I didn't have this one (I had a TK-90X, and a TK-95 after that; both could be described as improved Spectrum clones) but I briefly used one at a place where I took some programming course in the 80s (they had these, and Coleco Adam) and I loved it.
Last week I was at a conference at the Computer History Museum in Mountain View and saw one of this in real life for the first time in, easily, 38 years, and it brought tears to my eyes.
If you're ever near this place, I recommend a visit to this museum. It was my first time, but it won't be the last. I hope I can go back and take my kids there once. I left with a punched card with my login name punched into it, using the keypunch of the (working) 1401 that's on display.
Not the model though. The model really only takes input text and produces output text. Memory within a conversation is achieved by the harness adding the conversation (or parts of it) to the input text. The LLM itself has no memory, it’s the augmented system of several orchestrated LLM calls that does.
You know that thing about the answer to any headline that’s a question being almost always “No”?
I’ve been fully remote for 17 years and, before that, did short periods of remote work as early as 2003.
The first junior I remember being onboarded remotely (into an irc-mostly, voice call for ultra rare cases job) was me and it worked pretty well.
From about 2010 to 2012 I grew the support team of a database services company from 2 to about 20 people (the 2 at the start was counting me). A lot of the new hires were junior when it came too dabatases, some just junior period, and it worked great too.
There were some in person meetings but the vast majority of onboarding and training happened remotely.
Maybe people invested in office real estate are perhaps biased to get the wrong answer on this one.
If I had time now, I’d write an article titled “is the FT right about the cause of weak junior hiring?”
I don’t have time, but you can apply that saying about headlines to my imaginary article.
Besides the productivity boost (and I know you already mention a boost to your well-being), this is one of the simplest yet effective things you can do to improve your cardiovascular health. I had a heart attack at 40 and 30 mins a day is the minimum recommended, so 60+ is great.
They will also lie and produce output saying it is based on tool execution, without having actually used the tool.
Yes, another layer to cross-check, say, “in kubectl logs I see …” with an actual k8s tool call can help, that is, when the cross-check layer doesn’t lie.
For the time being, IMHO, human validation in key points is the only way to get good results. This is why the tools make experienced people potentially a lot more efficient (they are quick to spot errors/BS) and inexperienced people potentially more dangerous (they’re more prone to trusting the responses, since the tone is usually very professionally sounding).
Yup! That's what I was thinking about. In fact I did read this right before posting (though I had found it at https://doc.cat-v.org/bell_labs/utf-8_history) but only to validate that it had been in a NJ diner, so I missed my confusion of UTF-8 with Unicode.
On my 33mhz (I'm almost, but not quite sure about the frequency) 486 SX (yeah ...) it ran OK until the levels where you'd get a lot of monsters. In those, I had to zoom in to the smallest possible screen size and even then it was barely acceptable.
So while the video is impressive and I couldn't do something like this myself, I was glad when I saw how bad it ran, as that computer of mine would a little bit more than 30yo today, so to have that beat by a 40yo printer controller would make me think I could have done something to have it run better back then!
The code is only a (very important) part of this type of program. The samples are critical and (for the time being anyway) can't be generated by AI.
Especially important if you want orchestral instruments that sound realistic. Just think of the many ways that a single note can be played by a professional player and multiply that by the range of the instrument.
Edited to add: not orchestral instruments, and also not samples, but this gives an idea of the complexities of capturing the characteristics of an amplifier so that it can be modeled faithfully: https://neuraldsp.com/quad-cortex-updates/introducing-tina (I'm not related and I'm actually a Line6 customer, but I saw this at work in an interview by Rick Beato and though it was super interesting)
Thanks for the clarification, I guess my memory is very bad after all! :)
Do you remember if that was a recent addition?
Full disclosure: I was quite the newbie back then and most of what I "new" about SQL Server was what the more experienced coworkers told me. This was a very IBM-biased place so I'm not surprised they would have stuck to some old shortcoming, like people who still talk about bad MySQL defaults that have been changed for at least 10 years.
Up until that job (which was my second Actual Formal Job), all my DB experience had been with either dBase (I think III plus or IV) and access, so this was a whole new world with me.
It was through MS SQL Server that a colleague taught me about backups and recovery, after I ran an update in prod but forgot to include the where clause ... :)
My experience at the time was that it was perceived as not serious enough and lacking important features. If my memory isn't very bad, I believe as late as 2000 SQL Server still only supported AFTER triggers.
In my experience in the late 90s and early 00s, besides Oracle and Sybase, DB/2 and Informix were also regarded as good. Oracle was considered the best though.
You can still rip CDs with Apple Music. In fact, that's the only use I have for that app (I recently lost a hard drive with music and I'm in the process of backing up all my CDs again).
Say, to rebase upstream MySQL changes onto a fork in the monorepo (in a random, non-specific example)