I laughed when I read this, but there is something to it. I like to say "intellectual relaxation" or take a break. Sometimes getting up from your desk to do some mindless admin task like photocopy a document for HR can free up your mind. If we were line workers at a factory, this would be mandated breaks. Business/Financial newspapers and factory executives love the old quote: "With robots, they never need a break, never need holiday, and can work 24x7." With the advent of agentic LLMs, a tiny fraction of that reality is leaking into the white collar world.
It's definitely theoretically possible, but not there yet. I use cursor, claude (opus 4.7), and several proprietary LLMs/LLM frameworks at my job. The institutional knowledge I have wouldn't fit in the context window, and AIs lack my mental index/intuition of where to look for answers. When my AI makes a PR, I generally have to make some important changes, without which it's solution would be fundamentally broken. AI also cannot be trusted to make the right business tradeoff decisions.
Many things at my software engineering job are like this, which require constantly changing human institutional knowledge that is almost always undocumented, or changing so quickly that it isn't relevant anymore. By the time you decide to automate it, the process changes. Tribal knowledge used to be something I hated seeing senior engineers keeping to themselves, but now it seems like an asset.
It sounds plausible to me since this is pretty en par with most other engineering disciplines. I’m a civil engineer. My responsibility is ultimately mostly to produce a constructable plan set. I spend far less than 5% of my time drafting or modeling.
The least experienced developer writes the most code. Juniors would be spending whole day in the IDE, typing, testing, typing etc.
Senior developers will go to a park for a few hours, think, then come back spent an hour or less typing code that just works or write nothing at all, maybe even delete code.
Instead they might update documents, ask clarifications about found edge cases or errors in planning that were not considered.
Since software is in every industry of man, I think you'll need to mention which industry this perspective is coming from. This is definitely NOT the case in certain industries.
For those who claim to be developers who code no more than 5% of their time and resort to arguments like "we're already not writing machine code by hand for 50 years, how is AI different from a higher level language?", it's not commenting, it's shilling for the AI corpocracy on HN.
>> "we're already not writing machine code by hand for 50 years, how is AI different from a higher level language?"
I never got that argument. Compilers are formally proven, deterministic algorithms . If you understand what compiler does, you can have pretty good idea what it will produce. If it doesn't do that, its a bug. Definition of correctness is well defined by semantic equivalence.
LLMs are none of that. Its a fuzzy system that approximates your intent and does its best. I can make my intent more and more specific to get closer to what I want, but given all that is just regular spoken language its still open to interpretation. And all that is still quite useful, but I don't get the assembly language comparison here.
Because compilers are only deterministic when using ahead of time compilation, without profiling data, and always the same set of compiler flags.
Introduce dynamic compilation, profiling data, optimization passes, multiple implementations, ML driven heuristics, and getting deterministic Assembly output from a compiler starts to get harder to achieve.
You are right about that but that's talking about what you generate but not what the output does. My point is that the compilers still designed to preserve semantic equivalence. semantic equivalence makes sense here because there are semantics well defined for both input and output. That bit is supposed to be deterministic. If something breaks that that is a bug.
I just don't think comparing with compilers is a good argument.
and hence reading code is unnecessary because how well LLMs understand and converts my prompts is almost equivalent to how well compilers can understand programs and turn into assembly. The prompts carry equal amount of ambiguity as the prompts I would write to define the behavior and want.
If you spend 95% of your time on that stuff, you better be working on like critical infrastructure where nothing can go wrong, otherwise you are in an incredibly dysfunctional company.
I agree it would be absurd for it to take 95% of your time.
I have, however, seen that it takes a lot more time than one would think.
I did some contracting work for a severely dysfunctional meeting heavy organization and it was about 2 hours of meetings for every hour of real technical work!
Even when it’s not dysfunctional, you spend a lot of time on communication and reading stuff other people wrote (including code). It’s very rare to work in isolation.
I guess it depends on what you feel coding is. To me it's the architecture planning and reading other people code, not just writing code. If we say it's just typing, then 95% is not absurd no
> it depends on what you feel coding is. To me it's the architecture planning and reading other people code, not just writing code
And that would be where we disagree. I don’t read code to look at code. When I’m reading code, I’m looking for the contracts to follow when interacting with a system. It would be nice if it were documented, but more often than not you have to rely on code.
It’s very rare that I plan with a technical mindset. Yes I use the jargon, but it’s all about the business needs. Which again create contracts.
Same with writing code. Code is like English for me. If I don’t have a clear idea on what to write, I stop and do research (or ask someone). But when I do, it’s as straightforward as writing a sentence.
Huh? So you you don't research if something is technically feasible before you promise your stakeholders a delivery time/ price estimate?
We all do the same stuff, the disagreement would just be what you feel coding is and if you think technical work is the same thing or a superset. If you as software dev aren't hands on with planning or working more than 5% of your time, you are basically a PO with a programing hobby
> So you you don't research if something is technically feasible before you promise your stakeholders a delivery time/ price estimate
I believe 99% of requests are not about what’s technically feasible. And the rare time I encountered one of those, my answer has mostly been “you don’t have enough resources to try solving that problem”.
If you know your fundamentals well, very often you will find the same common blocks everywhere. People much more smarter than me has solved a lot of fundamental issues and it’s rare that I see a business request that doesn’t reuse the same familiar stuff.
That’s why coding is mostly boring. You follow the same pattern again and again. But what dictates the flows are the business parameters. And that’s why most senior spend so much time gathering good requirements. Because the code is straightforward after that.
Ah yes agreed, if it's more than 90% it just signals to me that a developers skills are probably being wasted too much on business/coordination stuff.
But i guess if we mean actual time tapping your keyboard making code, then it's true some days for senior+ devs, but definitely not technical work overall.
Yes, far better than how the UK runs its state pension system.
The Australians seem to have the best model overall though. Mandatory payments in to private investments has made them very wealthy.
The UK system takes the national insurance contributions of workers but doesn’t invest them in anything on behalf of the individual. So despite decades of payments you technically have nothing at the end and survive on the goodwill of the government and current taxpayers. That works right now because of the population pyramid.
>The UK system takes the national insurance contributions of workers but doesn’t invest them in anything on behalf of the individual. So despite decades of payments you technically have nothing at the end and survive on the goodwill of the government and current taxpayers. That works right now because of the population pyramid.
That's how Social Security works in the United States as well.
I believe it’s tied to your earnings in the US though, which it isn’t in the UK.
I also have a number of qualifying years in the UK when I didn’t work, and for decades you could buy a year contribution for about £150. The payout is £12,500 per year.
>That works right now because of the population pyramid.
Is it really a pyramid if the base is less wide than the top? I guess it would be an upside down pyramid, but not very useful for the intended purpose then.
I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code. Turns out there’s a GitHub Copilot CLI!
I thought I was pretty familiar with available options, but no one in my circles ever mentions this product. It doesn’t seem to have much mindshare.
I'm curious about the opposite: Why would anyone use the CLI when, at least with Copilot, the VSCode plugin is super tightly integrated with VSCode, meaning the agent can see everything I can see. There's no mismatch in linter calls where I can see a lint in the ide that the agent can't find for example. I've had this problem even using CC in their VSCode extension, so I can't imagine it's not an issue in the CLI as well.
I use the Claude Code VSCode plugin for 80% of my work.
I prefer it because I can look at the code (although not as often anymore) and config (very often!) easily.
It also lets me jump to previous conversations easily.
There are a few cases where the CLI makes sense. One big one is if you are running multiple simultaneous sessions on a remote server using Tmux to have them preconfigured when you reconnect is nice.
The vs code integration is pretty slick. I can copy and paste function names into the prompt and it automatically turns them into these `#sym:` reference objects that I presume populate the context window with metadata about the function and where it lives. It knows what file I'm currently looking at as I jump around in the code, and that automatically gets loaded into the context. I can also drag and drop folders or specific files for context into the sidebar.
It's a lot of stuff that makes me have to type less into the prompt, since it's already getting so much info from my editor
I've used it quite a bit. There are a lot of AI terminal coding products and this is another one. It works well, handles sub-agents without issue and does a reasonable job operating in the Copilot ecosystem. It handles mid-task questions and such we well.
But its a really good UI for agentic coding. Not sure why more people don't use it. I've tried the others and keep coming back to Copilot chat. It's a really good tool. Which is why the rugpull on pricing is so concerning.
I’m actually trying to move back from the Claude Code style, I feel like it’s easy to become distant from your own code, and I am feeling uncomfortable with that.
I’ve “vibe-coded” some projects and when I start to find issues or go to refactor them I don’t have that memory of why decisions were made, because many decisions were never made.
The other cool thing is Copilot SDK, so you can build agentic capabilities into apps, or build tools, that leverage the agent harness of the Copilot CLI:
> I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code.
I use Claude Code, but I kept my Copilot subscription around mostly for really cheap usage of other models when I need to try a different one (which appears to be ending, in a sense) and also the autocomplete in Visual Studio Code which was really great across a bunch of files, I could make changes in one file and then just tab through some others.
I wonder what other good autocomplete is out there.
> also the autocomplete in Visual Studio Code which was really great across a bunch of files <...> I wonder what other good autocomplete is out there.
I am in the same boat. I tried looking for tab/auto-complete implementations ~ a year ago and it was pretty disappointing. If that has changed, would love to know!
Windsurf has free unlimited tab complete, I use it as an IDE alongside Claude Code and it works pretty well. I think Google's Antigravity also has free tab complete but no idea if its any good.
Yeah, I've been using it heavily at work since the beginning of January (and have a personal Anthropic sub to compare to). Copilot CLI is pretty good, honestly. Most new features in Claude Code get cloned by Copilot CLI within a couple weeks. Claude models seem mildly more clumsy in that harness than the one they're trained on - subjective guess around 20% more turns for an equivalent task - but it's not a noticeable difference in the final output.
That's espoused as the big reason for the price increase: most Copilot subscribed developers it seems have moved to "agentic usage" with the CLI and Cloud-based agents.
Which feels a bit like a kick in the pants for me as a developer that was primarily using Copilot for VS Code ghost text and very rarely used the Chat sidebar much less "agentic" tools.
Copilot Pro sort of made sense for my personal account when amortized across a year, but I don't want to "waste" $10/month on credits I won't use most months.
I'm just so confused why people aren't just using ghostty/kitty/terminal.app and claude code. Compared to the other approaches I've tried, it's by far the most effective way to get performance from opus 4.6/4.7
I don't know about others, but I use Copilot more often than other apps because of its tight integration with the VS Code itself where I still spend most of my time working on other things while letting AI do some task that I decided to delegate to it.
Liability. I work in a tightly regulated market and I cannot use any service due to be in touch with PII data. If it was my option I would just use OpenCode with deepseek, it is more than enough for the majority of my tasks.
Search has become so bad that I also struggled to find Claude Code alternative and made my own tight (not editors, not plugins, not agents, strictly similar to Claude Code CLI) list: https://github.com/omarabid/cli-llm-coding
The list is not long but there are quite a few options. Even Grok has its own CLI!
The reality is, even though a CLI prompt looks very simple, it's a very complex piece of software. I personally use Claude Code (with GLM) and anything else I have tried was significantly inferior (with the exception of opencode).
The sidebar method is better than agents for many reasons.
The main one is I'd rather review the limited changes of a one shot prompt than ton of changes made by an agent.
Another being the massive use of tokens agents force you into. As the context window fills up, the quality goes down, and the token usage spikes as the model traps itself into low confidence reasoning loops.
I tried the VS Code + Copilot sidebar approach a few months ago. It was definitely rough around the edges compared to Cursor/Claude. In our corporate environment, we weren't even able to use frontier models.
I've been using it for a few months because Copilot was the only AI blessed by our corporate overlords. It's not bad, I would say it's about 80% as capable as Claude Code, which I've used extensively on personal projects. However CC was recently approved, and I'm betting that with these changes to Copilot pricing we'll end up dropping it like a hot potato.
Nah, though I imagine every insurance company that did these sorts of mail shots got the same stuff back.
Apparently there used to be people who would buy the policy purely for the parker pen. They'd get the pen then cancel the policy for a full refund. There's some weird people out there.
I had this experience starting a new company recently.
Every single SaaS product seemed to have a dozen onboarding floating modals that need to be dismissed. It would have been impossible to read them all. In most cases I had used the product a lot before but I simply had a new corporate email so they thought I was a new user.
So if any said anything important I wouldn’t know because I had to dismiss them all.
Does this cancel out the high energy prices people in the UK have been paying for the past decade+? Part of the reason the bills are high is because they subsidise the installation of renewable generation.
Is it obvious? Do you have the figures? One month of £1bn savings vs what cost over the last 20 years?
I support the roll out of renewables. I worked for a renewable energy company in the UK.
I just think a lot of consumers (industry and residential) have had to pay a very high price to get to the current situation. And lots of foreign private and other institutional (the Crown) organisations have benefitted most from this fast but expensive transition.
Thanks - I wish it could be drilled into by category, i.e. what are the stats for categories of import (filter out sports, crypto, etc). My worry is the average could appear rosier if the share of trivial events are high.
As a viewer I don’t think this is in my interest as I think they will get a lot less prestige guests now. They have interviewed some huge names recently.
May I ask if you could estimate how you spend the other 95% of the time?
reply