LLMs are the first technology I've experienced where there's a lot of top-down pressure to adopt it ASAP. Most other technologies in my career, like VCS or static analysis or whatever else were championed by colleagues or peers.
there's a lot of top-down pressure to adopt it ASAP
Because you're training it to replace you. MBA's have found a tool that finally allows them to cut out pesky intellectuals and creatives and they're chomping at the bit to make that a reality. Look around, dark enlightenment is being embraced/tolerated at the top.
It's all about the story that's sold to the higher ups. The higher you go up the corporate ladder, the vaguer the understanding of the technology. The big boss hears from a Microsoft salesman that AI = you can fire 20% of your workforce, but never questions exactly how that works. They probably never got sold static analysis in that way. That was just some kind of tool that somehow helps with that mumbo jumbo that developers spend all day typing. There's no story there that inspires a manager. AI = cut costs is music to the ears of the board. So then pressure gets applied to those lower down.
Something similar was going on with cloud a few years ago. The story was if you get cloud you can get rid of those expensive infrastructure people and it will all be so much more reliable. So the big boss gets a cloud strategy and foists it on those lower down. There's also pressure to be an on-trend boss. If all the other boss' are getting into it, then you need to as well.
I’m seeing this second hand at the Fortune 500 my spouse works for.
They are an enterprise SaaS company in SV. Where they have machine learning software that they have been selling for more than a decade, it’s all been rebranded as AI. That’s fair enough from a sales perspective, I guess. What’s odd is that their C-suite and SVPs are pressuring everyone to use LLMs everywhere, for pretty much everything, and none of them seem to understand why it’s only that level of employee that’s seeing any benefit or expressing any interest. My spouse has reported that the running joke across the company is that the executives have jobs that can be done by LLMs, but no one else does. The ICs could not be less interested, and even if they were, legal promulgated a policy against actually putting anything confidential into any LLM other than Copilot in Azure, which the whole workforce reportedly only really uses for summarizing the increasing number of meetings that are perceived as a waste of IC time. A lot of those meetings are “let’s use AI”.
I've seen it before. Adopting Windows Server or IIS. Choosing Oracle for your RDBMS. That kind of thing. It has all the hallmarks of a decision based on a salesguy's pitch with no technical evaluation.
Not that surprising. There's nothing in it for me:
If I get more productive, I don't get paid more. If I get less productive, I get told I'm not using it correctly. Worst case: I help train the LLM to eventually take my job.
I think LLMs are different because they are such a powerful developer technology. The most powerful I have seen in 25 years.
Some engineers went from useful to useless almost overnight. Someone who is stubbornly stuck in their ways and refuses to use AI tools, and yet spends days failing to achieve what Claude Code can do in a few requests, well... what's going to happen to them? All of a sudden there's no work they can do, because the easy problems got solved by the AI.
So it ends up being the manager's job, either find something that this person can do, train them up to use modern technology, or fire them.
The answer, well documented in the article, is yes.
While the article presents cases that appear the be problematic in the particulars, I think coming to the conclusion that bosses/managers shouldn't be pushing or mandating the use of AI tools in general is incorrect.
It's quite possible that any one new AI tool is wrong, but it is unlikely all of them are. A great historical analogies are the adoption of PCs in the 80s and the adoption of the internet/web in the 90s. Not everything we tried back then was an improvement on existing technologies or processes but in general if you weren't experimenting across a broad swath of your business you were going to get left behind.
It's easy to defend the utility of these tools so long as you caveat them. For example, I've had a lot of success in AI driven code generation for utility scripts, but it is less useful for full fledged feature development in our main code base. AI driven code summarization and its ability to do coding standards enforcement on PRs is a huge help.
Finally, I find the worries in the article about using these tools on sensitive data or scenarios such as ideation to be rather overdrawn. They are just SaaS services. You shouldn't use the free version of most tools for business purposes due to often problematic licensing, but purchasing and legal should be able help find an appropriate service. After all, if you are using google docs or Microsoft 365 to create and store your documents why would (at least with some due diligence that they don't retain or train on your input) you treat Gemini or Copilot (or their other LLM options) as presenting higher legal peril?
> It's quite possible that any one new AI tool is wrong, but it is unlikely all of them are.
All can absolutely be wrong at the same time, but the tool isn't the main issue IMHO. Its the user.
For simple generic stuff its not an issue, but where you need an expert, it has to be an expert in that field who uses the AI. So you know what is wrong.
A good recent example is the OpenAI Academy. Clearly the site content is generated by ChatGPT, and completely misses the point of the areas it claims to be training you in.
> A great historical analogies are the adoption of PCs in the 80s
Another historical analogy is Scientific Management, pushed top down and widely adopted by the industry. It has many flavors and all of them were wrong.
We have samples in basically any direction one would like to argue for. Historical precedence isn't a good argument IMHO.
> I think coming to the conclusion that bosses/managers shouldn't be pushing or mandating the use of AI tools in general is incorrect. It's quite possible that any one new AI tool is wrong, but it is unlikely all of them are.
If the tool is good, then management won't need to mandate it. People will be tripping over themselves to get access to the tool that helps them to do their job better. So perhaps you're right that some of the tools will be good (though I personally haven't yet had that experience), but I think that it is incorrect for managers to push for (let alone mandate) tool usage. Measure the result, not the path an employee takes to get there. If Bob uses AI tools to great effect, but Alice is doing just as well as him without using said tools, it's a mistake to force her to change her workflow thinking that the tools will be just as good for her as for Bob.
I work at MSFT. There’s top down pressure to use LLMs everywhere. At this point, if you can convince your management about using LLMs anywhere, they would happily head nod and let you go do that. And management themselves are not that technical wrt LLMs, they are being fed the same AI hype slop that we are fed.
Most of these efforts have questionable returns and most projects will usually involve increasing test coverage or categorising customer incidents for better triage, apart from these low hanging fruits not much comes out of it.
People still play the visibility game though. Hey, look at what we did using LLMs. That’s so cool, now where’s my promotion? Business outcomes wise, there’s some low hanging fruits that have been plucked but otherwise it doesn’t live up to the hype.
Personally for me, it is helpful in a few scenarios,
1. Much better search interface than traditional search engines. If I want to ramp up on some new technology or product, it gives me a good broad overview and references to dive deep. No more 10 blue links.
2. Better autocomplete than before but it’s still not as groundbreaking as AI hype hucksters make it out to be
3. If I want to learn some concepts (say how ext4 FS works), it can give a good breakdown of the high level concepts and then I go need to study and come back with more Q’s. This is the only genuine use case that I really like. Where I can iteratively ask Q’s to clarify and cement my understanding of a concept. I have used Claude code and ChatGPT for this and I can barely see any difference between the two.
I have a similar mandate and a similar take, but slightly different use cases.
As to the search engine, my searches are often very narrow, like I want to recall a specific message from a mailing list, so I don't use that too much. On the other hand, I found Google's NotebookLM to be really good at recalling concepts from both source code and manuals (e.g. processor manuals in my case).
Code generators are incredible refactoring machines. In one case (not so easy to reproduce in general, but it did work) Claude Code did a Python to decently idiomatic Rust conversion in a matter of minutes; it added mypy annotations to 2000 lines of Python code (with 90% accuracy) in half an hour and got the entire job done with my assistance in about an hour. For the actual writing and debugging where the logic matters they're still not there even for small code bases (again 2000 lines of code ballpark). They're relatively good at writing and debugging testcases but IMO that's also where there's a risk of copyright taint. Anyhow it's something I would use maybe 2-3 times a month.
In one case I used it for natural language translation, with pretty good results, but I knew both languages because I needed to check the result. Ask it first to develop a glossary and then to translate.
For studying they're interesting too, though for now I have mostly tried that outside work. At work, Google Deep Research worked well compared to the time it takes and it's able to find a variety of sources (including HackerNews comments in one case :)) which is useful for cross-checking.
If anything my bosses are pointing out how LLMs and AI tools are not worth investing in. They have been fairly opposed to them and only have become more so as time goes on. But we are known to think different.
I think Siri is bad because it’s an expert system. As good as LLMs are, they waste a lot of time with how wordy and inconsistent they are. Apple has been churning out polished versions of viable tech for decades and only now are people intimating Apple isn’t capable and that the tech is ready. My guess is that the tech isn’t ready and when it is, Apple will knock it out of the park.
My company literally built its own ChatGPT Enterprise wrapper and forced us all to make X prompts in it per week. If we don’t meet that quota, our immediate leader will “strongly suggest” we do it or we might get a bad performance review eventually. It’s also tied to our yearly bonus now.
The most important advice for people in this situation, from the article:
> I’d say my overarching advice, based on how difficult tech recruitment is right now, is to sadly play along. But — and I cannot stress this enough — make sure you document everything.
> What I mean by that is every single time AI tools cause problems, slow-downs and other disappointing outcomes, document that outcome and who was responsible for that decision. Make sure you document your opposition and professional advice too.
Personally, I would just add a warning to be careful to blame the tool, not the person. Otherwise, you will be seen as the "bad" person in the story even if your report is technically correct.
It's easy to blame person in a roundabout way. I have to do this all the time. Instead of saying "The suggestion from Bob was wrong and now we are in trouble" go with "We complied with the suggestion from email thread at this date and time and now it seems are in trouble". Whoever doesn't care, Bob is covered. Whoever does care, will find the info they need, but that's on them now.
Whatever I do for money isn't a huge part of my identity, so telling a boss (if/when I find myself in that situation) to stuff it with the AI nonsense isn't going to be difficult. Decoupling one's self-worth from the job makes it much easier to roll with being fired.
"Playing along" is a great way to be part of someone else's potentially-harmful project. Consider your values, and don't cross those lines. If the boss is upset about it, they have options. I don't do their work for them.
Collective action with your fellow workers against enshittification is a humanist way forward.
I have used my company LLM thingy. Able to summarize and document code leveraging remarks and general code behavior just because LLM just ingested the full python docs.
About generating things well... it just copypastes the same snippets you could find on stackoverflow, including bugs - if the task you throw at it has already been answered.
For complete and complex code... well it spews out the same useless advice you could get from a drunk non expert person while sitting at the bar.
Issue is... LLMs are too big to fail, everyone just poured billions in this huge statistics bean counter, and... someone has to justify those expenses at board meetings.
> This is the thing about AI tools. They are by design going to honour your prompt, which often results in your AI tool agreeing with you, even if you’re wrong.
LLMs augment the input with their trained data. LLMs don't inherently agree if you set up context correctly for analysis.
I've arrived at the conclusion that the top-down push without adequate upskilling creates bad experiences and subpar results. It's like adopting a new methodology for something without actually training anyone on the new methodology, it leaves everyone scrambling trying to figure it out often with poor results.
I find LLMs to be a great multiplier. But that multiplier will take whatever you put in context. If one puts in bias and/or fragmented mess, it's far more difficult to steer the context to correct it than it was to add it to begin with.
Among many small examples at my job, an incident report summary used to be hand written with a current status and pending actions. Then it was heavily encouraged to start with LLM output and edit by hand. Now it’s automatically generated by an LLM. No one bothers to read the summary anymore because they’re verbose, unfocused, and can be inaccurate. But we’re all hitting our AI metrics now.
I don't care about software craftsmanship anymore in my day job. 2 years ago I would have been deeply upset about this sort of thing but now I'm intentionally apathetic.
Once I changed my goal from maximizing code quality to maximizing billable hours, I feel a lot more optimistic about the future and these AI tools are going to create so many opportunities for me.
I found it hard to compete with junior developers in my last job (before AI was mainstream) in terms of volume of code because some of them would write 1000 lines of low quality code per day... Now I can also do this. It gives me a lot of surplus energy to figure out how to play politics and shift blame... It gives me an actual competitive upper hand over the juniors. I can out-compete them both in meeting/debate and code/feature volume. I talk wisely and code foolishly. Win win. I couldn't do this before because I was writing the code myself and I had essentially lost the ability to write high volumes of dirty code. I had a kind of analysis paralysis due to trying to solve problems in an optimal, minimalist, most reliable way. No longer a problem. Bugs are a problem for someone else.
I have so much more time to think about career strategy now. I managed to avoid being assigned to any difficult projects... I feel bad for the other people who try to go above and beyond and end up wedging themselves into a difficult situation where the software is down all the time and they have to take the blame... The AI never gets the blame.
I hated playing politics before but AI has made playing politics a necessity. It's like the more apathetic you are about your output, the better off you are.
The ironic thing is that I know it's possible to produce high quality code with AI. I've had some really positive experiences with Claude Code on side projects... But that doesn't align with the reality of 99% of software projects. The foundation is not set up right to get these kinds of results. I could set up the foundation correctly but I'd have to be present in the project since the beginning and I'd have to be given a lot of decision power; but I never get such opportunities. Bad foundations beget bad code, especially with AI because the AI never gets the idea to refactor... If your codebase is unmaintainable, it will hallucinate dirty code which doesn't work. It keeps coming up with more and more hacks... Then it delivers hacks on top of its own hacks.
Yes. My workplace soft-enforced it as well. I like using LLMs, but sparingly. I consider myself a better writer and find their tone bland and lifeless. Other than some minor proofreading, I almost never use them to generate text.
For coding, unless I’m writing trivial RPC endpoints, editing docs, or writing tests for an already hardened API, I find agents a complete waste of time. So my usage is mostly limited to chat sessions.
To use up the quota, I apply the provided tokens to a few personal projects here and there, but no one can make me push an actual production CL with them unless I find it useful myself.
After one of my client forced all employees and contractors to use AI, my boss, who was previously reasonable, started:
- Regurgitating AI crap to every answer, often just replying with a ChatGPT / Claude screenshot
- Not being able to explain code but "don't worry, I got Claude to generate some tests and the tests pass"
- Introducing random bots in slack and github which print tons of noise humans just skip through because they're not accurate enough.
The effect on the team of developers with various level of experience started showing up as well:
The application architecture turn into a horrible mess, it's worse than junior engineers.
The application started exhibiting tons of hard to debug issues, because the generated code was too low level and not covering corner cases.
Every attempt of the AI engineers to fix the issue generated one more class wrapping the existing codebase - with a fix which never worked (eg. ConnectionManagerWithTimeouts).
Eventually we basically had to rewrite the application, throwing away most of the code twice. One to just get something working with the existing architecture without crashing every hour and then another to use a framework and eliminate the last bugs occurring every once and then.
LLM needs to be in incredibly capable hands in order to be used safely and engineers will have to fight their instinct and not get swayed by the LLM telling them they're right.
Makes sense. I have seen far too many coworkers dismiss AI completely without trying it for their job.
At this point, you need to learn what AI can and cannot do, for the same reason you need to keep up with new versions of whatever framework you use. Since AI develops so fast (e.g many image use cases that AI would be terrible at 4 months ago, they now do perfectly), you need to repeat that exercise frequently.
There are 4 problems with adoptions as I see them:
1) Hype. Some people overhype what AI can do, which causes people to dismiss them when they don't immediately work;
2) Plenty of people don't like to change what they do/feel threatened by change. Doubly so when that change is perceived (real or not) to impact their job.
3) AI is weird and so it sometimes fails spectacularly at simple things, while it works very well at more complex things;
4) People use ChatGPTs free model or other AIs that are free. These are older/less powerful models, which means people end up with wrong expectations of what they can and cannot do.
5) Who likes to be told what to do? Especially by a clueless boss.
Where I running a company, I would ensure that my employees had access to a top of the line model and cursor/windsurf. I would monitor usage and have a talk with those whose usage was drastically lower than their peers.
However it would be a talk only - with the aim of figuring out why AI did not work for that employee, and what we could do to fix it.
> The CTO at my previous job tried Claude Code and really liked it so he said that all the devs had to use Claude Code in our work
Imagine this sentence with "Claude Code" replaced by anything else and "CTO" and "devs" replaced by more generic terms like "boss" and "employee". It's just "The boss tried Tool X and liked it so he said all the employees have to use it". It just seems to me that that is a bad way to make decisions regardless of what tool we're talking about or even what industry we're in. It's certainly possible it could make sense with a few more steps in there ("the boss tried this and liked it because X so he said we have to work on using it in way Y to accomplish Z"). But the way this is described sounds like a fire-and-forget mentality where the boss tells people to do a thing a certain way and that's the extent of his involvement, which seems stupid.
Where I work it's a solution looking for a problem, and we're heavily encouraged to implement it even if there's no real problem for it to solve, because "we're obligated to give it a try".
We have a new performance review and new company values at work (~1000 enployees) which are heavily leaning towards us being required to use “AI” (LLM’s) at work.
There seems to be some magical thinking at work that simply uttering the words “AI” and “automation” will somehow render us more productive.
In their defence, your boss is probably getting pressure from the board level to define their AI strategy. A partner at a VC company sitting on that board is in turn getting pressure to find synergies between their larger portfolio companies and shiny new AI investments.
I find it interesting why this article dropped so quickly from the front page. Is there a way for privileged users to actively push an article away from it?
I'm getting emails from the CEO that we need to take advantage of these tools and strong recommendations to try them out and leave feedback.
But my refusal to be AI-assisted at work is viewed as "healthy skepticism" by upper management. And my colleagues who have tried the tools are not particularly impressed.
54 comments
[ 3.0 ms ] story [ 65.2 ms ] threadHaving to do this says a lot about how fragile the state of AI/LLM hype.
Something similar was going on with cloud a few years ago. The story was if you get cloud you can get rid of those expensive infrastructure people and it will all be so much more reliable. So the big boss gets a cloud strategy and foists it on those lower down. There's also pressure to be an on-trend boss. If all the other boss' are getting into it, then you need to as well.
They are an enterprise SaaS company in SV. Where they have machine learning software that they have been selling for more than a decade, it’s all been rebranded as AI. That’s fair enough from a sales perspective, I guess. What’s odd is that their C-suite and SVPs are pressuring everyone to use LLMs everywhere, for pretty much everything, and none of them seem to understand why it’s only that level of employee that’s seeing any benefit or expressing any interest. My spouse has reported that the running joke across the company is that the executives have jobs that can be done by LLMs, but no one else does. The ICs could not be less interested, and even if they were, legal promulgated a policy against actually putting anything confidential into any LLM other than Copilot in Azure, which the whole workforce reportedly only really uses for summarizing the increasing number of meetings that are perceived as a waste of IC time. A lot of those meetings are “let’s use AI”.
It’s absolutely insane.
It's probably because there is little or no known long term consequences of using AI. In management circles AI is the magic solution to all problems.
Hence its safe to push onto users.
If I get more productive, I don't get paid more. If I get less productive, I get told I'm not using it correctly. Worst case: I help train the LLM to eventually take my job.
Some engineers went from useful to useless almost overnight. Someone who is stubbornly stuck in their ways and refuses to use AI tools, and yet spends days failing to achieve what Claude Code can do in a few requests, well... what's going to happen to them? All of a sudden there's no work they can do, because the easy problems got solved by the AI.
So it ends up being the manager's job, either find something that this person can do, train them up to use modern technology, or fire them.
While the article presents cases that appear the be problematic in the particulars, I think coming to the conclusion that bosses/managers shouldn't be pushing or mandating the use of AI tools in general is incorrect.
It's quite possible that any one new AI tool is wrong, but it is unlikely all of them are. A great historical analogies are the adoption of PCs in the 80s and the adoption of the internet/web in the 90s. Not everything we tried back then was an improvement on existing technologies or processes but in general if you weren't experimenting across a broad swath of your business you were going to get left behind.
It's easy to defend the utility of these tools so long as you caveat them. For example, I've had a lot of success in AI driven code generation for utility scripts, but it is less useful for full fledged feature development in our main code base. AI driven code summarization and its ability to do coding standards enforcement on PRs is a huge help.
Finally, I find the worries in the article about using these tools on sensitive data or scenarios such as ideation to be rather overdrawn. They are just SaaS services. You shouldn't use the free version of most tools for business purposes due to often problematic licensing, but purchasing and legal should be able help find an appropriate service. After all, if you are using google docs or Microsoft 365 to create and store your documents why would (at least with some due diligence that they don't retain or train on your input) you treat Gemini or Copilot (or their other LLM options) as presenting higher legal peril?
How so? I have access to a huge number of these tools and they're all pretty similar.
All can absolutely be wrong at the same time, but the tool isn't the main issue IMHO. Its the user.
For simple generic stuff its not an issue, but where you need an expert, it has to be an expert in that field who uses the AI. So you know what is wrong.
A good recent example is the OpenAI Academy. Clearly the site content is generated by ChatGPT, and completely misses the point of the areas it claims to be training you in.
Another historical analogy is Scientific Management, pushed top down and widely adopted by the industry. It has many flavors and all of them were wrong.
We have samples in basically any direction one would like to argue for. Historical precedence isn't a good argument IMHO.
If the tool is good, then management won't need to mandate it. People will be tripping over themselves to get access to the tool that helps them to do their job better. So perhaps you're right that some of the tools will be good (though I personally haven't yet had that experience), but I think that it is incorrect for managers to push for (let alone mandate) tool usage. Measure the result, not the path an employee takes to get there. If Bob uses AI tools to great effect, but Alice is doing just as well as him without using said tools, it's a mistake to force her to change her workflow thinking that the tools will be just as good for her as for Bob.
Most of these efforts have questionable returns and most projects will usually involve increasing test coverage or categorising customer incidents for better triage, apart from these low hanging fruits not much comes out of it.
People still play the visibility game though. Hey, look at what we did using LLMs. That’s so cool, now where’s my promotion? Business outcomes wise, there’s some low hanging fruits that have been plucked but otherwise it doesn’t live up to the hype.
Personally for me, it is helpful in a few scenarios,
1. Much better search interface than traditional search engines. If I want to ramp up on some new technology or product, it gives me a good broad overview and references to dive deep. No more 10 blue links.
2. Better autocomplete than before but it’s still not as groundbreaking as AI hype hucksters make it out to be
3. If I want to learn some concepts (say how ext4 FS works), it can give a good breakdown of the high level concepts and then I go need to study and come back with more Q’s. This is the only genuine use case that I really like. Where I can iteratively ask Q’s to clarify and cement my understanding of a concept. I have used Claude code and ChatGPT for this and I can barely see any difference between the two.
This is my balanced take.
As to the search engine, my searches are often very narrow, like I want to recall a specific message from a mailing list, so I don't use that too much. On the other hand, I found Google's NotebookLM to be really good at recalling concepts from both source code and manuals (e.g. processor manuals in my case).
Code generators are incredible refactoring machines. In one case (not so easy to reproduce in general, but it did work) Claude Code did a Python to decently idiomatic Rust conversion in a matter of minutes; it added mypy annotations to 2000 lines of Python code (with 90% accuracy) in half an hour and got the entire job done with my assistance in about an hour. For the actual writing and debugging where the logic matters they're still not there even for small code bases (again 2000 lines of code ballpark). They're relatively good at writing and debugging testcases but IMO that's also where there's a risk of copyright taint. Anyhow it's something I would use maybe 2-3 times a month.
In one case I used it for natural language translation, with pretty good results, but I knew both languages because I needed to check the result. Ask it first to develop a glossary and then to translate.
For studying they're interesting too, though for now I have mostly tried that outside work. At work, Google Deep Research worked well compared to the time it takes and it's able to find a variety of sources (including HackerNews comments in one case :)) which is useful for cross-checking.
That's the whole point
> I’d say my overarching advice, based on how difficult tech recruitment is right now, is to sadly play along. But — and I cannot stress this enough — make sure you document everything.
> What I mean by that is every single time AI tools cause problems, slow-downs and other disappointing outcomes, document that outcome and who was responsible for that decision. Make sure you document your opposition and professional advice too.
Personally, I would just add a warning to be careful to blame the tool, not the person. Otherwise, you will be seen as the "bad" person in the story even if your report is technically correct.
"Playing along" is a great way to be part of someone else's potentially-harmful project. Consider your values, and don't cross those lines. If the boss is upset about it, they have options. I don't do their work for them.
Collective action with your fellow workers against enshittification is a humanist way forward.
About generating things well... it just copypastes the same snippets you could find on stackoverflow, including bugs - if the task you throw at it has already been answered.
For complete and complex code... well it spews out the same useless advice you could get from a drunk non expert person while sitting at the bar.
Issue is... LLMs are too big to fail, everyone just poured billions in this huge statistics bean counter, and... someone has to justify those expenses at board meetings.
Using AI powered tooling is one thing, better IDE workflows, writing and voice recognition, translations and so forth.
Copying text around into and out of a chat window is worse than just writing COBOL, and at least COBOL is deterministic.
Its not forced per-se but its definitely being heavily encouraged.
LLMs augment the input with their trained data. LLMs don't inherently agree if you set up context correctly for analysis.
I've arrived at the conclusion that the top-down push without adequate upskilling creates bad experiences and subpar results. It's like adopting a new methodology for something without actually training anyone on the new methodology, it leaves everyone scrambling trying to figure it out often with poor results.
I find LLMs to be a great multiplier. But that multiplier will take whatever you put in context. If one puts in bias and/or fragmented mess, it's far more difficult to steer the context to correct it than it was to add it to begin with.
Once I changed my goal from maximizing code quality to maximizing billable hours, I feel a lot more optimistic about the future and these AI tools are going to create so many opportunities for me.
I found it hard to compete with junior developers in my last job (before AI was mainstream) in terms of volume of code because some of them would write 1000 lines of low quality code per day... Now I can also do this. It gives me a lot of surplus energy to figure out how to play politics and shift blame... It gives me an actual competitive upper hand over the juniors. I can out-compete them both in meeting/debate and code/feature volume. I talk wisely and code foolishly. Win win. I couldn't do this before because I was writing the code myself and I had essentially lost the ability to write high volumes of dirty code. I had a kind of analysis paralysis due to trying to solve problems in an optimal, minimalist, most reliable way. No longer a problem. Bugs are a problem for someone else.
I have so much more time to think about career strategy now. I managed to avoid being assigned to any difficult projects... I feel bad for the other people who try to go above and beyond and end up wedging themselves into a difficult situation where the software is down all the time and they have to take the blame... The AI never gets the blame.
I hated playing politics before but AI has made playing politics a necessity. It's like the more apathetic you are about your output, the better off you are.
The ironic thing is that I know it's possible to produce high quality code with AI. I've had some really positive experiences with Claude Code on side projects... But that doesn't align with the reality of 99% of software projects. The foundation is not set up right to get these kinds of results. I could set up the foundation correctly but I'd have to be present in the project since the beginning and I'd have to be given a lot of decision power; but I never get such opportunities. Bad foundations beget bad code, especially with AI because the AI never gets the idea to refactor... If your codebase is unmaintainable, it will hallucinate dirty code which doesn't work. It keeps coming up with more and more hacks... Then it delivers hacks on top of its own hacks.
For coding, unless I’m writing trivial RPC endpoints, editing docs, or writing tests for an already hardened API, I find agents a complete waste of time. So my usage is mostly limited to chat sessions.
To use up the quota, I apply the provided tokens to a few personal projects here and there, but no one can make me push an actual production CL with them unless I find it useful myself.
The effect on the team of developers with various level of experience started showing up as well:
The application architecture turn into a horrible mess, it's worse than junior engineers. The application started exhibiting tons of hard to debug issues, because the generated code was too low level and not covering corner cases.
Every attempt of the AI engineers to fix the issue generated one more class wrapping the existing codebase - with a fix which never worked (eg. ConnectionManagerWithTimeouts).
Eventually we basically had to rewrite the application, throwing away most of the code twice. One to just get something working with the existing architecture without crashing every hour and then another to use a framework and eliminate the last bugs occurring every once and then.
LLM needs to be in incredibly capable hands in order to be used safely and engineers will have to fight their instinct and not get swayed by the LLM telling them they're right.
At this point, you need to learn what AI can and cannot do, for the same reason you need to keep up with new versions of whatever framework you use. Since AI develops so fast (e.g many image use cases that AI would be terrible at 4 months ago, they now do perfectly), you need to repeat that exercise frequently.
There are 4 problems with adoptions as I see them:
1) Hype. Some people overhype what AI can do, which causes people to dismiss them when they don't immediately work;
2) Plenty of people don't like to change what they do/feel threatened by change. Doubly so when that change is perceived (real or not) to impact their job.
3) AI is weird and so it sometimes fails spectacularly at simple things, while it works very well at more complex things;
4) People use ChatGPTs free model or other AIs that are free. These are older/less powerful models, which means people end up with wrong expectations of what they can and cannot do.
5) Who likes to be told what to do? Especially by a clueless boss.
Where I running a company, I would ensure that my employees had access to a top of the line model and cursor/windsurf. I would monitor usage and have a talk with those whose usage was drastically lower than their peers.
However it would be a talk only - with the aim of figuring out why AI did not work for that employee, and what we could do to fix it.
Imagine this sentence with "Claude Code" replaced by anything else and "CTO" and "devs" replaced by more generic terms like "boss" and "employee". It's just "The boss tried Tool X and liked it so he said all the employees have to use it". It just seems to me that that is a bad way to make decisions regardless of what tool we're talking about or even what industry we're in. It's certainly possible it could make sense with a few more steps in there ("the boss tried this and liked it because X so he said we have to work on using it in way Y to accomplish Z"). But the way this is described sounds like a fire-and-forget mentality where the boss tells people to do a thing a certain way and that's the extent of his involvement, which seems stupid.
I'm not sure why we have that obligation.
There seems to be some magical thinking at work that simply uttering the words “AI” and “automation” will somehow render us more productive.
But my refusal to be AI-assisted at work is viewed as "healthy skepticism" by upper management. And my colleagues who have tried the tools are not particularly impressed.