by cryptophreak on 2/4/25, 4:06 PM with 417 comments
by wiremine on 2/4/25, 6:13 PM
I just finished a small project where I used o3-mini and o3-mini-high to generate most of the code. I averaged around 200 lines of code an hour, including the business logic and unit tests. Total was around 2200 lines. So, not a big project, but not a throw away script. The code was perfectly fine for what we needed. This is the third time I've done this, and each time I get faster and better at it.
1. I find a "pair programming" mentality is key. I focus on the high-level code, and let the model focus on the lower level code. I code review all the code, and provide feedback. Blindly accepting the code is a terrible approach.
2. Generating unit tests is critical. After I like the gist of some code, I ask for some smoke tests. Again, peer review the code and adjust as needed.
3. Be liberal with starting a new chat: the models can get easily confused with longer context windows. If you start to see things go sideways, start over.
4. Give it code examples. Don't prompt with English only.
FWIW, o3-mini was the best model I've seen so far; Sonnet 3.5 New is a close second.
by taeric on 2/4/25, 4:47 PM
In large, I assert this is because the best way to do something is to do that thing. There can be correspondence around the thing, but the artifacts that you are building are separate things.
You could probably take this further and say that narrative is a terrible way to build things. It can be a great way to communicate them, but being a separate entity, it is not necessarily good at making any artifacts.
by themanmaran on 2/4/25, 5:19 PM
Agreed that copy pasting context in and out of ChatGPT isn't the fastest workflow. But Cursor has been a major speed up in the way I write code. And it's primarily through a chat interface, but with a few QOL hacks that make it way faster:
1. Output gets applied to your file in a git-diff style. So you can approve/deny changes.
2. It (kinda) has context of your codebase so you don't have to specify as much. Though it works best when you explicitly tag files ("Use the utils from @src/utils/currency.ts")
3. Directly inserting terminal logs or type errors into the chat interface is incredibly convenient. Just hover over the error and click the "add to chat"
by croes on 2/4/25, 4:22 PM
So you either need lots of extra text to remove the ambiguity of natural language if you use AI or you need a special precise subset to communicate with AI and that’s just programming with extra steps.
by matthewsinclair on 2/4/25, 5:48 PM
I think about this like SQL in the late 80s. At the time, SQL was the “next big thing” that was going to mean we didn’t need programmers, and that management could “write code”. It didn’t quite work out that way, of course, as we all know.
I see chat-based interfaces to LLMs going exactly the same way. The LLM will move down the stack (rather than up) and much more appropriate task-based UX/UI will be put on top of the LLM, coordinated thru a UX/UI layer that is much sympathetic to the way users actually want to interact with a machine.
In the same way that no end-users ever touch SQL these days (mostly), we won’t expose the chat-based UX of an LLM to users either.
There will be a place for an ad-hoc natural language interface to a machine, but I suspect it’ll be the exception rather than the rule.
I really don’t think there are too many end users who want to be forced to seduce a mercurial LLM using natural language to do their day-to-day tech tasks.
by spolsky on 2/4/25, 5:34 PM
If you asked me two or three years ago I would have strongly agreed with this theory. I used to point out that every line of code was a decision made by a programmer and that programming languages were just better ways to convey all those decisions than human language because they eliminated ambiguity and were much terser.
I changed my mind when I saw how LLMs work. They tend to fill in the ambiguity with good defaults that are somewhere between "how everybody does it" and "how a reasonably bright junior programmer would do it".
So you say "give me a log on screen" and you get something pretty normal with Username and Password and a decent UI and some decent color choices and it works fine.
If you wanted to provide more details, you could tell it to use the background color #f9f9f9, but a part of what surprised my and caused me to change my mind on this matter was that you could also leave that out and you wouldn't get an error; you wouldn't get white text on white background; you would get a decent color that might be #f9f9f9 or might be #a1a1a1 but you saved a lot of time by not thinking about that level of detail and you got a good result.
by jakelazaroff on 2/4/25, 4:23 PM
Something like tldraw's "make real" [1] is a much better bet, imo (not that it's mutually exclusive). Draw a rough mockup of what you want, let AI fill in the details, then draw and write on it to communicate your changes.
We think multi-modally; why should we limit the creative process to just text?
[1] https://tldraw.substack.com/p/make-real-the-story-so-far
by Edmond on 2/4/25, 4:22 PM
Here is an example of our approach:
https://blog.codesolvent.com/2024/11/building-youtube-video-...
We are also using the requirements to build a checklist, the AI generates the checklist from the requirements document, which then serves as context that can be used for further instructions.
Here's a demo:
by ajmurmann on 2/4/25, 5:16 PM
The mode that I've found most fruitful when using Cursor is treating it almost exactly as I would a pair programming partner. When I start on a new piece of functionality I describe the problem and give it what my thoughts are on a potential solution and invite feedback. Sometimes my solution is the best. Sometimes the LLM had a better idea and frequently we take a modified version of what one of us suggested. Just as you would with a human partner. The result of the discussion is better than what either of us would have done on their own.
I also will do classical ping-pong style tdd with it one we agreed on an approach. I'll write a test; llm makes it pass and write the next test which I'll make pass and so on.
As with a real pair, it's important to notice when they are struggling and help them or take over. You can only do this if you stay fully engaged and understand every line. Just like when pairing. I've found llms get frequently in a loop where something doesn't work and they keep applying the same changes they've tried before and it never works. Understand what they are trying to do and help them out. Don't be a shitty pair for your llm!
by mlsu on 2/4/25, 6:17 PM
It will slowly grow in complexity, strictness, and features, until it becomes a brand-new programming language, just with a language model and a SaaS sitting in the middle of it.
A startup will come and disrupt the whole thing by simply writing code in a regular programming language.
by sho_hn on 2/4/25, 4:39 PM
https://aider.chat/docs/usage/watch.html
How jarring it is & how much it takes you out of your own flow state is very much dependent on the model output quality and latency still, but at times it works rather nicely.
by fny on 2/4/25, 5:44 PM
I think it's more ideal to have the LLM map text to some declarative pseudocode that's easy to read which is then translated to code.
The example given by Daniel might map to something like this:
define sign-in-screen:
panel background "#f9f9f9":
input email required: true, validate-on-blur: true
input password required: true
button "Sign in" gradient: ("#EEE" "#DDD")
connect-to-database
Then you'd use chat to make updates. For example, "make the gradient red" or "add a name field." Come to think of it, I don't see why chat is a bad interface at all with this set up.by karmakaze on 2/4/25, 5:00 PM
The example shows "Sign-in screen" with 4 (possibly more) instructions. This could equivalently have been entered one at a time into 'chat'. If the response for each was graphic and instantaneous, chat would be no worse than non-chat.
What makes non-chat better is that the user puts more thought into what they write. I do agree for producing code Claude with up-front instructions beats ChatGPT handily.
If OTOH AI's actually got as good or better than humans, chat would be fine. It would be like a discussion in Slack or PR review comments.
by quantadev on 2/4/25, 5:14 PM
1) The first thing to improve chats as a genre of interface, is that they should all always be a tree/hierarchy (just like Hacker News is), so that you can go back to ANY precise prior point during a discussion/chat and branch off in a different direction, and the only context the AI sees during the conversation is the "Current Node" (your last post), and all "Parent Nodes" going back to the beginning. So that at any time, it's not even aware of all the prior "bad branches" you decided to abandon.
2) My second tip for designs of Coding Agents is do what mine does. I invented a 'block_begin/block_end' syntax which looks like this, and can be in any source file:
// block_begin MyAddNumbers
var = add(a, b)
return a + b
// block_end
With this syntax you can use English language to explain and reason about extremely specific parts of your code with out expecting the LLM to "just understand". You can also direct the LLM to only edit/update specific "Named Blocks", as I call them.
So a trivial example of a prompt expression related to the above might be "Always put number adding stuff in the MyAddNumbers Block".
To explain entire architectural aspects to the LLM, these code block names are extremely useful.
by tiborsaas on 2/4/25, 5:19 PM
I've tested a few integrated AI dev tools and it works like a charm. I don't type all my instructions at once. I do it the same way as I do it with code. Iteratively:
1) Create a layout
2) Fill left side
3) Fill right side
4) Connect components
5) Populate with dummy data
> The first company to get this will own the next phase of AI development tools.
There's more than 25 working on this problem and they are already in production and some are really good.
by deeviant on 2/4/25, 4:47 PM
Everything else, is just putting layers, that are not nearly as capable at an LLM, between me and the raw power of the LLM.
The core realization I made to truly unlock LLM code assistance as a 10x + productivity gain, is that I am not writing code anymore, I am writing requirements. It means being less an engineer, and more a manager, or perhaps an architect. It's not your job to write tax code anymore, it's your job to describe what the tax code needs to accomplish and how it's success can be defined and validated.
Also, it's never even close to true that nobody uses LLMs for production software, here's a write-up by Google talking about using LLMs to drastically accelerate the migration of complex enterprise production systems: https://arxiv.org/pdf/2501.06972
by xena on 2/4/25, 5:09 PM
Last night I wrote an implementation of an AI paper and it was so much easier to just discard the automatic chat formatting and do it "by hand": https://github.com/Xe/structured-reasoning/blob/main/index.j...
I wonder if foundation models are an untapped goldmine in terms of the things they can do, but we can't surface them to developers because everyone's stuck in the chat pattern.
by spandrew on 2/4/25, 11:05 PM
The company I work for integrated AI into some of our native content authoring front-end components and people loved it. Our system took a lot of annotating to be able to accurately translate the natural language to the patterns of our system but users so far have found it WAYYY more useful than chat bc it's deeply integrated into the tasks they do anyway.
Figma had a similar success at last year's CONFIG when they revealed AI was renaming default layers names (Layer 1, 2, etc)... something they didn't want to do anyway. I dare say nobody gave a flying f about their "template" AI generation whereas layer renaming got audible cheers. Workflow integration is how you show people AI isn't just replacing their job like some bad sci-fi script.
Workflow integration is going to be big. I think chat will have its place tho; just kind of as an aside in many cases.
by ingigauti on 2/4/25, 8:45 PM
Then having ai generate code for my project didn't feel good either, I didn't really understand what it was doing so I would have to read it to understand, then what is the purpose, I might as well write it.
I then started playing, and out came a new type of programming language called plang (as in pseudo language). It allows you to write the details without all the boiler code.
I'm think I've stumbled on to something, and just starting to get noticed :) https://www.infoworld.com/article/3635189/11-cutting-edge-pr...
by benatkin on 2/4/25, 4:29 PM
by bangaladore on 2/4/25, 7:16 PM
In a real-world scenario, we begin with detailed specifications and requirements, develop a product, and then iterate on it. Chat-based interactions might be better suited to this iterative phase. Although I'm not particularly fond of the approach, it does resemble receiving a coworker's feedback, making a small, targeted change, and then getting feedback again.
Even if the system were designed to focus solely on the differences in the requirements—thus making the build process more iterative—we still encounter an issue: it tends to devolve into a chat format. You might have a set of well-crafted requirements, only for the final instruction to be, "The header should be 2px smaller."
Nonetheless, using AI in an iterative process (focusing on requirement diffs, for example) is an intriguing concept that I believe warrants further exploration.
by yapyap on 2/4/25, 7:59 PM
That’s the thing about language, you CAN’T program in human language for this exact reason, whereas programming languages are mechanical but precise, human languages flow better but they leave wiggle room. Computers can’t do jack shit with wiggle room, they’re not humans. That’ll always remain, until there’s an AI people like enough to have it’s own flair on things.
by r0ckarong on 2/4/25, 4:31 PM
by PaulHoule on 2/4/25, 8:24 PM
So far as this article is concerned (not the many commenters who are talking past it), "chat" is like interacting with a shell or a REPL. How different is the discussion that Winograd has with SHRDLU
https://en.wikipedia.org/wiki/SHRDLU
with the conversation that you have with a database with the SQL monitor really?
There's a lot to say for trying to turn that kind of conversation into a more durable artifact. I'd argue that writing unit tests in Java I'm doing exploratory work like I'd do in a Python REPL except my results aren't scrolling away but are built into something I can check into version control.
On the other hand, workspace-oriented programming environments are notorious for turning into a sloppy mess, for instance people really can't make up their mind if they want to store the results of their computations (God help you if you have more than one person working on it, never mind if you want to use version control -- yet, isn't that a nice way to publish a data analysis?) or if they want to be a program that multiple people can work, can produce reproducible results, etc.
See also the struggles of "Literate Programming"
Not to say there isn't an answer to all this but boy is it a fraught area.
by cheapsteak on 2/4/25, 5:44 PM
English behaviour descriptions -> generated tests
Use both behaviour descriptions and feedback from test results to iterate on app development
by Bjorkbat on 2/4/25, 8:12 PM
Absolutely insane that all the doors unlocked by being able to interact with a computer graphically, and yet these people have visions of the future stuck in the 60s.
by furyofantares on 2/4/25, 4:41 PM
by bcherry on 2/4/25, 6:02 PM
Chat is an awesome powerup for any serious tool you already have, so long as the entity on the other side of the chat has the agency to actually manipulate the tool alongside you as well.
by reverendsteveii on 2/4/25, 4:55 PM
by anarticle on 2/4/25, 5:10 PM
I think this post shows there could be a couple levels of indirection, some kind of combination of the "overarching design doc" that is injected into every prompt, and a more tactical level syntax/code/process that we have with something like a chat window that is code aware. I've definitely done some crazy stuff by just asking something really stupid like "Is there any way to speed this up?" and Claude giving me some esoteric pandas optimization that gave me a 100x speedup.
I think overall the tools have crazy variance in quality of output, but I think with some "multifacet prompting", ie, code styling, design doc, architect docs, constraints, etc you might end up with something that is much more useful.
by hoppp on 2/4/25, 5:17 PM
So I completely agree with this. Chat is not a good UI
by Vox_Leone on 2/4/25, 5:18 PM
Example of a Structured Pseudo-Code Prompt:
Let’s say you want to generate code for a function that handles object detection:
'''Function: object_detection Input: image Output: list of detected objects
Steps: 1. Initialize model (load pretrained object detection model)
2. Preprocess the image (resize, normalize, etc.)
3. Run the image through the model
4. Extract bounding boxes and confidence scores from the model's output
5. Return objects with confidence greater than 0.5 as a list of tuples (object_name, bounding_box)
Language: Python'''
by xenodium on 2/4/25, 8:11 PM
Been experimenting with the same approach but for "paged shells" (sorry for the term override) and this seems to be a best of both worlds kinda thing for shells. https://xenodium.com/an-experimental-e-shell-pager That is, the shell is editable when you need it to be (during submission), and automatically read-only after submission. This has the benefit of providing single-character shortcuts to navigate content. n/p (next/previous) or tab/backtab.
The navigation is particularly handy in LLM chats, so you can quickly jump to code snippets and either copy or direct output elsewhere.
by muzani on 2/4/25, 11:36 PM
Chat is also iterative. You can go back there and fix things that were misinterpreted. If the misinterpretation happens often, you can add on another instruction on top of that. I strongly disagree that they'd be fixed documents. Documents are a way to talk to yourself and get your rules right before you commit to them. But it costs almost nothing to do this with AI vs setting up brainstorming sessions with another human.
However, the rational models (o1, r1 and such) are good at iterating with themselves, and work better when you give them documents and have them figure out the best way to implement something.
by vismit2000 on 2/5/25, 5:47 AM
by fhd2 on 2/4/25, 5:53 PM
It has features to add context from your current project pretty easily, but personally I prefer to constantly edit the chat buffer to put in just the relevant stuff. If I add too much, Claude seems to get confused and chases down irrelevant stuff.
Fully controlling the context like that seems pretty powerful compared to other approaches I've tried. I also fully control what goes into the project - for the most part I don't copy paste anything, but rather type a version of the suggestion out quickly.
If you're fast at typing and use an editor with powerful text wrangling capabilities, this is feasible. And to me, it seems relatively optimal.
by weitendorf on 2/4/25, 6:54 PM
Many developers don't realize this but as you go back and forth with models, you are actively polluting their context with junk and irrelevant old data that distracts and confuses it from what you're actually trying to do right now. When using sleeker products like Cursor, it's easy to forget just how much junk context the model is constantly getting fed (from implicit RAG/context gathering and hidden intermediate steps). In my experience LLM performance falls off a cliff somewhere around 4 decent-sized messages, even without including superfluous context.
We're further separating the concept of "workflow" from "conversation" and prompts, basically actively and aggressively pruning context and conversation history as our agents do their thing (and only including context that is defined explicitly and transparently), and it's allowing us to tackle much more complex tasks than most other AI developer tools. And we are a lot happier working with models - when things don't work we're not forced to grovel for a followup fix, we simply launch a new action to make the targeted change we want with a couple clicks.
It is in a weird way kind of degrading to have to politely ask a model to change a color after it messed up, and it's also just not an efficient way to work with LLMs - people just default to that style because it's how you'd interact with a human you are delegating tasks to. Developers still need to truly internalize the facts that LLMs are purely completion machines, that your conversation history lives entirely client side outside of active inference, and that you can literally set your conversation input to be whatever you want (even if the model never said that) - after that realizing you're on the path towards using LLMs like "what words do I need to put it in to get it to do what I want" rather than working "with" them.
by jimlikeslimes on 2/4/25, 5:12 PM
by fullstackchris on 2/5/25, 11:37 AM
However, even with a "docs as spec" pattern, how can you control the actual quality of the code written? How maintainable will it be? If the spec changes (read: it _will_ change constantly), is it easy enough to refactor? What about tests? I also shrink in fear at the complexity of docs that could be _exactly_ captured as code... "well we almost always do it this way, but this one time we do it this way..."
by lcfcjs6 on 2/4/25, 4:29 PM
by sprucevoid on 2/4/25, 7:27 PM
- intellisense in the inputbox based on words in this or all previous chats and a user customizable word list
- user customizable buttons and keyboard shortcuts for common quick replies, like "explain more".
- when claude replies with a numbered list of alternatives let me ctrl+click a number to fork the chat with continued focus on that alternative in a new tab.
- a custom right click menu with action for selection (or if no selection claude can guess the context e.g. the clicked paragraph) such as "new chat with selection", "explain" and some user customizable quick replies
- make the default download filenames follow a predicable pattern, claude currently varies it too much e.g. "cloud-script.py" jumps to "cloud-script-errorcheck.py". I've tried prompting a format but claude seems to forget that.
- the stop button should always instantly stop claude in its tracks. Currently it sometimes takes time to get claude to stop thinking.
- when a claude reply first generates code in the right sidebar followed by detailed explanation text in the chat, let some keyboard shortcut instantly stop the explanation in its tracks. Let the same shortcut preempt that explanation while the sidebar code is still generating.
- chat history search is very basic. Add andvanced search features, like filter by date first/last message and OR search operator
- batch jobs and tagging for chat history. E.g. batch apply a prompt to generate a summary in each selected chat and then add the tag "summary" to them. Let us then browse by tag(s).
- tools to delete parts of a chat history thread, that in hindsight were detours
- more generally, maybe a "chat history chat" to have Claude apply changes to the chat histories
by azhenley on 2/4/25, 4:46 PM
by andix on 2/4/25, 7:43 PM
1. Ask AI to generate a spec of what we're planning to do. 2. Refine it until it's kind of resembling what I want to do 3. Ask AI to implement some aspects from the spec
by michaelfeathers on 2/4/25, 9:56 PM
I used this single line to generate a 5 line Java unit test a while back.
test: grip o -> assert state.grip o
LLMs have wide "understanding" of various syntaxes and associated semantics. Most LLMs have instruct tuning that helps. Simplifications that are close to code work.
Re precision, yes, we need precision but if you work in small steps, the precision comes in the review.
Make your own private pidgin language in conversation.
by yawnxyz on 2/4/25, 5:35 PM
Emails are so similar to Chat, except we're used to writing in long-form, and we're not expecting sub-minute replies.
Maybe emails are going to be the new chat?
I've been experimenting with "email-like" interfaces (that encourage you to write more / specify more), take longer to get back to you, and go out to LLMs. I think this works well for tools like Deep Research where you expect them to take minutes to hours.
by nimski on 2/4/25, 5:35 PM
Chat is single threaded and ephemeral. Documents are versioned, multi-threaded, and a source of truth. Although chat is not appropriate as the source of truth, it's very effective for single-threaded discussions about documents. This is how people use requirements documents today. Each comment on a doc is a localized chat. It's an excellent interface when targeted.
by foz on 2/4/25, 5:30 PM
Like with any coworker - when ideas get real, get out of chat and start using our tools and process to get stuff done.
by a3w on 2/4/25, 4:24 PM
Perhaps I should comment all todos and then write "finish todos" as the always-same text prompt.
by chaisan on 2/5/25, 1:25 AM
by firefoxd on 2/4/25, 6:44 PM
For writing, the canvas interface is much more effective because you rely less on copy and paste. For code, even with the ctrl+i method, it works but it's a pain to have to load all other files as reference every single time.
by nayuki on 2/4/25, 6:05 PM
by josefrichter on 2/4/25, 4:39 PM
It's not really a conscious choice, but rather a side effect. And we already see the trend is away from that, with tools like chatGPT Canvas, editors like Windsurf, etc.
by jfkrrorj on 2/4/25, 4:45 PM
AI in many levels is more capable than human programmer, in some it is not. It is not supersmart. It can not hold entire program in its head, you have to feed it small relevant section of program.
》 That’s why we use documents—they let us organize complexity, reference specific points, and track changes systematically.
Extra steps. Something like waterfall...
by tgraf_80 on 2/4/25, 5:42 PM
by kmarc on 2/4/25, 5:32 PM
di(
Yet, millions of programmers use their mouse to SELECT first something visually and THEN delete whatever was selected. Shrug.I won't be surprised if chat-based programming will be the next way of doing stuff.
by grumbel on 2/4/25, 6:23 PM
by darepublic on 2/4/25, 6:36 PM
by daxfohl on 2/4/25, 10:15 PM
by cratermoon on 2/5/25, 3:22 PM
by orand on 2/4/25, 8:43 PM
by randomNumber7 on 2/4/25, 9:14 PM
- Speed up literature recherche
- replace reading library documentation
- generate copy pasta code that has been written often before
by ansonhw on 2/4/25, 7:31 PM
by ypyrko on 2/4/25, 5:36 PM
by m3kw9 on 2/4/25, 4:22 PM
by whatsakandr on 2/4/25, 5:01 PM
by suralind on 2/4/25, 11:06 PM
by jes5199 on 2/5/25, 12:12 AM
by RyanAdamas on 2/4/25, 7:13 PM
by nbzso on 2/4/25, 7:04 PM
by 6h6j65j76k on 2/4/25, 5:44 PM
But that is true? Devs spend more time in meetings than writing code. Having conversations about the code they are going to write.
by sramam on 2/4/25, 9:21 PM
by anoncow on 2/4/25, 5:34 PM
by remoquete on 2/4/25, 5:45 PM
Back to... programming languages? :)
by icapybara on 2/4/25, 6:36 PM
Theoretically maybe, but chat windows are getting the job done right now.
by arnaudsm on 2/4/25, 6:33 PM
It could be quite fun !
by empath75 on 2/4/25, 4:20 PM
by mhh__ on 2/5/25, 2:52 AM
by notatoad on 2/5/25, 4:59 AM
chat is so drastically far away from my workflow that it doesn't feel like my workflow is wrong.
by Apocryphon on 2/4/25, 5:10 PM
So, something like Gherkin?
by shireboy on 2/4/25, 5:46 PM
For higher-level AI assist, I do agree chat is not what makes sense. What I think would be cool is to work in markdown files, refining in precise plain english each feature. The AI then generates code from the .md files plus existing context. Then you have well-written documentation and consistent code. You can do this to a degree today by referencing a md file in chat, or by using some of the newer tools, but I haven't seen exactly what I want yet. (I guess I should build it?)
by kristofferR on 2/4/25, 8:33 PM
by gunalx on 2/4/25, 9:57 PM
Its a problem of programming languages and definitions.
by kerblang on 2/4/25, 4:29 PM
by aantix on 2/4/25, 7:33 PM
by indymike on 2/4/25, 4:40 PM
by jillesvangurp on 2/5/25, 7:51 AM
The challenge is that I haven't seen anything better really.
Lately the innovation comes mainly from deeper integration with tools. Standalone AI editors are mainly popular with people who use relatively simple editors (like VS Code). VS Code has a few party tricks but for me swapping out Intellij for something else on a typical Kotlin project is a complete non starter. Not going to happen. I'd gain AI, but I'd loose everything else that I use all the time. That would be a real productivity killer. I want to keep all the smart tooling I already have and have used for years.
There are a few extensions for intellij but they are pretty much all variations of a sidebar with a chat and autocomplete. Autocomplete competes with normal autocomplete, which I use all the time. And the clippy style "it looks like you are writing a letter" style completions just aren't that useful too me at all. They are just noise and break my flow. And they drown out the completions I use and need all the time. And sidebars just take up space and copying code from there back to your editor is a bit awkward as UX
Lately I've been using chat gpt. It started out pretty dumb but these days I can option+shift+1 in a chat and have it look over my shoulder at my current editor. "how do I do that?" translates into a full context with my current editing window, cursor & selected text, etc. all in the context. Before I was copy pasting everything and the kitchen sync to chat gpt, now it just tells me what I need to do. The next step up from this is that it starts driving the tools itself. They already have a beta for this. This deeper integration is what is needed.
A big challenge is that most of these tools are driven to minimize cost and context size. Tokens cost money. So chat GPT only looks at my active editor and not at the 15 other files I have open. It could. But it doesn't. It's also unaware of my project structure, or the fact that most of my projects are kotlin multiplatform and can't use JVM dependencies. So, in that sense, every chat still is a bit ground hog day. It's promise to "remember" stuff when you ask it too is super flaky. It forgets most things it's supposed to remember pretty quickly.
These are solvable problems of course. But it's useful to me for debugging, analyzing, completing functions, etc.
by stevage on 2/4/25, 9:59 PM
by proc0 on 2/4/25, 5:14 PM
In its current form LLMs are pretty much at their limit, barring optimization and chaining them together for more productivity once we have better hardware. Still, it will just be useful for repetitive low level tasks and mediocre art. We need more breakthroughs beyond transformers to approach something that creates like humans instead of using statistical inference.
by jpcom on 2/5/25, 12:30 AM
by karaterobot on 2/4/25, 5:07 PM
First of all, most people can't write extremely complex applications, period. Most programmers included. If your baseline for real programming is something of equivalent complexity as the U.S. tax code, you're clearly such a great programmer that you're an outlier, and should recognize that.
Second of all, I think it's a straw man argument to say that you can either write prototype-level code with a chat UI, or complex code with documents. You can use both. I think the proposition being put forward is that more people can write complex code by supplementing their document-based thinking with chat-based thinking. Or, that people can write slightly better-than-prototype level code with the help of a chat assistant. In other words, that it's better to have access to AI to help you code small sections of a larger application that you are still responsible for.
I'd be more interested in reading a good argument against the value of using chat-based AI as another tool in your belt, rather than a straight-up replacement for traditional coding. If you could make that argument, then you could say chat is a bad UI pattern for dev tools.
by synergy20 on 2/4/25, 5:59 PM
or the other way around,give AI a design doc and generate what you want,this is still chatting, just more official and lengthy
by nektro on 2/5/25, 5:50 AM
by debacle on 2/4/25, 9:09 PM
by kordlessagain on 2/4/25, 6:39 PM
by Havoc on 2/4/25, 6:13 PM
I don’t buy that a document could capture what is needed here. Imagine describing navigating through multiple levels of menus in document form. That sounds straight up painful even for trivial apps. And for a full blown app…nope
There is a whole new paradigm missing there imo
by williamcotton on 2/4/25, 9:04 PM
Vague and prone to endless argument?
by fragmede on 2/4/25, 5:58 PM
Eh, that's just copium because we all have a vested monetary interest in them not being useful for "anything real", whatever that means. If it turns out that there useful for "real things", then then entire industry would get turned on its head. (hint: they're useful for "real" things), though putting the entire codebase into the context window doesn't currently work. Aider works past this by passing the directory tree and filenames as context, so the LLM guess that /cloud/scope/cluster.go is where the cluster scope code lives and ask for that specific file to get added to the context and you can ask it to add, say, logging code to that file.
by tommiegannert on 2/4/25, 6:04 PM
We play around with LLMs to build a chat experience. My first attempt made Claude spew out five questions at a time, which didn't solve the "guiding" problem. So I started asking it to limit the number of unanswered questions. It worked, but felt really clunky and "cheap."
I drew two conclusions: We need UI builders for this to feel nice, and professionals will want to use forms.
First, LLMs would be great at driving step-by-step guides, but it must be given building blocks to generate a UI. When asking about location, show a map. When deciding to ask about TIN or roof size, if the user is technically inclined, perhaps start with asking about the roof. When asking about the roof size, let the user draw the shape and assign lengths. Or display aerial photos. The result on screen shouldn't be a log of me-you text messages, but a live-updated summary of where we are, and what's remaining.
Second, professionals have incentive to build mental model for navigating complex data structures. People who have no reason to invest time into the data model (e.g. a consumer buying a single solar panel installation in ther lifetime,) will benefit from rich LLM-driven UIs. Chat UIs might create room for a new type of computer user who doesn't use visual clues to build this mental model, but everyone else will want to stay on graphics. If you're an executive wondering how many sick days there were last month, that's a situation where a BI LLM RAG would be great. But if you're not sure what your question is, because you're hired to make up your own questions, then pointing, clicking and massaging might make more sense.
by randomcatuser on 2/4/25, 6:09 PM
doc=programming in a DSL? / (what was that one language which was functional & represented in circles in a canvas?)
by newsyco21 on 2/4/25, 4:35 PM
by SuperHeavy256 on 2/4/25, 8:16 PM
by waylonchang on 2/5/25, 1:28 AM
by waychang on 2/5/25, 1:28 AM
by h1fra on 2/4/25, 4:42 PM
by dehugger on 2/4/25, 7:49 PM
Writing a crud web API? Great! Writing business logic for a niche edge case in a highly specialized domain? Good luck.
by sebastianconcpt on 2/4/25, 6:24 PM
by thomastjeffery on 2/4/25, 5:38 PM
What's a good interface?
There are a few things we try to balance to make a good UI/UX:
- Latency: How long it takes to do a single task
- Decision-tree pathing: How many tasks to meet a goal
- Flexibility/Configurability: How much of a task can be encapsulated by the user's predefined knowledge of the system
- Discoverability: What tasks are available, and where
The perfect NLP chat could accomplish some of these:
- Flexibility/Configurability: Define/infer words and phrases that the user can use as shortcuts
- Decision-tree pathing: Define concepts that shortcut an otherwise verbose interaction
- Latency: Context-aware text-completions so the user doesn't need to type as much
- Discoverability: Well-formed introductions and clarifying questions to introduce useful interaction
This can only get us so far. What better latency can be accomplished than a button or a keyboard shortcut? What better discoverability than a menu?
The most exciting prospect left is flexibility. Traditional software is inflexible. It can only perform the interaction it was already designed with. Every design decision becomes a wall of assumption. These walls are the fundamental architecture of software. Without them, we would have nothing. With them, we have a structure that guides us along whatever assumptions were already made.
If we want to change something about our software's UI, then we must change the software itself, and that means writing. If NLP was a truly solved problem, then software compatibility and flexibility would be trivialized. We could redesign the entire UI by simply describing the changes we want.
LLMs are not even close. Sure, you can get one to generate some code, but only if the code you want generated is close enough to the text it was already trained on. LLMs construct continuations of tokens: no more, no less. There is no logic. There is no consideration about what is right or wrong: only what is likely to come next.
Like you said,
> You can’t build real software without being precise about what you want.
This is the ultimate limitation of UI. If only we could be ambiguous instead! LLMs let us do that, but they keep that ambiguity permanent. There is no real way to tie an LLM back down to reality. No logic. No axioms. No rules. So we must either be precise or ambiguous. The latter option is an exciting development, and certainly offers its own unique advantages, but it isn't a complete solution.
---
I've been thinking through another approach to the ambiguity problem that I think could really give us the expressive power of natural language, while preserving the logical structure we use to write software (and more). It wouldn't solve the problem entirely, but it could potentially move it out of the way.
by talles on 2/4/25, 9:11 PM
by mehakkar3006 on 2/5/25, 11:05 PM