> bitnet.cpp is the official inference framework for 1-bit LLMs (e.g., BitNet b1.58). It offers a suite of optimized kernels, that support fast and lossless inference of 1.58-bit models on CPU and GPU (NPU support will coming next).
Log Base 2 of 3 = ~1.5849625, so that's the limit to how well you can pack three-state values into bits of data.
For something more practical, you can pack five three-state values within a byte because 3^5 = 243, which is smaller than 256. To unpack, you divide and modulo by 3 five separate times. This encodes data in bytes at 1.6 bits per symbol.
But the packing of 5 symbols into a byte was not done here. Instead, they packed 4 symbols into a byte to reduce computational complexity (no unpacking needed)
Yeah, "1.58 bit" is 1 trit with three states, since log2(3)≈1.58.
So it's not a inference framework for 1-bit models (two states per parameter) but for 1.58 bit models (three states per parameter). Annoying that they try to mix up the two.
> So, assuming the source material is correct and electrons indeed have mass, SSDs do get heavier with more data.
That is definitely wrong! No way the source material has more electrons. The only way it could do that is by being charged.
Richard Feynman, The Feynman Lectures:
"If you were standing at arm's length from someone and each of you had one percent more electrons than protons, the repelling force would be incredible. How great? Enough to lift the Empire State Building? No! To lift Mount Everest? No! The repulsion would be enough to lift a "weight" equal to that of the entire earth!"
See, now, if this was Reddit...this is the opportunity for a yo momma joke. But here we are on HN, so I'll just point out that this is the opportunity for a yo momma joke.
Gforth is free and well rounded so I'd recommend that if you want to experiment with Forth. It is not very fast though, SwiftForth with optimised subroutine threading will be a lot faster. I haven't tried SwiftForth though as you have to pay for it and it is x86 only.
We use this to store encrypted file names and using base32768 on providers which limit file name length based on utf-16 characters (like OneDrive) makes it so we can store much longer file names.
Have been surveying Computer Science courses at university with my son recently. All the ones we looked at had a compulsory ethics module which shows the direction things are headed at least.
Mine had one over a decade ago. After graduating, the industry decided that developing everything we just got done establishing was unethical, was the hot topic to innovate for the next decade. I never worked at any of those places and still got burned ethically in much more indirectly unethical product streams in the finance and insurance sectors. To be honest, if there is really good money to be made at this point, there is a safe bet that if you dig deep enough, there is an unethical core to it. Most of my peers assuaged themselves with some variant of "I'm a programmer, not an ethicist, and philosophy doesn't put food on my table. So sadly, the problem seems much more systemic and a priori to the capitalistic optimization function.
I wonder how many programmers working today are coming through universities though? I'm self-taught, most of my programmers friends are as well, same with most of my colleagues back when I worked. I can remember maybe the name of 3-4 people in total, out of maybe ~30 or so, who went to university for computer science before they started working.
In my experience CompSci ethics modules are about hacking or mishandling user data or code theft... i.e. things that companies don't want their employees doing.
I've yet to see an ethics module that covers ethics from the perspective of ethics over profit.
Whereas an accountant is taught that they should resign rather than get involved in unethical practices, like profit manipulation for example. I interview people with ethics questions. I discussed them frequently when training.
I refused the pressure to be unethical when I was pushed, even when I knew I would be fired (which I was). I was able to discuss it with old mentors, who made time to meet with me, even when I hadn't worked at their company for years.
Lastly I disclosed why I was fired at interview for a new job (without the confidential details), and was hired partly on the strength of it by a person who had been through much the same.
And I didn't learn it at University, I learnt it on my professional qualification, that was around 3 years long and was postgraduate level, although had non-degree based entry routes for technicians. It also required a wide range of supervised experience.
This was not at all the ethics program that was taught in my university computing ethics course. They did indeed cover the societal and moral responsibility of software developers. This was way back in 2002.
> Some believe AI Super-intelligence is just around the corner (for good or evil). Others believe we're mistaking philosophical zombies for true intelligence, and speedrunning our own brainrot
Not sure which camp I'm in, but I enjoyed the imagery.
Eneloops are fantastic. I've been using them for 10 years.
They have very low self discharge so you can keep them on the shelf for a year and they will still have 70% charge which is very different to most rechargeable batteries.
They don't leak unlike alkaline batteries and you can run them to 0% charge without damaging them unlike Lion batteries.
The terminal voltage is only 1.2v so there is the occasional thing they don't work well in.
That said my collection of eneloops get much less use than they used to as everything comes with built in lithium batteries and a usb charger nowadays. That is very convenient until the moment the lithium battery dies...
Changing the lithium battery in things sucks. Firstly it's really difficult to get a good one and not a crappy knock off. Secondly modern stuff isn't meant to be repaired. Every time I pick up my kindle I shed a small tear when I see the screen damage I caused changing the battery.
So, I buy stuff with AA or AAA batteries if possible and if I want it to last for more than 3 years.
The whole "everything I own just has a built in Li-ion battery" thing came to mind when I ordered my first set of Eneloops ahead of last week's snowstorm in North America - most things I own have their own batteries with a charger (right down to this decorative squishy whale lamp), and I can count on two hands the things that use AAAs and one hand the things that use AAs. It's wild.
Most of them are about 2000 mAh. Other NiMH batteries can have, say, 2700 mAh. So even though the latter have a higher discharge rate - after 6 months of storage the latter still has more juice.
The benefit with the 2700 mAh, of course, is if you're using when full, you can use it for much longer.
If they cost the same, I could see the hype. But most people are still better off with regular NiMH AA batteries.
Eneloop Pro cells have a rated capacity of 2500mAh.
I can't think of any good applications for conventional NiMH cells any more - they're dominated by LSD NiMH cells in low-discharge applications, by lithium primary cells in ultra-low-discharge applications and by the various lithium secondary chemistries in high-discharge applications.
One bit or one trit? I am confused!
reply