I really don't like this. The code I write between commits is my thinking. I think by writing some code out, deleting it, writing again. The code I write that's shipped in commits is written for others to understand, and is a product of that writing for thinking process.
I don't want my thoughts to be serialized, version controlled and publicly accessible.
This was my problem with JJ as well.
I don't want everything in between to be versioned. I'm not even sure, every intermediary state between a commit is relevant or useful.
This is why I use rebase before PRs, and despise squash. You are not going to remember why you wrote that code that way 2 years from now and all we'll have to understand bugs and identify Chesterton's Fence situations is the deltas and the commit history. If you squash them I have 400 lines of code you 'wrote' all at the same time and only have the feature request it was assigned to as context. Thanks for nothing.
The worst actor would write a new module and check nothing in until it kinda worked. I think it went along with the fragile ego that had people sneaking around fixing bugs in his code without talking to him about it first. He wrote convoluted code that exhibited Kernighan's law and he was about 10 years too senior to still be doing that shit. He bragged about how 'powerful' his code was as if that was a compliment instead of a harbinger. Many times I found bugs in code from the initial commit. Just... give me something man. Anything. Fuck.
Just because you tried random shit until you found the problem doesn't mean you have to fess up to it. You can tell any story you want that gets us from point A to point B now that you know point B is attainable. You can rearrange the commits the way you would have written it if you knew exactly what needed to be done. Drop 90% of the code you wrote and immediately deleted again, anything that doesn't support that narrative.
In law enforcement you have something called Parallel Construction. You can know a suspect is guilty by knowing facts that are not admissible in court. So you need to rediscover those same facts by the book. Grab his trash on trash day. Interview neighbors. Get enough circumstantial evidence to get a search warrant, then go find that evidence again.
If it's part of the rebase. If commit 3 in the chain doesn't compile and half the tests fail, that's not very useful for someone trying to figure out how an ancient bug nobody noticed for a year got into the code. If commit 4 fixes it, then 3+4 -> 3.
But if you add a function in commit 2, then realize three hours later than you just need to call another function, I don't need to see the function arrive, get used, half the tests written, the whole thing be deleted, and then a couple commits that do it the easy way and only needs 3 tests added.
That fakeout will be in the code tree forever and each new dev who is trying to grok the code will see that the line of whitespace on line 35 came along with a function that doesn't exist anymore and only really ever did on David's machine.
Probably coulda used an example that isn't itself a fourth amendment violation that essentially requires perjury to accomplish. Also less euphemistically called evidence laundering. Not really a neutral example.
A more charitable case is that the source cannot be disclosed because it's an undercover agent or informant. What the parent describes is indeed evidence laundering.
you can see that every file has a code block header with a UUID and the AI that was attributed to it. With the UUID, I can tell exactly how the code came about.
What they are working on will be more useful for AI code provenance. It is only a matter of time before you are expected to show your chats with AI as part of the code review and for performance reviews.
So I don't see human collaboration being the main use case. I see tracking, studying and improving the Human-AI relationship...and seeing if somebody should be promoted or not.
An interesting take I've heard is, we will have a token/impact stat where if you spent a shitload of tokens to produce the same impact as somebody else who spent a lot less, you will be the prime candidate for layoffs and/or less pay. This is why I think AI code provenance will become a serious thing in the future.
Audit purposes for sure. How was this code/concept generated, what were the prompts/requirements, what thinking did the model complete, can this be replicated or repeated, etc.
A vendor conference I was at a few weeks ago focused heavily on this, for most of their Agentic workflow items. How can you show the AIs work, prove what it did was within guidelines, then audit the process and result.
Like, if your system has an AI backed federated search for documents and you ask it a question about those documents, you need an audit trail of the ask, what documents were referenced, and what was returned to the user.
Then if wrong information was supplied that can be pinned down and explained in case of lawsuit or other need.
Fully agree, very icky surveillance vibes. In particular:
> DeltaDB breaks your work into a stream of fine-grained deltas. Where Git captures a snapshot at each commit, DeltaDB captures every operation in between and gives each one a stable identity.
I was curious about giving Zed a try, now that it has an emacs keymap. Not anymore. This is such a horribly invasive feature, I absolutely do not want my colleagues reviewing every single intermediate edit, down to the keystroke, that went into the commits I publish for review.
Before I put a PR up for review, I'll sometimes edit my commit history a little bit in magit to make it more linear and digestible--maybe update descriptions, squash some adjacent commits together, etc. This just throws that whole aspect of the job out the window and says "hey, colleague, hoover up this firehose of deltas and enjoy it".
And what the hell does this even mean?
> What we're really after is simple: the conversation with the agent becomes the only conversation you need to have.
The more I think about it and your comment the more I wish it was local only. It could be useful to analyze your editing habits and interactions with AI but I want that for my own benefit not random coworkers
Yeah I really would not want all my convos synced up like this. I've been working on a personal version of something like this but it keeps all your conversations and notes in a separate Git repo that you control that's totally separate from the project repo - it's entirely designed for personal use:
I would probably only use these data if I had some script I could run in it for analysis, and even then optimizing my AI chats is probably pretty low value.. certainly not enough to justify the cost of analysis, unless there's something that provides actionable insights with almost zero time investment.
Producing code edits itself is such a small fraction of the job I'm really not convinced these are valuable data. Mostly the effort and time goes into reading and understanding code, not producing or transforming it.
No, you are paid to deliver. Whether you do that by thinking + hand-coding or just vibe-coding, or handing the task description to Cthulhu and waiting for him to materialize the solution on your disk, is immaterial.
Unless of course you also have to explain your thinking and problem-solving process in meetings, which happens quite a lot the more senior you get.
Everyone probably has thought "what was the person thinking when they wrote this". Now you know that they probably didn't think (and since 2025 or so, it might not even have been a person).
Don't be afraid to show your thoughts when asked to. The best developers are those that can express their thoughts clearly at any stage throughout their process. This is one of the skills that shows to me the level of experience a developer has.
Nobody needs to know quite how messy the process of making the sausage is though. There are steps that provide information, even about how the bugs got in there. But not every thought needs to be expressed.
One of my professors in undergrad said: the most dangerous mathematicians are the ones that begin the proof with "Consider a case ...". He said that these mathematicians are the ones who don't share anything about how they got the case and they end up projecting this sort of "magician aura". I don't know how accurate that assessment it, but I think it captures something that never sat well with me.
In my life, I've never liked people who deliberately polish up their articulation to the level that it obfuscates how they arrived at that understanding (whether it's academics or engineers). They might not do it for attention and they might not be doing it knowingly. IMO, they are taking away the opportunity of learning from the people they are talking to. For me the conversation is one sided. I'm there to listen, but rarely can I ask questions, give feedback or grow from where they have possibly reached.
Ugh, let's take a step back and make a distinction:
I don't need your fluff. No one cares how you arrived writing another crud line to save an object to database or sent yet another AJAX call.
If you wrote some genuine great compression algorithm that's a different take on compression, I would like to see step by step reasoning and eventual dead ends.
I shared my thoughts in the context of someone saying that one should be able to share your line of thinking when asked to. Whether "when one asked to" applies to keystroke by keystroke blockhained versioning isn't my point of discussion.
I get it, that the overall discussion is about DeltaDB. I'd say interesting concept to toy with. I'd pay more attention to "micro commits" as the idea more than the keylogger.
I think a good argument ad absurdum for this is to look at how some recipe sites give the entire genealogical history of the author and an anecdote about how their gammy met Theodore Roosevelt and he stole her pen. Three pages later I discover I need to go to the grocery store because the recipe requires sour cream. And the store is closed so I need a different recipe.
Don't fucking do that. Do something way less than halfway to that line.
I personally stumble upon many topics where I only care about the what.
In that case all the theory is just a distraction I'm just wading through to get to the point.
If it's optional, then looking into the how and why is certainly nice, but it should be part of an appendix or a commentary and not interspersed within the proof unless an uncommented version exists.
"So I was thinking about #####'s Law this morning, and I realized that #######'s Theorem might apply if we do this and ignore that. Then I went up a blind alley and stopped for coffee and realized I was overcomplicating it, we don't care about negative or imaginary solutions."
They don't need to know I was brushing my teeth and thinking of bacon and an argument I had with my spouse right before I thought that though. Or how rude the customer in front of me was to the new barista who was just trying her best.
OTOH polishing up the presentation can really improve the experience of a first-time reader of the work (e.g. your code reviewers). If the polishing is done with good intent and proficiency you can make something that was very convoluted and difficult to arrive at digestible with far less effort. It also aids your own understanding: "If we can reduce it to the freshman level that means we really understand it" (or similar I didn't look up the exact quote, attributed to Feynman). If you're polishing something up to make it understandable that's totally different than polishing it up to make yourself seem smart, right?
All the things in between my commits is a messy soup. Looking there is not useful to anyone. I rewrite my history with git rebase so each commit is small and atomic. The story I create with my commits is what explains why things are as they are, it doesn't matter if it's the true chronological story on how it actually happened.
I agree with the author that reviewing pull requests is too late. The problem with pull request is that they make it hard to review individual commits since they are geared towards reviewing the result of an entire branch at once. But the answer is not to share all the noise, it should be to encourage small atomic commits so you can review the early work before the entire feature/fix is complete.
This just sounds to me like "frequent auto-commits" with less trust in git. git can handle frequent auto-commits just fine. If you want to "rollup" frequent auto-commits into "cleaner" top-level commits but also keep all the point-in-time "conversation" of your auto-commits, then `git merge --no-ff` from time to time and use tools like `--first-parent` to focus in on "top-level" commits over "conversation" commits.
The git backend already has a ton of "delta DB optimizations" (in git packs and other tools) and it's really just the git frontend that needs a little massaging (`--first-parent`, primarily) and the vast world of "subway-diagram-first/only" Git UIs that should have more "drill down" `--first-parent` counterparts simply because so many people find the subway diagrams ugly/confusing/distasteful.
A software team’s job is to collaboratively learn an effective model for operating in a domain. They express that model and those learnings in code, tests and associated documentation. So on the one hand I wholeheartedly agree that pull requests and code reviews fatally undermine this process, but immediately recoil that we’re creating yet more secondary processes and artefacts to distract ourselves. This stuff should all be evident from your codebase. It’s not an extra thing. It’s not a bunch of commit messages or ADRs. If your codebase isn’t entirely self explanatory (both the what and the why) to both humans and AIs you’ve failed and will spend your whole life creating more and more process to manage that failure.
I have an uneasy feeling in my stomach because i know anthropic or openai acquiring zed is inevitable. They have too many good ideas and their software is too good.
Ya, their coding harness is way better than Claude code, but because it’s directly using the clause api it’s way more expensive. Rolling it into the family would make it product-class-defining.
> I have an uneasy feeling in my stomach because i know anthropic or openai acquiring zed is inevitable. They have too many good ideas and their software is too good.
Why stop at zed? The trillion dollar investment AI companies have amassed was nominally for datacenters, but as those costs rise and completion timelines extend past the typical business planning horizon, it becomes more efficient to put the money to work elsewhere. You can buy whatever you want with a trillion dollars.
There are so many early-stage startups also competing in this space right now. I’ve been on the interview circuit the past few weeks and talked to at least two. It’s going to be stiff competition for any of these tools to get well-established enough to be successful at a large scale.
I can’t help but feel like it is all enabling a level of developer surveillance with which I am deeply uncomfortable, though.
Where are these developers who are willing to talk to you about your code and hash out ideas? Everyone I’ve ever worked with is always too busy, too many other priorities. Sounds nice though.
Unrelated: I started using Zed a few weeks ago and really love it, it is everything it promises to be.
A commit is useful because you cleaned it up first. The messing around in between is where you try things and delete the dead ends and most of it is meant to be thrown away. Saving every change and every agent message keeps all that junk around instead.
Google has been doing this for maybe a decade now with citc [0]. I don't know when Gemini is actually going to be taking advantage of this, but I do know that google has essentially a full history at "Ctrl-S" granularity, from ~every developer that works there, for at least 10 years now.
If Gemini seems stupid nowadays, it's only because they're being stingy with compute allocation.
If that works for someone, great. I would never pay for it, and would run not walk away from any shop that asked me to use it. The mess in-between working software is not something I care to have version controlled.
Zed are doing a great job at making the exact things developers don’t want. It’s clear that their focus is on raising funding rather than making a great text editor.
I don’t see the value proposition here. I’ve seen roughly this feature proposed by multiple companies, and absolutely none of the have given a convincing reason for the technology to exist.
I'd have to see the implementation in the editor to really know if I'd want it or not, but my initial impression is that I don't think it's something I would use. I've worked on a couple of products with in-editor collaboration built in (not text editors, but, content creation tools). For the most part collaboration just wasn't something users wanted that much. The problem was that the in-editor collaboration was only really useful for things that were 100% on the web, it never caught on with the more file-centric workflows. It works in something like Google Docs because it's natural that your data is in the cloud there, but for something that lives on your filesystem it feels like it takes extra steps that most people aren't going to be willing to do (especially if they have to use an editor different from their normal one)
Many years ago I helped a friend with some draft that had to do with him finishing his doctorate. It ended up being maybe 50-60 between three of us and as fine grained as possible.
That agents need something "beyond git" is lost on me but it keeps coming up. For one subject the tool doesn't matter -- in fact it can be obtuse like git or bash and it's fine because agents will handle it. Then for another thing the story is opposite.
The interesting question is whether agents actually benefit from finer-grained history or whether they're better served by semantic checkpoints – "this works, this doesn't" rather than line-by-line diffs. Git wasn't designed for that, but neither is anything else yet.
Like most/all engineers (I assume), I have entire notebooks that are filled front to back and in the margins with scratch. Random calculations and notes, equations running into each other, any piece of information that was too fleeting to stay in my head or overspill from an already full working memory.
Most engineers I know keep such notebooks when they're full, because it's kind of neat to open a random page and see a brain dump scrawled out. Nobody has ever looked at another engineer's scratch notes and drawn useful insight. Frankly, I wouldn't even want to other than for novelty.
It sounds like they ship a fairly minimal and efficient set of tools or text editor. This would be unwanted bloat by those choosing them for those advantages.
so i think the thing that everyone building these git alternatives is missing is a multi-repo story - unless the expectation is that everyone is going to start operating out of monorepos
i've settled on all of this context attached to issues in a project management system and referenced from commits
it works just fine - its not like your agent cannot read your issue tracker
I've built some skills to help work with multiple repos, but it's really annoying how e.g. repo-specific .claude/ configs are only read when you start the agent in the repo folder. There's a ton of low hanging fruit to improve dev experience.
Like others, I really like Zed as a text editor. I was a Sublime user before but I switched to Zed because it was even faster and cleaner than Sublime. The multi-buffer editing is a nice step up of multi-cursor editing. I just want a fast text editor with a few conveniences for development.
I don't care about any of the collaboration or AI tools that they are working on and I hope that these efforts don't detract from the text editor. At the very least I hope I can hide these features in the UI and continue to ignore them. I also hope that these features don't lead to a bloated, sluggish text editor like VSCode and IntelliJ.
The approach sounds interesting but perhaps still too focused on current development paradigms.
Where shared collaboration I think becomes much more interesting is when agents start behaving like a member of your team. Not someone you delegate to, but someone that actively seeks out conversations and interacts with others as though they're just another member of the team.
You shouldn't be aware that version control is happening behind the scenes, you're just interacting with a persistent persona through Slack/teams/notion/figma/linear/etc with something that's logging issues, fixing bugs, writing PRDs, designing mockups, etc. The only point at which you know there's version control is when you're asked to review something before it's integrated.
I would be interested in a clear statement about how this scales. I've not used this workflow myself, but I've seen teams that did it. Whether they got huge benefits out of it I don't know, but I do know that watching them, I was not jealous of what I saw. If I make a change, and I run some tests that were passing a moment ago, and they fail, and the reason why they failed is that Bob hit "save" on his editor (or his editor autosaves) and he made a syntax error in a shared library, and this happened often... I would go insane. I cause enough problems for myself without other people's problems actively intruding at uncontrolled times into my tests.
AI's code writing velocity makes this even worse, there's no way I can be simultaneously working on a code base while an AI agent is running around it doing something else.
It feels like maybe there's a ghost of an idea here about how to get the best of both worlds, but I'm not sure I follow the throughline on it.
This seems like such a weird article. Software is made at build time, which is usually due to a cicd process in place, which is dependent on committing code. It works. There’s no magical way of avoiding the idea that code usually exists in a commit someplace without proper justification.
This seems like a great way to facilitate data gathering for improving LLMs coding performance.
If previously you needed to take action 1, 2, 3 to go from state A to B, all you saw was the change from A, B. Now you see intermediates 1, 2, 3 and can train the models to skip straight to B with the added context of the intermediate states.
Well shoot, they beat me to the punch. I’d been circling around something like this, just not collaborative and obviously more thought out than my random experiments. Minus the collab portions I’m interested to see how it compares to jujutsu
I can see how this is a great building block for what Zed is doing around collaboration.
One thing I would really love to see, however, is a way to better attach code review comments to the specific version of code they were left on. I find it quite difficult to do with git and github, considering that commit hashes change every time one is forced to rebase (say, for example, to handle a merge conflict).
Do you expect DeltaDB to help address this problem?
I guess this partly explains why Zed (still) lacks a PR review flow, let alone a coherent one, despite some interest [1]. Pretty much the only reason I’m still with JetBrains.
This is what I’ve been thinking of ever since coming across Ramp’s agent[1] and other orgs trying to build their own remote dev envs with agents at the center. Makes so much more sense in this kind of a dev spree to not wait for a PR. Because all actors end up repeating the same thing with their own agents which can be totally avoided. Conversations and code to jump in and out of, to collaborate etc seems like the right way to go.
What is apparent to me is that we are moving towards dark factories if the promise of LLMs writing most of the code is fulfilled. So this means that it is less about conversations and it is more about iteration.
Just a stream of thoughts: if git commits were a list of sequential primitive changes instead of diff snapshots, conflict resolution would be trivial in most cases.
Not without cons of course: commit byte size, public WIP work and leaked secrets/unwanted edits.
I hate software tools now. I really do. A hammer would never ask you to think about it constantly. If you think about your hammer it’s because something is wrong with it.
It's not just tools. Pretty much all software is like that.
The problem is, is that it works, if you assume "working" means the software sellers get wealthy.
There's a reason that most waitstaff wear black. They should blend into the background, and not be what the folks at the table are talking about. In rare instances, restaurants exist, where the waitstaff is the service.
In software, though, you're being served by a waiter wearing a clown suit, screaming slogans at you, and serving you lukewarm, pre-chewed goo.
I use OpenBSD as a daily driver (but could use Alpine or VoidLinux too) and my setup is pretty much silent. No notifications, no rainbows of colors, no glitz. Let’s take mail. I use a combination of mutt hto directly connect via imap) and fdm/mu4e (to have them locally). I”m not interested in having counters or notifications for any of those.
The “calm technology” book has an handful of advices, but one of the best example is the xbiff program. It switches picture when you have new mail on your local spool.
I totally see where they are coming from, jitsu, too is making every change its own artifact. And it plays hard to the "faster shipping" that especially AI-driven teams are pushing hard, and find the review process to be the next big bottleneck now (as I just saw with my last client as well).
There has also been a lot of discussion about the value of the peer review process recently, in in general. But I wonder if this isn't all going into the wrong direction. Quite honestly, even with the previous review and discussion system of (squashed) commits how often did you really use git blame and opened up the original PR discussion of that changed line that caused the bug? And how often did it help you beyond learning it was done by that rockstar developer who has left long ago? And that the discussion on that PR was a point in time and the code around it has evolves beyond that and it would need looking at another 10 PR discussion to get the entire context.
What I am saying is that git (and before that Subversion and CVS) has a full history is so that it can resolve the latest state. Period. We made that commit ritual somewhat of a hallmark in putting more supposed meaning into recording ever more in that history. But we rarely stopped to check if that is all that useful. Recording even more, all the time, reminds me of these work group meetings that have minute records of every bike shedding meeting word said by everyone, that, honestly, no one ever looks at after the next meeting ever again. I don't think there is value in minute record keeping, it becomes too much noise that just makes it harder to parse. Now also adding all AI conversation and agent thinking to that tree? What's the value of that in like 3 months down the line? I don't see it.
I'm glad to see this feature and looking forward to see how it evolves.
Many of the product decisions that Zed's made caused me to switch to Zed for my daily driver IDE (previously JetBrains). The recent AI agent threads and improvements around diffs really solidified the move.
I'll probably get more hate for saying this but fine: I use Zed 50% of the time (the other 50% dedicated to vim) for two reasons:
1. It is fast and snappy. Nothing comes even close besides vim (and I don't mind going full time to it if I have to)
2. The ability to completely shut off and block any slop machine features from interfering with my workflow or leak code back to sloppenai, sloppus or any other self-installed-worst-security-practice-backdoor garbage.
Having said that, I hope they don't remove that ability in the future and enforce the "slop is so good man, you should try it" philosophy.
I am happy about even though I've never tried gram, because if zed goes to shit there will be an alternative, which hopefully pressures zed to stay sane
> I also object to making myself and my work depend on paying a subscription fee to anyone. I don't want an outage at Anthropic to affect my ability to do my work. I think it is a grave mistake to build anything on such shaky foundations as the sustainability and profit margins of the AI industry.
That’s not the premise people are building on though.
Any particular model almost doesn’t matter at this point. Harnesses are built around them. OpenAI and Anthropic are basically interchangeable in an open-source harness like OpenCode; the switching cost is virtually zero. Local models are improving rapidly and are already “good enough” for many use cases already. The bet that LLMs will continue to exist as an algorithm is pretty solid.
Oh, that's a breath of fresh air. And they are on codeberg. Nice! Thank you!
Edit: After further inspection, I think I'm jumping ship before it's too late. And I'll look, see if there's a way to lend a hand or two when I have time!
> Before agents, it was easier to believe that the ceremony of trading comments on snapshots was an effective way to collaborate on software,
I’m highly skeptical of this claim. For any complicated feature, there’s always a design doc (or an RFC, or a wireframe) and that’s what people used for discussion. Discussion in a PR are mostly about whether to accept the code, reject the feature, or provide feedback about alternate implementations. It’s not for pair programming or directing design.
Collaborating together in a research lab (brainstorm session) is not the same as asking feedback for a journal article (PR). What is described in the article is pair programming with extra steps.
I really like Zed. It's customisable enough for me to make it look how I want, it's faster than every other editor I've tried (scrolling is silk, zero lag anywhere), it has enough features that I don't need an IDE (debugger, refactoring tools), and it generally gets out of my way.
I also like the AI tools, the inline assistant is good and the agent is also pretty nice and well integrated into the editor without it being the focus point. I'm not against using AI but I certainly don't use it as much as a lot of people do.
That being said, I really dislike this recent push towards becoming more like a cursor wannabe. They have a new (for now) opt-in default layout that almost hides the editor panel in favour of the agent threads and agent panels. And now this. I don't want to switch editors, but if they keep pushing a different workflow from what I use it might send me back to Jetbrains...
Next up: a kit to wire up a developer's brain activity direct to the AI companies, so not a single part of their thought process remains in control of the developer, in order to feed the models. You must brainsignalmaxx 8 hours a day or be fired.
Your boss will also demand that you narrate your thought process 8 hours a day. Speakmaxxing will be required to keep your job. There will be a "silence" wall of shame for the non-compliers.
With LLMs now being responsible for the physical typing of code and mundane plumbing tasks, this is a wise direction to go into
Our human ability is not defined by our _absolute_ output, but, by the quality of the _delta_ applied to an engineering artefact
Great engineers obsess over every keystroke
With LLMs, a much smaller number of keystrokes can create a much larger and more positively impactful delta
Every delta to the codebase can tell us some informational property about the behaviour of the system and storing that information WILL prove to be useful in the future
I don't want my thoughts to be serialized, version controlled and publicly accessible.
https://www.nature.com/articles/s44222-025-00323-4
but feels like, I'm in the minority.
The worst actor would write a new module and check nothing in until it kinda worked. I think it went along with the fragile ego that had people sneaking around fixing bugs in his code without talking to him about it first. He wrote convoluted code that exhibited Kernighan's law and he was about 10 years too senior to still be doing that shit. He bragged about how 'powerful' his code was as if that was a compliment instead of a harbinger. Many times I found bugs in code from the initial commit. Just... give me something man. Anything. Fuck.
Just because you tried random shit until you found the problem doesn't mean you have to fess up to it. You can tell any story you want that gets us from point A to point B now that you know point B is attainable. You can rearrange the commits the way you would have written it if you knew exactly what needed to be done. Drop 90% of the code you wrote and immediately deleted again, anything that doesn't support that narrative.
In law enforcement you have something called Parallel Construction. You can know a suspect is guilty by knowing facts that are not admissible in court. So you need to rediscover those same facts by the book. Grab his trash on trash day. Interview neighbors. Get enough circumstantial evidence to get a search warrant, then go find that evidence again.
But if you add a function in commit 2, then realize three hours later than you just need to call another function, I don't need to see the function arrive, get used, half the tests written, the whole thing be deleted, and then a couple commits that do it the easy way and only needs 3 tests added.
That fakeout will be in the code tree forever and each new dev who is trying to grok the code will see that the line of whitespace on line 35 came along with a function that doesn't exist anymore and only really ever did on David's machine.
https://github.com/gitsense/gsc-cli/blob/main/internal/cli/r...
you can see that every file has a code block header with a UUID and the AI that was attributed to it. With the UUID, I can tell exactly how the code came about.
What they are working on will be more useful for AI code provenance. It is only a matter of time before you are expected to show your chats with AI as part of the code review and for performance reviews.
So I don't see human collaboration being the main use case. I see tracking, studying and improving the Human-AI relationship...and seeing if somebody should be promoted or not.
An interesting take I've heard is, we will have a token/impact stat where if you spent a shitload of tokens to produce the same impact as somebody else who spent a lot less, you will be the prime candidate for layoffs and/or less pay. This is why I think AI code provenance will become a serious thing in the future.
A vendor conference I was at a few weeks ago focused heavily on this, for most of their Agentic workflow items. How can you show the AIs work, prove what it did was within guidelines, then audit the process and result.
Like, if your system has an AI backed federated search for documents and you ask it a question about those documents, you need an audit trail of the ask, what documents were referenced, and what was returned to the user.
Then if wrong information was supplied that can be pinned down and explained in case of lawsuit or other need.
Seems like you copied the ellipsified version, so what we get is https://github.com/gitsense/gsc-cli/blob/main/internal/cli/r
> DeltaDB breaks your work into a stream of fine-grained deltas. Where Git captures a snapshot at each commit, DeltaDB captures every operation in between and gives each one a stable identity.
I was curious about giving Zed a try, now that it has an emacs keymap. Not anymore. This is such a horribly invasive feature, I absolutely do not want my colleagues reviewing every single intermediate edit, down to the keystroke, that went into the commits I publish for review.
Before I put a PR up for review, I'll sometimes edit my commit history a little bit in magit to make it more linear and digestible--maybe update descriptions, squash some adjacent commits together, etc. This just throws that whole aspect of the job out the window and says "hey, colleague, hoover up this firehose of deltas and enjoy it".
And what the hell does this even mean?
> What we're really after is simple: the conversation with the agent becomes the only conversation you need to have.
Lmao. No. Wrong.
https://github.com/modulecollective/moe
Producing code edits itself is such a small fraction of the job I'm really not convinced these are valuable data. Mostly the effort and time goes into reading and understanding code, not producing or transforming it.
Unless of course you also have to explain your thinking and problem-solving process in meetings, which happens quite a lot the more senior you get.
It makes one vulnerable though, that's for sure. Psychologically I mean.
In my life, I've never liked people who deliberately polish up their articulation to the level that it obfuscates how they arrived at that understanding (whether it's academics or engineers). They might not do it for attention and they might not be doing it knowingly. IMO, they are taking away the opportunity of learning from the people they are talking to. For me the conversation is one sided. I'm there to listen, but rarely can I ask questions, give feedback or grow from where they have possibly reached.
I don't need your fluff. No one cares how you arrived writing another crud line to save an object to database or sent yet another AJAX call.
If you wrote some genuine great compression algorithm that's a different take on compression, I would like to see step by step reasoning and eventual dead ends.
I get it, that the overall discussion is about DeltaDB. I'd say interesting concept to toy with. I'd pay more attention to "micro commits" as the idea more than the keylogger.
Don't fucking do that. Do something way less than halfway to that line.
They don't need to know I was brushing my teeth and thinking of bacon and an argument I had with my spouse right before I thought that though. Or how rude the customer in front of me was to the new barista who was just trying her best.
I agree with the author that reviewing pull requests is too late. The problem with pull request is that they make it hard to review individual commits since they are geared towards reviewing the result of an entire branch at once. But the answer is not to share all the noise, it should be to encourage small atomic commits so you can review the early work before the entire feature/fix is complete.
The git backend already has a ton of "delta DB optimizations" (in git packs and other tools) and it's really just the git frontend that needs a little massaging (`--first-parent`, primarily) and the vast world of "subway-diagram-first/only" Git UIs that should have more "drill down" `--first-parent` counterparts simply because so many people find the subway diagrams ugly/confusing/distasteful.
Why stop at zed? The trillion dollar investment AI companies have amassed was nominally for datacenters, but as those costs rise and completion timelines extend past the typical business planning horizon, it becomes more efficient to put the money to work elsewhere. You can buy whatever you want with a trillion dollars.
I personally want better read-only code tools, or maybe the return of UML?
I can’t help but feel like it is all enabling a level of developer surveillance with which I am deeply uncomfortable, though.
Unrelated: I started using Zed a few weeks ago and really love it, it is everything it promises to be.
If Gemini seems stupid nowadays, it's only because they're being stingy with compute allocation.
0 - https://en.wikipedia.org/wiki/Piper_(source_control_system)
That agents need something "beyond git" is lost on me but it keeps coming up. For one subject the tool doesn't matter -- in fact it can be obtuse like git or bash and it's fine because agents will handle it. Then for another thing the story is opposite.
Most engineers I know keep such notebooks when they're full, because it's kind of neat to open a random page and see a brain dump scrawled out. Nobody has ever looked at another engineer's scratch notes and drawn useful insight. Frankly, I wouldn't even want to other than for novelty.
i've settled on all of this context attached to issues in a project management system and referenced from commits
it works just fine - its not like your agent cannot read your issue tracker
I don't care about any of the collaboration or AI tools that they are working on and I hope that these efforts don't detract from the text editor. At the very least I hope I can hide these features in the UI and continue to ignore them. I also hope that these features don't lead to a bloated, sluggish text editor like VSCode and IntelliJ.
Where shared collaboration I think becomes much more interesting is when agents start behaving like a member of your team. Not someone you delegate to, but someone that actively seeks out conversations and interacts with others as though they're just another member of the team.
You shouldn't be aware that version control is happening behind the scenes, you're just interacting with a persistent persona through Slack/teams/notion/figma/linear/etc with something that's logging issues, fixing bugs, writing PRDs, designing mockups, etc. The only point at which you know there's version control is when you're asked to review something before it's integrated.
This benefits those who make the machines you have conversations with and those that invest in them.
AI's code writing velocity makes this even worse, there's no way I can be simultaneously working on a code base while an AI agent is running around it doing something else.
It feels like maybe there's a ghost of an idea here about how to get the best of both worlds, but I'm not sure I follow the throughline on it.
If previously you needed to take action 1, 2, 3 to go from state A to B, all you saw was the change from A, B. Now you see intermediates 1, 2, 3 and can train the models to skip straight to B with the added context of the intermediate states.
I didn't see anything in there about branching which seems like a big thing not to have.
One thing I would really love to see, however, is a way to better attach code review comments to the specific version of code they were left on. I find it quite difficult to do with git and github, considering that commit hashes change every time one is forced to rebase (say, for example, to handle a merge conflict).
Do you expect DeltaDB to help address this problem?
> I have never been a big fan of pull requests.
I guess this partly explains why Zed (still) lacks a PR review flow, let alone a coherent one, despite some interest [1]. Pretty much the only reason I’m still with JetBrains.
[1]: https://github.com/zed-industries/zed/discussions/34759
Excited to see what Zed team comes up with! :)
[1]: https://builders.ramp.com/post/why-we-built-our-background-a...
Gives jj vibes, interesting!
Not without cons of course: commit byte size, public WIP work and leaked secrets/unwanted edits.
The problem is, is that it works, if you assume "working" means the software sellers get wealthy.
There's a reason that most waitstaff wear black. They should blend into the background, and not be what the folks at the table are talking about. In rare instances, restaurants exist, where the waitstaff is the service.
In software, though, you're being served by a waiter wearing a clown suit, screaming slogans at you, and serving you lukewarm, pre-chewed goo.
The “calm technology” book has an handful of advices, but one of the best example is the xbiff program. It switches picture when you have new mail on your local spool.
A table saw does tho.
Not convinced DeltaDB is useful for humans directly but assume new and interesting interfaces can arise from the abstractions it provides.
Agree that PRs/snapshots is an antiquated way of sharing information.
There has also been a lot of discussion about the value of the peer review process recently, in in general. But I wonder if this isn't all going into the wrong direction. Quite honestly, even with the previous review and discussion system of (squashed) commits how often did you really use git blame and opened up the original PR discussion of that changed line that caused the bug? And how often did it help you beyond learning it was done by that rockstar developer who has left long ago? And that the discussion on that PR was a point in time and the code around it has evolves beyond that and it would need looking at another 10 PR discussion to get the entire context.
What I am saying is that git (and before that Subversion and CVS) has a full history is so that it can resolve the latest state. Period. We made that commit ritual somewhat of a hallmark in putting more supposed meaning into recording ever more in that history. But we rarely stopped to check if that is all that useful. Recording even more, all the time, reminds me of these work group meetings that have minute records of every bike shedding meeting word said by everyone, that, honestly, no one ever looks at after the next meeting ever again. I don't think there is value in minute record keeping, it becomes too much noise that just makes it harder to parse. Now also adding all AI conversation and agent thinking to that tree? What's the value of that in like 3 months down the line? I don't see it.
Many of the product decisions that Zed's made caused me to switch to Zed for my daily driver IDE (previously JetBrains). The recent AI agent threads and improvements around diffs really solidified the move.
Honestly all of this drives me back towards nvim or notepad sometimes.
I have had a jetbrains subscription since pycharm came out, and the killer feature was always the visual debugger. Seems nearly quaint now.
What specific things do you like about zed?
1. It is fast and snappy. Nothing comes even close besides vim (and I don't mind going full time to it if I have to)
2. The ability to completely shut off and block any slop machine features from interfering with my workflow or leak code back to sloppenai, sloppus or any other self-installed-worst-security-practice-backdoor garbage.
Having said that, I hope they don't remove that ability in the future and enforce the "slop is so good man, you should try it" philosophy.
I am happy about even though I've never tried gram, because if zed goes to shit there will be an alternative, which hopefully pressures zed to stay sane
> I also object to making myself and my work depend on paying a subscription fee to anyone. I don't want an outage at Anthropic to affect my ability to do my work. I think it is a grave mistake to build anything on such shaky foundations as the sustainability and profit margins of the AI industry.
Someone actually sensible, excellent.
Any particular model almost doesn’t matter at this point. Harnesses are built around them. OpenAI and Anthropic are basically interchangeable in an open-source harness like OpenCode; the switching cost is virtually zero. Local models are improving rapidly and are already “good enough” for many use cases already. The bet that LLMs will continue to exist as an algorithm is pretty solid.
Edit: After further inspection, I think I'm jumping ship before it's too late. And I'll look, see if there's a way to lend a hand or two when I have time!
I’m highly skeptical of this claim. For any complicated feature, there’s always a design doc (or an RFC, or a wireframe) and that’s what people used for discussion. Discussion in a PR are mostly about whether to accept the code, reject the feature, or provide feedback about alternate implementations. It’s not for pair programming or directing design.
Collaborating together in a research lab (brainstorm session) is not the same as asking feedback for a journal article (PR). What is described in the article is pair programming with extra steps.
I also like the AI tools, the inline assistant is good and the agent is also pretty nice and well integrated into the editor without it being the focus point. I'm not against using AI but I certainly don't use it as much as a lot of people do.
That being said, I really dislike this recent push towards becoming more like a cursor wannabe. They have a new (for now) opt-in default layout that almost hides the editor panel in favour of the agent threads and agent panels. And now this. I don't want to switch editors, but if they keep pushing a different workflow from what I use it might send me back to Jetbrains...
Your boss will also demand that you narrate your thought process 8 hours a day. Speakmaxxing will be required to keep your job. There will be a "silence" wall of shame for the non-compliers.
Our human ability is not defined by our _absolute_ output, but, by the quality of the _delta_ applied to an engineering artefact
Great engineers obsess over every keystroke
With LLMs, a much smaller number of keystrokes can create a much larger and more positively impactful delta
Every delta to the codebase can tell us some informational property about the behaviour of the system and storing that information WILL prove to be useful in the future