Vibe coded plugins

[] Peter:H []

Microsoft github Copilot dropped Claude Opus 4.5, 4.6 from Pro "Plans". Pro+ can use Opus 4.7 with a 7,5 cost-factor Seems to be a coordinated attack on our purses. I'll give Mistral a try and if it works I'm going to cancel github copilot

[whassup]

Stay tuned.

This.

[ghettosynth]

Microsoft github Copilot dropped Claude Opus 4.5, 4.6 from Pro "Plans". Pro+ can use Opus 4.7 with a 7,5 cost-factor Seems to be a coordinated attack on our purses. I'll give Mistral a try and if it works I'm going to cancel github copilot

https://opencode.ai/zen

Zen gives you access to a curated set of AI models that OpenCode has tested and benchmarked specifically for coding agents. No need to worry about inconsistent performance and quality, use validated models that work.

With opencode they regularly rotate through different providers who give free access for some time. The real value with opencode though is that you can define per agent model use which gives you tighter control. Since it's open source and highly configurable, you can build a long term workflow around it. You can get started with $20 to see what api costs of the various lower cost models might eat up for your development projects. It gives you insight right in the TUI as to what your ongoing costs are each session.

[Urs]

Obviously, if they ever want to become profitable, they have to charge big bucks. Currently the gamble is a) make people depend on this stuff and b) nobody can offer a cheap solution. If anyone coughs up a decent LLM that runs locally, all these big AI companies are bankrupt in an instance. So they need to inflate their importance, to retain their value and to try and test waters with profitable pricing. Not even 100 bucks a month is gonna make them profitable, more like as much as that dev costs that you're trying to replace.

[ghettosynth]

Obviously, if they ever want to become profitable, they have to charge big bucks. Currently the gamble is a) make people depend on this stuff and b) nobody can offer a cheap solution. If anyone coughs up a decent LLM that runs locally, all these big AI companies are bankrupt in an instance. So they need to inflate their importance, to retain their value and to try and test waters with profitable pricing. Not even 100 bucks a month is gonna make them profitable, more like as much as that dev costs that you're trying to replace.

I think that's a little too much of a neatly packaged narrative. Local LLMs will not compete with frontier models any time soon at what they do best, they simply are not going to be able to run in the memory that most people can afford, even assuming that you can get your hands on one. The discussion on reddit hits a little closer to reality from my POV. The enterprise will pay the API rates and, long term, those are the customers.

Over time, inference costs will both go up, and go down. This is already happening. Larger frontier models mean they go up, gains in efficiency and the need to use existing hardware combined with the fact that older models aren't useless, and competition, mean that they go down.

The mid tier models that are still too large for most users, but cost significantly less, are perfectly fine for many coding tasks. Probably the limitation that you run into is whether they are sufficient for the kind of highly autonomous tasks that some people have set their sights on, and that's less likely. Local models are fine already for low level tasks. Hence, the challenge is in building workflows that deploy models effectively to the needed scale. This favors an interactive approach over a highly autonomous one.

Competition cannot be ignored, especially at the mid-tier. The Chinese labs have open sourced many of their models so that providers can run them at scale. This is for their own reasons, and one shouldn't expect the practice to continue.

In my view the discussion around this in the public sphere is overly simplified. What cannot be lost sight of is that coding is a done deal in terms of usefulness, how it's applied is really the open question, how it fits into workflows, how it maximizes the human labor around it.

[akadajet]

I use LLM tools, but I try my best to not be a "vibe coder" about them. I actually read the output and make adjustments.

[akadajet]

I think of vibe coders as people who just blindly make slop putting in the minimal amount of effort.

[vitocorleone123]

I am definitely not putting in the minimum effort into the vst3 MIDI editors I’ve designed! I think of it more like outsourcing development while I design, manage, and test. And, soon, market.

[_leras]

Let's see ... I did not plan to derail the good discussion in this thread to my specific example, but share my subjective opinion and approach to the vibe-coding subject. Nevertheless I now kind of feel inspired to start the journey and see how far I can get with my example Stay tuned.

I don't think you derailed anything. I think people read the hype and think we're further ahead than we are.

Pretty sure if you type 'make me vst clone of a Moog synth', then you won't very far...

[oobesan]

Why hasn’t a mod moved this stupid thread to the AIcirclejerk forum where it belongs?

[_leras]

Because I don't want to single out developers when I don't know for a fact how the product was developed. I see some basic commonalities though:

• One-man company

• First release by the developer

• Claims to have created the plugin for personal use

• Niche product, non-emulation, low ROI potential

• Somewhat immature UI

• Product updated almost daily based on the feedback of users and testers

Not saying these developers don't have some technical acumen. But bringing new, novel products to the market in this age with such a rapid pace of updates would indicate to me that likely most of the plugin is coded by AI.

Isn't that list like anyone's first plug in?

[jamcat]

* The Alternative: Everyone's new best friend is Codex. A whole sub-discussion broke out, with many users saying it's a great replacement with more generous limits for the same $20 price tag.

Isn't Codex and ChatGPT the same thing as of ChatGPT 5.4?

[Tiles]

Obviously, if they ever want to become profitable, they have to charge big bucks. Currently the gamble is a) make people depend on this stuff and b) nobody can offer a cheap solution. If anyone coughs up a decent LLM that runs locally, all these big AI companies are bankrupt in an instance. So they need to inflate their importance, to retain their value and to try and test waters with profitable pricing. Not even 100 bucks a month is gonna make them profitable, more like as much as that dev costs that you're trying to replace.

I think that's a bit too absolute. I have good experiences with Google Gemma offline. But it's not Claude by any means. For me Gemma is a complement to Claude, not a replacement.

Same goes for the offline AI movie makers. And here the gap has become bigger again, not smaller. Wan has stopped to offer the newest model as freeware. And the only other powerful open source model that you can use offline, LTX Video, is not on par with it.

So the market looks good for the commercial Wan version. And people really pays for it. I see it happen in the discords where i am active in. It's a little bit like Open Source vs proprietary solutions. Open Source will you just bring so and so far. Money is the limiting factor. And people want the best.

Of course, dependency-building is real. Companies like Microsoft, Google, and Amazon have done this for years. AI is following the same playbook: cheap entry, then monetization. And that's what we see. The companies starts to monetize now. The price for Claude is exploding as we speak about it.

But a strong local LLM wouldn’t instantly bankrupt big AI companies. Models are only one part of the stack. Training, infrastructure, scaling, compliance, and enterprise support still cost a lot and are hard to replace locally.

[] Peter:H []

Obviously, if they ever want to become profitable, they have to charge big bucks. Currently the gamble is a) make people depend on this stuff and b) nobody can offer a cheap solution. If anyone coughs up a decent LLM that runs locally, all these big AI companies are bankrupt in an instance. So they need to inflate their importance, to retain their value and to try and test waters with profitable pricing. Not even 100 bucks a month is gonna make them profitable, more like as much as that dev costs that you're trying to replace.

But a strong local LLM wouldn’t instantly bankrupt big AI companies. Models are only one part of the stack. Training, infrastructure, scaling, compliance, and enterprise support still cost a lot and are hard to replace locally.

Strong local LLM and strong Hardware. In the course of recent events, where github crippled it's copilot "plans" for individuals I looked into none-apple unified-memory hardware. And there's for instance ryzen al max+ 395 which can have 110 of 128 GB mem for models when run under linux. I think with 110 GB you can run rather good models even better ones if you apply optimizations, but what is the price tag? Now it's about speculating about a few developments in different areas ... dram price, improved competition in unified-memory offerings, llm subscription price, llm resource optimization research ... depending on your assumptions, it might turn out that investing in the right hardware might be worth it. I think prices for llm "plans" will get up considerably, that makes buying capable hardware an option.

[Tiles]

Unfortunately hardware prices have increased dramatically in the last months. I would need more Ram, but that's out of reach now.

I think we will see both. It is not either or.