Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GitHub Copilot Workspace: Technical Preview

github.blog

Post reply on HN
Print view View original

61–70 of 333 posts

Re: GitHub Copilot Workspace: Technical Preview

#61

Post by lispisok »

studenthrow3831 wrote: post #3

Student here: I legitimately cannot understand how senior developers can dismiss these LLM tools when they've gone from barely stringing together a TODO app to structuring and executing large-scale changes in entire repositories in 3 years. I'm not a singulatarian, but this seems like a brutal S-curve we're heading into. I also have a hard time believing that there is enough software need to make such an extreme prod…

>they've gone from barely stringing together a TODO app to structuring and executing large-scale changes in entire repositories in 3 years

First they are definitely not currently as capable as you say. Second there is a misconception that the rise of LLMs has been exponential but the curve is really logistic and we've hit the flat tail hard imo. Where is ChatGPT5? All the Coding AI tools I've tried like Copilot either havent gotten better since release or seemingly gotten worse as they try to fine tune them. Third there is ton more to being a software engineer than writing ReactTodoAppMVCDemo which many responses have been talking about.

Re: GitHub Copilot Workspace: Technical Preview

#63

Post by cedws »

davidbarker wrote: post #60
cedws wrote: post #47

Earlier quoted context omitted.

Looks interesting, but I don't really want my code to go via some unknown company. As far as I can tell in "Privacy Mode" code still goes via their servers, they just promise not to store anything (with the caveat that OpenAI retain stuff for 30d).

They give you the option to use your own OpenAI/Anthropic/Azure API keys, but in all honesty, I don't know if they still gather information about your code even using your own API keys. You could use something like Little Snitch (on Mac) to check if it makes any calls to their servers. They also allow you to override the URL for the OpenAI models, so although I haven't tried, perhaps you can use local models on your…

Unfortunately, it looks like code still goes via their servers:

https://cursor.sh/privacy

> Even if you use your API key, your requests will still go through our backend!

> That's where we do our final prompt building.

Re: GitHub Copilot Workspace: Technical Preview

#64

Post by davidbarker »

cedws wrote: post #63
davidbarker wrote: post #60

Earlier quoted context omitted.

They give you the option to use your own OpenAI/Anthropic/Azure API keys, but in all honesty, I don't know if they still gather information about your code even using your own API keys. You could use something like Little Snitch (on Mac) to check if it makes any calls to their servers. They also allow you to override the URL for the OpenAI models, so although I haven't tried, perhaps you can use local models on your…

Unfortunately, it looks like code still goes via their servers: https://cursor.sh/privacy > Even if you use your API key, your requests will still go through our backend! > That's where we do our final prompt building.

Ah, that's unfortunate.

Re: GitHub Copilot Workspace: Technical Preview

#65

Post by piva00 »

michaelmior wrote: post #31
wilsonnb3 wrote: post #17

Earlier quoted context omitted.

> I also have a hard time believing that there is enough software need to make such an extreme productivity multiplier not be catastrophic to labor demand. Every single time a change like this happens, it turns out that there is in fact that much demand for software. The distance between where we are now and the punch card days is greater than where we are now and the post-LLM days and yet we have more software devel…

> Someone is still gonna have to convert what the business wants into instructions for the LLM It seems like with GitHub is aiming for is a future where "what the business wants" can just be expressed in natural language the same way you might explain to a human developer what you want to build. I would agree that right now, LLMs generally don't do well with very high-level instructions, but I'm sure that will improv…

> It seems like with GitHub is aiming for is a future where "what the business wants" can just be expressed in natural language the same way you might explain to a human developer what you want to build.

We've been there before with 4GL in many forms, they all failed on the same principle: it requires reasoning to understand the business needs and translate that into a model made in code.

LLMs might be closer to that than other iterations of technology attempting the same but they still fail in reasoning, they still fail to understand imprecise prompts, correcting it is spotty when the complexity grows.

There's a gap that LLMs can fill but that won't be a silver bullet. To me LLMs have been extremely useful to retrieve knowledge I already had (syntax from programming languages I stopped using a while ago; techniques, patterns, algorithms, etc. that I forgot details about) but every single time I attempted to use one to translate thoughts into code it failed miserably.

It does provide a lot in terms of railroading knowledge into topics I know little about, I can prompt one to give me a roadmap of what I might need to learn on a given topic (like DSP) but have to double-check the information against sources of truth (books, the internet). Same for code examples for a given technique, it can be a good starting point to flesh out the map of knowledge I'm missing.

Any other case I tried to use it professionally it breaks down spectacularly at some point. A friend who is a PM and quite interested in all GenAI-related stuff has been trying to hone in prompts that could generate him some barebones application to explore how it could be used to enhance his skills, it's been 6 months and the furthest he got is two views of the app and saving some data through Core Data on iOS, something that could've been done in an afternoon by a mid-level developer.

Re: GitHub Copilot Workspace: Technical Preview

#66

Post by StefanWestfal »

studenthrow3831 wrote: post #3

Student here: I legitimately cannot understand how senior developers can dismiss these LLM tools when they've gone from barely stringing together a TODO app to structuring and executing large-scale changes in entire repositories in 3 years. I'm not a singulatarian, but this seems like a brutal S-curve we're heading into. I also have a hard time believing that there is enough software need to make such an extreme prod…

My two cents: I worked in a different engineering field before transitioning to Software Engineering because "coding" was and is what we need to solve problems, and I got the hang of it. A few years in, I spend little of my day actually writing code but more time in meetings, consoles, documentation, logs, etc. Large language models (LLMs) help when writing code, but it's mostly about understanding the problem, domain, and your tools. When going back to my old area, I am excited about what a single person can do now and what will come, but I am also hitting walls fast. LLMs are great when you know what you are doing, but can be a trap if you don't and get worse and worse the more novel and niche you go.

Re: GitHub Copilot Workspace: Technical Preview

#67

Post by michaelmior »

skydhash wrote: post #58
michaelmior wrote: post #31

Earlier quoted context omitted.

> Someone is still gonna have to convert what the business wants into instructions for the LLM It seems like with GitHub is aiming for is a future where "what the business wants" can just be expressed in natural language the same way you might explain to a human developer what you want to build. I would agree that right now, LLMs generally don't do well with very high-level instructions, but I'm sure that will improv…

SQL and python are arguably the languages closest to English, and even then getting someone to understand recursion is difficult. How do you specify that some values should be long lived? How do you specify exponential retries. Legalese tries to be as specific as possible without being formal and even then you need a judge on a case. Maybe when everyone has today’s datacenter compute power in their laptop.

> arguably the languages closest to English

Yes, but they're not English. All the concerns that you mention are ones that I think LLM development tools are aiming to eliminate from explicit consideration. Ideally, a user of such a tool shouldn't even have to have ever heard of recursion. I think we're a long way off from that future, but it does feel possible.

Re: GitHub Copilot Workspace: Technical Preview

#68

Post by idan »

studenthrow3831 wrote: post #3

Student here: I legitimately cannot understand how senior developers can dismiss these LLM tools when they've gone from barely stringing together a TODO app to structuring and executing large-scale changes in entire repositories in 3 years. I'm not a singulatarian, but this seems like a brutal S-curve we're heading into. I also have a hard time believing that there is enough software need to make such an extreme prod…

The opposite, we see these tools as mechsuits to help developers, and particularly newer developers, to do things that they would struggle to do.

Power tools did not result in fewer buildings built. I mean I guess some early skyscrapers did not benefit from modern power tools. But I don't think any construction company today is like "nah we'll just use regular saws thanks".

The allergy to hype is real; I don't think this or any tool is a magic wand that lets you sit back and just click "implement". But the right UX can help you move along the thought process, see solutions you might not have gotten to faster, and iterate.

Re: GitHub Copilot Workspace: Technical Preview

#69

Post by michaelmior »

piva00 wrote: post #65
michaelmior wrote: post #31

Earlier quoted context omitted.

> Someone is still gonna have to convert what the business wants into instructions for the LLM It seems like with GitHub is aiming for is a future where "what the business wants" can just be expressed in natural language the same way you might explain to a human developer what you want to build. I would agree that right now, LLMs generally don't do well with very high-level instructions, but I'm sure that will improv…

> It seems like with GitHub is aiming for is a future where "what the business wants" can just be expressed in natural language the same way you might explain to a human developer what you want to build. We've been there before with 4GL in many forms, they all failed on the same principle: it requires reasoning to understand the business needs and translate that into a model made in code. LLMs might be closer to that…

I agree that we're far off from such a future, but it does seem plausible. Although I wouldn't be surprised to find that when and if we get there, that the underlying technology looks very different from the LLMs of today.

> something that could've been done in an afternoon by a mid-level developer

I think that's pretty powerful in itself (the 6 months to get there notwithstanding). I expect to see such use cases become much more accessible in the near future. Being able to prototype something with limited knowledge can be incredibly useful.

I briefly did some iOS development at a startup I worked at. I started with literally zero knowledge of the platform and what I came up with barely worked, but it was sufficient for a proof of concept. Eventually, most of what I wrote was thrown out when we got an experienced iOS dev involved. I can imagine a future where I would have been completely removed from the picture at the business folks just built the prototype on their own. Failing that, I would have at least been able to cobble something together much more quickly.

Re: GitHub Copilot Workspace: Technical Preview

#70

Post by sqeaky »

cryptoz wrote: post #24

This is quite similar to my idea and prototype I applied to YC with just last week. Uh-oh! Haha. I did call out GitHub as a primary competitor so I did see this coming - and I guess on second looks I'm still quite differentiated. All these new tools will overlap some but also have their own pros/cons since there are so many different ways to approach building an AI code tool! Just yesterday I put up a cheap $5 linode…

Good luck! I hope to see many competitors in this space because hegemony in this space would inhibit the software industry as a whole. Figure out your differentiators and figure what copilot sucks at. Succeeding here can be done.
Post reply on HN