status_quo69

LLMs are computers pretending to be human, even when their intelligence surpasses their originating species, why shouldn't they benefit from the same kind of structure and flexibility that ruby and rails provide?

Rails isn't dead yet or else I'd be surprised considering I write rails enterprise every day with LLMs who seem to do a perfectly fine job at writing the framework code and libraries. They're absolutely atrocious at testing. I'm also not particularly worried yet because my money making machine has no bottlenecks that dropping to a lower language would solve, and the loss of productivity for humans or LLMs by dropping to a lower level language would be staggering. Turns out monkeypatching in ruby is actually a huge benefit in the age of LLMs vs having to rewrite or fork everything if there's a bug or bad behavior by a lib.

I mean, sure we can assume that (although the Anthropic article from the other day say that claude mostly builds Claude) but this is the Claude Code team which, I again am assuming so I could be wrong, focuses on the TUI/CLI interface to those models, not necessarily the models themselves. That coupled with the fact that this is how a large portion of their clientele interact with their moat, it would be odd to me if they didn't apply significant work towards this thing to keep the hype or productivity train moving

Fair enough, I did invent the 1000 a day because the article was vague, I'll edit my submission accordingly to also be vague

No wonder most people at Anthropic believe our jobs are in jeopardy, they've already made a system wherein the machines are treating them like pets.

I really don't get the tendency here from Anthropic to describe what's most likely churn? A high pressure amount of changes/PRs a day on a TUI to wrap a while loop? I'm not going to be a reductionist here so I'm going to cut it off there, I'm positive that the complexity of handling multiple terminal emulators is something that needs some handling (even if doing their own rendering engine was a self made problem). Yes I've vibe coded my own harness from scratch without writing any of it, I still know how it works.

I'm also not sure that they addressed point 2 in any actual capacity? They say "the system verifies itself" which is cool and is something that automated testing has done for decades, but they eschew the mental model of how the system works as a whole and prefer instead to have the LLM focus and summarize a slice of the system when they need to know about it or ask for it. But.... That's not knowledge? That's just search a la google? And thus seems to be subject to the same problems where the quality of your search dictates the quality of your knowledge that you surface at any given time?

N.B. Changed "1000 PRs" to "A high pressure amount" to be vague since the article is vague here on what CC undergoes.

No mention of lines of code changed or total LOC, what percentage of this is bugfixes vs features?

Personally I see it in the tendency for solutions that are absolutely bonkers in their structure and form.

I don't even mean that from an aspect of "doesn't look nice from a technical standpoint" I mean

- need to track progress and/or defect rate for a component of the system, invent solution (we pay for jira). This was an executive - folks proposing rewrites of front ends both in design or technology stack because "Claude can do it in 3 days", quoting Claude for the "market research" (completely fabricated metrics and studies). Current tech stack is perfectly fine and there's no monetary or engineering pressure to changes - QA has Claude qa the code only, hands ticket back to dev with qa notes straight from Claude instead of.... Automating a way to have the llm do this in a reproducible manner

I'm going to sound incredibly naive but is this RSI? I mean by this definition isn't the training process that produced the hugging face RSI since a model was judging the alignment and ability of the swarm of agents to accomplish a given task? If that fits the definition (I think it does) is this just a generalized form of that model of execution?

True, but I don't know. I'm not going to gatekeep the transmission of information (how could anyone?) but there's a certain level of ennui here for me and the easiest tell comes from the writing style. I still think there would be something lost in "claude write me a blog, don't use load-bearing".

We invented algorithms and machines that ostensibly understand and interpret our words and can make things happen, which is a revolution of understanding. From a legibility perspective, these machines now can "understand" me just as well as if I were writing like Gene Wolfe, yet instead of embracing the idea of "wow I can talk to a computer" we have an overweight on "wow, I can have computer talk to everyone else on my behalf!".

Even if the transmission of information here was perfect I'd still question the utility of outsourcing your brain on writing a very short article compiled from release notes.

Where were the facts invalidated? Why can't a reader despise a writing style that LLMs converge towards?

I think you're right on your last point, btw, nobody will care and we'll be the poorer for it. Just like how McDonald's and Walmart have driven out and undercut localized taste and culture in the US, so it shall be with writing of all kinds. It's happening right now and readers are getting used to reading the intellectual equivalent of a Big Mac.

What does an advance in music even look like? Shifting tastes for pop music? Or new techniques? New music theory? Or just experimentation?

Considering all music is subjectively influenced by the culture in which it's born (see the difference between Asian traditions of music, European traditions of music, African traditions, and traditions of the Americas) not even all of those have a given structure that is present today like the typical 4/4 and have polyrhythmic and multitonal structures by design. The fact that everything on the radio has converged towards 4/4 165bpm major chord progressions is evidence of that cultural phenomenon.

I use fable 5.1 (tried opus but it lied to me 3 times in quick succession and ignored me in another)

This is weird to ask because I feel like of course the model isn't omniscient? Isn't the whole point of iterating on a plan to assess impact, risk, know your (the user) variables, user impact, product impact, etc for making a change? I cannot count the times even in the past few months where I start a conversation with my C suite because their desired outcome would have a potential negative impact elsewhere for other products or users.

Is this just not something that comes up at Anthropic?

Rails dev here, still using it, still plan on using it, use agents, don't see them making food gun decisions with rails apis just code.

Rails comes with the console by default, so production inspection of a bug is something the ai can do without a ton of bs to be set up, and that console can be forced to be auditable (using standard gems) or read only (using flags).

I used to love rust, still do sorta, but yeah I think this is really not a good argument against a tool that operationally is better at most things at being observed or poked or prodded or tested in production

The same way we did before. Very very simple code submission (most people using AI use it even though we call out we're going to ask them later to modify later without AI tools writing code for them) then pairing interview where we ask some basic "are you actually at the level you say you are" question, then ask them to extend their program submission with:

- engineers on the call as pairing assistants - google, ai tools, whatever for libraries, syntax, etc. we tell the candidate directly that it's impossible for us to gsther signal on how they think about problems if they ask Claude to just whip them up a solution - the expected output - their own unit test suite

Every candidate that has submitted an ai submission thus far has failed because they have literally no idea where to go. I've interviewed dozens at this point. I'm not saying "they're unfamiliar with the structure", I'm saying "they cannot actually break down the problem even verbally".

It doesn't matter their pedigree or past experience on their resume, if they used AI to generate they don't seem to be able to resurrect the skills that actually matter for the thing, engineering and product work.

Note because I know folks hate code submissions. It's not hard. We give a CSV with 3 columns, 10 lines. Do some basic mapping and some structuring, some basic data modeling. We only expect about 1 actual class or struct. Then unit tests and it should run in the terminal. Max submission length with verbosity has been a java program at something like a hundred lines total if that, most folks complete the submission in an hour or two. Extension is that we modify one of the rules and extend the CSV by 5 lines.

I seem to still be getting good signal from this, since the engineers that I've hired off of this have been fantastic with or without AI tooling immediately in their hands during the day.

I agree with the idea that it's not probable but I do think it's important to point out-- we only need these elements to run with the bandwidth they have and the token rate because we want to see things in human-scale time. But slow things down to a 1tok/sec doesn't matter to this hypothetical anti-aligned LLM. Time is, after all, relative, and it's not like LLMs give a shit how long something takes. They don't have squishy stupid organs that fail after a certain amount of time, or those pesky glands that emit impatience hormones.

But yeah you're not going to be able to shard out the terabytes of Fable weights that are needed to run inference without addressing some fundamental physics problems.