kamranjon

You’re out of the loop if you don’t think m series chips with unified memory aren’t one of the best platforms for running local inference

Gain some readers for… Wikipedia? I just learned about this and thought it was interesting given the current situation with RAM pricing. You’re free to draw your own conclusions.

How does something with 232 comments and 207 points over just 2 hours get pushed to the 3rd page in hacker news? I’m just really curious how it works, like why would something with so much engagement be push down so quickly?

I think one of the more ominous things to see in recent years was all of the tech execs at the presidential inauguration, after having collectively donated several million dollars to the inauguration fund. So if we go with that list, which happens to overlap with many of the circular deals we’ve seen in the AI space recently, you’d have people like: Sam Altman, Jeff Bezos, Elon Musk, Mark Zuckerberg, Tim Cook, Sundar Pichai and Sergey Brin

I also wouldn’t be surprised if memory providers weren’t intimately involved, as they’ve been caught price fixing in the past: https://en.wikipedia.org/wiki/DRAM_price_fixing_scandal

I really think this is a coordinated effort to restrict computing capacity for individuals and force adoption of centralized AI - I think there already is evidence of this from the moves OpenAI had made to lock up memory and gpu markets.

Damn I was considering an m5 max with 128gb just a few days ago and it was 5099 and now it’s 6699 - 1.6k increase - definitely a massive increase and has dissuaded me - this is pretty insane.

Seems possible it’s both this time: https://kotaku.com/valve-says-the-companies-making-ram-give-...

Yes apologies, Hackernews was just an example, you can do this with any website - it’s just a simple benchmark I like to use for testing vision models.

A fun thing I do with Qwen 3.5 0.8b is to take a screenshot of the Hackernews homepage and ask it to give me a JSON representation of the data and it does surprisingly well. With a well structured prompt I think it could be made to be pretty reliable tool for that type of task out of the box.

I have! I recently compared Gemma 1b to ModernBERT Large for a binary classification task and ModernBERT was the clear winner. It learned faster and performed the task better by a significant margin by the end of training. It seems the bidirectional encoder only architecture works really well for classification tasks, and I think it is related to being bidirectional whereas decoder only models like Gemma (or Qwen) can only “look backwards”. I used a mixture of FFT and LoRA as well as a mixture of CE Loss and SupCon Loss.

While you may be correct in some of these assumptions, the repo is open source, contains actual schematics and code, appears to have real community contributions, and also happens to exists now, which is something that can't be said for the original post this was a response to. I was hoping to just provide some reference to existing work that I found interesting since it is what I think is a much more traditional type of detection system.

thank you!

For those interested there is an open source radar based phased array project here: https://github.com/NawfalMotii79/PLFM_RADAR

I may have even discovered it from hacker news, I forget now.

Can you tell me more about this?

Does anyone understand how Meta is able to spend so much money on AI with basically no AI product to speak of? Especially after sinking billions of dollars into a failed VR product? I just don't really understand why they are investing in data centers, I don't know of any actual product they offer that anyone is seriously considering using in the space.

a pretty fun and quick tests i do with vision models is to screenshot the hackernews homepage and ask the model to return a json representation of the screenshot - qwen 3.5 0.8b did surprisingly well at this.

To me this sounds like tailscale - does anyone have any insight into how what this is doing is similar or different?

I do quite a bit of 3d printing of functional parts and am trying to understand how this would fundamentally differ from printing at 100% infill? What type of part requirements are causing failures when something is just solid? Just curious what problem space this is targeting?

I was using the 8 bit quant and no reasoning - it’d make mistakes but then fix them at a speed that was impressive - it also was like incredibly tenacious and would honey badger its way around any issues it hit. My second best was Qwen 3 coder next - I did play with 3.5 and 3.6 (both moe and dense variants) but always seemed to go back to GLM 4.7 8 bit mlx variant. I have 128gb mbp so I’ve migrated to Deepseek v4 flash for everything now and haven’t looked back but if a new GLM flash model came out I’d be very excited.

Crossing fingers for a 5.2 flash release - it’s been a while but I still feel like 4.7 flash is one of the strongest local coding models

I have been using deepseek v4 flash as my main model for everything ever since dwarf star came out. I run it on my M4 Max MacBook Pro with 128gb of memory. I run it usually as a server and connect to it over tailscale with my coding machine and use the Pi coding agent. It’s a big leap over using the Qwen models though it doesn’t have vision - so I still will run those when I use vision. GLM 4.7 flash was my previous go to for coding but I’ve completely switched to deepseek for all non-vision things.

“even though you're delivering code at a good pace, you're taking too long to deliver those Design Docs. Are you using AI? You should use more AI.”

This here is the crux of it I think… it’s often promoted that AI will give us the time to do the “real” engineering work of designing systems and really serving the user, but in practice all I’ve seen is further attempts at optimizing every last process with AI - just homogenizing every product and feature into slop.

It feels like every leader has been to some talking points boot camp where they’re incentivized to apply pressure to every part of their process - sort of a desperate attempt to justify the costs they’re incurring. I think we will look back at this and see how obviously short sighted it was.

There have been plenty of motorized lenses in the past that relied on the micro-contoller inside the camera body for control. What does having the controller live on the lens permit that the pattern we've used for years doesn't afford?

My info might be a few years old, but I think he uses it on his laptop machine.

https://lwn.net/Articles/903023/

"On a personal note, the most interesting part here is that I did the release (and am writing this) on an arm64 laptop. It's something I've been waiting for for a _loong_ time, and it's finally reality, thanks to the Asahi team. We've had arm64 hardware around running Linux for a long time, but none of it has really been usable as a development platform until now."

I believe it's what Linus Torvalds uses.

I have heard he would write entire books out on index cards and then rearrange them as part of his creative process. I love that.

Apparently Vera (his wife and editor) would carry a handgun in her purse to protect them on their travels. My wife recites this fact often.

I interestingly came to a pretty similar conclusion about Kerouac around the same time as this reviewer, and brought it up in my review of Henry Miller’s book on Big Sur (which I can highly recommend): https://kamranjon.com/writings/henry-miller-and-the-big-sur....

Just curious - are you at all concerned about the legal implications of ai-generating property listing videos?

not really the same - that's sort of the nice thing about ipads and some business laptops is you can add a sim card and use them anywhere. The MNT Pocket Reform has this but the waitlist is months out.

if only it supported sim cards