We all need to use nuclear, bio and cybersec terms in all our code to make low quality filtering like this untenable. When you can't work on a resume that has cybersecurity or biology terms in it or reply to a job opening that includes them because the "AI" filtering is so bad that it confuses these for threats, that deserves a collective response, particularly to an IPO'ing company that claims they'll make workers obsolete in two years.
The best agent framework is Pi (pi.dev). It is minimal and doesn't assume a use case, runs fine interactively or non-interactively, has an active community building with it and supports everything you need to build whatever kind of agent you want with plugins.
Ethan is a booster but I wouldn't call him a shill. He cites data and mostly in a fair way, though you could argue the sources he chooses to focus on are biased.
Opus 4.7/4.8 often over-engineers on my setups, plus:
- It talks a LOT more like GPT models. You know: wrinkle, shape, gate, coarse, scope, gap, path, production-ready-workflow-of-the-day, and so on -- "that's expected, a consequence of the previous like-driven workflow". If I wanted to get a headache using AI I would have gone with GPT in the first place!
- It outputs text in a much harder way to follow along. I can't exactly say what it is. Maybe a bit of everything? Bolds are missing, bullet points are gone, paragraphs are bland and too long, and it doesn't feel like a model programming with me, but rather a somewhat full of themselves grandpa developer looking down on me. It's very weird to describe this, but it is definitely how I feel.
Granted this can totally be because of the way it reacts to the prompts now. We've got a rather large corpus of skills and "rules and good practices" that Opus 4.6 responded to great, and maybe the new models just get turned into this when fed with them....I don't know.
Either way, with Opus 4.6 being as good as it is, I need Fable to be a significant step up to justify a price increase. if it can get me to babysit opus a little bit less on some stuff, it might be worth it. Otherwise, I'm very happy with Opus 4.6 and hope they don't deprecate it.
I'd argue that 4.8 is a straight downgrade. For every type of task I've tried. It's been a gambit at this point. If 4.6 quits being available, I'm out at this point.
Reading so many contrary positions about which model is better or worse shows how difficult it is to measure intelligence based on personal experiences. Of course, benchmarks try to make the process as objective as possible, but they often don't correlate with our personal experiences.
The other day 4.6 was fantastic for x task. Today, 4.6 overengineered everything and I had to revert all my changes. When evaluating models, perhaps it makes sense to consider luck as an ingredient before reaching any personal conclusion.
Yes but there’s a reason we don’t evaluate these models this way and instead do it as carefully and thoughtfully as we can at scale. Human evaluations are important but they are an absolute minefield of footguns. 4.8 is not a downgrade from 4.6 there is an insane amount of hard data that contradicts this.
Again correct but it overstates the issue. I can say labs don’t want this. This happened arguably unintentionally in Metas llama 4 release, it went horribly, heads rolled, and like several billion dollars were paid for new talent and the org that built llama 4 was destroyed.
Evals come from a million places and new evals and robust perturbations of existing evals abound. They test a variety of tasks in a variety of ways. All of them individually are flawed. Taken together the aggregate signal is highly useful as you more or less marginalize over a lot of different things. Not to mention these companies have plenty of proprietary internal measurements, they build benchmarks themselves to probe their models and then also have flywheel traffic and A/B tests.
You are right to call out benchmarks but to dismiss them or not take them seriously is a mistake.
Listen, you can say “but benchmarks, the benchmarks!” all day long, but consumer know when we are being sold a lemon. If it can’t do the most basic of things at least as good as it used to, this is table stakes. Nevermind that if you can’t do the basic stuff, how on earth can you be trusted with more?
And you can say “If it can’t do the most basic of things at least as good as it used to, this is table stakes” all day long while people point you to much better evidence to the contrary too, I’d rather be on the other side of that.
Listen. I don’t care about evidence. I care about my lived experience for the product I paid for. I used the new product. It’s actively terrible. To the point of not being usable. We’re all ancedata, but what is “better evidence to the contrary”? The known and game-able benchmarks that they know they need to win at, so they train it to. It’s all he said, she said, which is the only reason we keep having this conversation.
Yea but it’s not right? You or I or the myriad of other institutions inside and outside of academia can probe these models with an evolving landscape of evaluation sets, even those unavailable to the developers. It’s just ignorance to claim benchmarks are somehow useless or all being gamed. You choose your tools in the way you want, but just don’t call it somehow better than a myriad of more carefully constructed setups and scaled evaluations.
Actually anecdata I gather on my job from myself and coworkers is the only benchmark I trust anymore, because it so heavily diverges from the “benchmarks”.
I would encourage you to look into the open evals of some of these benchmarks (find one that actually is open-data, this is itself a good challenge), read the results generated and assess them for yourself.
This is what myself and my coworkers (and many other people in this thread) are doing on a daily basis with real stakes and real tasks – which these benchmarks are all aiming to be a proxy for. There's a real, tangible [cost]benefit to [not] using the highest-ROI models and harnesses.
The people with real incentives and skin in the game are telling you that the data diverges from "the data".
I don't mind if you don't take it seriously, our jobs are more important to us than a benchmark is.
But I wouldn't opt-out of using your own eyes and the eyes of others so easily, especially when there are literally hundreds of billions of dollars in invested capital with an interest in a certain outcome... this is how you end up in "Emperor's New Clothes" situations.
Investigating on your specific use cases, codebases, workflows and tasks is important, there is nothing wrong with this and in fact it’s more important than benchmarks if you can do it well but the point is that is very hard and easy to totally fool yourself and go down a suboptimal path. I understand that people are going to do it regardless, I certainly do. And I have looked at more raw benchmark data than I can really even stomach, I can see annotation data in my dreams now.
Eyes and ears of others is incredibly important. But you still seem to think somehow benchmarks is part of some giant conspiratorial cabal. You have institutions without ANY skin in the game making extremely high quality benchmarks. Consider in academia there is little else to do outside of partnerships with these companies. But benchmarks you can do completely independently and with university grant level money (it costs maybe $10-100k for a reasonable benchmark in many cases). Not only that, “real tasks” are what many benchmarks measure. You have these companies with extremely good logging and well scaled measurements to really look at what works and what doesn’t.
At this point I have a workflow that is fairly rote. I've yet to use a model newer than 4.6-1M-XHIGH that I trust to earn a higher ROI on that workflow, and not for lack of trying!
I personally don't believe in any sort of cabal (Occam's Razor hasn't let me down yet). Ultimately, I don't really care *why* they're wrong as much as I care *that* they have diverged from my rubber-meets-the-road measures of value.
That is concerning to me, because people are investing 100s of B's of capital based on the putative RoI putatively available to people like ourselves. When the benchmarks support this RoI thesis, but none of the anecdata does... that's really concerning!
Re: academics, I don't think any of the data academics have access to are good proxies for the work real people are doing. And for the data that are good proxies, the model labs certainly have access to the same data, and therefore the benchmark performance against those data is irrelevant.
I am in full support of custom workflow benchmarks, and choosing the best model for your use case to balance performance and expense. Thats just good operating behavior, but the problem is the foot guns and biases people have that they are convinced they dont even if they understand on an intellectual level that everyone else has them
> but none of the anecdata does... that's really concerning!
But see this is not really true -- adoption, subjective benchmarks, verifiable benchmarks, task-dependent performance, internal product metrics, living benchmarks, all point in a pretty consistent direction. Anecdata is not the plural of data. An anecdote is like a case study. It's there to motivate the things we already have which is a huge amount of performance measures for a variety of different tasks.
> Re: academics, I don't think any of the data academics have access to are good proxies for the work real people are doing.
But this isn't really true either -- you can get this data from a variety of sources that are licensable or open source, or data that you can commission. You can critique any one methodology for this but a blanket "they are hamstrung" is not really fair or accurate.
> And for the data that are good proxies, the model labs certainly have access to the same data, and therefore the benchmark performance against those data is irrelevant.
But this is also not true -- you can have exclusive license agreements, data you hold close to the heart, or data to measure models that haven't had access to it because that data was created after these models were released.
There are plenty of problems in model measurement but the answer is not to just abandon it to be cavemen with zero respect for rigor and the biases we have to be subject to as human beings.
"Carefully and thoughtfully" is antithetical to the approach to benchmarks these days.
Maybe back when this was a scientific endeavor; not now when enormous, enormous amounts of capital are on the line. Along with an entire cult's chosen eschatology.
You can call it a cult but it’s several thousand skilled workers who know what they’re doing, by and large, most of whom have a PhD and know how science and statistics work. Benchmarks are incredibly hard, and any PR or comms department at any company is going to obviously want to make things as rosy as possible, but beneath this are earnest, expensive efforts to get good quality measurements. The better you can do this the better you can compete. If you want to make a modeling decision you run an ablation, and the quality of that decision is only as good as your measurements.
The cult in this case is TESCREAL, not everyone working on AI. Last I checked not all the "several thousand skilled workers" in AI subscribe to TESCREAL ideology, although it has been a while since I've been to the Bay. Maybe things have changed since my time at Berkeley, and Dario's belief that he will eventually be made immortal by mind uploading is more widespread.
Otherwise we agree that benchmarking is hard, the benchmarks contain hard problems, and that there are many hard working people trying to accurately gauge what is going on. It is getting harder to watch though as all that is on the line taints the overall endeavor.
No it’s: evaluating these systems are complex and there’s a reason why sociology, cognitive psychology, medicine, etc are all done in careful double blind conditions with pre registered tests. It’s not that humans are not smart enough, as I said human evaluations are incredibly important. And yet they are a minefield of biases you have to worry about and correct for.
- evaluations need to be done at the same time to avoid drift in your bias
- you need to worry about your test set: which questions are you asking? How many of them? Are they representative of your work?
- which one did you do first? Raters have a tendency to bias in one direction or another
- you also know the label! You know which model is which! This biases your assessment…
And on and on and on. Careful science exists for a reason.
There is no data that I would trust that contradicts it.
Frankly I don't give a damn about data that could be made up on the spot or appears to be scientific or meaningful while it's not at all clear how it was made (up).
Claude was heavily lobotomised for my work starting somewhen in February.
I talked to friends and people I know and trust and many felt the same. (I didn't ask them whether they felt like I did, but what they felt, how happy they were with agentic coding etc.)
I quit my abo in March and talked to said friends who are still on a plan just last week: they are still not happy, but company pays so whatever...
That’s ok but at what point is this getting into conspiracy territory? You have just said there is nothing you would believe to the contrary, but then by definition that’s not exactly a very thoughtful or insightful position.
I never said that I am not willing to believe the contrary.
I am not willing to believe the contrary from strangers on the interwebs or PR departments of companies who want to sell me something.
If people I genuinely trust tell me about their experiences, I am willing to try again.
But yes, if it doesn't work for me (for whatever reason, could be that I am holding it wrong), then I can accept that it works for everyone but me and still not use it.
Also "scientific" doesn't mean what it used to mean. When the n is small or it's just anecdotes (I am aware of the irony) blown out of proportion I really can't take the data and conclusions seriously
N isn’t small, science means what it’s always meant, statistics is a thing, and what you’re describing is just putting your trust in a very poor quality benchmark. You said you would not trust any data that indicates something that contradicts your opinion. Benchmarks are not PR they are designed by a variety of institutions completely outside the control of frontier labs. Again congratulations on your conspiracy theory.
Lol. If you're doing anything non trivial that's not a CRUD webapp but e.g. some physics simulation or high performance GPU code any and all models I've tried suck.
They are not just leagues behind what experts would code, they are not even playing the same game.
Which is to be expected, as there isn't so much physics or high performance gpu code available as there is for your typical CRUD API and JS frontend.
I can attest to this, I had a very simple 20-line shader that I asked Claude to do a basic 90-degree rotation on it, and it just completely got it wrong. Frequently adds pointless abstractions / intermediate variables even when I tell it explicitly not to in the system prompt. I can go on and on, these things just don't understand architecture. And why would they? They were trained on text.
There is something remarkable about turning speech into code (don't need to hunch over a keyboard nearly as much these days, can just talk into a mic) and it's good for first drafts / exploring ideas. But it's obvious to anyone that's paying attention we're hitting the top of the S-curve. It's no wonder the IPOs are around the corner. I mean even Dario admitted he doesn't know how they're gonna substantially increase the context window size. That says a lot.
That being said I think the harnesses are only getting better. And maybe we will get multi-modal models that understand architecture eventually. But the growing-the-blob-of-text training method that's being used now appears to be getting diminishing returns
Subs lose money on individuals to get those individuals to force their companies to pay for the corporate plan. The economics are bad, but so are the economics of grocery stores selling Milk and Bananas at a loss to drive traffic, which they basically ALL do.
I pay a lot but barely use it except for some intense days, where the lower plans would have throttled me in like 30 minutes. API billing is still more expensive. If you want to not pay much, go to openrouter and use chinese models. They are cost efficient.
Retain and hire the engineers who don’t require heavy use of AI to deliver value? The current SWE job market speaks for itself. Where will you go where they will let you burn up tokens in a high cost of capital macro?
ZIRP (zero interest rate policy) is over, software engineers no longer call the shots now that there isn’t vast amounts of capital chasing yield, and that capital bidding up salaries and keeping the labor market for engineers tight.
If you are x more productive with generative AI, very shortly you are going to have to prove it with a token budget (or, if you’re lucky, an org willing to spend for on prem hardware for capped token cost, fixed capex vs uncapped opex).
The comparison is not SWE vs SWE with AI. It is SWE vs SWE with AI with a constrained token budget ($x/month) delivering the same value at the same or lower cost. If you cannot prove that you are wildly (vs marginally) more productive with the AI, why would they pay for it? Prove it.
> The comparison is not SWE vs SWE with AI. It is SWE vs SWE with AI with a constrained token budget ($x/month) delivering the same value at the same or lower cost. If you cannot prove that you are wildly (vs marginally) more productive with the AI, why would they pay for it? Prove it.
> That is the real content of the Uber story, and it is why filing it under "budgeting discipline" misses what is actually unfolding across half the engineering organizations in the country right now. They ran the same experiment Uber ran, most of them without Uber's $3.4 billion R&D cushion to absorb the surprise, and almost none of them having modeled the heavy-user tail or instrumented the gap between tokens consumed and value shipped. The reckoning will arrive for each of them on their own fiscal calendar, and the first instinct will be the wrong one. The tool is too good to abandon, the bill is too large to absorb, and the only durable resolution runs through a question the entire rollout was designed to defer.
> You cannot get labor-replacement economics out of a tool you deployed as a labor supplement, and the bill comes due before anyone is willing to admit which one they actually bought.
I don't think they'll phase out subscriptions ever, their whole play has been to drive demand from the bottom up. Get engineers hooked on building with claude at home, then get them to demand the ability to use it at work, and bend over their employer with no lube.
They'll probably tighten the quotas to reign in whales though.
This doesn't match my experience, in academia I saw ~40-45% utilization NVIDIA GPU clusters that went 6 years with <20% failure rate. Might be a TPU thing?
I'm FAR form an expert on this, but I believe that the operating costs such as power + cooling form a big part of the lifecycle. I have no doubt that at some point within the 6 years that are being booked, that replacing entire working racks won't be more cost efficient.
That is current practice, yes. The economics of replacing racks then selling the old ones to people who will salvage and resell working components works out better than trying to repair/retrofit in place.
Inference has been going down in price on a cost/intelligence basis. If you don't need the smartest model, there are plenty of good Chinese models that are dirt cheap.
It supports his point that they're planning to massively overbuild compute, which was already well supported by the financials. A lot of that planned compute buildout can be walked back though, and the technology is unquestionably useful in moderation, so it's not the catastrophe he suggests, and his hyperbole is part of what makes me dislike him even if there are elements of his foundational argument I agree with.
Open models lag the frontier ~3-6 months, though they're likely smaller than frontier models as well so that lag might not be fully real. Qwen 3.6 27B is very usable for average coding, and Gemma4 31b is very usable for day to day tasks.
The problem there isn't the models, it's consumer hardware. Even 16GB cards aren't the norm, and even with massive improvements in per-parameter performance we probably still need 48GB memory to get models that feel smart enough to trust.
“Average” is also doing terrible things there. The “average GPU” is probably the integrated graphics on the CPU of a laptop.
If you scoped it to “average gaming desktop”, double digit VRAM is pretty normal at this point. If costs came down, I imagine the higher end GPUs would start including enough VRAM for 30B-ish models.
reply