jj4211

I've had LLM generate so many web sites about various random animals I've crammed into a prompt. No one wants web sites about those random animals, but my management is pleased at my token utilization.

Can do my real work and get praised for my actual productivity, and burn the tokens to get praised on AI adoption...

Ah yes, some random intern suddenly has 'credit' for almost all the codebase because they ran a linter with different settings than previous linter settings..

Well, they are, but not for the takeaway the article gives. The article is so close, but fails to extract the accurate conclusion.

First are what he calls the "lazy" engineers — workers who rely heavily on AI to write code, answer questions, prepare updates, and complete tasks with minimal engagement.

Then there are the "craftsmen," experienced engineers who bear the burden of understanding, reviewing, and fixing the growing flood of AI-generated code.

This is accurate. You have a set of "developers" who just need to make a good showing on the telemetry, whether it's "tokens used" and/or prominence in commit activity. They are not held to account on actual productive outcomes, just that they supervised a credible volume of AI activity. If the AI generates code and tests and the AI is satisfied that the code passes the tests, then their job is done. You have another set of developers that have to live with the nightmarish consequences of the first, because they just generated a pile of shit that would have been better not to exist at all.

'The craft they loved is dead'

Wrong takeaway, the craft is alive, but mismanagement is diluting it with bullshit.

Incidentally, this isn't new, but the magnitude is new. I have had significant segments of my career consumed by management insisting that I somehow make the bottom dollar offshored developers "productive", and similar pattern, if they "looked busy", management was happy, and management didn't care about whether the work was useful, because frankly they couldn't tell. They could tell if some volume of "stuff" was happening and they just settled on that, and if the "stuff" alienated customers, well that was the fault of those "craftsmen" for failing to properly manage the output from the "lazy" engineers.

Yeah, my manager expressed satisfication with me being one of the people using my quota of tokens.

I have generated so much throwaway content that never gets used and only gets deleted to burn the tokens to avoid getting the "you aren't using AI enough" talk. The fact they can see my actual productive output and believe AI is involved shows how utterly disconnected the metric is from reality.

Heh, a few weeks back a new project manager at my work held a meeting about an upcoming project, and half the team was able to say the timeline was workable, but the specifics the project manager laid out would lead to disaster, and we just had to adjust the strategy, but still have same time and same cost. We spelled out exactly what would go wrong and how, based on previous attempts to do it the way he said. It was scheduled to be a weeklong project, which would have been a fine timeline.

He got stubborn, insisted that based on his research his approach was right, and while he would have us on standby in the unlikely event of a problem, he would largely outsource the project to a company that agreed with his plan.

So the project started Monday, and based on past experience we expected to be called into action on Tuesday morning and have to hustle, or maybe Tuesday end of day and really get overworked to close it in time. So Friday comes along and we are shocked that it must be going ok since we hadn't heard anything. 4pm rolls around, the project manager calls us in a panic saying it's all gone nowhere, zero progress made, and he has escalated to make sure we take over and now we had to make the Monday morning deadline, or our asses are screwed. Everyone worked their asses off, a couple didn't sleep the whole weekend.

So in a followup call, the project manager said "no one could have predicted it would go so badly", and then an email came out from executive team congratulating the project manager for making the project work despite challenging circumstances.

Yeah, and for those who don't know, the rationalization output of the LLM is just so pursuasive. It sounds quietly confident and rattles off things that sound like real details.

People are believing the LLM output over actual human experts and the human experts have to expend non-trivial effort trying to disprove an LLM output before they can get on with the business of doing it right.

It's obnoxious enough to try to use for myself, sometimes useful but obnoxious to review the code and just constant screwups except for exceedingly boilerplate stuff or stuff that can take some sloppiness (e.g. LLM can make it easy to indicate some variables to get from argv and do the tedium of that plus help text plus man page edits and generally do that fine). Even if it doesn't screw up obviously, if the code is verbose, I know a screw up is lurking and just ditch it and do it myself.

However, the real pain comes in as other people use it. Just today someone had an issue and normally they'd ask a developer for help and offer debug appropriate information and/or access. However, they "just had Claude do it, even used Opus 4.8 to make sure it's good" and it generated a very verbose report on the issue, why it went wrong, and the appropriate change to make it work. Very detailed and the explanation sounded quite reasonable. Problem was that it was horribly and absolutely wrong, a fiction of a rationalization over a bad code change. It made a change that happened to appear to work for him, but in reality it replaced a failure due to unrecognized data to silent corruption of the data in a facet the user specifically did not care about. Claude claimed it was correctly mapping the unrecognized data correctly, but it just made up a completely untethered conversion based on nothing. Now I could tell the explanation and code change was bullshit at a glance, but it became an argument because the user wouldn't give me actionable debug details because "he already had Claude fix it". I had to keep trying to find holes in the Claude rationalization that the user would also recognize, and he sided with Claude four times until the fifth problem in Claude's explanation finally stuck (it asserted that the problem was due to running a specific outdated version of a specific software, problem being that specific version never even existed, and the minimum "good" version was 10 years old and the version the user was running was about a month old).

I don't understand how people get this far and still don't understand that AI is much better at sounding plausible than being correct.

An example I can think of is IBM.

In 2010, the CEO proclaimed an earnings target for 2015 based on nothing at all. There was no plan or reasonable expectation, just a flashy number. Investors ate it up and stock went up. No shortage of white men eager to be at the helm of a company that seemed to be on top of the world.

He promptly left the company and handed it over to a woman. As one could predict, a hollow wish about earnings without an actual plan failed to actually deliver. To a lot of folks who understood the nuance, they called it from the moment he said it.

However, the news coverage was basically that she failed to execute on his "plan".

Now she wasn't amazing leadership or anything, but neither was he. However he got to be celebrated as a strong leader mostly on the back of hollow promises and she got to be blamed for the fact it was hollow.

I suppose I can't prove that it was because she was a woman that she got to be the fall person, but I am at least sure they could have found a willing white dude to be the fall guy at least.

It means that if you are so obsessed with protecting a user from making an informed decision about their own security, then you could gracefully degrade in your 'horribly insecure context' instead of just bombing out completely.

You cannot start a car with a suction cup.

I can't start my car with my car's app either.

If you really want to be picky about it, block out the unlock feature and any potential 'phone as key' functionality. Leave starting the air conditioning and information.

I get the touchscreen thing, though I really don't mind a touchscreen so long as I have real hard controls for volume and heating/cooling. GPS without touchscreen is painful.

But the backup cameras, absolutely those need to be a thing. They have saved lives. People have arguably never been driving 'just fine', just getting closer to just fine. Backup cameras reduced child death from being backed over by 80%.

The apps that didn't work well were not due to lack of lock-in.

The apps didn't work well due to lack of maturity in the platform. This app is not failing because the OS is somehow '2013-like', it is failing because Android app developers are going all-in on lock-in.