cruffle_duffle
If I can shit out some vibe-coded thing over a weekend that does <useful thing> it means anybody else with the same tool can do the same thing. There is no value in the actual code itself -- the faster these models run, the more true that statement will become. In the not too distant future these things will run directly on silicone and operate at thousands of tokens a second. That means they'll iterate and explore something far faster than they do now -- mistakes they make will be fixed in seconds or perhaps milliseconds.
That being said my <useful thing> is likely only "60%" of what makes an actual product somebody would buy. It's finding the other 40% that is something I don't think that LLM's can solve. That requires human judgement and then knowledge of how to communicate that vision to your agents.
In short, I believe the challenge is knowing what to build that others would find useful. It's easy to crank out <useful thing> for yourself. Making something others can use is not. And that is, possibly, the bottleneck. But once these are fast and dirt cheap, (and that is a when not an if) the cost to iterate to something workable goes way down. Of course mumble mumble "no silver bullet" / "mythical man" month mumble mumble and all that soo......
I dunno... doesn't directly answer your question but it is what came to my mind anyway.
I’m just about to tell opus or sol to go make a custom watch holder that fits kiddo’s bed frame. I’ll give it the measurements of the bed frame. It already knows the dimensions of my printer. I’m prepared for it to absolutely fail at the task, but honestly I wouldn’t be surprised if it came up with something I can send to my printer and just work. Part of me wonders if I could instruct it to simply generate the raw gcode and skip the entire stl -> slicer pipeline. Probably would blow its context window though.
If you haven’t used these tools, you really don’t understand their capabilities. It’s pretty crazy stuff.
Token guy’s role gets even better when you let them recommend entirely different hospitals for a given step in patient care (OpenAI models, random shit on openrouter, etc).
I have a “token guy” in my workflow and the dude has looked at underlying tool call patterns, cache use, etc and made all kinds of recommendations to the clients calling it all on their own and completely unprompted—just following a simple role.
I’d experiment with expanding their role into prompt engineer or some kind of specialist on that domain.
Basically, this sounds interesting in theory but it’s useless for me. I even tried letting codex go over it and it wasn’t clear without additional prompting what was supported.
Sorry.
Seems like it based on my first session. It still does the whole “bury the important thing in a pile of words” coupled with the “it might actually be important” thing… so basically you never really know what it’s talking about.
Honestly I trust opus so little that the entire “opus” brand is completely tarnished. Its writing style is so god awful that it needs more than just a point release. Either dump the name and ship a different model entirely or at minimum call it “opus 6”. Calling it 5.5 makes it sound like it’s basically a continuation of the same garbage output that 5.1 had but with some minor adjustments. And based on my single first test, that is what it appears like to me.
Bingo! And on this axis many SOTA models fail miserably. These things are acting on my behalf under my direction. All the supposed intelligence in the world means fuck-all if nobody can understand it.
And like somebody else said… when meat-based humans talk like Claude does, it almost always means they either don’t understand what they are talking about, or are actively trying to conceal something and are a fraud. Not always, but almost always.
So yes. Absolutely assume AI agents acting on behalf of their humans are finding all the token efficient ways to get at your sites data.
I've cut token counts significantly for some of my MCP servers by returning csv or tabular data or sometimes even just good old fashioned plaintext "template style" words. JSON is highly repetitious. In some instances like 30% or more of the token response was just boilerplate JSON.
Once you realize that to an LLM JSON is just a stream of tokens no different than XML, markdown, or a punctuation-less stream of conciseness.... you start to realize there is no reason not to just make up your own crazy formats. The LLM isn't throw parsing errors if your MCP is returning something other than JSON. The LLM is smart enough to figure it out. It's far more important to make sure your tool descriptions and parameter descriptions properly "market" what it is you do (and dont do) than return syntactically correct JSON.
Same with versioning.... it's all ephemeral. There is no backwards compatibility to speak of with MCP (at least for tooling "contracts".... what the tool actually does however... that is a different level and is product requirements not API contracts).
I remember discovering this when i wrote my first MCP server. It was like "huh? why would they do that? what use case did they have in mind?". We've spent decades making "internet shit" as stateless as possible on the backend because making it stateful is expensive and complex if you want to have any reasonable scalability. I mean good luck trying to host a stateful service on any kind of commodity serverless "scale-to-zero" infrastructure here in 2026.
Maybe it's because these AI-labs are used to statefulness. I mean LLM-based sessions are hugely stateful if you want any kind of reasonable caching to happen and caching is the only way you can economically scale out LLM's. Seen from that perspective it kind of makes sense why they'd look at MCP and think "hey, why not make this stateful on the backend as well". Statefullness just part of their DNA.
Photoshop basically destroyed a centuries-old artistic tradition of depicting six gyoza on a white plate. Now with all this AI nonsense, you might as well just make your food at home rather than stare at yet another “six finger hipster eating a vegetable spring roll” cranked out by ChatGPT.
And yes the output of these researchers are highly sensitive to prompting. Left to their own devices the LLM will often ship some very biased prompts to its deep research agents loaded with pre-conceived ideas rather than letting the agents uncover things themselves. Then all the agents do is confirm what the prompt told them to rather then “think independently”. (Very similar to open ended interview questions rather than asking yes/no questions)
It’s is far better to spend a session writing writing the research prompt itself.
All of this takes time and tokens of course…
It just depends on what you are building. Some things are more more forgiving to muck-ups than others and if the LLM screws the pooch, oh well! Roll forward.
Sites that do this irritate me so much. Ones that try to block pasting and stuff… like somebody intentionally baked that into the site. Who? And what was their rationale? Are they really so arrogant to think people are going to carefully type in some elaborate password not once but twice?
That and blocking paste in fields like bank account numbers and stuff.
Surely somebody here has been asked to implement these mis-features. Please explain what went through the heads of the people responsible for it?
The harness is great, but opus is such an arrogant little prick that spews out unintelligible word salad. Opus 5 is so bad at communication it amazes me that somebody green-lit it. It's absolutely awful.
The fact that this isn't an acknowledged regression (and Fable 5.1 isn't much better) leads me to believe that people at Anthropic actually like Opus 5's output.
I wouldn't call copy & paste code from a webui of chatgpt either traditional or boring. I'd call it tedious, error prone and guaranteed to get poor results. There is much better tooling and harnesses to leverage now.
In other words, it ain’t you. It’s the model. It’s just genuinely bad.
Then you switch to ChatGPTs lineup and realize how things can actually be better. It took about a week to really get the feel for how to use their models… then I basically switched. I’ll check in every now and then when they actually make a deal about how opus “now makes sense”.
But honestly I’m half convinced Anthropic actually prefers the output of opus 5. I dunno why, but how else could you explain how such a thing got shipped? I mean somebody in the pipeline had to say “dude this model doesn’t make sense, you think we should fix it?” Right? Like it’s a pretty massive drop in quality for such a major brand in this space, you know? How did it make it out the door?!?
Humans love a good end-of-world story. Always did, always will. It’s when they want the world to conform to their ungrounded irrational bat-shit crazy nonsense that I get a worried. Especially when people in power actually take their nonsense seriously.