roger_

M5 Max 128GB for $1k?

Would putting this between a small model and an agent like Hermes improve performance?

Ha! I spent time also hacking together Armbian on an old A20 TV box.

Claude was definitely helpful the second time around to help with the DTS.

Looking forward to testing this!

Is full 3D acceleration eventually possible and how's battery live?

I love how easy AI makes it to hack devices that otherwise wouldn't be worth the time.

Have they published?

So this is PCA in kernel space?

Really good job!

Do you have more of these writeups?

Thanks, I was looking through the article for exactly that. Does it lock on to a configuration of stars?

Really curious how they did this mechanically.

I can see valid uses of this but I also feel like a probabilistic calculator would be more useful.

e.g. the result for the 1 / [-1, 2] example doesn’t tell you how likely each value is and it clearly won’t be uniformly distributed (assuming the inputs are).

DFlash immediately came to my mind.

There are several Mac implementations of it that show > 2x faster Qwen3.5 already.

Looks great but even $40 for a perpetual license with only two years of updates still seems excessive.

Taskbar, Uber, etc. cost less and have unlimited updates. How is this better?

Here's my (hopefully) intuitive guide:

1. understand weighted least squares and how you can update an initial estimate (prior mean and variance) with a new measurement and its uncertainty (i.e. inverse variance weighted least squares)

2. this works because the true mean hasn't changed between measurements. What if it did?

3. KF uses a model of how the mean changes to predict what it should be now based on the past, including an inflation factor on the uncertainty since predictions aren't perfect

4. after the prediction, it becomes the same problem as (1) except you use the predicted values as the initial estimate

There are some details about the measurement matrix (when your measurement is a linear combination of the true value -- the state) and the Kalman gain, but these all come from the least squares formulation.

Least squares is the key and you can prove it's optimal under certain assumptions (e.g. Bayesian MMSE).

Skimmed this but don't have an intuitive understanding of why this works and how temperature and truncation factor in.

I was expecting something about the morphological erosion operator but this was pretty cool.

Some of the techniques here seem to be motivated by physical processes (e.g. rain). I wonder if that could be taken further to derive the whole process?

Criticize gatekeeping all you want, but I feel it’s safer to recommend a Mac or iPhone to an older, non-technical person than the equivalent Windows / Android machine.

And I’m still able to install any app I want with minimal fuss.

Why no need to make it the default? I’m all for rethinking legacy decisions.

It helps 99% of the user base and the security risk seems negligible.

Can anyone explain why Mamba models start with a continuous time SSM (and discretize) vs discrete time?

I know the step isn’t fixed, also not sure why that’s important. Is that the only reason? There also seems to be a parameterization advantage too with the continuous formulation.

Great idea and seems quite obvious in hindsight.

Is it guaranteed to have the same effect on vanishing gradients though? What if it put weight 1 on a layer that had a tiny gradient?

MLX support on Macs was the main reason for me.

Appeal to nature.

Async embedded is something that's always made sense to me and I've been awaiting a long time for it to happen.

But what's the overhead price with Embassy?

I know what you mean, but the issue the parent comment brought up is real and "bad" chats can contaminate future ones. Before switching off memories, I found I had to censor myself in case I messed up the system memory.

I've found a good balance with the global system prompt (with info about me and general preferences) and project level system prompts. In your example, I would have a "Python" project with the appropriate context. I have others for "health", "home automation", etc.

Disable memories so each chat is independent.

If you want chats to shared info, then use a project.

In one case it was a urinary flow test (uroflowmetry). The results go to a lab and then the doctor gets the summary. Was able to diagnose the issue, prevalence, etc. and educate myself about treatment and risks before seeing a doctor. Papers gave me distributions of flow by age, sex, etc. so I knew it was out of range.

In another case I uploaded a CSV of CGM data, analyzed it and identified trends (e.g. Saturday morning blood sugar spikes). All in five minutes on my phone.

If they pepper it with warnings and add safe guards, then I'm fine.

I think they can design it to minimize misinformation or at least blind trust.

I’ve had mixed experiences with doctors. Often times they’re glancing at my chart for two minutes before an appointment and that’s the extent of their concern for me.

I’ve also lived in places where I don’t have a choice in doctor.

I understand enough about these systems to know they’re not perfect but I agree some people might be misled.

But I don’t know if I should be denied access because of those people.

Yes, if it will help me and others. Again, I understand those who disagree.

I’d rather have all my health data be used in a way that can actually help me, even with a risk of a breach or misuse, than having it in a folder somewhere doing nothing.