SparkyMcUnicorn

That's not the one I'm referring to. See my other comments or your sibling comment.

Gemini 2.5 Pro is the highest ranking model on the aider benchmarks leaderboard.

For benchmarks, either Gemini writes code that adheres to the required edit format, builds successfully, and passes unit tests, or it doesn't.

I primarily use aider + 2.5 pro for planning/spec files, and occasionally have it do file edits directly. Works great, other than stopping it mid-execution once in a while.

My link is to the benchmark results _over time_.

The main leaderboard page that you linked to is updated quite frequently, but it doesn't contain multiple benchmarks for the same exact model.

Aider has one, but it hasn't been updated in months. People kept claiming models were getting worse, but the results proved that they weren't.

The Aider discord community has proposed and disproven the theory that 2.5 Pro became worse, several times, through many benchmark runs.

It had a few bugs here or there when they pushed updates, but it didn't get worse.

It's a bit dated now, but it would be cool if people submitted PRs for this one: https://aider.chat/docs/leaderboards/by-release-date.html

30-A3B is a really impressive model.

I throw tasks at it running locally to save on API costs, and it's possibly better than anything we had a year or so ago from closed source providers. For programming tasks, I'd rank it higher than gpt-4o

My guess is that Orbstack might switch to using this, and it'll just be a more competitive space with better open source options popping up.

People still want the nice UI/UX, and this is just a Swift package.

Just close the loop and give it direct access to your console logs in chrome and node, then it can do the "full manual debugging" on its own.

It's not perfect, and it's not exactly cheap, but it works.

Install Roo Code (or Cline) in VS Code, plug in your API key(s), and start using them. It's that simple.

My most recent flow is very similar, but I use AiderDesk[0] instead of Prompt Tower for easier creation/editing of plan files.

AiderDesk lets you save snapshots of a point in time, so I create "presets" to restore sets of context files and/or conversation history (you can restore one or both), which is a really nice bonus. You can also add/remove context as needed without the manual copy/pasting work when I forget to include something or accidentally included too much. Its VS Code extension makes adding/removing files from context seamless.

[0] https://github.com/hotovo/aider-desk

Definitely concerning, although I'm having trouble finding anything in the codebase to support this.

This paper even seems to contradict aspects of the project's no tracking stance. If someone told me this paper was for a different (but similar) project, I'd believe it after looking at the two side by side.

Would definitely want this to be addressed before I'd consider using it.

Packages can get hijacked too.

100% sure. Tested in the Anthropic workbench[0] to double check and got the same result.

The web interface has a prompt that defines a cutoff date and who's president[1].

[0] https://console.anthropic.com/workbench

[1] https://docs.anthropic.com/en/release-notes/system-prompts#c...

Interesting. It's claiming different knowledge cutoff dates depending on the question asked.

"Who is president?" gives a "April 2024" date.

Although I believe it, I wish there was some observability into what data is included here.

Both Sonnet and Opus 4 say Joe Biden is president and claim their knowledge cutoff is "April 2024".

Yep!

I'm aware of the tradeoffs here. For home assistant specifically, there's two options if you want to stay on the path of first-class support. Run it bare metal or in a VM.

Going a different path isn't a bad choice, or even a big downgrade.

I had fun with all the different ways of running home assistant 6+ years ago, and then decided to embrace a solution that required the least fuss and would hold up long term. I'm happy with my choice, and it gave me exactly what I was expecting.

One-click updates work from the dashboard? I don't think they would.

Of course you don't have to go the proxmox route, but it is an easy route.

I have a proxmox cluster, so moving things between machines, high availability, and backups are a breeze. Had an SSD go bad a few months ago, and I just moved everything on that machine to another node until the new drive showed up. It was a pleasant experience.

There's plenty of other ways to achieve this, but this is what I chose and I'm happy with it. It's simple, and I can manage everything from my phone if needed.

Proxmox adds very little overhead. I'm running dozens of things alongside HAOS.

The OS is the path of least resistance and gives you the best experience for low maintenance.

https://community-scripts.github.io/ProxmoxVE/scripts?id=hao...

If you choose your hardware wisely, it's incredibly reliable. Most of my house is automated and I all I've done with it in the past year is add a few devices.

I will admit that low (or zero) maintenance is easier to accomplish if you know your way around a little bit.

I did this a few months ago for a test home assistant setup and had the opposite experience. I forget exactly what I did, but it involved editing a file with the wifi settings and dropping my public ssh key on the SD card before installation.

If you use ethernet, no editing required. Web interface goes into setup mode automatically.

Worth noting that ha cli[0] (ha>) does have a `network` command to configure this as well.

[0] https://github.com/home-assistant/cli

You have that backwards. The leaderboard results have the thinking model as the architect.

In this case, o3 is the architect and 4.1 is the editor.

It's also missing the documentation part. Without additional context, method/type definitions with a short description will only go so far.

Cherry picking a tiny example, this wouldn't capture the fact that cloudflare durable objects can only have one alarm at a time and each set overwrites the old one. The model will happily architect something with a single object, expecting to be able to set a bunch of alarms on it. Maybe I'm wrong and this tool would document it correctly into a description. But this is just a small example.

For much of a framework or library, maybe this works. But I feel like (in order for this to be most effective) the proposed spec possibly needs an update to include little more context.

I hope this matures and works well. And there's nothing stopping me from filling in gaps with additional docs, so I'll be giving it a shot.

MLC[0] indicates that it can run models in the 8B range on iOS, but 1-3B sounds more reasonable to me.

[0] https://llm.mlc.ai/docs/deploy/ios.html#bring-your-own-model

Maybe because it's not exactly NAT64, even though it has the same goal?

Anyone know of any coding agents that support video inputs?

Web chat interfaces are great, but copy/paste gets old fast.

I ran them again several times to make sure the results were fair. My previous runs also had a different 30B model loaded in the background that I forgot about.

LM Studio is an easy way to use both mlx and llama.cpp

anemll [0]: ~9.3 tok/sec

mlx [1]: ~50 tok/sec

gguf (llama.cpp b5219) [2]: ~41 tok/sec

[0] https://huggingface.co/anemll/anemll-DeepSeekR1-8B-ctx1024_0...

[1] https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Lla...

[2] (8bit) https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-8B-...

M4 Max with 128GB of memory.

The 1B model works on iPhones[0].

See my other comments. anemll appears to use less memory.

[0] https://huggingface.co/anemll/anemll-llama-3.2-1B-iOSv2.0

See my other comment for results.

mlx is much faster, but anemll appeared to use only 500MB of memory compared to the 8GB mlx used.