SparkyMcUnicorn

I ran R1-8B for both anemll[0] and mlx[1][2] models on an M4 Max.

Prompt: "Tell me a long story about the origins of 42 being the answer."

anemll: 9.3 tok/sec, ~500MB of memory used.

mlx 8bit: 31.33 tok/sec, ~8.5GB of memory used.

mlx bf16: 27.17 tok/sec, ~15.7GB of memory used.

Memory results are from activity monitor across any potentially involved processes, but I feel like I might missing something here...

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

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

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

"deepfaking" video[0] and voice is relatively easy these days, and is definitely being employed by some of these candidates. Lower the "webcam" quality a little bit, and it can be difficult for many interviewers to notice something is off.

[0] https://github.com/hacksider/Deep-Live-Cam

You can use 0.6B for speculative decoding on the larger models. It'll speed up 32B, but slows down 30B-A3B dramatically.

This is pretty much exactly what I was going to build. It's missing a few things, so I'll either be contributing or forking this in the future.

I'll need a way to extract data as part of the tests, like screenshots and page content. This will allow supplementing the tests with non-magnitude features, as well as add things that are a bit more deterministic. Assert that the added todo item exactly matches what was used as input data, screenshot diffs when the planner fallback came into play, execution log data, etc.

This isn't currently possible from what I can see in the docs, but maybe I'm wrong?

It'd also be ideal if it had an LLM-free executor mode to reduce costs and increase speed (caching outputs, or maybe use accessibility tree instead of VLM), and also fit requirements when the planner should not automatically kick in.

And stellar OCR performance. Flash 2.0 is cheaper and more accurate than AWS Textract, Google Document AI, etc.

Not only in benchmarks[0], but in my own production usage.

[0] https://getomni.ai/ocr-benchmark

I've seen evidence that suggests this is false, and that it's more likely that cursor degraded the experience in their context window to save on costs.

The date stamped models haven't had any evidence of ever changing or degrading, to my knowledge. Aider did a test for this as well.

Fully agree. My analogy fits here too.

As annoying as it is when the human support tech is wrong about something, I'm not hoping they'll lose their job as a result. I want them to have better training/docs so it doesn't happen again in the future, just like I'm sure they'll do with this AI bot.

My understanding is that long context models can create embeddings that are much better at capturing the overall meaning, and are less effective (without chunking) for documents that consist of short standalone sentences.

For example, "The configuration mentioned above is critical" now "knows" what configuration is being referenced, along with which project and anything else talked about in the document.

I had the same thought, although voyage is 32k vs 128k for cohere 4.

Anecdotal evidence points to benchmarks correlating with result quality for data I've dealt with. I haven't spent a lot of time comparing results between models, because we were happy with the results after trying a few and tuning some settings.

Unless my dataset lines up really well with a benchmark's dataset, creating my own benchmark is probably the only way to know which model is "best".

If I were to recommend one to someone today, I might pick RooCode. I'd suggest checking out boomerang mode and RooFlow on GitHub.

Here are some others that I've tried and could recommend, in no particular order:

- https://github.com/ai-christianson/RA.Aid

- https://github.com/anthropics/claude-code

- https://github.com/block/goose

- https://github.com/hotovo/aider-desk

I've also created a few "agents" to do specific tasks using Probe[0] as an MCP server, although I'm sure you could create a full-fledged agent with it if you wanted to.

[0] https://github.com/buger/probe

You excited me for a minute, but that list isn't anywhere close to comprehensive.

It depends on the task, and prompting feels different.

I've found that sonnet is possibly better at starting things from scratch and frontend code, while Gemini has been able to one-shot difficult problems and fix bugs that sonnet has struggled with.

Switching between them is a frequent occurrence for me.

It might be relevant that I've completely stopped using Cursor in favor of other tools/agents.

> get bought by Azure

Please no.

For the times I need this type of tool, I sometimes reach for repomix[0]. It does have a webpage[1], although I've never used it.

It supports XML, Markdown, and plain text. XML is the default, and I've found it to perform well, but it might be a good tool to test differences for yourself.

If you're looking for an alternative to repoprompt, I'd recommend checking out Aider's copy-paste mode[2].

[0] https://github.com/yamadashy/repomix

[1] https://repomix.com/

[2] https://aider.chat/docs/usage/copypaste.html

I thought Gmail didn't support js execution. Did Google make an exception for themselves?

This addresses questions I had about certificate generation across nodes/regions.

Slightly off topic, but something nice about Caddy is that it automatically falls back on ZeroSSL (if you have an email address defined) when you hit letsencrypt rate limits. I have a case where more certificates for a root domain are needed than LE is capable of providing, and this fallback solves for the rate limit problem seamlessly.

None of these require anything more than read access, and are for human assistance rather than replacement. If people are blindly accepting CDK updates without review, they probably shouldn't be in charge of IaC at all.

No coins or weapons, but this has something at least: https://youtu.be/F6wJcwQmz4Y

opnsense + ctrld[0] + unbound works great and automatically upgrades upstream requests to DoH (etc.)

Was using NextDNS for a while, but stability and performance was a common issue. I like the idea of something like pihole, but ControlD is good, works anywhere, and is easy to manage.

[0] https://github.com/Control-D-Inc/ctrld/wiki/pfSense-and-OPNs...

Have you tried "Let's get this production ready" as a prompt for this or any other coding tasks?

Sometimes when I ask for "production ready" it can go a bit too far, but I've found it'll usually catch things like this that I might miss.

I go for smart tv's that can be dumb. As long as it reliably uses my input each time it starts and doesn't try to overlay anything, that's all I need.

Once or twice a year I'll go trough firmware update notes, connect it to the internet if there's things that can improve my "dumb" usage (fixes/improvements to refresh rate, Dolby xyz, etc.), then disconnect it from the internet again.

This should work in ComfyUI, stable-diffusion-webui, etc. You can run it locally if you have a decent GPU or an M* macbook.

Also supports the web, and it looks like they might intend to support more platforms.

`readonly platform: 'Android' | 'iOS' | 'macOS' | 'pc' | 'headless';`

> Not only is the core engine of Lynx framework-agnostic, but it's also agnostic to host platforms and rendering backends ... expand to even more platforms, such as Desktop, TV, or IoT devices.

Yes.

In the past week I created two apps within 3-5 iterations. No manual edits (other than .env).

Another one I thought was pretty cool was handing it some API docs for something, then had it build a UI and admin interface from scratch.

The dev agent tools these days aren't half bad, and they're getting better.

Trust, but verify. We should always read and understand what gets spit out anyways.

When doing something with any risk potential I first ask the model for potential risks with the output, and then I manually read the code.

I also "recreated" this tool with Sonnet 3.7. The initial bash script worked (but was slow), and after a few iterations we landed on an fclones one-liner. I hadn't heard of fclones before, but works great! Saved a bunch of disk space today.

The reason 3.5 (as the editor) makes more sense to me is the edit format success rate (99.6% vs 3.7's 93.3%).

Flash 2.0 got 100% on the edit format, and it's extremely cheap, so I'm pretty curious how that would score.

It looks like Sonnet 3.7 (extended thinking) would be a better architect than R1.

I'll be trying out Sonnet 3.7 extended thinking + Sonnet 3.5 or Flash 2.0, which I assume would be at the top of the leaderboard.