GitHub Copilot: The Agent Awakens • Hacker News Board

Re: GitHub Copilot: The Agent Awakens

#221

Post by scarface_74 » Fri, Feb 07, 2025, 3:37 AM UTC

bradchris wrote: post #219

scarface_74 wrote: post #218

Earlier quoted context omitted.

> A big role of full time employees is not just to code, but to interact to various degrees with PMs/Sales/Customers/the rest of the company stakeholders to varying degrees. That’s true the further you get up in your career. But most of the time, it is: - junior developers get tasks where both the business problem and technical solution is well defined and they need a lot of handholding and supervision - mid level de…

Ok, except I guess I would say your definition of mid-level and junior both fall under what I would consider “junior”— maybe I would call what you define as “junior” as “intern” ? I don’t see how LLMs completely eliminate anyone who is doing anything more than simply pulling well-defined tickets off a board

While I’ve spoken to people at other BigTech companies about their leveling guidelines, the only one that is still in my personal possession after I left is Amazon’s :).

But this is a high level industry summary.

https://www.levels.fyi/blog/swe-level-framework.html

Re: GitHub Copilot: The Agent Awakens

#222

Post by scarface_74 » Fri, Feb 07, 2025, 3:52 AM UTC

wholinator2 wrote: post #148

throw83288 wrote: post #115

Earlier quoted context omitted.

So Work=0.1^Ct where C is the development pace. Everything points to the C of AI being large. How quickly does Work become a rounding error? Sure, C=log(t), but it could also be C=ke^t. Everything to me feels like it's the latter, I really want to be wrong.

Cmon man, look at nature, exponential curves almost never are actually exponential. Likely it's the first part of a logistic curve. Of course you can sit here all day and cry about the worst outcome for an event in the long list of things no one can predict. It sounds like you've made your mind up anyways and refuse to listen to reason, so why keep replying to literally everyone here telling you that your buying into…

People keep telling students basically to “think happy thoughts” and are not being honest with them. The field is contracting today while more people with experience are chasing fewer jobs and then AI is hallowing out the low end.

Every single opening gets 1000s of applicants within the first day. It’s almost impossible to stand out from the crowd if you are either new to the industry or have a generic skillset.

Re: GitHub Copilot: The Agent Awakens

#223

Post by throw83288 » Fri, Feb 07, 2025, 4:43 AM UTC

scarface_74 wrote: post #205

throw83288 wrote: post #200

Earlier quoted context omitted.

> a senior developer who operates at a certain level of scope, impact and “dealing with ambiguity” and can deliver business value? Is there any chance for me (a student) to become like this? I'm fine with coding changing (I just love computing) but I'm scared of the entirety of the field being completely torched.

Please take my advice with a huge grain of salt. It’s been literally decades since I was an entry level developer. I try my best to keep my ear to the ground and look through the eyes of people at all levels of the industry. Part of my job is mentorship as a “staff software architect” at a consulting company. What would I do these days? I would stay in computer science and if possible get an MBA. I dropped out of gra…

As far as I've heard, MBAs have also become completely saturated as well. Out of the frying pan into the fire.

I get you're trying to be "consoling", but frankly the bajillion pivot ideas, hopium arguments, endless counterarguments, and other indirection is why I think there's nothing optimal that can be done. All I can do is go through the motions with my current internship and major and rely on Christ rather than this fickle world. I made the wrong choice. Nothing that can be done.

Re: GitHub Copilot: The Agent Awakens

#224

Post by scarface_74 » Fri, Feb 07, 2025, 4:45 AM UTC

throw83288 wrote: post #223

scarface_74 wrote: post #205

Earlier quoted context omitted.

Please take my advice with a huge grain of salt. It’s been literally decades since I was an entry level developer. I try my best to keep my ear to the ground and look through the eyes of people at all levels of the industry. Part of my job is mentorship as a “staff software architect” at a consulting company. What would I do these days? I would stay in computer science and if possible get an MBA. I dropped out of gra…

As far as I've heard, MBAs have also become completely saturated as well. Out of the frying pan into the fire. I get you're trying to be "consoling", but frankly the bajillion pivot ideas, hopium arguments, endless counterarguments, and other indirection is why I think there's nothing optimal that can be done. All I can do is go through the motions with my current internship and major and rely on Christ rather than t…

I got nothing then

Re: GitHub Copilot: The Agent Awakens

#225

Post by crakhamster01 » Fri, Feb 07, 2025, 5:00 AM UTC

I keep going back and forth on whether Agents are good for the software development discipline.

While I think it's extremely short-sighted that we continue to push full steam ahead on AI automating away jobs, I can't deny that LLMs have given my development flow a decent productivity boost. 80% of the time, my workflow with Cursor looks similar to the golden path depicted in this blog post - outline the changes I want made -> review the code -> suggest edits/iterate -> ship it. There's undoubtedly a class of problems where this feature can slot in and start chipping away immediately.

The other 20% of the time, Cursor will hit a wall and is unable to complete the task through just prompting. It will either introduce a subtle bug in its logic, or come across an error that it incorrectly diagnoses. These stumbles can happen for a variety of reasons:

  1. Poorly documented code - the LLM infers the wrong responsibility for a piece of code, or is led astray by old comments

  2. Misleading or unhelpful errors from 1st/3rd party libraries

  3. Task is too complex - perhaps I asked for more than I should have

In any case, the "self-healing" functionality that Agents rely on to iterate is often insufficient. Prompting for a fix usually just leads me in circles or further down the path of a bad solution. In these instances, I have to drop the coding assistant altogether and do things the old fashioned way - gain a sufficient understanding of the code and figure out where the LLM went wrong (or just write the solution from scratch).

I guess going back to my initial point, it feels like the easy answer is that Agents are good if you're a senior/experienced developer. This means that in the short-term the demand for junior engineers will dry up, since we have Agents to do the rote work, but doesn't this mean that we're effectively choking out the pipeline for experienced devs? Though they're low in complexity/value, the tasks we will handoff to Agents are immeasurably useful for building software fundamentals.

It seems like in 2025 we've suddenly forgotten about "teaching a man to fish"...

Re: GitHub Copilot: The Agent Awakens

#226

Post by danielbln » Fri, Feb 07, 2025, 6:19 AM UTC

username223 wrote: post #124

gokhan wrote: post #49

Earlier quoted context omitted.

As a programmer, I think it's easier to replace programmers than many other occupations. We work with generic DSLs called programming languages, which can be easily handled by AI. Most of what we produce can be easily parsed by an AI. We work in small, incremental tasks which is perfect for AI. Most of us use the same set of tools. What else? UI is similar everywhere, especially in LoB apps. What we produce is repeti…

Sure, programmers work in formal languages, but they can't make mistakes. If your marketing copy contains a few lies or your hero image has a guy with six fingers, the recipients will still infer the intended meaning (or ignore it). If your program has a subtle bug, the computer will faithfully do the wrong thing. LLMs are great for tasks where small mistakes don't matter, and useless for ones where they do. Generati…

Of you think most programmers won't mess up 10 lines every 10000 lines then I have a bridge to sell you.

Re: GitHub Copilot: The Agent Awakens

#227

Post by semi-extrinsic » Fri, Feb 07, 2025, 6:21 AM UTC

azan_ wrote: post #185

semi-extrinsic wrote: post #119

Earlier quoted context omitted.

It sounds like you've never used a welding torch, installed a kitchen sink, or done similar blue collar work. These jobs will never be replaced by robots, or by a non-trained person wearing a headset.

> It sounds like you've never used a welding torch, installed a kitchen sink, or done similar blue collar work. These jobs will never be replaced by robots, or by a non-trained person wearing a headset. Why do you think they will never be replaced by robots?

One thing is as sibling post commented, the complexity of such jobs are staggering from a robotics point of view.

The other thing is that the salary of the plumber or welder is in the range $20/hr to $40/hr. Can you make a general purpose, agile robot function at a total cost of ownership that's substantially lower than this?

Re: GitHub Copilot: The Agent Awakens

#228

Post by mhast » Fri, Feb 07, 2025, 7:53 AM UTC

throw83288 wrote: post #215

BarryMilo wrote: post #195

Earlier quoted context omitted.

Do you not see the contradiction in what you're saying? "LLMs will replace us all... I can't find a use for LLMs". Get to replacing us already. It's a gold rush, someone needs to get rich selling shovels.

I'm not trying to express that, it's that I can't find many meaningful projects with them.

If you can't figure out something innovative and ground breaking to do as a project try to copy something which already exists and you find useful.

It can be an app or website, or something else. By trying to replicate something you will begin to learn and understand what the limitations and opportunities in these tools are.

Developers hanging out on eg Hacker News are the very tip of the wave. 99% of all developers don't visit sites like this. It will take a long time (10+ years.) before AI moves through all fields. Companies which are inherently software focused will be first. But that's not where the long tail of software is.

The long tail is in companies which does no software currently. Places where software is something you buy to do inventory or keep track of invoices or time sheets.

I use these tools everyday now and they are both magically awesome and stupendously stupid, sometimes in the same reply. Things always change less in the short term (say less than 5 years), and more in long term (10+ years). Just as it was with smartphones, arguably the most recent big revolution which is now integrated everywhere.

To reiterate. To find a project chose something which already exist and you use and have use of. Then copy it. Then improve it (to better for your use.)

Personally I think this new revolution will make software exponentially more abundant and infinitely customizable and adaptable to individuals. My guess is also that in 10-20 years we will have more people than today who do "programming".

A lot of tasks which are now hacked together in Excel sheets will be a lot easier to make into "proper" programs.

In this world the people who know enough about many things to efficiently use agents to accelerate them will be the most valuable.

Re: GitHub Copilot: The Agent Awakens

#229

Post by singularity2001 » Fri, Feb 07, 2025, 9:03 AM UTC

lolinder wrote: post #20

theasisa wrote: post #10

Earlier quoted context omitted.

I tried Cursor a couple of years ago and wasn't impressed - has it improved a lot? I only use autocomplete, not the chat function and at the time found CoPilot superior.

I tried it last month on a medium size personal project and was blown away by the autocomplete. I'd previously staunchly refused to try it on the grounds that I'm too productive in IntelliJ, but at this point I'm most likely going to start paying for both. I don't know if I'm ready to use it as a daily driver, but there are certain kinds of tasks—especially large refactors—where its ability to rapidly suggest and acc…

did you use intellij with copilot auto complete before that?

Re: GitHub Copilot: The Agent Awakens

#230

Post by singularity2001 » Fri, Feb 07, 2025, 9:11 AM UTC

simonw wrote: post #79

dabinat wrote: post #5

I use Copilot and it’s a game-changer for my productivity, but I really wish it was capable of natural language searching. So for example I could ask it “show me all places in the code where x is assigned a value but a flush() command is not immediately issued afterwards”.

I answer those kinds of questions by piping my entire codebase into a large context model (like Claude or o3-mini or Gemini) and prompting it directly. Here's a recent example: files-to-prompt datasette tests -e py -c | \ llm -m gemini-2.0-flash-exp -u \ 'which of these files contain tests that indirectly exercise the label_column_for_table() function' https://gist.github.com/simonw/bee455c41d463abc6282a5c9c132c...

doesn't it get very expensive quickly if you don't use prompt cashing