Ask HN: Who has seen productivity increases from AI
I would love examples of positions and industries where AI has been revolutionary. I have a friend at one of the largest consulting firms who has said it'd been a game-changer in terms of processing huge amounts of documentation over a short period of time. Whether or not that gives better results is another question, but I would love to hear more stories of AI actually making things better.
23 comments
I'm an engineer in the oil and gas industry. Some of my job involves messy judgement calls that I would never involve LLMs in, but some of my job boils down to integrating different items of data that exist across documents, drawings and a few databases that are in different formats and don't cross reference each other. At times I have used LLMs as a kind of "highly enhanced search engine" to do semantic search across documentation of every different types. My alternative was opening each document and using ctrl+f, along with my intuition of knowing what document titles to search for.
For a more concrete example, I have an interface to the data that comes from every sensor on the oil processing facility. It has a built in "AI" (I try not to use that term!) but it has a feature where I ask how to process data in plain language and it'll give me the calculations, then it'll also provide a plain language summary of all the calculations I conducted. That saved me 10 hours of work.
I am a negative nancy on LLMs in general but I still passionately believe that they're a tool which every white collar employee will need to learn to use effectively.
I cringe when I hear engineers say "I didn't know the answer so I asked ChatGPT" but I also do worry that I could be significantly outperformed by another engineer with 10 years less experience in engineering and 1 year more experience in judicious use of LLMs.
For a more concrete example, I have an interface to the data that comes from every sensor on the oil processing facility. It has a built in "AI" (I try not to use that term!) but it has a feature where I ask how to process data in plain language and it'll give me the calculations, then it'll also provide a plain language summary of all the calculations I conducted. That saved me 10 hours of work.
I am a negative nancy on LLMs in general but I still passionately believe that they're a tool which every white collar employee will need to learn to use effectively.
I cringe when I hear engineers say "I didn't know the answer so I asked ChatGPT" but I also do worry that I could be significantly outperformed by another engineer with 10 years less experience in engineering and 1 year more experience in judicious use of LLMs.
What's an example of a messy judgement call?
Sorry I didn't see your comment until now, apologies for the late reply.
A classic messy judgement call would be:
1. Input information includes some word of mouth info that I have no reason to doubt, but also absolutely no way of verifying against field data
2. A single piece of equipment is not functioning - the plant is reasonably safe to operate with the failure, but how safe? Are the other relevant protective layers in place and effective in the relevant scenarios?
3. If I decide to implement a really robust and good-quality solution that'll stand the test of time, will it actually take so long to implement that I would have been better off with something simpler but less robust?
4. Is my decision making process clearly communicated enough for the decision makers involved? Which installation manager is on shift?
5. If the regulator audited my decision making process would they raise a recommendation? So what if there's a 0.1% risk that they'll raise a recommendation as long as people are safe?
These kinds of thought processes are where I add value as an engineer in a way that's irreplaceable by LLMs. I just hope that LLMs can really improve how quickly I can access data to make my human decision-making better based in fact.
A classic messy judgement call would be:
1. Input information includes some word of mouth info that I have no reason to doubt, but also absolutely no way of verifying against field data
2. A single piece of equipment is not functioning - the plant is reasonably safe to operate with the failure, but how safe? Are the other relevant protective layers in place and effective in the relevant scenarios?
3. If I decide to implement a really robust and good-quality solution that'll stand the test of time, will it actually take so long to implement that I would have been better off with something simpler but less robust?
4. Is my decision making process clearly communicated enough for the decision makers involved? Which installation manager is on shift?
5. If the regulator audited my decision making process would they raise a recommendation? So what if there's a 0.1% risk that they'll raise a recommendation as long as people are safe?
These kinds of thought processes are where I add value as an engineer in a way that's irreplaceable by LLMs. I just hope that LLMs can really improve how quickly I can access data to make my human decision-making better based in fact.
Me! I'm in a leadership position at a marketing agency. I'm a self-taught programmer (just found it interesting and started on my TI-83 in high school doing BASIC and then some Z80 assembly) and studied creative writing in college.
I used to hand-write a lot of the scripts we used to automate processes (mostly Google Apps Scripts and Python). It was very helpful to us but slow going and sometimes frustrating.
In the past, I was limited by my programming knowledge and ability. Now I am limited by my knowledge of our business and ability to explain what I want to happen.
With LLMs, I can generally 1-shot most things I want to automate, including things that would've taken me days to figure out in the past. I generally just say that I'll spend 30 minutes on something and more often than not, it's either done or very close at the end of that time.
I used to hand-write a lot of the scripts we used to automate processes (mostly Google Apps Scripts and Python). It was very helpful to us but slow going and sometimes frustrating.
In the past, I was limited by my programming knowledge and ability. Now I am limited by my knowledge of our business and ability to explain what I want to happen.
With LLMs, I can generally 1-shot most things I want to automate, including things that would've taken me days to figure out in the past. I generally just say that I'll spend 30 minutes on something and more often than not, it's either done or very close at the end of that time.
I find that I mostly use AI to compensate for Google search becoming unusable.
I use it more and more as advanced API documentation and writing code snippet. It made me a little bolder, and reduced friction from coding. I will test Claude Code in the upcoming weeks.
I use it more and more as advanced API documentation and writing code snippet. It made me a little bolder, and reduced friction from coding. I will test Claude Code in the upcoming weeks.
As a founder, it helps me keep track of things by understanding our user session recordings and telling me where exactly they're going wrong.
Three key things in coding which has helped recently:
1) Debug mode from Cursor. It highlights all possible hypotheses based on the steps to reproduce and whatever you know from the code. It slaps down logs and tests all these hypotheses at once.
2) Log reading, though this is a variant of the above. At times we get massive logs, like 10k lines for a bug. We go through what may possibly cause the bug - what screens, threads, race conditions, improper handling of ANY_KEY or default or something. Then we ask Opus to compare the logs to see if it matches the story. I ain't manually checking Line 314, Line 500, Line 44 to see if it shows in logs, but AI is just great at piecing together this whodunit.
3) Finding documentation. So there's been a few bugs unsolved around how very specific Android devices handle wifi reconnection or bluetooth or Ethernet. Instead of making assumptions, I can dig into the AOSP source code directly and see the conditions that trigger these. AI will even point out the exact lines.
1) Debug mode from Cursor. It highlights all possible hypotheses based on the steps to reproduce and whatever you know from the code. It slaps down logs and tests all these hypotheses at once.
2) Log reading, though this is a variant of the above. At times we get massive logs, like 10k lines for a bug. We go through what may possibly cause the bug - what screens, threads, race conditions, improper handling of ANY_KEY or default or something. Then we ask Opus to compare the logs to see if it matches the story. I ain't manually checking Line 314, Line 500, Line 44 to see if it shows in logs, but AI is just great at piecing together this whodunit.
3) Finding documentation. So there's been a few bugs unsolved around how very specific Android devices handle wifi reconnection or bluetooth or Ethernet. Instead of making assumptions, I can dig into the AOSP source code directly and see the conditions that trigger these. AI will even point out the exact lines.
I work in cloud consulting + app dev. I’m always responsible for the full life cycle of a project - team lead, discovery sessions with the clients, working with them through the project and handoff.
I can now do everything by myself on most projects. Up to ones that would have taken at least 2-3 other people before. Before I would have had to delegate it just because I couldn’t do it all by myself on time. I know how to develop (professionally did 30 years) and I know cloud (professionally for 8 years). I just didn’t have time.
On the other hand, I haven’t done web dev in a decade. But I can vibe code an internal website for operations and authentication via Amazon Cognito. It’s just a free feature that I give them even when it’s not in the contract
I can now do everything by myself on most projects. Up to ones that would have taken at least 2-3 other people before. Before I would have had to delegate it just because I couldn’t do it all by myself on time. I know how to develop (professionally did 30 years) and I know cloud (professionally for 8 years). I just didn’t have time.
On the other hand, I haven’t done web dev in a decade. But I can vibe code an internal website for operations and authentication via Amazon Cognito. It’s just a free feature that I give them even when it’s not in the contract
Has it affected how you calculate/charge your rates in any way?
I don’t have a good comparison of what rates are like for independent consultants. I have only worked full time salaried for consulting departments - at AWS directly (Professional Services - full time blue badge RSU earning employee) and now for a third party AWS partner.
But let’s say I make more as a staff consultant considering my actual billed hours, benefits (401K, health insurance, paid time off etc) than I would make working independently and I always know how much I’m going to get paid.
But I will say in today’s cloud consulting environment, it is a race to the bottom unless you can lead projects and even then there are relatively few high paying (over $200K with benefits and 80% utilization not including PTO) outside of being a full time employee working in the consulting divisions of AWS, Google (they had an RTO mandate so I ignored recruiters) or maybe Oracle.
The only reason they can justify American rates for someone like me who knows cloud but who is mostly a developer is because I can lead projects and churn out code quickly - thanks to AI.
But let’s say I make more as a staff consultant considering my actual billed hours, benefits (401K, health insurance, paid time off etc) than I would make working independently and I always know how much I’m going to get paid.
But I will say in today’s cloud consulting environment, it is a race to the bottom unless you can lead projects and even then there are relatively few high paying (over $200K with benefits and 80% utilization not including PTO) outside of being a full time employee working in the consulting divisions of AWS, Google (they had an RTO mandate so I ignored recruiters) or maybe Oracle.
The only reason they can justify American rates for someone like me who knows cloud but who is mostly a developer is because I can lead projects and churn out code quickly - thanks to AI.
These "questions" just seem like a trap/disingenuous. The people "asking" commonly have decided the answer and are just looking to reinforce that by playing "No true Scotsman" with other people's productivity increases.
We have software written in what is essentially an obsolete platform (a RAD solution from the 1990s). We've been slowly hand rewriting it into React. A single developer converts one a week by hand, and the results are good (we still have over a hundred).
We've developed an internal tool to turn the software's definition files into a single XML, and then are feeding it through a multi-process Codex pipeline (multiple instruction files, that product intermediaries), which ultimately outputs a "90%" working React page.
Our developers then PR the output, fix/adjust/test, and release. We've gone from one per week for a single developer, to roughly three per week without a marked quality loss.
It is taking care of most of the repetitive parts of the work, and allowing our developers to spend more time in technology they're familiar with (React) instead of legacy stuff. Is it perfect, no, but the cost/benefit is clear.
We have software written in what is essentially an obsolete platform (a RAD solution from the 1990s). We've been slowly hand rewriting it into React. A single developer converts one a week by hand, and the results are good (we still have over a hundred).
We've developed an internal tool to turn the software's definition files into a single XML, and then are feeding it through a multi-process Codex pipeline (multiple instruction files, that product intermediaries), which ultimately outputs a "90%" working React page.
Our developers then PR the output, fix/adjust/test, and release. We've gone from one per week for a single developer, to roughly three per week without a marked quality loss.
It is taking care of most of the repetitive parts of the work, and allowing our developers to spend more time in technology they're familiar with (React) instead of legacy stuff. Is it perfect, no, but the cost/benefit is clear.
What you described is that your team is producing software. That is exactly what a team of developers did 10 or 20 years ago.
Doing X in Y days does not mean much, same thing happened with Ruby on Rails.
Doing X in Y days does not mean much, same thing happened with Ruby on Rails.
I got same results in few minutes and cents that I got from a lawyer in few days and thousand dollars.
Surely you realize that the same result does not mean that the process leading to the conclusion was the same.
Even if the processes were identical, using an LLM arguably likely broke any privilege and confidentiality protections you might previously have enjoyed. See, e.g., United States v. Heppner (S.D.N.Y.) (recent case of first impression).
So there's that.
I was also going to comment separately that LLMs have been great to identify opposing counsel whom I suspected of not litigating in good faith / actually reading the cases and applying them. Now the confabulated cases are proof positive!
Even if the processes were identical, using an LLM arguably likely broke any privilege and confidentiality protections you might previously have enjoyed. See, e.g., United States v. Heppner (S.D.N.Y.) (recent case of first impression).
So there's that.
I was also going to comment separately that LLMs have been great to identify opposing counsel whom I suspected of not litigating in good faith / actually reading the cases and applying them. Now the confabulated cases are proof positive!
I didn’t understand half of your comment and guess what - I don’t care.
As long as it is not a privacy sensitive, high stakes or human factor related matter, I choose LLM any time of day.
All the bullshit lawyers made up to protect their own job I deeply can not care less about.
As long as it is not a privacy sensitive, high stakes or human factor related matter, I choose LLM any time of day.
All the bullshit lawyers made up to protect their own job I deeply can not care less about.
It’s made search a little easier. I don’t know if it’s enough to call it a productivity gain, but more of a quality of life improvement, surfacing more pages with added context in the results.
I imagine content farms, social media bots, at-scale scam operations, blackhat seo, busywork homework, social media propaganda botnetworks have all seen a significant boost of oroductivity.