1. Personally - these are for apps I use like Codex. This one is pretty simple
2. For our product (which is mostly an internal tool, with minimial customer facing UI) - we have a skills factory that lets our employees create reusable instructions for our internal agent (fully custom built harness with routing). User describes the task and writes the instructions (can attach reference files, etc.). Custom skills and their versions are stored in postgres, with attachments in file storage. The assistant loads those instructions and references when it uses the skill. To update a shared skill, users edit a draft, test it, and submit it for review. Once approved, that version becomes live. Skills can learn or rewrite themselves from conversations (cuts a new draft and prompts the user if they want to update/improve the skill).
EDIT: I should say that our employees have a particular set of expertise and knowledge that make skill sharing insanely useful. Which is why I took the time to build this out. It's helped reduce manual work, and has increased our AI usage drastically. We also measure AI output (in terms of quality) and the reduction in slop has sky rocketed.
Not a great argument. You're describing what LLMs are trained to do, not necessarily every capability or property produced by that training. We do not know what conditions are sufficient for consciousness. Any statements saying it's not possible for consciousness to emerge (or that it hasn't already) in such a system is no more established than claiming that it has
Consciousness can factually exist even though another being’s subjective experience is not directly observable. we can measure its behavioral and neural correlates and infer consciousness from them but we cannot directly observe the experience itself or conclusively prove its absence in another being. I don't think we'll ever get to a point where subjective experience in another is directly observable. But I don't think we need that to happen to prove consciousness factually exists.
For anyone on the fence, I was a hardcore CC user since it was released. I made the full switch to 5.6 sol and Codex about a month ago.
It's the better experience. The limits are way higher (I almost never burn through my $200/m plan), and the output is better than Opus 4.8 (Opus 5 is completely unusable for me).
I had a very similar experience. The combination of Claude going down frequently and the hype around open source models was enough for me to swap my harness from Claude Code to OpenCode. Then it became trivial to test out different models, and now the rest is history. Currently using OpenCode with z.ai and Codex $200/m plans for all my work atm.
Also try a hundred bucks on Openrouter and experiment with the Chinese models like Deepseek, Kimi, GLM. $100 will last a good while, and it isn't tied to a single model or provider.
I burned though $50 on OpenRouter in three days, this is with selective model routing based on tasks/strict AGENTS.md. A Claude 20x sub running Sonnet 5/Opus 5 and light Fable usage gets me through the month. It wasn't even close even though I wish it were.
Agreed, it's not SOTA but I'm happy with the trade off. My preference is to use mostly local models and augment them with Deepseek. It's been a pretty great equilibrium.
I have the $100 plan on OpenCode Zen, which I imagine is similar. I burn through it ridiculously fast. I haven't run the numbers, but it felt about the same as an Anthropic $20 plan
As a counter point, I've been using CC since December and I've never really had any problems with it. It's always done exactly what I want it to and I continue to be blown away by its capabilities, regardless of which of their models I've used. I also almost never burn through my $200/m plan.
Yep, switched from CC to 5.6 Sol. Not using it in my day job (only side projects), but I can crank 5.6 Sol on Extra High "fast mode" and never have to worry about credits.
Not doing crazy multi agent swarms, but have yet to hit any limits during pretty intense weekend sessions.
Antigravity (at least Antigravity CLI which I use) is the CLI coding (?) agent harness from Google that uses Gemini as the model. There are other models available but I haven't tried them.
> This all sounds good on paper but I have an honest question: How do you develop judgement as a junior or mid level engineer to critique AI if you yourself never learned that skill by making mistakes in the first place?
Shouldn't this be a pretty fundamental part of a degree? If you're in school right now for a CS or SWE related degree, I would imagine you're learning how code actually works, how the math actually works, etc.
If an LLM makes a mistake, you should be able to call that out.
You might be arguing that a junior engineer may not have the RWE to make those judgements, but by the time you've completed a 4-5 year degree, you should have done the following:
- Completed coursework where you've learned the fundamentals of programming and software engineering
- Have done dozens of projects (building everything from basic web apps to more advanced pieces of software) where you've seen what works, what doesn't work, etc. This also gives you real world exposure to the latest and greatest frameworks, tools, etc.
- Have done several (at least 2) internships where you've worked at a real company writing real code, and have seen/been mentored into what AI is good at, where it fails, etc.
> Shouldn't this be a pretty fundamental part of a degree? If you're in school right now for a CS or SWE related degree, I would imagine you're learning how code actually works, how the math actually works, etc.
CS and SWE are radically different subjects. Science vs Engineering. I'm sure it varies by school but many CS grads will have approximately zero exposure to engineering concepts or any of the latest and greatest frameworks.
In Europe, informatics degrees, or more specifically technical informatics or engineer-informatics is often the right balance if you want math but also practice. In Hungary they call it engineer-informatics, but officially translate it as Computer Engineering, but it's not really heavily about how to engineer a computer as a machine, though there is some architecture and flipflops etc. of course in the curriculum. It's housed under the electrical engineering faculty, and not under the math and natural sciences faculty, but they still have rigorous discrete math, analysis, graph theory, optimization theory etc. classes.
It always baffled me that US colleges seem not to offer such programs. Or maybe they do, they are just not prestigious enough.
SWE was a sub degree within the school of CS where I went to school. You could do CS with a minor in SWE and get the best of both. I suppose if you're just doing a straight ticket CS degree you'll have little exposure to engineering (I believe when I went to school 13 years ago there were 3 courses over 4 years on engineering within a pure CS degree).
That's sort of broadly true of engineering (including CS if they call it that) at major research universities. They don't really teach practical basics but mostly expect you to pick it up on your own if you don't already know it. They may use Python as a relatively accessible tool in an algorithms class but they mostly don't teach it.
> internships where you've worked at a real company writing real code
Though why would companies make that nontrivial investment if they can use AI instead?
Furthermore, I’d argue that learning good judgement takes around a decade of full-time dev experience. In particular, experiencing the long-term consequences of one’s design and implementation decisions. You don’t get that just by college education and a few internships.
These are two completely different types of mistakes.
LLMs are, by now, pretty good at not making "code doesn't work" and "math doesn't work" kind of mistakes when writing code. These are also "easy" things to get good at, you learn what each part does, understand the abstractions, ensure it makes sense, and go on with your day. Unit testing helps here.
LLMs are not that good at not making "this works wrong" kind of mistakes. Maybe the code compiles and does what it has to do, but maybe it's 2 lines of code with 6 lines of comments (looking at you Claude), maybe it defines three helper functions it doesn't really need, maybe it does something "here" when it should be doing that something "there" instead, maybe it finds itself in a framework and completely disregards how the framework is supposed to do things, etc etc. These are harder things to get good at and you WILL end up with an unreadable mess if you disregard caution and let the LLM go at it.
> but maybe it's 2 lines of code with 6 lines of comments (looking at you Claude), maybe it defines three helper functions it doesn't really need, maybe it does something "here" when it should be doing that something "there" instead, maybe it finds itself in a framework and completely disregards how the framework is supposed to do things, etc.
I would fully expect a brand new college grad to call these things out. There's nothing in here that requires extensive experience to understand. These are basic principles any SWE should know. Nonsense comments are common sense to pull, if the LLM is pulling in a framework you should look into that framework and understand how it works. Don't know the framework during the interview? Say that. Tell the person interviewing you "looks like it's pulling in XYZ. I'm not entirely familiar with that. I understand at a high level what it's doing, but I'd want to dig deeper and understand if the LLM is doing this part right"
> Shouldn't this be a pretty fundamental part of a degree? If you're in school right now for a CS or SWE related degree, I would imagine you're learning how code actually works, how the math actually works, etc.
I’m sorry to say but colleges don’t teach the kinda stuff you’re gonna use day-to-day. They’ll focus on architecture, data structures, algorithms, databases, OS internals theory that only a minuscule number of systems engineers would get to work on. They won’t teach you that in your /login endpoint, if a user is not found, you should verify the password against a pre-computed dummy hash so the response delay matches a real user account workflow to avoid timing attacks. You only learn this on the job under the supervision of a senior mentor.
Depends a lot on the college. I had electives like Java phone app development (J2ME, dating myself), DirectX game engine development, web development, alongside the more theoretical math classes.
> They won’t teach you that in your /login endpoint, if a user is not found, you should verify the password against a pre-computed dummy hash so the response delay matches a real user account workflow to avoid timing attacks. You only learn this on the job under the supervision of a senior mentor.
Actually this kind of timing attacks were taught to me in information security and cryptography class, alongside other side channel attacks.
Agreed. I’m not insinuating that colleges should teach all of these extremely domain-specific things. Most of the stuff would indeed be unnecessary for most students depending on what they do in the future. But that is my original point that you learn this on the job while making a mistake and being corrected by a senior mentor. AI wouldn’t call this out by itself unless you give it a very specific prompt which you also can’t write if you don’t know that these kinds of problems happen in the first place.
Not so sure. AI can flag non-best-practices like this quite easily from generic prompts. You just have to remember to ask it. I mean general review / audit prompts, like "are there any correctness bugs, or things that don't match best practices for efficiency, maintainability, security, or typical UX/API expectations?" And frontier models often flag such things. And it's just getting better. I would not confidently state "AI will never be able to X".
Sure you can do it but I think you’ll agree with what I said about ownership. Whenever I ask AI to audit my codebase, it always plays it safe and gives me a few options to choose from. It just won’t give me a concrete answer unless I insist it to give only one but then I wouldn’t trust that answer that I basically forced out of it.
Yes, I don't know how a junior brain would read it. I already have a map of roughly what kinds of things exist out there, even if my picture is blurry, so when I read a suggestion from the AI, I have a rough ballpark of what it means, and can ask followups and triage what is promising. If you never developed this, probably as a junior you have very little to base such decisions on.
Exactly, I absolutely agree that AI supercharges seniors with judgment and conviction. They’re no longer constrained by how fast they can type.
But this is about what comes next i.e. seniors of tomorrow
If AI gets so good at software that vibe coding is the new norm and as Elon Musk says that AI will generate the machine code directly without any intermediate compilation or interpretation then we wouldn’t need seniors or juniors but I don’t think that’s coming any time soon, if at all!
I doubt AI will ever generate straight up machine code directly as a best practice. Hierarchical, high-level abstractions help even AI reasoning. Re-reading some undifferentiated machine code make it harder to reason about even for the AI, while having a compact representation expressing the high level algo is much more efficient.
Regarding seniors of tomorrow, I think there are big differences between good devs and sub-mediocre ones, and I think their proportion won't change, and the good ones will have the drive to use the AI to learn and understand because they are simply curious and want to know. But most programmers don't care at all, and that's the case even today, and they will be less forced to learn. But I don't think we will lose much with this. Important things are held up by a small proportion of engineers, a Pareto-like principle.
> I’m sorry to say but colleges don’t teach the kinda stuff you’re gonna use day-to-day.
This was not my experience in college. A lot of it was VERY applied. Granted, that was 13 years ago.
But even so, if you're at a college where you feel that you're not getting enough exposure, that's why I also called out "Have done dozens of projects" and "Have done several (at least 2) internships".
Question: for a smaller SaaS tool, or even internal product. If a team doesn't want to manage AWS or another IaaS provider, what are the best alternatives for the following
I have read plenty of snark about them on HN, but I found their product incredibly useful, well-designed, and easy to work with. If I was building a new startup from scratch, I'd definitely be giving them a look.
I'm sure there are plenty of the like 1,000 AWS products that DO has no viable competitor for, but for what they do offer, they're great.
If you are unable to use IaaS directly. You need to accept that your service might be down.
Even if you use AWS and the like, if you aren't building your app with redundancy across multiple AZs, then you'll have some downtime occasionally.
And even if you do build redundancy with multiple AZ, some services might fail anyway as AWS is not entirely isolated. So you might have downtimes.
So just accept downtimes and use the best tool for you (unless they are really bad, like GitHub level bad).
If you cannot accept any downtime, you'll have to spend millions of dollars and months of work to have the confidence to expect no downtime. Something like Netflix's chaos monkey and infrastructure would be enough.
The advantage of going with AWS is that when us-east-1 goes down, half the internet goes down so you don't have to defend why you had a service outage.
Yup. I don't know enough people at giant companies to know how many actually do this though. Not just talking having 2 AZs, I'm talking about ability in a DR scenario to fail over, within 5-10 minutes, to a different cloud provider, e.g. AWS → Hetzner, or GCP → Azure.
My gut feeling is that the number of significant applications that have this capability can probably be counted on two hands. Especially since a lot of the largest footprints of software stacks running in the cloud belong to Google and Microsoft, who I'm pretty sure do not replicate their services into someone else's cloud.
Before the cloud it was commonplace to have redundant data centres from two or more colocation provider companies. Similarly, Internet uplink diversity was commonplace.
An intermediary can provide value but there’s also a risk so I’d consider why you don’t want to use AWS, GCP, etc. directly. All of the major cloud providers have services which are only slightly harder than what Railway does but allow you to grow into more advanced things as your needs expand without adding a third-party who controls your features, security, and availability.
As an example, I note that GCP responded within 7 minutes according to their timeline. If you’d been using Cloud Run, that would have reduced downtime by over 7 hours — and there’s a good chance that you never would have gone down in the first place if the unknown trigger event was related to other customer activity or something odd Railway did.
There’s also a complexity factor: note how much complex infrastructure they mentioned having to fix that you wouldn’t need for your own account. That code does useful things, I’m sure, but it’s also a lot of moving parts which a hosting provider needs and you don’t – this outage took everyone down, whereas individual AWS or bare metal users would’ve otherwise been unaffected. There isn’t a global optimum which is the same for everyone but I think developers are prone to wildly over-estimating how much time they save by removing a couple of deployment steps relative to the direct costs and the less obvious costs of working within someone else’s environment.
This entire thread illustrates why you don’t want Google in any critical part of your business. AWS, sure. Azure? Maybe. I’m not familiar with Azure, but if I have to pick one, it’s AWS.
I love Fly, but their docs are.. tough. They've had multiple iterations of the control plane API, and it's very hard to do things the "correct" way with conflicting official docs.
Depending on exactly what you're building, all of these things sounds like one VPS. A bit of maintenance/security burden managing the machine if you're not used to it but as the others have said: Next.js can be selfhosted, unless you need the serverless/edge stuff; then I would go to Cloudflare Workers.
But really any service (or even on-site hosting) can have downtime, if that's not acceptable then I suppose building/using a tool that can be distributed between multiple hosts located in different geographical areas is the best option.
Haven't used railway but my understanding is they are something similar to Heroku. Fly.io has been pretty great for tiny projects in that niche.
For Vercel if your nextjs site can be compiled statically you could probably throw it up on almost anything. We've self hosted before which is pretty straightforward but you lose a lot of the image optimization stuff unless you go deep into setting up open next.
I love how everyone in Silicon Valley acts like Microsoft doesn’t exist.
Azure!
It’s the enterprise cloud with enterprise support. They won’t randomly pull the plug on your account, unlike companies that have a wildly different cultural background:
Google - ad tech (you’re the product)
Amazon - shop front (you’re a comptetitor)
Oracle - lawyers (you’re a future lawsuit for license extortion)
Generally agree. I use AI very heavily, but rarely am I letting it actually think for me. It's a tool that reduces the time it takes for ideas in my head to manifest into reality. If you don't have those ideas, or a poor understanding of the system the AI is working on, you're going to produce slop. If you can't recognize this slop, you're more susceptible to having psychosis.
Having tried most (all?) of the commercially available + open source options, and even tangential competitors like CC, Conductor, Antimetal, etc. I haven't found anything that's close to the experience of Cursor. The harness they've built is incredible.
I'd even go so far as to say that any competitors that are direct (windsurf, kiro, etc.) aren't even in the same universe. Cursor is just so much better, faster, has better features (plan and debug mode), and squeezes much better results/code out of the same models. They absolutely have some secrete sauce that the other options just don't have.
Cursor is my favorite of the VS forks. Agree that it delivers better plans than others. I prefer using Claude in Cursor over CC CLI when I am heads down going through bugs. I am disappointed in how "little value" in token use Cursor provides compared to others.
Not quite first party, but composer 2 is far superior to grok for coding. Unless you're eluding to them using SpaceX infra to train their own model vs. using grok
1. Personally - these are for apps I use like Codex. This one is pretty simple
2. For our product (which is mostly an internal tool, with minimial customer facing UI) - we have a skills factory that lets our employees create reusable instructions for our internal agent (fully custom built harness with routing). User describes the task and writes the instructions (can attach reference files, etc.). Custom skills and their versions are stored in postgres, with attachments in file storage. The assistant loads those instructions and references when it uses the skill. To update a shared skill, users edit a draft, test it, and submit it for review. Once approved, that version becomes live. Skills can learn or rewrite themselves from conversations (cuts a new draft and prompts the user if they want to update/improve the skill).
EDIT: I should say that our employees have a particular set of expertise and knowledge that make skill sharing insanely useful. Which is why I took the time to build this out. It's helped reduce manual work, and has increased our AI usage drastically. We also measure AI output (in terms of quality) and the reduction in slop has sky rocketed.
reply