Just wait until inference becomes 10x faster/cheaper.
2x, not 10x: coding with LLMs in 2026
121–130 of 260 posts
Re: 2x, not 10x: coding with LLMs in 2026
#122I think it matters a lot to understand the kernel of AI in the past 10 years. It's the only way you can truly appreciate the exponential. It looks flat unless you see the dramatic leap in transformer architecture and scaling laws and reinforcement learning. I know it's been hyped to death but you have to see how those things are truly exponential at their core to appreciate how it's not just walking up one step or two steps faster and how it's walking up buildings etc.
Re: 2x, not 10x: coding with LLMs in 2026
#123I'm probably a weird outlier. Coming from academia, it's ranging from 1x to infinity-x (as in, certain tasks wouldn't even be touched if not for AI). For stuff that I'm used to (R) I can write nice and compact spaghetti (long %>% pipes). I'm not comfortable when a working script doesn't fit the screen (plus a few scrolls max). My style is probably easy only to me. When I teach, I don't teach it in particular. AI give…
Interesting. Framing it as 1x to infinity-x matches my experience too. I've have good success with it reproducing papers with existing code, but not such much with one-shotting new code. Do you have any particular setup for this i.e. special validation prompts or multi-agent stuff or do you just ask something like reproduce paper X until it gets results Y?
I used to chat with "the paper" until both me and the AI were roughly on the same page at what's happening. It's important, as a "full read" of the "full paper" is not always used efficiently if not discussed, I've noticed. I then ask it for a detailed step by step realization (of the algorithm, or the paper/section/whatever), plus some context. Then I pass this as a prompt to a fresh new instance, and task it to implement in full based on the description (Gemini Deep Think was quite good with that).
Now I feel like 5.6 Sol Ultra is capable of doing roughly the same with its Agents, so it's getting easier in my experience. With the Codex, it can adjust or correct until the output is suitable.
I'm sure it depends on the field and the method.
Re: 2x, not 10x: coding with LLMs in 2026
#124Sometimes small tools are all you need, even if they're slop.
About a month ago I pirated an Argentinian movie and the only subtitles available in my language were out of sync and at a different speed/framerate so adjusting for delays wasn't enough. I was unsuccessful at fixing it with VLC and every other "online tool" I could find.
Knowing a srt file is just text with timestamps I vibe coded a python script to take in sample times throughout the movies so it could recalculate the rate and shifts and replace them in the file. It worked on the first try using only the deepseek web chat interface and my terminal.
Without AI I theoretically could have sat down with a pen and paper to figure out the math adjustment, then looked up python input handling syntax which I already forgot, typed something out and then hammered it into shape through trial and error over a few hours. But the friction and time investment of doing that would have been so great I would have just given up on watching the movie instead.
Re: 2x, not 10x: coding with LLMs in 2026
#125I was saying it's a 3x multiplier for me during the Opus 4.8/GPT 5.5 era. Now, I don't think I'm notably faster than that, but I'm taking on projects I couldn't have done without LLM assistance, and not just because of time, but because of knowledge.
It's become a lever that allows me to learn as I go while tackling problems that are far beyond what I already know. I know a lot, I've programmed for decades in a lot of languages in a lot of domains. But, with the current crop of the best LLMs, I can reach for bigger problems...and actually make progress. I've read books about DSP for audio, and have done little toy projects in the past, but with LLMs, I'm building complex and working synthesis engines in a couple of weekends. That would have required a few months of study and experimentation before. So, that's a huge multiple. Like 100x.
For things in my area of expertise? Probably still just 3x, maybe 4x, because it makes fewer mistakes I have to fix in code review. It still writes terrible docs, as the post mentions...they don't understand user desire, so they simply can't write documentation for a user to actually use. You can't prompt them to make really good docs, but I can usually prod their docs into coherence a bit faster than writing it myself.
But, it's the lever for doing things I've never done that is such an addictive thing. Which, I imagine is how non-technical people feel shipping their first web app or whatever with these things. For basic work, I think we're at a point where almost anyone can use an LLM to make working software (not necessarily secure or stable software, but working). But, I think we're also at a point where an expert can make that lever really do something, and I hope that means we'll begin to see extremely ambitious new software in addition to all the throwaway junk that's been proliferating at a frightening pace.
In short: Maybe don't make another fucking "memory" system for your chatbot so they can be your friend who remembers your birthday, and instead work on something meaningful.
Re: 2x, not 10x: coding with LLMs in 2026
#126While I agree with the premise, I think this angle only applies on work one was going to do no matter what. The real power of these tools is that there are so many ideas people would like to try, but never have the time or motivation to pursue. So the comparison is not only "built with and without LLM" but "would you even build this if you didn't have the LLM?". The gap in productivity in this case is much more wide.
Yes but I’ve seen some devs waste a lot of time using AI to build something that was a bad idea to begin with. Without AI they might have first spent more time validating the idea was worth it.
For me it is exactly the opposite. AI makes it so easy to create test fixtures and run experiments that I now spend much more time validating ideas than I could ever afford to do before. When I write code for production, it's not "this ought to work", it's "here are the figures showing how well this works, on this dataset", where the dataset is also much larger than anything I would have used before, because I used AI to generate the tools which collected and organized it.
Re: 2x, not 10x: coding with LLMs in 2026
#127[flagged]
Re: 2x, not 10x: coding with LLMs in 2026
#128I'm probably a weird outlier. Coming from academia, it's ranging from 1x to infinity-x (as in, certain tasks wouldn't even be touched if not for AI). For stuff that I'm used to (R) I can write nice and compact spaghetti (long %>% pipes). I'm not comfortable when a working script doesn't fit the screen (plus a few scrolls max). My style is probably easy only to me. When I teach, I don't teach it in particular. AI give…
Infinity is a ridicilous statement. Even the work that wouldn't have been done before is now time away from work that needs to be done. At worst it's negative x.
It's a division by zero. We can change infinity to 'undefined'
Re: 2x, not 10x: coding with LLMs in 2026
#129I'm probably a weird outlier. Coming from academia, it's ranging from 1x to infinity-x (as in, certain tasks wouldn't even be touched if not for AI). For stuff that I'm used to (R) I can write nice and compact spaghetti (long %>% pipes). I'm not comfortable when a working script doesn't fit the screen (plus a few scrolls max). My style is probably easy only to me. When I teach, I don't teach it in particular. AI give…
I have many problems dealing with statements that use Infinity, especially if I think about the math or physics implication of that :)
Re: 2x, not 10x: coding with LLMs in 2026
#130I'm probably a weird outlier. Coming from academia, it's ranging from 1x to infinity-x (as in, certain tasks wouldn't even be touched if not for AI). For stuff that I'm used to (R) I can write nice and compact spaghetti (long %>% pipes). I'm not comfortable when a working script doesn't fit the screen (plus a few scrolls max). My style is probably easy only to me. When I teach, I don't teach it in particular. AI give…
I would have said 0x to infinity-x. Some stuff I think it would be a good idea and waste a bunch of time and then give up and just write it myself. And like you say, there’s lots of stuff I would just never have done, basically any kind of front end for example, if I didn’t have an LLM to do it. I also do mostly research or one-of script development, for which I find LLMs less useful, as I want simplicity and I want…
For me, I find the biggest benefit is that I actually do all the things I know I’m supposed to do, but usually don’t bother if I’m the only audience.The first time a model asked me, “do you want me to wrap this in an easy to use CLI?” I was like, “Yes. Yes I do.”