I get kicks out of developing novel solutions to unsolved problems. I might be wrong, but I like to think it's the kind of stuf LLMs are not yet great at. Call me selfish, but I now have absolutely zero interest in sharing my findings with the world because of LLMs, since once I do, LLM companies can start selling my work.
LLMs evened the battlefield for all that has been developed so far, but what about future discoveries? I think it remains to be seen if LLMs can come up with these novel solutions rendering innovators useless. If not, then LLMs might be crippled by software becoming more closed in the future.
There was a level of patent/copyright protection, there was a level of awareness of stealing, a company could be taken to a court of law. A writer of a PhD thesis could show that someone lifted huge parts of a dissertation without attribution. All that is gone.
They could, but I could always hope they wouldn't, and most of the time I wouldn't be disappointed. AI companies don't even have means to attribute to someone the solution they're selling.
Effective use of frontier models involves "driving the build". They can't really do it themselves, human is still involved to tie break complicated tradeoffs and take on deployment, security, financial and other risks. The innovators are not becoming useless, they are offered a choice to 10x their output and speed and quality of execution. Execution still matters, it just became different.
Not centauring for your AI or training dataset. Sorry. What the current AI industry comes down to is a tollbooth on " interesting" research work, where the people hosting the infra are just waiting for monetizable opportunities to float in to be scooped. The absolute same business model as Amazon Basics as applied to intellectual work.
Let it rot, and the bubble pop from distrust. The people running this shit have absolutely no one's best interests in mind, and I'm not about to sacrifice my expertise to keep your ledgers looking promising.
Good list. And if you support both portrait and landscape, you need to support twice that, and you need to be able to switch on the fly.
But for the apps that don't currently support both orientations, it's possible they don't have support for switching aspect ratio on the fly. This design might help with those apps not completely breaking when folding/unfolding.
Yeah to be fair I'm all for the 1/sqrt(2) aspect ratio, I just think it's funny that it's treated as some kind of holy design constraint when they're all over the map on phone sizing :)
And the extra irony is I'm currently shipping an app that doesn't deal well with different size phones (just pads the edges), heh, so do as I say, not as I do.
I'm a coder and a daily driver of ChatGPT and I don't really know what Sol is or how I can access it. I follow discussions here and know of its excistence, but I can't remember its relation to other models or tiers by heart. What ever the default thing the website is giving me has been good enough for my needs for years.
That's what I do. But I don't write enough code these days to spend time on anything more complicated. For the odd shell script or python program I have ChatGPT write it, then I copy/paste it, make it actually work, and deploy it.
Lots of companies have security policies that forbid installing a local agent/harness. So employees still only have access to the web chat interface, and copy paste content yes.
Yes! Though I mainly use it for one-shotting full files, so I'm not copying over small diffs. I did try copilot at some point in vscode, but after a week or so I felt like it was slowing me down. Something I knew would be a simple and quick change was now slow since I didn't have enough ownership of the code to do it quickly myself, and prompting for the change took longer than changing it manually if I had had full ownership. So I went back to coding by hand things that aren't a one-shot file.
Thanks for the reply. I'd suggest you improve your workflow with better tooling like Codex or Claude code, and if you have some kind of weird constraint they'll easily follow an AGENTS.md with that.
Copy pasting ai code from the browser is not perfectly acceptable. The only reason to do it is because you haven’t tried agents. It shows a lack of imagination and a willingness to settle for inferiority.
“Terrible inefficient workflows”, like having to prompt Claude to change one or two words in a file rather than just manually editing it, because it gets confused when the file changes out from under it and will later overwrite the changes?
That's what I used to do (copy and paste from web ui and with copilot in my code editor).
I started using claude code a couple of months ago and now almost exclusively use it. My workflow is mostly creating a md file describing what I want then telling claude code to look at it and implement it. I simply use the default model 95% of the time.
Using ChatGPT.com for coding tasks has to be a joke in 2026, right? I assumed everyone is using Codex/Claude Code/OpenCode/Pi by now if you’re using these tools.
Yeah! Seems like that's the assumption in this ongoing discussion, so I thought my point of view would be interesting. I'm not saying I'm doing it the right way, possibly I'm being very stupid, but the point here is that it's not just the "normies" who don't know what Sol is.
Lots of large enterprises and organizations have security policies that make it very hard to self-install these tools and haven’t figured out they should add them to their approved software list yet. Federal government in particular is still very slow
This might depend what you mean by code. In the sysadmin space I ask a web interface for poweshell to generate specific information multiple times a day, and pasting it into a terminal is really the cleanest workflow on a corporate machine where you have to justify running an executable.
There are ways to make copy-paste ChatGPT usable with things like AI Badger. Also, ChatGPT web is unmetered, so you can do review/planning work and only delegate the actual implementation to Codex.
I code by hand (sigh) and copy paste some snippets once in a while from a web chat LLM. This also means every single line of such snippets gets reviewed and anaylzed. Usually, these are some arcane Win32 API usage where otherwise I'd have to dig into old forums. I don't want to become a prompt engineer so I stay away from "agentic" workflows.
As an engineer, the enjoyment from the process is not less important, or should I say more important, than the enjoyment from shipping a final product.
That's probably the single thing that I found to be biggest difference between those who are very pro AI vs. those more reserved or negative.
A colleague has openly stated that he very much does not care for the process, he just want the product. Where I only care about the product in the sense that I have to, because it's my job and for hobbies I pretty much only care about the process. So an LLM pretty much doesn't make sense, because it help by removing the part I care about. They are good tools for debugging and if you're hopelessly stuck on a detail of some weird and obscure API or configuration.
As an engineer, I know what I'm getting paid to deliver is the work product. I don't have too much attachment to any particular set of tools to get there.
If we follow that logic, I believe it means that you would be quite happy in the Product Owner role, or in the role of a client who outsources the actual development activities. I wouldn't.
Its mostly used with Codex so you can give finer control over task delegation.
its definitely worth learning the different strengths and weaknesses for each one and which you should use for planning/building/testing/documentation updates etc.
No point in burning excessive tokens using a high tier/expensive reasoning model like Sol/Xhigh if just updating a readme file.
I'm not sure about the listed facts... I opened up Helsinki where I live, and got this: "Every residential building has a sauna -- it's not a luxury, it's basic infrastructure, like having hot water."
Certainly it's very common, but certainly not "every" residential building has one.
I looked at some other cities I know a bit. Those "facts" are typical AI slop. Not always completely wrong, but stereotypes you will easily find on the internet.
It might be fine for pure navigation to addresses but business addresses and other metadata are often missing or out of date because people only bother to update Google Maps. Also that’s where all the reviews are.
I don’t want to use two different apps for navigation vs finding businesses.
I've used it almost exclusively since around 2019, but that has been in the Bay Area and LA area where the data has presumably been very good relative to other areas.
I had the same knee-jerk reaction. "Did I read that correctly?"
But yeah, I guess it can be used to increase certain aspects of quality by letting them go wild. But I think I mostly hear about security or crash issues. In my experience they don't outweigh the number of other issues they cause. Like UI bugs. I've seen more than one service constantly rolling out features that are completely broken, just to have a completely new, still broken, solution available the next day.
Ok but a task that works fine on qwen 397b can be finetuned on qwen9b. But in every case so far when building the eval for evaluating the traces I’ve discovered a better prompt that closes the gap better than the finetuning.
Compromise is the last tool in the designer's toolbox. Before settling with a compromise, the designer should relentlessly try to narrow down what the problem being solved is, while trying out all possibilities, until it looks like there is no solution that matches exactly the problem (or they run out of time).
More often than not, a compromise means that the problem wasn't scoped out well enough, and it ends up partially solving a problem that didn't need solving.
Edit: I do think it's about trade-offs though. Choosing to solve one problem, acknowledging another problem will be disregarded.
That’s not my experience. Every decision favors one use case or principle over another. The best design is when you have the vision and confidence to say what’s important for each those decisions and what you’re ok leaving out.
It depends on how you define compromise. Is it a compromise to consider things like screen size, load wait times, or the average expected user's attention span? If so, then design is a big juggling act of compromise.
If not, and compromise is just stuff like "add yet another floating ad footer to make the c-suite happy," then absolutely it should be the last tool.
I disagree, we are faced with "true dilemmas" every day numerous times, not even limited to work and creation — e.g. you might be going to a shop to buy crackers, and while you have a preferred brand, a new entrant might be half the price — do you use the opportunity to try the new one, or get the tried-and-true?
These types of things are ubiquitous when designing and building products — how do you communicate to your users, how much you communicate and how much you automate, how do you target slightly different users (new customer vs old one used to old patterns in your product), what use-case is the most important one (by customer type, customer numbers, and how commonly they do something...), etc.
LLMs evened the battlefield for all that has been developed so far, but what about future discoveries? I think it remains to be seen if LLMs can come up with these novel solutions rendering innovators useless. If not, then LLMs might be crippled by software becoming more closed in the future.
reply