's not deterministic, and also your harness is wrong, and also you didn't do it when US users were offline, and also you got it wrong, and also we don't care about your results, the hivemind is speaking louder than you (also our bots are spamming more than you and drowning you out).

This very website has, at all times, a group of people saying "<Previous model> was never good enough for coding, but <current model> is the best thing and a game changer!" while the other goes "<current model> bad, <previous model> was better!". It's all vibes.

Those software settings are being changed in real-time by an algorithm and those algorithms are being tweaked and A/B tested daily by the ~~performance~~ revenue optimization teams. On the hardware side the footprint a particular model is running on is materially changing, growing or being re-distributed across DCs ~weekly.

GPT5.6-Sol on Max thinking just became regarded as of a few days ago.

The boosters will tell me it’s my fault for using such an old, cheap out-of-date low quality near useless wish.com model (that was SOTA and better than human coders one month ago).

The cycle repeats.

Thanks for your insight

Sam Altman believes he can train a model entirely on synthetic data, which he admits would not have human world knowledge but is interesting none the less, which likely led to their mathematical models.

Overall models have become cheaper to run and smarter per token.

They might but multiple competitors engaging in ongoing deception as an intentional corporate strategy isn't required to explain what we're seeing. It's entirely possible to get the same clearly unethical outcome without any employees knowingly participating in an explicitly unethical plan of record.

Instead it happens without overt coordination when individuals and groups within an org each pursue their local metrics and incentives. In isolation, no individual action seems obviously unethical on its own. They just look like 'optimizing performance', 'maintaining ASP or ARPU targets' or 'achieving operating margin', etc. Customers are still getting deceived and receiving less for their money than they think. The difference is most of the people involved in enabling it get to not feel bad about themselves.

We are being A/B tested on and there is nothing you can do about it.

Oh, yes there is. DeepSeek 4.1 Flash on max thinking can simply be dropped into Claude Code. Close your eyes as the chain-of-thought traffic scrolls by and you can easily fool yourself into thinking you're still running Opus, in terms of both cognition and throughput.

To be fair, matching Opus's throughput costs about as much as a new car, but cars suck nowadays and you didn't want a new one anyway, right...? Failing that, rent a cloud server, one that you control.

https://www.tomshardware.com/features/crucial-p2-ssd-qlc-fla...

https://www.tomshardware.com/news/wd-blue-sn550-ssd-performa...

https://www.tomshardware.com/news/adata-switches-nand-on-sx8...

Otherwise, even if one company did something like this, everyone would notice because the other companies would be pulling ahead. Are all the AI developers coordinating a "dumbing down" of models? i.e. Are Open AI, Anthropic, Google, Meta, DeepSeek, Mistral, xAI, and so on, all working together?

So we might be approaching some limit (the "there's only so round a sphere can get" argument). But I very much doubt there is some massive conspiracy between all the AI developers.

That was one of the main points of the movie 'A Beautiful Mind' - that actors can coordinate without any explicit communication.

Fable seems to be following the same enshittification arc of other Anthropic models.

Generally OpenAI seems to be taking the opposite approach with an increasing improvement over time. As sad as I feel to say this, open ai seems to have the right strategy. Making your product worse over time rarely plays well with customers. At this point it feel often hard to justify using Anthropic for anything. I generally like Anthropic better as a company and they really had the initiative and advantage and customer good will, then proceeded to squander it faster than a cigarette company or the Sacklers could have.

I have no way to judge myself. I don’t even use them. But it’s interesting to follow by just reading stories and comments

On the other hand, New Coke was a resounding success. Well, it, itself wasn't, but in the aftermath, Coke outsold Pepsi 2:1.

I would do a test to verify my suspicions.

I’m actually so far removed from this tech that I couldn’t run such a test myself lol

Not unless your competitors do the same, or else you will only be perceived as falling behind others.

If one lab makes a breakthrough, can the other labs just distill to bear parity anyways and then set a new baseline industry wide.

I’m quite ignorant on this topic, so if any of this sounds moronic, forgive me

A very small raise in prices may cause a very large loss in customers that you risk never getting back.

For example if customers figure out that the Chinese models are just as good, they are gone because they are so much cheaper.

This isn't what we see in benchmarks.

there's also probably load balancers that downgrade models during high use.

Yes, the AI technology is known primarily for how stagant it is.

I have no idea, just had a thought and put it out there

https://thedailywtf.com/articles/The-Speedup-Loop

Surely you're not talking about the AI industry. Astra was released less than 3 weeks ago, and Fable-level models became public only 6 months ago. The rate of change is dizzying.

Change is fast and abundant, and at the same time, it is hilariously more mundane than the dangerous-AGI-in-six-months tune we've been reading daily for years.

I would define it as a quick-moving market, but not nearly moving enough for the fantastic claims they make to justify ever-increasing funding.

The best case scenario is a situation like Y2K: a ton of people coordinate and work hard to produce no perceptible change, because unlike catastrophe, averting catastrophe feels boring.

Absolutely none of this points to “stagnant.” Stagnant is a terrible description of the AI industry.

The harnesses have improved somewhat, but the code produced on large or legacy code bases is still very average and I still see similar mistakes made that I saw back a year ago (although less now that harnesses have become better at steering).

For my use cases, we are definitely on the flatter part of the curve at the moment.

I’m not saying it is stagnant. I’m saying for a hypothetical industry that was (maybe that fits AI, maybe not, I have zero authority to say myself)…

Are you really saying AI is a stagnant industry?

Unlike you, however...

Kind of reminds me of that, but with more smoke and mirrors

HN: Well, must be a stagnant industry...

In addition to the dozens of opaque model parameters and hardware variables that can nerf or buff model intelligence, speed and profit, there's also the very real possibility that models aren't just training on benchmarks but could be evaluating if they are being benchmarked in real-time and applying more resources adaptively. 'Driver optimizations' that detected benchmarks in real-time were deployed in the first 'GPU Wars'.

> I have no idea is the actual frontier is stagnating.

Like a lot of complex, rapidly evolving tech, the truth is it's probably rapidly accelerating on some measures for a few and stagnating on many others for most - hence the divergence in user reports. It's depends on how you use it, for what problems, how rigorously you assess the output and whether you happen to be on a server bank, RAM pool or shard at this moment which hasn't yet been sufficiently 'cost optimized' by the margin algorithms. They don't call them load balancers anymore. They're Margin Balancers.

Could be what happens next, though.