It definitely changed how I get into flow state for me. But music still works, if not even better when coding with AI (listening to: techno, electro, edm). Generally my flow is to sit down, make a small plan of what I will work on, fire off 2 agents to work on different parts of the code that are lower hanging fruits (takes 2-10 mins for them to complete). Then while this is busy, map out some bigger tasks.
Agents finish, I queue them up with new low hanging fruits, while I architect the much bigger tasks, then fire that off -> Review smaller tasks. It really is a dance, but flow is much easier achieved when I do get into it; hours really just melt together. The important thing to do is to put my phone away, and block all and any social media or sites I frequent, because its easy to get distracted when agents aren just producing code and you're sitting on the sidelines.
That's not the same flow state experienced by programmers.
While programming, it's possible to get into a trance-like state where the program's logic is fully loaded and visible in your mind, and your fingers become an extension of your mind that wire you directly to the machine. This allows you to modify the program essentially at the speed of thought, with practically zero chance of producing buggy code. The programmer effectively becomes a self-correcting human interpreter.
Interrupting someone in this state is incredibly disruptive, since all the context and momentum is lost, and getting back into the state takes time and focus.
What you're describing is a general workflow. You can be focused on what you're doing, but there's no state loaded into memory that makes you more efficient. Interruptions are not disruptive, and you can pick up exactly where you left off with ease. In fact, you're constantly being interrupted by those agents running in the background, when they finish and you give them more work. This is a multitasking state, not flow.
So the article is correct. It's not possible to get into a flow state while working with ML tools. This is because it is an entirely different activity from programming that triggers different neural pathways.
Nice way of gatekeeping "flow state of programmers", considering I have been a software engineer for the past 2 decades. So I have experienced quite some flow states in my life, and this is no different.
Not gatekeeping, just pointing out that they're different activities with very different experiences. What you're talking about and what the article is talking about are different things.
When using ML tools you have no deep understanding of the behavior of the program, since you don't understand the generated code. If you bother to review the code, that is a huge context switch from anything you were doing previously. This doesn't happen during deeply focused programming sessions.
You may have been a software engineer for decades without ever experiencing the programming flow state. I'm not passing judgement.
I have been in software for 20 years, and was just about to quit 2-3 years ago because of how mundane things became. And now I am actually loving it again because of AI. I'd say, AI writes 95% of my code, and I use it for 75% of the decisions during working on a project.
I am under MUCH more pressure to deliver more in shorter periods of time, with just me involved in several layers of decision making, rather than having a whole team. Which may sound scary, but it pays the bills. At one company I contract with, I now have 2 PMs; where I am the only dev on a production app with users, shipping new features every few days (rather than weeks).
It feels more like performance art, than it even feels like software development at this point. I am still waiting for some of my features to come crashing prod down in fantastic fashion, being paged at 3am in the morning; debugging for 12 hours straight because AI has built such a gigantic footgun for me.... but it has yet to happen. If anything I am doing less work than before - being paid a little more, and the companies working with me have built a true dependency on my skills to both ship, maintain and implement stuff.
Would you mind sharing your setup (LLM model, IDE, best practices)? Personally, I'm struggling to get value out of Continue.dev in VSCode (using Gemini 2.0 Flash by default, with the option to switch to more advanced models). I still revert to pasting code into ChatGPT chat window (using the website), frequently.
Are you using agentic features, given that you have not just one but two PMs?
The biggest tip I can give you is to stay in the framework you are most convenient in, and have the most experience. Start building stuff the way you would by yourself, but then start delegating the repetitive tasks to an agent. My best recommendations would be using Cursor in Agent mode, and switching to VScode in Agent mode when your credits with Cursor run out. The reason why I like Cursor more is because of the Checkpoints. And VScode Copilot Agent checkpoints suck; but you can still use git to create your own checkpoints (git add / git stash, etc).
I don't even use completions, really just agent mode. I do planning, wireframing, creating specs all with agents. Even small MVPs created in 5 minutes, deployed in 10, during a meeting to just brainstorm. As for the models. Go with Claude 3.5 or 4.0, GPT5. Use sequentialthinking and Taskmaster MCP. I could write a book about it... but the best way to go about it is to dive into, get frustrated, push through and then learn it the hard way. I started delegating a lot of my programming work the day ChatGPT came out; just copy and pasting, and since that day, my reliance on AI has just been increasing, and I have been getting better at it (and now I am at this stage.. with 2 PMS).
Also to add another point is that if you felt like an agent did not help you correctly, or way overshot, did too much edits, etc. Go back to the original prompt, rephrase it - sometimes you need 1-2 times. Sometimes the model just don't work for your workflow. It can become quite delicate.
One of the bigger things is when you introduced some bug, start working backwards with the agent, simplifying whatever you build to its bare necessities, and the moment it dissapears, start a new chat, and build it back up to what it was before (in the desired non-bugged state). This often works if you then also switch to a completely different model.
Not OP, but regarding your situation, I suggest moving to an agentic solution instead of “copy-pasting to GPT” — this will boost your coding productivity. There are several tools available, and to each their own, but try out Claude Code.
I was thinking of doing something similar. I think I’m well positioned for this as I have a natural ability to juggle many contexts, I used to run a software agency, and I’m pretty good at architecture early on which means solutions come out more robust and flexible. I have had really good experience with AI tools and I’m constantly evolving my workflows.
I land most of my clients by maintaining my blog and a github with open source projects. I have build a lot of general purpose MCPs and quite some tools, which are all written by Claude (3.5 and 4.0) and now GPT5. On my blog I just blog together with AI. It sounds silly, yes... I don't want to share it here publicly, but it looks good and it gets me people in my inbox (email/linkedin).
So I post on LinkedIn & Reddit, and I am not doing it in a spammy way. Do some outreach through LinkedIn and post on YCombinator on my personal account on the monthly who's hiring/freelancing posts. But a lot of the traffic I get comes from organic search and reddit -> clients. I had a client who told me they found me on Twitter; but I never even posted there, so someone reposted an article.
I usually do write the whole article. Maybe I spend an hour on it. Sometimes even much longer. And then I have a way of rewriting it with AI to improve the writing style. Which I then proof read and keep improving. I do this because reading back what I wrote feels even worse than listening to my own voice. It just gives me a visceral reaction lol. But yes this worked. I always wanted to write and blog prior to AI, but my aversion to proofreading my own writing stopped me from doing so over a decade (dozen actual genuine attempts).
Google does not mind. I rank quite highly for some niche keyword on LLM programming.
1. I love AI/ML hearing about stuff, and seeing it boom this much is great.
2. I really do enjoy working with LLMs and seeing what they can do.
3. It is quite amazing what non-technical people can now do with AI Assisted coding.
4. Working with LLMs within IDEs is getting quite good too.
I understand that there are still people who are not buying it, but quite honestly, its becoming harder to side with them. I have been in Software for 2 decades, and this "craze" has given me the most amount of enjoyment I have gotten since I figured out how to build a website sometime in my teens!
I think you're missing the general complaint. It's not against any of the the things you like, it's the over-saturation of not only AI, but a very small segment of AI.
Perfectly fair. It's not like YouTube is some free open source platform. Infra needs to be paid, creators need to be paid, they have a whole eco-system. Why not just pay for premium if you use it that much?
I still don't understand how cursor is making any money at all. I spend so much time inside cursor, that I am spending 10-20$ per day on additional requests. Now if I connect model provider APIs to windsurf, I'd be spending upwards of 100$ due to amount of tokens I use through the API per day. And if I connect my own API key to Cursor, I immediately get rate limited for any request, because I go well above 50 per minute. And I did try claude code, but its just not on par with my experience with Cursor.
I could probably go much lower, and find a model that is dirt cheap but takes a while; but right now the cutting edge (for my own work) is Claude 4 (non-max / non-thinking). To me it feels like Cursor must be hemorrhaging money. The thing that works for me is that I am able to justify those costs working on my own services, that has some customers, and each added feature gives me almost immediate return on investment. But to me it feels like the current rates that cursor charges are not rooted in reality.
Quickly checking Cursor for the past 4 day period:
Requests: 1049
Lines of Agent Edits: 301k
Tabs accepted: 84
Personally, I have very little complaints or issues with cursor. Only a growing wish list of more features and functionality. Like how cool would it be if asynchronous requests would work? Rather than just waiting for a single request to complete on 10 files, why can't it work on those 10 files in parralel at the same time? Because now so much time is spend waiting for the request to complete (while I work on another part of the app in a different workspace with Cursor).
> I still don't understand how cursor is making any money at all.
They don't make any money. They are burning VC money. Anthropic and OpenAI are probably also not making moeny, but Cursor is making "more no money" than others.
For OpenAI: short answer is no. From what I've seen, their biggest expense is training future models. If they stop that (putting aside the obvious downsides) they'd still be in the hole for a few billion dollars a year.
edit: Well, if they shed the other expenses that only really make sense when training future models (research, more data, fewer employees ..) they would be pretty close to break even.
The market for AI-assisted development is exploding and token costs are plummeting all the time. It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.
There is no loyalty or lock in though. There is little real uniqueness. And everyone in AI is trying to make everyone else on AI the "commodity complement"
It's like a horse race.
But yeah enjoy the subsidies. It's like the cheap Ubers of yesteryear.
This is exactly it. Selling the output of a LLM is going to an incredibly cut-throat and low-margin business.
The more interesting, novel, and useful work you wrap the LLM in the more defensible your pricing will be.
That said I think this can describe a lot of agentic code tools - the entire point is that you're not just talking to the raw LLM itself, you're being intermediated by a bunch of useful things that are non-trivial.
I see this with Anthropic most - they seem to have multiple arms in multiple lines of business that go up the abstraction ladder - Claude Code is just one of them. They seem to also be in the customer service automation business as well.
[edit] I think a general trend we're going to see is that "pure" LLM providers are going to try to go up the abstraction ladder as just generating tokens proves unprofitable, colliding immediately with their own customers. There's going to be a LOT of Sherlocking, and the LLM providers are going to have a home field advantage (paying less for inference, existing capability to fine-tune and retrain, and looooooots of VC funding).
This may be old fashioned thinking and the automated loom might come get me but I think traditional software products with enthusiastic customers, some kind of ecosystem will benefit with AI being used.
However they will benefit in a way like they benefit from faster server processors: they still have competition and need to fight to stay relevant.
The customers take a lot of the value (which is good).
While there is a lot of fear around AI and it's founded I do love how no one can really dominate it. And it has Google (new new IBM) on it's toes.
It's hard to add sophisticated abstractions though, because they are all selling text by the pounds (kilos?). So it feels the same as vendor lock for a cucumber seller, doesn't it? The seller can sell you an experience that would lock you in, but aside from it there is no moat since anyone can sell cucumbers.
To try and give examples: an autonomous agent that can integrate with github, read issues, then make pull requests against those issues is a step (or maybe two) above an LLM API (cucumber seller).
It doesn't take much more of a stretch to imagine teams of agents, coordinated by a "programme manager" agent, with "QA agents" working to defined quality metrics, "architect" agents that take initial requirements and break them down into system designs and github issues, and of course the super important "product owner" agent who talks to actual humans and writes initial requirements. Such a "software team system" would be another abstraction level above individual agents like Codex.
When people talk about how sophisticated hierarchical agent swarms will be built up that perfectly reflect existing human social structures I'm reminded distinctly of all the attempts to build DDD frameworks for modeling software, and then the actual result is that software went in the opposite direction - towards flattening.
As native LLM task completion horizons increase another order of magnitude, so much of this falls out of the window.
This exactly. I built CheepCode to do the first part already, so it can accept tasks through Linear etc and submit PRs in GitHub. It already tests its work headlessly (including with Playwright if it’s web code), and I am almost done with the QA agent :-)
You nailed it. I imagine this is why OpenAI is looking to develop hardware. Right now, I have tabs open for ChatGPT, DeepSeek, and Gemini. I have zero loyalty to any of them. But if I owned a piece of hardware, I immediately am locked into that ecosystem.
My second bet is on Google (for general-purpose LLMs in general) - not because of any technical advantage, but because they have a captive audience of large organizations using GSuite that would be happy to just get Gemini on top to satisfy need for AI tools, instead of having to jump through the hoops of finding another provider. Sales, sales, sales.
Do you mean AWS? They're competing with half a dozen or more hyperscalers now. Cloud infrastructure components are so heavily commoditized now, many of them have open source solutions with compatible API's. (Think Minio)
There is a loyalty if they keep winning, if they stop running their competitors will beat them. I don't switch between cursor or windsurf daily, i keep cursor only even if windsurf has some marginal improvement in workflow as i know cursor will have them in short time. No need to switch, But if they stop improving they will get eaten away. They have already taken lot of developer market share away from vscode and vscode copilot.
Inference cost is plummeting. It’s like the cheap Ubers of yesteryear, if the cost of hiring a driver dropped by a factor of a thousand in the past three years.
I use Aider with Openrouter and I keep wondering about the pricing of LLMs after providers decide to be profitable. Can we still afford a model which knows Python, Java and how to disrupt snail biology without poisoning mammals?
Yes. It’s already profitable to run inference at today’s prices. AWS isn’t subsidising you when you buy compute from them. And inference cost is declining steeply.
> The cost of LLM inference has dropped by a factor of 1,000 in 3 years.
But, we need a future where unlimited inference, in parallel is profitable. It is not: even less than cloud compute (where it is terrible also), when I buy 500 flimflams for $50/mo, what did I buy exactly? As currently it seems to depend on the position of the moon: one time 10 prompts make what I want, sometimes 100 prompts keep looping over the same issue unable to fix it (like a typescript type issue which takes me 1 seconds, llms, the flagship ones, can easily burn 100 prompts and not fix it). I do very much NOT want to pay for those 100. I see 'vibecoders' aka people who cannot code, burn through all Tokens for the month without having anything working in a single day.
A bit of an out of context reply for me to jump in here, but in the abstract, it can be a reasonable question to ask if infinite usage is affordable. Maybe not infinite without constraints… but as an example from the past there are many mobile phone plans that have “infinite” calls and texts for an affordable monthly cost. There would’ve been a time where asking if unlimited calls would be affordable would’ve sounded insane, but now it’s fairly normal.
The answer to that depends on when the VC bubble bursts- if it lasts long enough costs will eventually drop far enough. Pets.com was a .com-boom era joke but today I actually buy my pet-food online and I'm pretty sure nobody is subsidising me doing that.
Commercial animal feed is subsidized. So are some forms of human food in many countries.
Pet food is not subsidized in my country nor the EU. If any countries do subsidize pet food, they are the exception. Maybe the US? Pet food is often manufactured from the waste of other processes, including the human food industry, but that is not a subsidiary.
I understand that it is not directly subsidized. However the sources it comes from while are the "waste" of a greater product. That greater product is heavily subsidized.
This also goes to a personal issue that why would you feed your pet a waste product. My dog gets food I cook for him just like myself. There are tons of crock pot recipes online for safe cheap high quality dog food.
That depends, there are specific diets for, say, cats, and it not only needs to be prescribed (to be able to purchase), but it costs a fuckton of money.
<rant>
Think of it like this: imagine if lactose-free or gluten-free food could be bought only with a prescription. Sadly the prices are already high as it is for gluten-free, but I would rather not get into the reasons here. :)
My girlfriend (LA, US) just left 1k USD on 2 visits to the vet with her cat, for some cheap ass antibiotics, and a "specific prescription-only food". Crazy. All that would have been "free" (not quite, but you know) or at a very low cost for humans around here, in Europe. Not the gluten-free food though!
> It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.
Switching costs are zero and software folks are keen to try new things.
The time it would take me to switch IDE and work process and learn the best prompting style and idiosyncrasies of a new model (and do some testing to build confidence) would be half a day, at very least.
That makes the opportunity cost of switching significant.
This won't create race conditions all of them will know the others live accept of commit to directory? or have to wait then hit enter on other tab instantly after?
I still worry this could cause issues even with file locking, if one is reading a file to understand how something is structured and another tab does a refactor on that moments later…
they say they are making hundreds of millions, but they never say how much of it is going to GPU cost. If I had to guess, they are burning everything and far from being profitable
If history has taught us anything, it's that unless full accounting data is released, there is a reason that full accounting data is not being released, and that reason would almost certainly paint the company in a bad light.
GPU type and utilization mean that the costs likely rise only logarithmically or sub-linear. If you commit to buying enough inference over long enough, someone can buy a rack of the newest custom inference chips and run them at 100% for you, which may be a lot cheaper per request than doing them on a cpu somewhere.
I disagree tbh. I mean, I accept that new silicon will have better power usage and probably be more efficient in terms of flops/Joule, but there would need to be a major technical breakthrought to get a logarithmic relationship between N requests and inference cost. N requests at P flops, still means I need C x P flops for C x N requests. A not-so-steep linear relationship is still linear.
I meant you're using Cursor (I assume Claud as the model) and also wondering that you go over monthly usages and then what product/project is that where you find it so useful that you have such a usage.
I have used or rather use Claud with CoPilot and I find it pretty useful but at times it gets stuck in niche areas.
