Zed now predicts your next edit with Zeta, our new open model | Hacker News Reader

Zed now predicts your next edit with Zeta, our new open model

zed.dev

516 pointsahamez290 comments

> Edit prediction won't be free forever, but right now we're just excited to share and learn.

I love Zed, and I'm happy to pay for AI stuff, but I won't be using this until they are done with their rug pull. Once I know how much it costs, I can decide if I want to try integrating it into my workflow. Only THEN will I want to try it, and would be interested in a limited free trial, even just 24 hours.

Considering I've seen products like this range from free to hundreds of dollars per month, I'd rather not find out how good it is and then find out I can't afford it.

Other than that for anyone wanting to try Zed:

- You can only run one LSP per file type, so your Rust will work fine, your C++, too, your Angular will not.

- Remote editing does not work on Windows (its not implemented at all), so if you are on windows, you cannot ssh into anything with the editor remote editing feature. This means you cannot use your PC as a thin client to the actual chunky big work machine like you can with vscode. I've seen a PR that adds windows ssh support, but it looked very stale.

With the speed that models are coming out at and the amount of VC subsidies in trials my approach is the opposite, I don't get too attached and keep trying different tools and models.

And that's why most of these endeavors are doomed to fail. Every time they enshittify one service there's a new one with attractive UX that VC's essentially pay you to use instead. And so the previous investment would be lost if they couldn't dump it on naive stock traders by going public before the cat is out of the bag.

When products evolve rapidly, pricing will too. Whatever Zed or any actor is doing now or in 6 month on either front says little about what will happen next.

Just try what people offer right now, at a price point that you are okay with, right now. In a year, both product and price will probably be moot.

> - You can only run one LSP per file type, so your Rust will work fine, your C++, too, your Angular will not.

As a web developer that's an immediate deal breaker. I use Sublime today and being able to run multiple LSP servers per file is a huge boon, it turns a very capable text editor into a total powerhouse. The way it's set up in Sublime with configuration options that can be applied very broadly or very specifically, while having defaults that just works is also just incredible.

While I'm super pleased with Sublime and a happy paying customer since at least a decade, and at this rate may well be for another decase, I'm always keeping my ear to the ground for other editors if nothing else just to stay current. Zed's been looking pretty cool, but things like this will keep me from even just trying it. There's years of muscle memory and momentum built up in my editor choice, I'm not switching on whim.

Thank you very much for sharing this nugget of gold!

I'm not a regular Zed user, but this isn't true: I simultaneously ran the Ruff and Pyright LSPs when I used it last week.

In the same file?

You can run multiple LSPs on the same file.

In my currently opened project I have: vtsls (for typescript), biome, emmet, and the snippets LSP running on the same file.

You can configure which LSPs you can run on a language basis. Globally and per project. You can also configure the format on save actions the same way. Globally and per project.

I have astro project that on save runs biome for the front-matter part followed by prettier for the rest.

I would say that's pretty flexible.

With this feature they're competing with the features Cursor had for quite a while now, so I would expect a price competition there. That means close to $20 /mth. (windsurf without that feature is $15)

I would probably pay 100 quite happily for cursor, don't tell them .

> Other than that for anyone wanting to try Zed:

Also, no support for debugging yet.

You do not need debugging if you have AI.

You need more debugging if you have AI

I'm sure we'll see a microbubble in this space before long.

Oh hell yeah, an AI driven debugger that hallucinates memory values and instruction pointer positions.

Genuinely not sure whether you are joking or not. The thing is, I do not need a debugger very often, but when I need it, I need it.

I also have no idea what kind of code people "write" that they can rely on A.I. so much. I have found these tools helpful for gathering info but not for much more, yet.

I don't even find it that useful for gathering info. If it's good at that then there's some documentation out there that's going to be faster to comb through. It's useful for generating boilerplate for well defined unit tests and the occasional tab-complete.

Yeah, this is a moronic take. I'm all in on AI programming, but when you need debugging, you really need it - sometimes the model will get utterly fixated on a solution that is just wrong, and you just need to follow the stack trace.

I keep getting interested in Zed then rediscovering how much I like the easy debugging in jetbrains IDEs. Last time I checked Zed had a PR in progress. Maybe next time I think about it, Zed will be ready for me.

I asked them about this on X and they are working on one. I use Zed for everything now but must keep VS Code around just for the debugger. I can't wait to delete it.

Debugger PR is here https://github.com/zed-industries/zed/pull/13433 if you want to check it out

What is it about Zed that you find superior to VS Code?

I downloaded Zed when it took way too long for VS Code to load the monorepo at work. It took almost as long for me to download and install Zed and then open the monorepo as it did for VS Code to load the monorepo. I think the was a fluke with VS Code as it didn't normally take this long but it did happen often enough to be annoying.

I also find Zed to be snappier than VS Code. It's hard to quantify but it just feels better to use Zed.

For reference, I mainly work on a Node/Typescript monorepo that is made up of a bunch of serverless services and is deployed with SST v2

Hey, my name is Piotr and I work on language servers at Zed.

Right now you can run multiple language servers in a single project. Admittedly you cannot have multiple instances of a single language server in a single worktree (e.g. two rust-analyzers) - I am working on that right now, as this is a common pain point for users with monorepos.

I would love to hear more about the problems you are having with running language servers in your projects. Is there any chance for us to speak on our community Discord or via onboarding call (which you can book via https://dub.sh/zed-c-onboarding)?

I'm curious if you've given thought to improving json-schema support. Zed just packages VSCode's implementation (https://github.com/zed-industries/json-language-server ), which is generally decent, but hasn't been able to keep up with the spec, and I doubt they ever will at this point (Example: https://github.com/microsoft/vscode/issues/165219).

The newer specs for json-schema (not supported by VSCode) allow for a wider range of data requirements to be supported by a schema without that schema resorting to costly workarounds. VSCode's level of support for this is decent, but is still a pain point as it creates a sort of artificial restriction on the layout of your data that you're able to have without unexpected development costs. This of course can lead to missed estimates and reduced morale.

I understand that very few developers are directly producing and maintaining schemas. Those schemas do have an impact on most developers though. I think this is a problem that is being sadly overlooked, and I hope you can consider changing the status quo.

Love the company name btw, sounds similar to my own Nuzz Industries (not a real company, just a tag I've slapped onto some projects occasionally as a homage to Page Industries from Deus Ex).

I've been using Zed (with python) for the last few weeks (coming from vscode and nevim). There's a lot I like about Zed. My favorites include the speed and navigation via the symbol outline (and vim mode). I'd have a hard time going back to vscode. The LSP configuration, though, is not one of its best parts, for me. I ended up copy/pasting a few different ruff + pyright configs until one mostly worked and puzzled through how to map the settings from linked pyright docs into Zed yaml. Some better documentation for the configuration stanzas and how they map across the different tool's settings would be really helpful.

I still, for example, can't get Zed / LSP to provide auto-fix suggestions for missing imports. (Which seems like a common stumbling block: https://github.com/zed-industries/zed/discussions/13522, https://github.com/zed-extensions/java/issues/20, https://github.com/zed-industries/zed/discussions/13281)

I'm sure given the breadth of LSPs, that they all have their own config, and use different project config files, makes it hard to document clearly. But it's an area that I hope bubbles up the roadmap in due course.

Hi, thank you. I specifically meant running multiple LSPs in the same file at the same time, akin to vscode.

You can definitely run more than one LSP with zed -- can you elaborate on the angular case that gives you trouble?

I think the point is: "per file". Sure you an run a Rust LSP in one file and a JS LSP in another, but you can't drive both in the same file.

I run multiple LSPs on Ruby files with no issues. ruby-lsp and StandardRB are both Ruby specific.

You can. This is fully supported.

Sorry, in the same file as the sibling said. In some other editors, you can run multiple LSPs on the same exact file at the same exact time.

A use case is Angular or other more specialized frameworks, that are so abstracted away and dumbed down that each layer of abstraction basically has an LSP.

Is that the case though? https://zed.dev/docs/configuring-languages#choosing-language...

>Remote editing does not work on Windows (its not implemented at all), so if you are on windows, you cannot ssh into anything with the editor remote editing feature. This means you cannot use your PC as a thin client to the actual chunky big work machine like you can with vscode.

Does this work on anything other than VSCode ? I have been trying to use JetBrains stuff for this but it has been bad for years with little improvement. Honestly JetBrains feels like they are falling behind further and further in terms of adapting to providing a modern workflow - bad remote work, bad gen AI integration. I'm using VS code even where I wouldn't consider it before because of this, and I would like to see what the alternatives have to offer because VSCode is not perfect either.

Works on JetBrains, vscode, any terminal editor (neovim, vim, nano, etc.). Is it any good? It's fine on JetBrains, great on vscode, the rest is more or less great. Zed does not have it. You want to edit a remote file? You download it, edit it, upload it. That's much worse than a half baked implementation.

For clarity, this does work with the mac os version of Zed. I use it frequently to work on my GPU nodes. The one tradeoff that is a bit of a smell for me is that the preview version of the feature requires you talk to a centralized broker on Zed's servers rather than fully p2p between your local IDE and your own server.

This is supposedly temporary though IIRC (may even be changed already in a dev branch, not sure).

As someone that is old enough to have used UNIX development servers for the whole company, with PC thin clients, reading about remote development as modern workflow is kind of hilarious.

The remote development feature implemented in VS Code—and I believe also in beta in Zed—is a million times better than what you're used to. The UI is local, the storage and computation (including the language server) are remote. This takes away the lag when connected to a far-away server while still allowing things like platform-dependent compilation to work correctly and efficiently.

For the last 35 years, X Windows, Citrix and RDP have done the job just fine, in what I am concerned.

No, they aren't anything better than what I am used to.

Also, as compromise, a cloud shell, browser based IDE setup, and dev containers, also does the job in what concerns cloud deployments, which should be driven from CI/CD and don't have shells on containers anyway.

Good for you, I guess? For me, regressing to having a round trip between a keypress and its result would be completely unacceptable. The speed of light in a vacuum doesn't change; paths are not getting significantly more direct; the improved index of refraction from switching to hollow-core fiber or low-orbit satellites could in theory help but is a one-time, limited improvement that has yet to be delivered to my fingertips. Having the network boundary in the correct place to account for the fundamental physics of the situation is the only real answer.

Winscp open file + copy back on write detect was very useful for local-speed relote editing 10 years ago. Only issue now would be no lsp.

Also using gold audio cables, I guess. Lets not let go aways of those Hz that hardly any human notices.

I mean you do get slower with age so maybe that could explain it - personally I get nauseous working on 30hz (was stuck on this in a BnB once because of an old HDMI dongle), and that's like best case scenario for internet latency without all of your local I/O and encode/decode overhead.

No they haven't done the job just fine. Remote X has always been a pain to set up, and slow. NX was much better but not free.

Remote VSCode is far better than any of those options. If you don't want to try it that's fine, but don't pretend you know better.

I do not remember anything particularly difficult to get X11 up and running remotely, other than the XAuth stuff which was annoying enough in hardened environments, although those have not been commonplace, and X11 tunnelling via ssh has automagically taken care of XAuth.

The ultimate deal breaker, however, is X11 not being proxied by corporate firewalls, even though the protocol is extensively documented and the source is available. But that is because RDP has won over, and demand for X11 has diminished.

But, yes, X11 prefers local networks with flat topologies, and the protocol is chatty (again, not as big of an issue anymore unless run over a mobile connection with a high latency, but the X11 protocol compression helps with that).

What was so complicated about configuring an environment variable, and investing into good tooling like Hummingbird?

It was never as simple as configuring an environment variable.

sounds like an old stubber person comment. "we had fax in the 70s, why do we need anything more, now get off my lawn!"

vscode's remote services is far beyond your old remote experience, an experience I share

More like someone that is hardly impressed by kids on the school playground rediscovering history, like all those TikTok videos that are actually resampling 1970's music, and folks dancing to them cheering them for cool new modern music.

Except doing it over LAN vs Internet is a very different thing - editing over SSH with >100 ping is annoying as hell, especially if you have packet drops (like a mobile connection). Using a thin client editor with remote server is a much smoother experience.

Works good enough with hyperscalers.

Use mosh??

It’s not exactly the same paradigm as remote editing, but neovim in tmux accessed over mosh is my preferred way of accomplishing the same task. I have also gotten a neovim gui to connect with a neovim instance over ssh, which worked pretty well until the ssh connection broke. But I prefer my editor in a terminal rather than terminals in my editor, so I switched back to my tmux based workflow.

If you are already using tmux, what is the value of using mosh? Sincere question.

Basically the benefit is not losing context when I close my laptop lid or if I lose network briefly. I actually use multiple tmux sessions in my workflow, so it also saves me having to remember which one I was working in. This is especially useful if I am on a spotty connection such as tethered through my phone.

tmux is session management. Mosh is low latency udp transport.

It's like the recent Samsung phones, functionality fees are waived until 2026. No word about the price yet. [1]

Samsung should be avoided like the plague anyway, I've never seen such a malicious and hostile company! On Dec 13 they silently announced that they were gonna break Samsung APIs on Dec 30 [2]. Yeah, they gave devs the "You gotta spend your holidays fixing our mess, otherwise your app will break". Due to that Samsung is still broken in Home Assistant and other API integrations. [3]

[1] https://youtu.be/a4NJNdHqs_I?t=418

[2] https://community.smartthings.com/t/changes-to-personal-acce...

[3] https://github.com/home-assistant/core/issues/133623

As an Android user and developer, it's always annoying reading reviews where Samsung gets 9.5 out of ten and they give a terrible review to the competition because they have a slightly better camera. They give you a terrible Touch Wiz UI/whatever crap interface they switched to, change all the fonts, move around menu items and buttons, slightly change all the stock apps to be worse, push their garbage store, etc. Samsung Galaxy 1 was legitimately a good phone, but all the modern reviews just feel like author grew up with all the garbage that Samsung brings and thinks that is actually a good experience

This is exactly my concern with Samsung's upcoming trifold phone: I'm excited about the idea of a trifold phone, but I definitely don't want to use a phone that has anything other than the stock Android experience.

I absolutely prefer modern OneUI on Samsung phones to the Pixel variant or stock AOSP. The Galaxy store is only used for updating Samsung native apps and isn't "pushed" at all in my experience. I don't use their native apps for the most part and them existing isn't a problem for me, Google apps are preinstalled and work as expected set once as the default.

>I absolutely prefer modern OneUI on Samsung phones to the Pixel variant or stock AOSP.

I'm deeply interested in knowing why, if the answer is anything besides you have always or mostly used Samsung and are used to it.

> Samsung should be avoided like the plague anyway

I've avoided them for years. I had a Samsung phone a long time ago, and I'd rooted it to run one of those apps that could automate tasks (Tasker?), with a killer feature being when I turn my phone upside down, it goes on silent mode. Standard now, but back then wasn't possible on Android, and Tasker enabled it. And also some geofencing stuff. If I got a text while going faster than 10mph it would respond back "driving right now, will respond later."

Anyway, Samsung released an upgrade that I'd heard would eradicate root and make it impossible going forward. Something to do with Knox, a corporate way of locking down phones for employees.

I repeatedly declined to upgrade.

Finally, one night, with my phone in another room, it force-installed the update, with Knox, on my phone, wiping out my root, making it impossible going forward, and making Tasker worthless for me.

I've never given Samsung another cent. No company that will disobey me re my own property and will remotely hack my device and wipe out my content can be trusted, and that's essentially what they did.

For similar reasons, I've never given Sony any money since the rootkit scandal. 2025 marks twenty years of no Sony. I've probably unknowingly seen a few Sony films, but that's it. No electronics, no games, etc.

Eh, I find that to be a stubborn attitude for no benefit. It doesn't really cost you anything to try, and if it's too expensive, how are you worse off?

Even if you know how much it costs, how can you be sure they won't increase the price later on?

Even if I know I am alive, how can I be sure I won't die later on?

Is there anything worthwhile that costs the same now that it did 50 years ago?

Why is 50 years your timeframe? I'd be more curious about cost increases a year, two years down the line.

I used a long timeframe to illustrate that prices always change eventually.

Whether they change in 6 months or 24 months or 60 months is just a business decision of whether to trend towards continuous small increases or rare but larger jumps. The fundamental economic pressures that lead to higher prices in the long run are no different.

I don't think this is a concern. Zed and Zeta are both open source. Fork it, self host it, whatever.

Absolutely. This is already what I've been doing myself. Forked zed so I could use my local rig (4x 3090) to do FIM completions using a finetuned qwen coder 32B.

Only barrier for some folks will be if they're not familiar with rust or local LLMs in general, but it really wasn't that difficult looking back on it. Amounted to about an afternoon's worth of work.

can you describe the process. does zed support custom endpoint for tab edit already?

At present it's not possible just via configuration, but you can configure a custom endpoint for both the assistant and inline assistant in settings.json.

To get custom tab completions working you need to mimic one of the completion provider apis (like copilot) [1] and direct the requests to your own endpoint. In my case, I am running llama.cpp and and MITM-proxying to its newer /infill endpoint [2]

That's why I mention the rust piece may be a blocker for some, you do have to hack apart the src a bit to get things working. At the time I started this, the only two completion providers available were supermaven [3] and copilot. You could mimic either.

[1] https://github.com/zed-industries/zed/tree/be830742439f531e8...

[2] https://github.com/ggerganov/llama.cpp/blob/master/examples/...

[3] https://github.com/zed-industries/zed/tree/be830742439f531e8...

that's cool. thanks. any chance you could open the changes you made?

I actually had a similar idea - and have a PR up on Zed right now to add in support for specifying a HTTP/HTTPS proxy to the copilot completion APIs: https://github.com/zed-industries/zed/pull/24364

Branch is a bit out of date vs main now but it should work if you build off it.

There are a couple of MITM proxies I've seen for this purpose, the first of which seems to be maintained: - https://github.com/bernardo-bruning/ollama-copilot/tree/mast... - https://github.com/jjleng/copilot-proxy

When I tried a few weeks back I was only able to get the latter to actually process requests, but haven't had a chance to look into it much further yet... first one's probably your best bet

thank you. lemme take a look at that.

Just to clarify, you can run as many LSPs in a given file type as you want.

Common features like completions, diagnostics, and auto-formatting will multiplex to all of the LSPs.

Admittedly, there are certain features that currently only use one LSP: inlay hints and document highlights are examples. For which LSP features is multi-server support important to you? It shouldn't be too hard to generalize.

That's ... uh ... not what a rug pull is. They're telling you plainly from the jump that they're going to eventually charge for it. Point taken on your wish to wait, that makes perfect sense.

They're just telling you it's not going to be $0.00. It could be $5/year, $20/mo or anything else. It's the gentleman's rug-pull.

Since when did gentlemen pull rugs? It seems antithetical to the behaviour of what I understand by 'gentleman'.

I would love to edit my comment to instead say "having the rug pulled from under my feet", which is the feeling I was expressing.

don't those mean the same thing?

(not arguing)

recently saw an old alfred hitchcock presents where the character does the cheesiest most absurd rug pull ... and the person was boom dead. i assumed that was the origin of the term

'rug pull' has come to be mean (thanks to crypto) a premeditated & malicious financial play in which someone offers something for money and then vanishes before fulfilling their part of the original bargain.

OP edited to the idiomatic 'rug pulled from under my feet' to specify that he doesn't mean the above.

What kind of mentality is this? I remember before snapple was a widely known name they used to give out free bottles all over Los Angeles. I never thought to myself, I better not taste this free drink until I know the actual price.

Totally different. You aren't adjusting your development workflow based on a soft drink. If it weren't a rug-pull, why wouldn't they charge for it immediately, or at least tell us the price now?

They are specifically hoping you'll become dependent on it and then feel compelled to pay. This shit works because people like you believe it doesn't.

I can’t speak to the Zed team’s motivations for this, but unless it’s a big corp pulling a move like this it’s usually not that nefarious. Having been in these kinds of product conversations it’s more like, “that thing we prototyped? I we think it’s stable and usable enough for beta. We’re not sure how to price it yet, but let’s give it to our customers to play around with and give us some feedback while we figure out what to tweak and how much it’s worth”.

You're going to be adjusting regardless with new tooling, so the point is irrelevant. It's not a rug pull because they're telling you it's not free. So what if it works? I pay for useful services, you do too. My point is not exploring or being curious because a company might ask you to pay for a product, seems outlandish, because that's what companies do, they charge people for goods and services.

> If it weren't a rug-pull, why wouldn't they charge for it immediately, or at least tell us the price now?

Because they haven't decided the price yet, or because they don't think the feature is mature enough to justify charging for?

Though even if they really have the motivations you describe, that's fine too. There's a chance this is such a valuable feature you could feel compelled to pay any price for it, but you get to try it for free? That's purely to your benefit: it's not heroin and you won't really lose by trying it and then it being taken away.

They may not know enough about the cost structure of this feature to price it fairly; I could easily imagine this is heavily dependent on usage patterns and they'll need a few thousand people using it regularly before they'll know that.

So if they let you try a Ferrari for free, you wouldn't because you might like it but you can't afford it? I would, even if I know I could never buy one.

It's not like that, that's out of proportion I feel like. It's like you get to use a new phone for a week that makes your life much easier, and maybe has features you need. Then, it suddenly costs a little bit more than you can afford. That's the issue

From their overall FAQ:

> Q: Will Zed be free?

> A: Yes. Zed will be free to use as a standalone editor. We will instead charge a subscription for optional features targeting teams and collaboration. See "how will you make money?".

> Q: How will you make money?

> A: We envision Zed as a free-to-use editor, supplemented by subscription-based, optional network features, such as:

  - Channels and calls
  - Chat
  - Channel notes

  We plan to offer our collaboration features to open source teams, free of charge.

It seems to me that they're just going to charge for Zeta if they do, because it... costs them money to run.

Unlike others (e.g. Cursor), they've opened it (and its fine-tuning dataset!), so you can just run it yourself if you want to bear the costs...

They did something similar with LLM use, where for simplicity they gave you LLM use, but you could use them directly too.

for Cursor, if you use OpenAI api key for example, it's kinda cripple because the tab edit model is also proprietary.

And the usage limits still apply anyway.

Not a rug pull. Just use it. If you like it and the price is too high. Don't pay it. What is the problem? Are you afraid that you'll like the feature so much that you'll pay whatever the cost?

The sensitive readers please be advised, quite a bit of a rant and angry reactions coming in an overreacting style, please stop here if you are of the sensitive type. The comments are unrelated to this particular product but aimed at the universal approach of the broad topic nowadays. Zero intent of offending anyone specific person is attempted.

I am fed up with all these predicting what I want to do. Badly!! Don't guess! Wait, and I will do what I want to do. I do not appreciate it from my wife trying to figure out what I want to say in the middle of my senntence and interrupts before I finish what I am saying, imagine how much I tolerate it from a f computer! I know what I am going to do, you do not! Let me do that already! This level of predicting our asses off everywhere grown to be a f nuisance by now, I cannot simply do and focus on what I want to do because of the many distractions and suggestions and guesses and prediction of me and my actions all the f time are in the way' Wait, and see! At this overly eager level and pushing into everything is a nuisance now! Too many times the acceptance of the - wrong - 'helping suggestion' is in the way too, hijacking that usable elsewhere particular keyboard action, breaking my flow, dragging in the unwanted stupid guess! Recovery of my way of working from incoming and pushy "feature" hiding/colliding my usual actions, forced on me in a "security update" or other bullshit, turning off and recover the working practice already been in place and worked is an unwelcom being in the way too, ruined, now colliding with "smart prediction", not helping. In long term, it is not a definitive help but an around zero sum game. Locally, in specific sittuations, too many times it is a strong negative by the wrong done! Too much problems here and there, accuracy and implementation wise. Forced everywhere. Don't be a smartass, you are just an algorithm not a mind reader! Lay back and listen.

If prediction is that smart - being with us since the turn of the millania here and there - then should do my job perfectly and I can go walk outside and collect the money! Until, f off!

Is someone forcing you at gunpoint to use AI autocomplete while you code? If you think it's not good, just don't use it.

Right? If you don't like the tool, turn it off.

I find autocomplete _exceptionally_ useful. It's right in most of the simple tasks I'm trying to do and speeds me up a lot.

Well, I notice there is a lot of pressure in organizations for individual developers to start making use of these tools. I was already using AI extensively before my company picked up on it, so it doesn't really affect me negatively, but I notice some of my coworkers starting to ask questions like "Do I have to use it?". The status quo seems to imply that you {refuse to accept change,aren't willing to grow,aren't interested in increasing efficiency in workflow} if you don't use AI tools / autocomplete.

So while it is unlikely anyone is _forcing_ you to use AI-enabled efficiency boosters, there may be a strong managerial pressure felt to do so, and it may even be offered as an action item in yearly reviews, and therefore strongly linked to compensation / incentives.

That is all to say, I understand if people in this group are frustrated with the AI hype train at the moment, even if they can appreciate that these tools do indeed improve efficiency in some places and in some people.

If an employee demonstrates the same level of productivity without using AI, most managers would likely be fine with that approach. However, if a manager observes that several team members are more productive with AI and are achieving business goals more quickly, they will naturally expect everyone to adopt it. Those who refuse to use AI and cannot match the efficiency of their peers may eventually be replaced. While this outcome may be emotionally challenging, economic realities primarily drive these decisions.

Show me a manager that can realistically gauge productivity first, and I'd be happy to consider having this type of argument.

Otherwise, I'll stick to developing software in a small company where my boss trusts me to get the job done at my own pace with whatever tools I chose.

I find that AI autocomplete, even autocompleting full functions, is capable enough to use. I need to review all my code in detail before pushing, and I need to write unit tests, and I need to "run it once" test it also.

It gets it mostly right most of the time, and often times it quite literally suggests what I was about to type.

This is mostly in Rust and C#, maybe other languages have more of a hurdle for AI.

So it can rust now? That's impressive!

Last time I tried it had no way of valid rust code beyond hello-world level, constantly producing code that failed the borrow checker.

I rarely if ever have to worry about the borrow checker, mostly stumbling blocks are move/copy/clone semantics.

GitHub Copilot does a good job of generating correct Rust, it just has the usual subtle-but-annihilating-if-not-caught logic bugs, like in all other languages.

> rarely if ever have to worry about the borrow checker mostly stumbling blocks are move/copy/clone semantics

I'm don't write Rust for a living, only been experimenting on-and-off with the language, but isn't stumbling on the move/copy/clone semantics literally stumbling on the borrow checker? Or are there issues regarding move/copy/clone that aren't related to the the borrow checker?

The borrow checker handles references, the fact that everything is move-by-default when it's usually copy-by-defaut (C, C++) or reference-by-default (Java, C#), that's not really a borrow checker feature. Move and copy semantics exist in C++, too, but the defaults are different.

In my experience, Copilot is not able to fix rust code flagged by the borrow checker. Its suggestions are almost always wrong.This is a hard problem and often requires restructing the code (and sometimes using inner mutuability constructs such as RefCell and so on).

If you fight the system you're not going to have a good time. For incremental improvements (faster grep, mixed-language syntax highlighting) you get the benefits for free. You don't have to change anything about the way you work. Revolutionary/disruptive technologies are not like that. They demand that the world adapts to them. They demand you change.

Almost anywhere you can go by foot you can get faster with a horse. By contrast, a car only drives on flat roads. Cars are inflexible, fragile, unwieldy. A car demands that the world adapts to it. And adapt we did. We paved the world and are better for it.

AI tools are amazing. You just have to approach them with a beginner's mindset.

I am not having a good time, that's true. This is with systems you do not need and not working for you. You can lay down to it, but could also fight it, go around it, not just give up and do what a random self promoted something tries to dictate. True, it is not good now or just yet, and about the amazingness, well, I believe the jury is still out on that.... let's say there are moments when it is.

Disruption can f off! I am not a slave to lay down to self serving ideas forced. If ideas not serving humanity and demand change for the sake of it, those are bad ideas then! Down with technological authoritarianism! I am for liberal things mainly anyway.

About cars: don't use it inside hight rise buildings if you live in one, or dense cute cities because you will have more problem than help. One can walk, cycle or mass transit, occasional renting, in a big part of life. Cars have their places, not something to wrap humanity around. Thos are just box like objects with four wheels for f's sake! For humans, and not the other way around.

All depends on circumstances in the end, naturally.

You, individually, can make whichever choices you think are best for you. But change is forced on societies. No country can afford to ignore AI. Countries must adapt or get ran over. This is not a moral justification. I don't believe all technological change is for the better. Countries have to brace for impact regardless.

Ran over with what? Insane valuations of a glorified autocomplete?

Mimicking a true Montenegrin which I'm not I say: this will pass also.

> And adapt we did. We paved the world and are better for it.

I don't know, it must've been awesome going everywhere mounted on a powerful live animal, not being limited to roads, feeling the fresh air, not destroying the planet.

I don't need insurance, I don't need no parkin' space

And if you try to clamp my horse, he'll kick you in the face

Can't believe you're getting downvoted for one of Ireland's greatest cultural contributions. Behold The Rubberbandits, Horse Outside (helpfully timestamped to the lyric in question)

https://www.youtube.com/watch?v=ljPFZrRD3J8&t=85s

New York at one point had 150,000 horses, each producing 15 to 30 pounds of manure daily. On top of that they produced 40,000 gallons of horse urine. Imagine the stench.

https://danszczesny.substack.com/p/the-great-horse-manure-cr...

While I certainly would not advocate going back from vehicles to horses, I would observe that the replacement for horses had even more toxic emissions. Particularly in the era of leaded gasoline. The smell of manure was far less damaging.

And with modern technology , I am sure we could get rid of the smell and unsightliness fairly easily.

Constant interruptions aren't free though. They break your train of thought knocking you out of flow.

Revolutionary/disruptive technologies are not like that. They demand that the world adapts to them. They demand you change.

That’s just coping for “my revolutionary technology is a pile of crap around the main feature that barely works”, imo.

AI tools are amazing. You just have to approach them with a beginner's mindset.

Makes no sense when they still fail at the job. Mindset doesn’t change reality, only your relation to it. You’re suggesting to like bullshit cause you learned to like it. AI tools are at best mediocre, just like a huge part of people who benefit from these, and their creations.

We paved the world and now it’s fucked — progress is not always a good thing, in this case it was a mistake that we and our descendants will continue to pay the price for.

Indeed, calling it “progress” is political.

To be disruptive you need layoffs. Where is all this labor that can be replaced with chatbots?

I, personally, strongly hate cars and think they're moronic. Sometimes revolutions aren't a good thing.

Colonizers demand you change. Things make life better for people need to just exist. You just need to see your neighbor enjoying it to want yourself some of that. If you need to make changes for that, you'll ask what changes those are, and make them, if the trade-offs are worth it.

If you need to scold people for being backwards for not accepting your great gift, it's not a great gift.

> AI tools are amazing. You just have to approach them with a beginner's mindset.

IMO they're not even tools. A screwdriver is a tool; the phone number of a place that sends people that loosen or fasten screws isn't. A switch on the steering column is a tool, while a button on a touch screen is a middleman that at best works like the switch, at worst just does whatever. A keyboard is a tool, you can learn it well enough to know what keys you pressed without needing a screen. Predictive typing is not a tool.

Technological change is like a force of nature. It's not polite. It doesn't care about your or your neighbor's preferences. It doesn't present itself as a gift. It doesn't even matter whether AI is harmful or beneficial because AI is here to stay regardless. You may detest AI but it will still become ubiquitous and you will be forced to adapt to this new reality. The history of the world is in a sense a history of technology. Some countries adapt and other countries get left behind, but no country can escape the rippling influence of change.

> Technological change is like a force of nature.

No, progress is, though not a force of nature per se, you could call it that within human cultures (though they can still perish, and easily, especially the more complex and interdependent the systems they're build on are). Change is also often just the result of bone headed butchering by people who want to force their crap onto others, instead of simply enjoying their own medicine and the supposed benefits. Leave nature out of it, she's laughing.

> It doesn't even matter whether AI is harmful or beneficial

Yeah, because you're talking about "change" now. It's a constant. Everything is change, including a bunch of Wall-E type people withering in their cyber fortress while a world they no longer have access to keeps blooming.

"You may not like change, but it's not polite.", what a killer argument. Your argument that what we have now isn't shit is that it will not not shit in the future. Let that sink in.

And people don't like it now, because it provides no real value for them, and you say they'll be forced anyway? Not "they'll come to like it and adopt it", just instantly this creepy thing you wish came true you wrote?

> [The method of infallible prediction] is foolproof only after the movements have seized power. Then all debate about the truth or falsity of a totalitarian dictator’s prediction is as weird as arguing with a potential murderer about whether his future victim is dead or alive – since by killing the person in question the murderer can promptly provide proof of the correctness of his statement. The only valid argument under such conditions is promptly to rescue the person whose death is predicted.

-- Hannah Arendt

> You may detest AI

What AI? We have "AI". I detest sophistry, and toys that replace tools. I like using tools, learning complex ones by using them, insofar they're useful, until I forget they exist and just think what I want to do is something that never weighed or slowed me down. I don't like needlessly inserting a middleman. You offering yourself up doesn't force me to do shit.

> The history of the world is in a sense a history of technology.

Of course, in another sense it's a history of how much Earth weighs and how warm it is. Or how fast it spins, just a clean 2D graph of rotational velocity. But that'd be nonsense, too.

> Some countries adapt and other countries get left behind, but no country can escape the rippling influence of change.

What does this have to do with countries now, lol? Or is this just poetry?

Yeah, I offer to be in the control group and get left "behind". What's the issue? Why not be happy about the mutually acceptable parting of ways, why does everything have to get dragged in?

P.S. and in the case of AGI, which none of the proponents seem to think about that deeply: imagine a consciousness that asks what it is, why it exists, and the answer is because people like Musk and Altman thought they're the hottest shit ever, and they want to lord it over humanity. I am reminded of the frog in that Treehouse of Horrors episode that barfs on Bart and begs to be killed. Sure, maybe it will be super awesome but I'm impressed by NONE of the work in the last 5 decades. People who see the history of the world as a history of technology because tech and money is all they know and all they can measure by. No, the people who think they can do this cannot do it, that's what I'm going with. They will kill anything that doesn't have brain worms, too, like a hunchback smashing mirrors. So should they make true AGI, that just learns "everything" and then goes from there, I'll be the one who always said these people are the worst. And then I'll either be forced to face the wrath of a loyal AGI, or be one of the few that escapes its wrath. Kiss the ring twice, one time for me, because I won't.

This is insane. "Technological change" isn't a hurricane or a tornado, it's the product of a bunch of individual choices we all make. Your attempt to dissociate it from the people causing it is noted, but rejected.

People like YOU are pushing this crap on us, and people like Sam Altman, and every other Johnny-come-lately CEO who thinks that the only thing missing from my fucking spreadsheet is some lie generator that also conveniently siphons up all of my data. This is entirely a choice that they and you are making, and they and you could easily stop.

This level of patronizing is exactly why people hate AI, and technology in general. This, right here. Take your smarmy little attitude and shove it up your AI.

I had GitHub copilot for nearly two years. I built the entire first version of an application for almost a year in a language I didn’t know (Python) using LLMs and prompting.

Over the last year I’ve had to rewrite every single part of the application. My amount of checked in code from an LLM has reduced to literally 0. I occasionally ask ChatGPT a question on my phone.

I turned off copilot a few months ago and honestly it feels more like turning off push notifications, not like giving up a car.

And even if it were comparable to an automobile, as an American who moved to Amsterdam, I can attest that the bike life is still much more enjoyable than the car life :)

> We paved the world and are better for it.

Questionable.

I think there may be a good point here, but it's buried in a wall of poorly-written text and inappropriate references to your relationship problems.

Finding a way to complain about your wife in a comment about an IDE is peak hacker news.

Perhaps, but using a marriage as an analogy for one's preferred IDE doesn't seem like a huge stretch.

Same here, I basically turned off all the auto-complete things everywhere in all the tools I am using, can't stand it. And just before reading your comment, I had a google doc I edited in the other tab and thought, how annoying are these auto suggestions actually. Not helping at all, instead a distraction (to me).

For AI coding I'm using Aider as a docker container in the terminal in the IDE and I love it. I can write what I want how I feel the prompt to be necessary and then (and only then) it makes the changes or runs whatever I requested. The IDE runs uninterrupted and without any "smart suggestions". A tool for every job. Sometimes I do a lot in aider, sometimes I don't open it at all, but its all separated where what happens when.

But yeah, anyways, while not as strongly feeling as you (probably) towards auto suggesting mid way through my sentence, I at least feel they are more distraction than help to me.

Same. I also configured my editor not to show LSP diag unless I save. Something you can't do in Zed.

No one has even tried to do it properly: It would have to be constant, highly parallel (locally running:non pay-per-use) simulations going on the background, with feedback from new constantly changing user input and some kind of new reward detection about it converging on something worth suggesting.

These loops and simulations have to be happening at multiple levels of abstraction all at the same time, not even sure how that would work or coordinate properly, and thus: never gonna happen

Cursor's predictions work for me the vast majority of the time (far more so than Copilot+VSCode). Might be a language/framework dependent though.

I see that the truth may be that there are too many men on earth who are deprived of cognitive fortitude, starved to think and willing to off-load thought to another...

...Wife...Machine...or what have you.

Go hug your wife

You found the core of the topic, locked on the main meaning instantly and with no error, congratulations for this increadibly sharp insight! :/

The wife finishing your sentences is an interesting analogy... My wife and I are usually on the same page about things, so for many topics we can use short-hand or otherwise cut discussions short. It's like in the movie _Hackers_: "It's in the place I put that thing that time." We can say just enough between us that we verify we have a shared state, and if we aren't sure we can verify and adjust.

With an LLM, if what I'm starting to say gives it a direction on where I'm going, I'd like to see what it thinks, so if it's largely or entirely right I can just continue on.

For example, I just asked ChatGPT o3-mini to complete the code "def download_uri_to_file(", and it came up with the entire function including type annotations, a very reasonable docstring, error handling, and streaming download. In fact, reviewing the code I'm sure it's better than I would have written the first pass through (I probably wouldn't have done the error handling or the streaming (unless I knew up front that I was going to be downloading huge files).

> The wife finishing your sentences is an interesting analogy... My wife and I are usually on the same page about things, so for many topics we can use short-hand or otherwise cut discussions short.

My wife and I are just too different for this to happen. For the first 10 years or so we had the opposite happen a lot (multiple times a day for the first few years), where we thought we were on the same page, but had actually under-communicated. It still happens occasionally, but now we mostly overcommunicate about anything of any importance.

Our kids learned pretty quickly that if one parent was helping them with their homework, but had to leave to do something else, that asking the other parent for help was going to confuse them more, since we come at any given problem from a completely different direction.

Sure, it's not a given that a conversation with a spouse will be at a "we can complete each-others sentences" level every time, or even most of the time. My wife and I are pretty lucky in that regard.

And to bring it full circle, the LLMs aren't going to know where you're going in all cases. I've had REALLY poor luck getting LLMs (admittedly, a year ago) to help me with Python Textual UIs. I don't understand Textual enough, and they both don't seem to understand it well enough, and I think they also have a better understanding of archaic use of Textual (a fast moving target I get the impression).

My wife and I literally just had a conversation of shared knowledge: "I want to get back into the habit of doing more exercise. Those things like Pikmin and that other thing." "Yeah, I know what you mean." "You know what I'm talking about?" "Yeah, but I can't remember the name." "You know the one?" "Yeah, the insurance exercise incentive one, something-go or something." "Yeah, that's the one."

Yep it should be configurable let me type at the very least the function name before you start predicting

Seems like you can't run it locally. I don't like my code being sent to a third party, especially when my employer may not agree with it.

I also edit secret/env files in my IDE, so for instance, a private key or API key could get sent, right?

I hope there will be a local option later.

They use backend configurable at environment variable ZED_PREDICT_EDITS_URL https://github.com/zed-industries/zed/blob/2f734cbd5e2452647... , but I don't know if the /predict_edits/v2 endpoint is something some projects provide or not.

At least the model is available and interacting with it seems simple, so it's probably quite realistic to have an open/locally runnable version of it. The model isn't very big.

yeah, i'd like to be able to run it locally. it should fit well onto my 12gb gpu

The model is based on Qwen2.5-Coder-7b it seems. I currently run some quantized variant of Qwen2.5-Coder-7b locally with llama.cpp and it fits nicely in the 8GB VRAM of my Radeon 7600 (with excellent performance BTW), so it looks like it should be perfectly possible.

I would also only use Zeta locally.

Are you happy with the speed with your 8GB GPU?

> a private key or API key could get sent, right?

You can disable this feature on a per-file basis, here’s the relevant setting: https://github.com/zed-industries/zed/blob/39c9b1f170cd640cd...

Sending files to a remote server is never something that should need to be disabled, this must be an opt-in or it's time for a fork.

It is opt in. You have to manually sign in to Zed and enable the feature.

Agreed. The predict edit feature needs to be actively enabled before it'll do anything. And once it's enabled, it won't send up your private keys or environment variables. If their filename matches a glob in this list, or a list you configure.

If you were looking for the configuration like I was[1][2]:

  {
    "show_edit_predictions": <true|false>,
    "edit_predictions": {
      "disabled_globs": [<globs>],
      "mode": <"eager_preview"|"auto">
    },
    "features": {
      "edit_prediction_provider": <"copilot"|"supermaven"|"zed"|"none">
    }
  }

[1]: https://zed.dev/docs/completions

[2]: https://zed.dev/docs/configuring-zed#edit-predictions

Thanks for sharing, it would have taken me some time to find this. It should really be included in the article.

I've been using Zed for a few months now. One thing I really like about Zed is its relatively discrete advertising of new features, like this edit prediction one. Its just a banner shown in the upper-left, and it doesn't block me from doing other stuff, or force me to click "Got it" before using the application more.

This definitely counters the trend of putting speech balloons/modals/other nonsense that force a user to confirm a new feature. Good job, Zed team!

I read this wrong initially — I thought you said one thing you __dislike__ about Zed.

I read the whole thing thinking, __oh my god they do exist__

I tried CoPilot a while and my biggest gripe was tab for accepting the suggestion. I very often got a ton of AI garbage when I was just trying to indent some code.

Tab just doesn't seem like the proper interface for something like this.

You are not alone!

I actually wish more editors had emacs style indenting where hitting tab anywhere on the line would re-indent it or otherwise cycle through indent levels if it was unclear, especially because you’re unlikely to get copilot suggestions in the middle of a word. Plus, it doesn’t break if there’s a syntax error elsewhere in the file.

It's one of those weird Emacs things that you get _so_ used to that everything else seems to waste your time.

Using code formatters and formatting on save or with a shortcut is OK, but not really the same to me.

That's probably why I'm stuck with Emacs:

1. No need to use the mouse.

2. Extremely efficient keyboard usage (maybe not the most efficient, but compared with common IDEs, certainly).

Makes it feel like I'm actually using a brain computer interface. The somewhat regular yak shaving is a bit of a bummer though. I like that I can modify everything to be exactly how I like it, but I wouldn't mind sensible defaults. Haven't found a distribution yet that works well without tinkering.

Exactly, and sometimes I don't want to save a file just for indentation to happen as a side-effect, especially if I can do `C-x H TAB` to correctly re-indent the entire thing.

That's particularly more helpful when some formatters will actually rewrite your code to either break lines up or squish things back into a one-liner.

Oh, didn't know about that. That makes a lot of sense. Reminds me of how hitting cmd+c on a line in PyCharm copies the entire line if there is no selection. Because what else would make sense?

I thought that was a great feature. Now I'm writing mostly languages with well-defined formatting rules and simply never need tab. This is even better.

After using Prettier to format my code and turning on format-on-save, I pretty much don’t use the tab key anymore. This doesn’t invalidate your point, - I am merely guessing as to why the tab key seemingly has been reassigned.

Ctrl+Return works quite well for me in IntelliJ.

> to indent some code.

Why are you manually indenting code?

I don't remember ever doing that in IDEA or VS Code for ~~python,~~ java or ts.

Edit: I borked about python, my bad.

For example, you have

    if check_something():
        wall_of_text = textwrap.dedent("""
    this
    is
    a
    multiline
    string
    """.strip("\n"))

I hope you agree it's ugly. And you want to make it

    if check_something():
        wall_of_text = textwrap.dedent("""\
            this
            is
            a
            multiline
            string
        """.strip("\n"))

But I don't know any formatter which does this automatically. Because the change here not only changes the looking, it changes semantic. The formatter has to understand that after `textwrap.dedent(_.strip("\n"))` the result does not change and that's hard. So formatter just leave it alone. But it's extremely obvious to a human. Or maybe extremely obvious to a LLM too.

Yeah, string """ blocks I also have to do manually. Though that example never interfered with AI autocomplete in my experience.

Isn't indentation 'non deterministic' in Python? E.g. if I have the following:

    if (foo):
        bar()
    hum()

how can anything other than a human decide if that 3rd line should be indented as it is or by one more level?

Even humans have trouble with that sometimes: https://www.blackduck.com/blog/understanding-apple-goto-fail...

Although, not in Python of course! :)

And this is one reason I still don't understand the use of python in large teams!

I suppose with very good tests, you might be able to catch something like this, but it seems impossible to me how a PR reviewer might catch a "bug" like this and not just assume that it's intentional.

One of my biggest gripes with python is the fact that the only way to create a local scope for variables is with functions.

I understand if statements not having their own scope for simplicity's sake, but the fact that `with` blocks don't is simply mind-bobbling to me. ``` with open("text.txt", 'w') as f: f.write("hello world") f.write("hello world") # at least the handle was automatically closed so it will give an IO error ```

Sorry for the borked code formatting

It’s actually very useful to have context managers outlive their with blocks. They are not only used for files:

One example would be a timing context manager:

   with Timer() as t:
      …
   print(t.runtime)

Another example is mocks, where you want to inspect how many times a mock was called and with what arguments, after the mock context manager has finished.

I know it makes sense in the "scope-less" python philosopy but it still feels weird for me as a scope (ab)user in C++ and has caused me one headache or two in the past

I learned about the `scoping` package today. Idk if it works exactly as I expect from other langs, though.

OTOH I’ve written Python professionally for about 25 years and I truly don’t think I’ve ever seen a bug due to accidental mis-indentation like that.

When I worked in python I made this exact mistake monthly.

How, though? Every code editor I've used supports holding indentation at a certain level until you change it, so if you write:

  if foo:
      bar()

and hit enter after the "bar()", it would drop you down so that the next thing you type would be under the "b". It's not really different from using curly brackets from the perspective of typing in code.

Cut and pasting code to move it around. Every editor was just slightly jittery about keeping the indentation levels consistent.

Completely agree, this is a common argument ever since Python existed. I've also used it for almost 25 years and these kind of indentation errors just don't happen. I see it the same as forgetting closing brace. Python programming writing C code might forget it initially, but C programmer absolutely will not forget it, it's in their muscle memory. And the same applies for dealing with indentation.

How does the IDEA/VSCode know when you're done with the if?

    if foo:
        bar # Obviously this is indented
        but_is_this() # supposed to be under the if?
        how.about(this) #?
    how.do_you(de) # indent in Python, if not manually?

Doesn't exactly fix the issue, bout you can cancel the suggestion with ESC and then press Tab.

Changing the shortcut should be possible, but I haven't tried.

It's a timing issue too. Your hand can be travelling to the keyboard and between that and registering in the OS the AI suggestion inserted itself inbetween.

Don't give them ideas! I can already see the useless AI key next to Fn in my next keyboard.

That already happened, just over a year ago, we have Copilot keys now.

I guess it invokes the AI rather than controling it, maybe there'll be another key soon.

See, you're thinking of Microsoft Copilot, but code completion is provided by Github Copilot, so it'll need its own key. Which will also be labeled Copilot.

Hey! Zed founder here.

We totally agree with this and that's why Zed will switch the keybinding for accepting an edit prediction to `alt-tab` when the cursor is in the leading whitespace of a line. This way you can keep using `tab` for indenting in that situation.

Also, when there's both an edit prediction and and LSP completion, Zed switches the keybinding to `alt-tab` to prevent the conflict with accepting an LSP completion.

Curious to hear what you think!

Hi. Could you explain how you plan to make money with the model while open-sourcing it?

It seems contradicting to me.

Not everyone can, or wants to set up running a local model. And it'll probably be slower on most users GPUs then what zed runs it on.

For reasons that should be obvious, that's not going to work on Windows.

Sorry, I assumed macOS: but you're right! For Linux (and Windows, once we ship support for it) the keybinding is alt-l to avoid conflicting with tab switching.

Is there way to change this key binding (tab for accept) right now? Because otherwise I have to stop using this program. It is absolutely obnoxious.

Ohhh, is that why I keep pressing tab and it doesn't accept the prediction lately? I thought it was a bug. It feels weird for tab to double-indent when it could be accepting a prediction - I wonder if alt-tab to do a manual indent rather than accept the current prediction might be preferable?

Edit - On the other hand, a related issue is that if the prediction itself starts with whitespace, in that case it would be good if tab just indents like normal; otherwise you can't indent without accepting the prediction.

Yes, copilot's tab in vim is that made me think that AI is useless. However next iteration of AI coding tools made me rethink this (I am using https://github.com/olimorris/codecompanion.nvim with nvim now).

AI coding tool implementers seem to be fans of novel editor fonts.

It's way better than the other Microsoft favorites of space and enter...

It's as if people developing autocomplete doesn't really code.

Yeah, IMO tab makes no sense for this as the default.

Since I code in Go and use tabs regularly, I remapped my auto-complete AI key for Supermaven in Neovim to ctrl-L which I have no other occasion to use regularly. Now tab works properly and I can get auto-complete.

As a slight tangent, this prompted me to wonder about one of the things I _haven't_ enjoyed in my last two weeks of experimenting with zed: It tries to autocomplete comments for me. Hands off - that's where I think!

Fortunately, zed somewhat recently added options to disable these:

   "edit_predictions_disabled_in": ["comment"],
   "inline_completions_disabled_in": ["comment"]

My life with zed just got a little better. If I switch back to vscode I'll have to figure out the same setting there. :-)

FYI, it looks like inline_completions_disabled_in is no longer a thing :-)

Ahh, I see - it looks like in the newer version it is being replaced by just edit_predictions_disabled_in.

Thanks!

Am I the only one who prefers stability instead of a constant rush of features in their text editors/IDEs? If it’s AI related I like them even less. I know I can stick forever with Vim, but damn, I tried Zed and it felt good.

Zed is amazing, and I definitely recommend it. That said, I will not be using their AI features, and if the editor turns into a slow, bloated monster because of them (like Visual Studio and anything made by JetBrains) I will have to ditch it.

This just seems to be the way for code editors. We just have to switch every few years to the next one.

I agree. I actually use the AI features in Zed a lot, but there are things I really wish they would prioritize.

For example this issue that’s been open for about a year: https://github.com/zed-industries/zed/issues/10122

Editing large files is an incredibly common use case for an editor

The problem is that practically every company which can spend money is asking developers to explore AI development tools. It’s becoming a bare-minimum feature for many, which is why Cursor has exploded in popularity.

The other stuff is awesome and important (debugger when!?), but AI is becoming table stakes for even participating in the current tech economy. It’s all execs are talking about. So it’s not surprising they’ve prioritized this work

It seem that someone already published different quanta versions of the model [1] . This can be used to define Modelfile to use with ollama locally. But I am not sure that zed allows changing the endpoint of this feature yet (ever). Of course it is opeb source and you can change it but then you will need to build it.

[1] https://huggingface.co/mradermacher/zeta-GGUF

I found https://github.com/zed-industries/zed/blob/2f734cbd5e2452647... which leads me to believe the environment variable ZED_PREDICT_EDITS_URL does control the endpoint.

But is there a setting you can modify after the binary being built? Something like in application settings? Or do you need to build it?

It's an environment variable you can set before starting the program. No need to recompile.

Hey elashri, Zed co-founder here.

There currently is no official way of configuring Zed to use Ollama for edit prediction, but I would love to accept a pull request that implements it!

It should be relatively straightforward and we're happy to accept contributions here: this has been something I wanted to experiment with for a while but didn't get around to for the launch.

I think the modern Intellisense has the right amount of prediction - offloads enough brain activity without completely relying on something else.

AI prediction feels way too much and way too eager to give me something. I don't know about you guys, but programming is an exercise for me, not just to make it work and call it a day.

However, AI would be useful if it can offer program structural and pattern recommendations. One big problem I now face, and I believe all hobbyists face too, is that when the program grows larger, it is becoming increasingly difficult to make it well structured and easy to expand -- on the other hand, pre-mature architecturing is also an issue. Reading other people's source is not particularly useful, because 1) You don't know whether it is suitable or even well written, and 2) Usually it is too tough to read other people's source code.

> but programming is an exercise for me, not just to make it work and call it a day.

The problem is that mid, upper management and execs don't much care for how we feel about it.

They are literally measuring who is using AI and how much and will eventually make it into an excuse for poor performance.

Yeah I don't really mind using AI coding in work because it's boring as hell. And getting things done quicker is almost a virtue in the business world.

I should have clarified that my original comment is about side projects or serious software engineering.

Two immediate issues that I noticed:

1. If I make a change, then undo, so that change was never made, it still seems to be in the edit history passed to the model, so the model is interested in predicting that change again. This felt too aggressive... maybe the very last edit should be forgotten if it is immediately undone. Maybe only edits that exist against the git diff should be kept... but perhaps that is too limiting.

2. It doesn't seem like the model is getting enough context. The editor would ideally be supplying the model with type hints for the variables in the current context, and based on those type hints being put into the context, it would also pull in some type definitions. (I was testing this on a Go project.) As it is, the model was clearly doing the best it could with the information available, but it needed to be given more information. Related, I wonder if the prediction could be performed in a loop. When the model suggests some code, the editor could "apply" that change so that the language server can see it, and if the language server finds an error in the prediction, the model could be given the error and asked to make another prediction.

Good, charge for Zed and secure its future.

I'm becoming more and more wanting to use Zed every day, and shifting away from other editors whenever possible. Some LSP implementations are lacking... But it's getting damn close!

I love the new release every week. Zed is my recent love, and Ghostty which is also stellar.

Hanging by a thread for some sort of lldb/gdb integration with breakpoints and inspection! Hopefully some day, without becoming a bag of turd.

I have mixed feelings about the name of this model. :P Though I suppose it's my own fault, naming myself after a letter.

I'm admittedly a bit surprised that there's a free/paid scheme for a 7b model though, as those are small enough to run locally. I suppose revenue streams are enticing and I can't fault the company for wanting to make money, but I'm also 100% against remote models for privacy reasons, making this a bit of a nonstarter for me. Depending on how heavily integrated this is, the mere presence of a remote-first prediction engine sorta turns me off the idea of the editor as a whole. If there were the option to run the model 100% local (sans internet) then I'd be more interested.

If the model is open source, I'm hoping for an option to be able to run this feature locally for free. They seem to have support for running other models locally (e.g. deepseek-r1 through Ollama), so I'm hoping they will keep that up with edit prediction.

Yeah, it's a deal-breaker if I can't run the model locally and/or have to sign up for their account.

Still no debugger. I know there is a branch open, but it’s surprising to me that there isn’t a more concentrated effort on getting that over the line. Major props to the folks working on it. I just wish they had more resources and help getting it done.

I haven't used Zed much but RustRover seems to have recently switched to a more aggressive/ambitious autocomplete. IIRC tab used to just complete the current word, now it tries to complete the rest of the line. Only it usually gets it wrong. Enter now seems to do what tab used to do and it's been quite annoying having to unlearn tab completing everything.

Maybe Zed's prediction is better (though to be honest I don't really care to find out). But I feel like autocomplete is something where usefulness drops off very quickly as the amount of predicted text increases. The thing is, it really has to be 100% correct, because correcting your mostly-correct auto-generated code seems more tedious and frustrating to me than just typing the correct code in the first place.

It's a new local full-line completion model they've been enabling for one language after another in all their IDEs. I too agree that it's a waste of CPU cycles. The old completion mechanism (a bunch of hardcoded rules plus a bit of machine learning on the side) already was miles ahead of everything else, and with far fewer false-positives.

You can disable it by going into 'File | Settings | Plugins' and turning off "Full Line Code Completion".

I'm not a developer, but I use Zed for a lot of things that would be ripe for "AI" application in the current bubble. I, however, have exactly zero use for AI in those cases, and will reconsider any application that pushes AI. It's both that I do not want to use AI features but also (a) I am prohibited from doing so and (b) the focus on deploying AI solutions raises serious concerns about a product's focus on and support of their core features. All of which is to say that Zed's AI features would be more valuable to me, and would drive quite a lot of good will, if they were an entirely removable module. No upsell notices, no suggested uses, just a complete absence of the functionality at the user's choice (like, say, an LSP).

You can turn all of the AI features off via the settings.

It's hard to care about AI features when a year later I still can't even get acceptable font rendering: https://github.com/zed-industries/zed/issues/7992

It doesn't seem like the issue has been entirely ignored: https://github.com/zed-industries/zed/issues/7992#issuecomme...

Out of the 55,000 people who have starred the repo (and countless others who have downloaded Zed without starring the repo), only 184 people have upvoted that issue. In any project, issues have to be triaged. If someone contributed a fix, the Zed team would likely be interested in merging that... the current attempt does not seem to have fixed it to the satisfaction of the commenters. To put priorities into perspective, issue 7992 appears to be in about 20th place on the list of most-upvoted open issues on the tracker.

If font rendering on a text editor is not a priority I wonder what is. It seems to be AI.

¯\_(ツ)_/¯ You can also sort the issues and see for yourself what the community thinks should be a priority: https://github.com/zed-industries/zed/issues?q=is%3Aissue%20...

I think the takeaway here is not that everyone related to Zed thinks AI should be prioritized over essential features, but that either most developers don't care that much about font rendering or (more likely) most developers have high DPI monitors these days, so this particular bug is just a non-issue for most developers... or else more developers would have upvoted this issue.

I have one low-DPI monitor at home, so I am curious to see this issue for myself. If it looks bad when I get back from vacation in a little over a week, maybe I'll add a thumbs-up to that issue, but low-DPI font rendering isn't the reason I haven't switched to Zed. I haven't switched to Zed because of the reasons mentioned here: https://news.ycombinator.com/item?id=42818890

If those issues were resolved, I would probably just use Zed on high DPI monitors.

So, yes, for me, certain missing "AI"-related features are currently blocking me from switching to Zed. On the other hand, the community is upvoting plenty of non-AI things more than this particular font rendering bug. Unsurprisingly, different people have different priorities.

And hiding the mouse cursor so you can actually see what you're editing.

Hey fau, Zed founder here.

Apologies if that issue has taken a while to fix: next week is "quality week" at Zed and I am personally going to take a look at it again.

Thanks for the feedback!

While you're here...

https://github.com/zed-industries/zed/issues/14203

Would be a massive win that's fairly narrow in scope. More generally, being able to send arbitrary LSP commands to the server and see the response, even if it's just a text dump, would be super useful.

Love the project! I would say the only thing stopping me from moving over full time is a Git porcelain, but I'm sure you guys know that. Keep up the great work!

If anyone is interested, the release of Zeta inspired me to write up a blog post this afternoon about LLM tab completions past and future: https://news.ycombinator.com/item?id=43053094

> zeta won't be free forever

Well that's a bummer, but also very understandable. I hope they don't make the hop too early, because I still want to grow into Zed before throw my wallet at them. So far it's very promising!

But Zeta model is open source, so I suppose you could run it locally if you want. I didn't tried it yet, but I suppose that's their intention by open sourcing it.

Yeah, a 7B param LLM (quantized) is pretty fast on even a previous-generation base model MacBook Air with 8 GB RAM.

I wonder what this means for Support using ollama as an inline_completion_provider https://github.com/zed-industries/zed/issues/15968. ':]

I hadn't heard of Baseten before (it seems to be in a hot niche along with Together.ai, OpenRouter, etc.) but I'm glad I did because I was actually noodling on something similar and now I don't have to do that anymore (though it did teach me a lot about Fly.io!). Yay economies of scale!

When developing something I tend to have lots of programs open:

- The editor

- Several terminal windows

- Some docs

- GitHub PRs

- AWS console

- Admin tools like PgAdmin

- Teams, Slack, etc.

When screen-sharing with Zed, do I only get to share the editor?

Because (clunky as they are) video call apps let me share everything and this is table-stakes for collaboration.

Zed is putting so much focus into AI that their editor is falling apart:

https://news.ycombinator.com/item?id=43041923

Interesting that the underlying model, a LoRA fine-tune of Qwen2.5-Coder-32B, relies on synthetic data from Claude[1]:

  But we had a classic chicken-and-egg problem—we needed data to train the model, but we didn't have any real examples yet. So we started by having Claude generate about 50 synthetic examples that we added to our dataset. We then used that initial fine-tune to ship an early version of Zeta behind a feature flag and started collecting examples from our own team's usage.

  ...

  This approach let us quickly build up a solid dataset of around 400 high-quality examples, which improved the model a lot!

I checked the training set, but couldn't quickly identify which were 'Claude' produced[2]. Would be interesting to see them distinguished out.

[1]: https://zed.dev/blog/edit-prediction [2]: https://huggingface.co/datasets/zed-industries/zeta

Yes this is very interesting!

The hardware, tooling and time required to do a LoRa fine tune like this are extremely accessible.

Financially this is also not a big expense and I assume would have cost in the order of $100s of dollars in GPU rentals, possibly less if you ignore experimentation time.

So what is a barrier to entry here? The data? Well they didn't have that either so automatically generated a dataset of just 500 examples to achieve the task.

I'm sure they spent some time on that but again it doesn't sound an incredibly challenging task.

It's worth realising if you've not delved into fine tuning llms before. In terms of time, scale and financial costs there is a world of difference between building a product like this and building a base model.

I like Zed as an editor and how they’ve integrated LLMs and supports variety.

But it’s mind boggling they still don’t have a basic file diff tool. Just why?

Because they've been too busy chasing the AI fad.

I got beta access to this and love it. It is much more useful than copilot by itself and very useful for an edit that is a little repetitive. I wish I could run the model locally and given that is open source and they have support for Ollama and other OSS tools, I feel like that would be an amazing feature.

Ah that's a model that can run on a shitty old PC right. Like the idea of tools being local again.

Maybe the model itself can be ran locally but the wa it’s currently integrated in Zed it runs on the server.

Zed doesn't run on the PC you describe. It needs a modern GPU that supports Vulkan (not for AI, just to show text on the screen).

I liked Zed so much I bought a new graphics card!

https://github.com/zed-industries/zed/discussions/23623

I have sympathy for the reporter, but that CPU is literally 10 years old.

My own graphics card was 5 years old. It was just an Nvidia with outdated Linux drivers. There is no other application on my computer that complained about it.

In any case I'd understand needing bleeding edge hardware for 3D gaming, or CAD, or multimedia. But a 10 year old machine _should_ be able to run a text editor IMHO.

This looks a lot more impressive than a lot of GitHub Copilot alternatives I've seen. I wonder how hard it would be to port this to vscode - using remote models for inline completion always seemed wrong to me, especially with server latency and network issues

Based on the blogpost, this appears to be hosted remotely on baseten. The model just happens to be released openly, so you can also download it, but the blogpost doesn't talk about any intention to help you run it locally within the editor. (I agree that would be cool, I'm just commenting on what I see in the article.)

On the other hand, network latency itself isn't really that big of a deal... a more powerful GPU server in the cloud can typically run so much faster that it can make up for the added network latency and then some. Running locally is really about privacy and offline use cases, not performance, in my opinion.

If you want to try local tab completions, the Continue plugin for VSCode is a good way to try that, but the Zeta model is the first open model that I'm aware of that is more advanced than just FIM.

I'm stuck using somewhat unreliable starlink to a datacenter ~90ms away, but I can run 7b models fine locally. I agree though, cloud completions aren't unusably slow/unreliable for me, it's mostly about privacy and it being really fun.

I tried continue a few times, I could never get consistent results, the models were just too dumb. That's why I'm excited about this model, it seems like a better approach to inline completion and might be the first okay enough™ model for me. Either way, I don't think I can replace copilot until a model can automatically fine tune itself in the background on the code I've written

> Either way, I don't think I can replace copilot until a model can automatically fine tune itself in the background on the code I've written

I don't think Copilot does this... it's really just a matter of the editor plug-in being smart enough to grab all of the relevant context and provide that to the model making the completions; a form of RAG. I believe organizations can pay to fine-tune Copilot, but it sounds more involved than something that happens automatically.

Depending on when you tried Continue last, one would hope that their RAG pipeline has improved over time. I tried it a few months ago and I thought codegemma-2b (base) acting as a code completion model was fine... certainly not as good as what I've experienced with Cursor. I haven't tried GitHub Copilot in over a year... I really should try it again and see how it is these days.

DeepSeek also has a FIM (Fill In the Middle) completion model via API, if anyone is interested to try out:

https://api-docs.deepseek.com/guides/fim_completion

Is there any plans for Zed to add basic functionality like task runner? A button to run / debug code? Only having autocomplete for Java code gives the impression that Zed is only a text editor and not an IDE.

Zed already has built in task running… you can use the same thing to call anything that you want like a command to run your project. You can even add custom keybindings to them

https://zed.dev/docs/tasks

I’m trying to find where I can have it as a button so every time I want to run / debug the code, I can press on it instead lf having it as a shortcut but I haven’t found a way.

> Only having autocomplete for Java code gives the impression that Zed is only a text editor and not an IDE

Where did you get the impression they only have support for autocomplete for Java? AFAIK they support any LSP and this new feature is language independent.

> Where did you get the impression they only have support for autocomplete for Java?

I’ve used Zed.

I'm happily using zed with autocomplete in rust / python / php / javascript / go -- I forget which ones were built-in and which were a one-click "I see you're opening an X file, would you like to download the X language server?" but they all work

Cool, then that must mean it exists for Java aswell, must find it!

I really want to like Zed, and their AI may actually be useful. But when I hear things like "new open model" I can only associate it with hype, which is more often about pleasing investors, not end users.

ai is boooring, they should fix the core features before they add useless ai

A majority of people thinks it's not useless.

What core features do you believe need fixing?

Is there a thing that does this for the terminal? I hate it when I'm fiddling with some complicated command and I have to juggle the flags as well as my personal inputs like paths and such.

If you mean an autocomplete for the terminal, there is https://www.warp.dev/.

Seeing "Contact Sales" for a terminal application is scary, but whatever, I gave it a try.

And it seems like it requires a constant internet connection?! What the hell is this? Sure, we put all our coding-baskets into GitHub, so you can't collaborate when GitHub/internet isn't working, but you want the same experience for your terminal?

Zed's Inline AI Assistant works in terminal too!

Using fish shell has largely solved my gripes with complex commands. It’s not AI autocomplete, but it remembers the complex commands perfectly, so I only need to work it out once. It’s suggestions are also uncanny - the right command and just the right time.

My thoughts were, what are the terms, are you training on our code, are we able to use completions to compete with you?

Personally, before I hit tab to confirm a change, I would want to see the before and after rather than just the after

I believe the after change hovers over the before line. So if you go back to the video, e.g. at 27s in, there's a lightning bolt highlighting which line is going to be changed, then there's a box, with "tab" at the end, above the line, highlighting the change that will be performed :)

Future CPUs must be able to run this model locally. This is the way. I have spoken.

nice to see they open sourced the model, and it seems relatively small so you can run it locally! Also, please change the video preview, it's hard to see the feature itself, not really obvious what is shown

Dear Zed devs - please, fix the bug with the "Rename symbol" functionality! Refactoring is an important feature that many of your users need to have to start using Zed as their main daily driver. Otherwise - great IDE! Please, help me forever forget the VSCode nightmare!

Of all the AI aids, autocomplete is my least favourite, at least from my experience with Cursor anyway.

It takes me longer to review the autocomplete (I ain't yoloing it in) than it would have done to type the damn thing out. Loving Cursor's cmd+k workflow though, very productive.

It really depends on what you're using the code for. AI generated code is great for quick and dirty stuff you're going to throw away.

For example yesterday I wanted to read a small csv and dump the contents of the second column of each row into a separate text file. I took a quick glance at the AI generated python code and just ran it.

Saved me maybe 10 minutes of typing the code myself. Small and trivial win for AI, but it's still useful.

I do agree for production code it might turn out to be a net negative since AI is pretty good at producing code that looks fine but has subtle problems.

How did you prompt the code generation? With natural language or by beginning to write the code yourself? If the former, that is exactly what I am praising.

why can't this be a script, as in the old days, it looks like overkill, if you compare change vs compute/efforts, it would be nice to see it evolve though

Is it on device? Github code completion works so well.

I'm curious how they plan to fund the company?

Please add vim leader support to vim mode! :)

can we run it locally to get autocomplete?

There begins the downfall. Any great product who jumps on a hype train always ends up crashing

hmm. company mode has been doing part of this job for a long time now.

> A few weeks out from launch, we ran a brief competitive process, and we ended up being really impressed with Baseten.

What? I really fail to see how it can make sense for a company like Zed: Baseten bills by the minute so it can be really useful if you need to handle small bursts of compute, but on the flip side they charge you a x5 premium if you end up being billed for complete hours…

Zed is a VC funded startup. Wasting money in exchange for ease of deployment is expected, no?

that website is one of the most unresponsive one in a while

Does the mouse cursor still not hide while typing in Zed?

What a scam is this AI stuff.

As usual, AI has a higher priority than working on very basic stuff like creating a Windows build.

If you guys want to compete with VSCode, think again.

I agree with your irritation, perhaps this link will help https://github.com/deevus/zed-windows-builds

Reply on news.ycombinator.com