Turns out that it's actually 2-bit and not 1-bit. It uses a combination of the EAR and MIC jacks, and when you output through the MIC jack, it slightly boosts the EAR jack. That gives you 4 possible levels, but they are not evenly balanced.
I gotta try this on real hardware, reading the original article it seems it's barely audible without an amplifier, unlike on an emulator.
Will post an update after I get back to my parent's home, where I have these beauties that launched me into computer science: https://imgur.com/a/GKrA0vX
I created a program like that back in the day. In my case, I recorded my voice through the "MIC" input and played it back through the TV. A few seconds of sampled audio required the entire available memory. While I fully understood what was going on, I thought it was magical that the Spectrum could "speak with my voice".
On ZX-Spectrum? There was no "MIC" input on ZX-Spectrum, not that I know of. I was in high school at the time with only a rough understanding of the hardware but it was kinda obvious to me that whatever I plugged into the casette input port was only read as "0" or "1" by the IN (forgot which) instruction. So not really possible to record voice at the time (didn't even know of PCM anyway), but I tried to decode Morse transmissions from radio. Never got anywhere with the software but my ear got good enough at some point that I could decode them myself by hearing :)
There was Z80 code to record from the MIC socket in one of the small books (Melbourne House maybe?) that came out in the early 1980s. I remember borrowing it from the library and it worked! Heavily quantized of course by the snapping to 2 or 0.
Also some games had sound samples, Activision’s Ghostbusters had a couple.
With proper volume control, even one bit could probably sound voice-like enough to a person in the 1980s.
It may be feasible to retrieve a pcm for a very short clip by playing the same recording at different levels which then could be converted on the same machine to a sequence of opcodes playing a more realistic voice.
Nope. I did the same thing myself with my Spectrum. I guess I was 14. I didn't have a microphone so I tried it with my dad's stereo headphones. I got the idea to try it after looking at the Z80 machine code of cassette saving and loading routines. Maybe it was also after ETX came out, first game I remember with speech samples. (Terrible game). I forget the details now but it's all about the timing between turning the mic/speaker on or off.
I have identical twin boys - when signing up for a bank account app and while completing all the associated hoops for ID, the second to do it got the rather dystopian error message 'an account has already been registered with this face'
I wonder how it's possible they didn't think about this problem. Considering everyone involved in building this feature, not a single person thought about the possibility of twins or simply people who are very alike? Also, can't people open multiple accounts?
I imagine that it probably went like this "what about identical twins?" / "Don't worry about that for now, we'll solve for it before we go live, but it's not in this milestone". Then someone pulls rank and decides the product is "good enough, we can't wait for all bugs to be fixed" and gets it deployed.
Just look at the "falsehoods programmers believe about ..." articles that pop up from time to time. There are plenty of complexities in the real world that people either don't know about, aren't front of mind when they are tackling a larger problem, or simply don't want to deal with in the here and now. Lost knowledge and new problems are part of the reason why programmers are discouraged from rewriting software from scratch and why businesses in general are hesitant to replace legacy systems that still work.
True, but many of those bits would be learned in some primary schools. It simply depends upon whether they are relevant where you live. That's especially true for names, which is culturally dependent. It is also true for the culturally dependent aspects of dates/time, though there are obscure bits that are only relevant to certain disciplines.
Based on my experience I assume someone mentioned the risk, others said it’s unlikely and only a small percentage, that we should care about the 99% for now. Then the issues isn’t revisited until someone impacted complains enough.
p99 is ok for web stuff, not for systems that impacts people in their life, but I’ve seen too often things being dismissed because they are edge cases (which becomes an actual real issue at the scale of a country of course)
What a pointless point to make. I’ll spell it out for you: related within a generation or two. There are no twins in my family that I know of for several generations, for example. Thus I would say I’m not related to any twins because I’m not in any meaningful way, even if there might be twins somewhere in my ancestral history.
If you mean that all humans are related, that is true but a completely useless fact in this context.
Because it probably does the dystopian tracking well enough, they don't care if it's not 100% accurate. Edge cases in the new world have no recourse, citizen.
Maybe they didn’t care. Identical twins are a fraction of a percent of the population. Excluding them (really, only half of them) from the customer base wouldn’t hurt noticeably.
What is strange is that even the best face recognition algorithms have false positive. So surely they must have planned for two unrelated customers resolving to the same embeddings.
One imagines they are far more worried about person X fraudulently claiming to be person Y than they are about occasionally causing problems for identical twins.
Obviously the solution is to accompany facial recognition with DNA tests.
And worse yet, you'll find the code peppered with comments relating to 'phase 3' and 'section 11', ephemeral stuff that had meaning in the moment but now enshrined forever. And what happens when the LLM stumbles on this and working off a whole different phase 3 or section 11?
I turned that to my benefit. I use that design-doc pattern where you first ask it to make a ticket with a formal section list (why, how, etc) and then I ask it to use comments with permalinks. I put it all in policy files. As a result, comments have clickable links to coherently worded tickets.
Still, this requires a second pass, typically. In its default-mode it often ignores the policies and does all the usual Claude stuff.
So the critique from the other story holds true - it really is just running the Linux driver inside docker, and (perhaps) dressing it up as something more - though I note the claim on this page is walked back from the original 'writing the driver'. On the other hand, it really did get him printing from his Mac, so in this new LLM results based world we live in it did its job admirably.
Nice, I will give this a go. This is what /resume should be but because sessions are tied to directories it's very easy to lose track of where you started what. Until now I've been asking Claude 'what was the conversation we had about [X]' which works but is clunky in the extreme.
Same thing bit me. What fixed it was keeping the session id next to the checkout instead of letting the directory be the key. Small local db, one row per worktree. Resume then doesn't care where I'm standing, it just continues that transcript. I've got worktrees I haven't touched in weeks that still pick up fine.
One thing I didn't expect: forking a conversation copies the history into a new transcript. So anything you count per transcript, cost in my case, double counts that prefix. Resuming the same id doesn't.
