A brief history of computers
lesswrong.com
lesswrong.com
Wonderful.
It's interesting to me that, in just a few months, I've already developed muscle memory for checking whether or not things I read online are machine-generated. The first thing I do on any website is search for "GPT", "Bing", and "AI" and stop reading if I find them.
Reading someone's writing is an exercise in trust. If they claim something, I have to be able to trust that they have done enough of their homework to back it up. If they cite a source, I have to be able to trust it says what they claim. Otherwise what's the point? If I can't rely on the author, then reading their writing requires checking everything they've said. Their writing is useless to me since I'll need to do my own research anyway.
If you write something and ask me to read it, you are asking me to trust that you have done the legwork. If you really just typed it into ChatGPT, that's more than just stupid. It's a betrayal.
That sentence was almost more rhetorical flourish than anything else, and while it didn't land particularly well for me personally, I find it _far_ weirder that the commenter I replied to went off about authorship and trust.
In contrast, GPT has no indication of where it got anything from, but even that's giving too much credit. It's just stringing words together. Whether it's wrong or right, I have no way of knowing. If I'm reading a website and I know the first thing it says is for sure correct (or incorrect) then I can judge the rest of it appropriately. That's absolutely not the case for GPT, because the correctness of each sentence is relatively (entirely?) statistically independent.
That's only the beginning of it. There are so many more ways that doing actual research is superior; there's just no contest.
I will be horrified if most searches are mediated by LLMs in the future and I will do my best to never use such a search product. (I use Kagi myself and haven't had a problem yet.)
I expect even searches that are not explicit, obvious LLMs like bing search to be using AI in the background.
I really don't understand your issue here. Your search is already algorithmically mediated. The algorithms are just getting better.
And all of this is of course a tangent on the original point which is that asking an LLM a related question is in _no way_ the same as having an LLM write the article.
Yes.
Though at least with Google, the author still gets a chance to somewhat evaluate the links they are given.
It's not like a one shot affair, and I feel that assumption is more common than would be generally useful.
Consider (from my own use of GBT):
Author queries Google, follows a few links, realizes they really need to take a step back and make new queries.
Author queries ChatGPT and essentially short cuts past that initial Google learning step I just mentioned above and follows with a combination of chats and Google searches.
End result is author ends up with a rough understanding and a set of information they can draw from and potentially continue to explore, depending on their needs.
There is also running queries on say, Yandex, Google, DuckDuck to see what each returns. I do that frequently.
ChatGPT mostly makes me a lot more productive. One can jump past that initial set of queries often necessary when jumping into unfamiliar territory.
Evaluating ChatGPT info works the same as using search does! Like doing multiple searches with various terms, exclusions, quotes and such, using multiple prompts can be powerful.
As others have said, the only real issue is trust.
Did the author realize something to the degree needed to speak competently, or just take what they were given and run with it?
These things being used like tools is no big deal. Like any good tool, there comes an obligation to use them correctly and with integrity.
Abstract algebra was a new snd developing thing back then, the idea that you can generalize from numbers and addition and multiplication to other structures that have something like numbers and addition and multiplication.
Boole found that if you take the two-element set {0, 1} and choose saturating addition as the addition-like operation and normal multiplication as the multiplication-like operation, you get an algebra (specifically a ring) that is isomorphic to propositional logic with its AND and OR operations.
So the idea that the number 1 can represent true and the number 0 false was Boole’s insight and the foundation of modern digital circuits.
They are flipping cool tbh. I wish more was done with the idea. Balanced ternary is beautiful.
Rumours? This is the internet. Rumour to research is a few clicks away.
They did exist: https://en.wikipedia.org/wiki/Setun
Very few masters theses have changed the world, but Claude Shannon's was one of them.
Unfortunately, I'm having trouble understanding though. The furthest I got with math is calculus. I'd love to hear a more ELI5 description of what you said. Something that doesn't require prior knowledge of abstract algebra. (I recognize that's a big/difficult ask. No obligation or anything ofc.)
Babbage had a chance to build his difference engine. The problem was that the engineering was a lot harder than he thought it was going to be, and he was a mathematician/economist, not so much a working engineer. The idea that if only he had had a better chance the difference engine would’ve been successful is just simply a misreading of what happened. It didn’t help that after the British government poured tens of thousands of pounds into the project Babbage suddenly decided to start pushing analytical engine before he had even finished the difference engine. That made it look like these were just wild, cockeyed schemes, when Babbage was supposed to be engaged in a practical, mechanical calculating project (to help reduce the labor expended on computing, for example, navigational tables).
This is true of many (most?) innovations. E.g. the steam engine: people knew about steam power, and built primitive steam engines. Watt succeeded eventually in manufacturing one that had the right mix of reliability, power, cost, maintainability to be widely useful. E.g. the jet engine : Whittle conceived of turbine aircraft power during WW1, but didn't succeed in manufacturing a viable engine and putting it in a plane until the end of WW2.
Some cultural factors like slavery meant they were less interested in e.g. labor-saving inventions. And there probably were health and life sciences concepts you could introduce--but might have limited ability to prove. But, for the most part, there are technology trees that you can't really shortcut and, even with the right high-level knowledge, it's hard to accelerate things too mych.
https://en.wikipedia.org/wiki/Per_Georg_Scheutz
It wasn't too hard for the time... Babbage just kept getting distracted.
The key issue in early computing, and the one most people miss, is how memory-limited computing was. IBM had electromechanical arithmetic working by the 1930s, and electronic arithmetic working before WWII. Programmability was still plugboards and cams, because there was nothing suitable for storing a program yet.
Take a look at the IBM 601.[1] IBM introduced this in 1931. Add, subtract, and multiply, a few registers, and programs on a plugboard. This was the first commercial programmable calculator.
But only a few registers of memory. That was the hangup. Everybody involved, especially J. Presper Eckert, who went on to work on ENIAC and UNIVAC, recognized that. All the approaches to memory required building wheels or relays or something for each digit or bit. They had arithmetic and control. If only there was some way to store lots of data...
Bulk memory suitable for program storage finally appeared with delay lines, magnetic drums, and storage CRTs. Eckert again.[2] Architecture was not the problem. Hardware was the problem. Computing needed hardware reliable enough that you could have a few thousand gates, and memory reliable enough to not drop bits. As soon as memory had been cracked, and there was somewhere to store the program, stored program computing took off.
Memory still had a huge cost problem. Memory was a million dollars a megabyte as late as 1970. But that's another story.
Gödel was a Lutheran Platonist who was personally morally opposed to Logicism and his contemporary mathematical program. He was an odd man, really, but he was in no respect a booster of, or a person working to promote Hilbert's program. He was tearing it down very deliberately.
- I think Mauchly and Eckert (of ENIAC) in the 1940's were unaware of Babbage (1810's)
- There was the (in)famous von Neumann paper describing the ENIAC and patent lawsuit - https://www.historyofinformation.com/detail.php?id=639 - this page says that "most likely" Von Neumann and Mauchly/Eckert developed it together
https://en.wikipedia.org/wiki/Honeywell,_Inc._v._Sperry_Rand....
Hm these sources are a bit vague -- my memory is that "The Dream Machine" was more critical of von Neumann. Basically it ended up that he put his name on work that wasn't really his, or he gave that impression.
i.e. the name "von Neumann architecture" doesn't give the proper credit
- Did they need Turing or Church to build a real computer? Probably not, I guess Babbage also proves that. Computation is just "out there"; it's part of nature; it's been discovered many times.
- That said, I would guess that Boolean logic is the most important theory/math in building a computer, though Babbage didn't have that either !!!
The computer wasn't really invented at all. It evolved from earlier things in a step-wise manner. There were computing machines for decades prior. E.g. before WW1 there were sophisticated gunnery computers that could fire shells taking account the vector velocity of a ship, wind, distance measured optically, movement of the attacking ship. Boolean logic was used in telephone switching systems. Boolean circuits already existed both in electromechanical form (relays) and electronics (vacuum tubes|valves). So when Turing decided he needed a machine to do so and so calculations on some kind of data, Flowers didn't need to invent Boolean logic nor design Boolean circuits -- those already existed off the shelf. Teleprinters existed. Paper tape existed.
As far as I remember, Woz's biography is good evidence that you don't need the idea of "boolean logic" to design circuits. That did come after the fact -- the way it's taught, not the way it was invented
I think he just said he figured it all out himself essentially, and often did a lot better than the pros. Some of his claims were suspect to me, but I do think his claimed ignorance of prior work is genuine :)
Shannon did come decades earlier though, so the designs of somebody who was influenced by Shannon probably influenced Woz. It's hard to tease apart, but I agree with "evolution" and "tinkering" as the main explanations.
The entertaining explanations are the ones that tend to stick in our minds, but they're not necessarily true
---
The other example I think of is when I look at Debian -- a lot of it is not at all what the Unix inventors had in mind. Debian/Linux basically looked at an artifact and then came up with their own divergent ideas around them
Likewise Woz probably looked at a lot of computers, but he didn't have much of an idea what the creators were thinking -- just the artifacts themselves
From the electrical engineering perspective, the earliest relay and vacuum tube computers were built out of elements developed for radio and telephone exchanges. Bell Labs developed the transistor with their phone system primarily in mind. Same with high speed digital data circuits. (Digital audio was demo'd in the late 1940s, deployed in the phone network in the late 50s).
It's not just the physical circuits; much of the theory, too. Claude Shannon was trying to optimize subnets of switches in the phone network, when he proved that binary switching logic is equivalent to Boolean algebra and so such systems could be described, manipulated, and optimized symbolically.
Similarly, both frequency and time division multiplexing date to the late 1800s with telegraphy. One of the first uses of vacuum tubes as switching elements was for multiplexing telegraph lines c. 1940 or so. (The terminology from that era is quite charming - modern Wi-Fi might be described as supersonic harmonic multiplexed radiotelegraphy.)
Indeed. A "main frame" was originally the housing used for the relay switches in the original telephone exchanges.
History and Architectures
Edited by Raúl Rojas and Ulf Hashagen
Specifically on the relation of telecommunications to computers I will toot my own horn and recommend my book, McDonald, "How the Telegraph, Telephone, and Radio Created the Computer."
I think the actual important part is being able to address and manipulate code like it's data somehow, rather than the specific architecture. Having two separate address spaces for code and data doesn't necessarily prevent that, though it's surely simpler with only one.
Thank you to the author for creating this! This style of super personal historical overview is very enjoyable. I like how the author says outright to take everything with a grain of salt and I like how they call out things about the narrative that don't make sense to them.
> "Why is it so helpful to have the data and instructions share the same memory? Well, for starters, it's simpler, cheaper, and it made execution faster because stuff is all closer together. It was also convenient to access stuff in a single address space rather than having separate schemes."
Why are ALL µprocessors today then using separate instruction and data cache? For early pre-"von-Neumann" computers, memory was an extremely precious resource and you wouldn't waste it some something as frivolous as almost-completely static code. Only once memory became large/cheap enough did the (not-so-surprising) idea of stored-program become practical. The benefits were many: self-modifying code (for good reasons very popular in those days, even much younger PDP-8 can't be used without it), moving more complexity of IO and bootstrapping into code, and eventually, code to make code.
Another point that is glossed over is that none of these technology changes were a slam dunk. For example, in the early days, transistors were FAR from the 1000X faster claim; they were limited, unreliable, and expensive and not all in the industry believe they would ever take over "proven technology like radio tubes". Same, it actually took a good while before silicon memory replaced core-memory. A more recent example: CMOS was originally the slow technology, only suitable for watches and cheap calculators.
Source: countless books on the early history of Z1-3, ENIAC, Cray, Intel, etc.
1. Computation before ~1800. Abacus, Napier's Bones, Slides rules, Pascal's Calculator, motivations from celestial navigation and astronomy.
2. Modern analog computers ~1900-1950. The author seems to refer to them as "math machines" and leaves it at that, without exploring much deeper than that they were used for besides calculating firing solutions for artillery. I think the author lacks a solid grasp of how mathematical tables were used from 1614 onwards, and that analog computers were used to create much more accurate and complex tables which could be used for more accurate firing solutions. And for other purposes as well, beyond code-breaking.
>"It's hard for me to wrap my head around the fact that early, pre-general purpose computers (~1890-1950s) weren't computers in the way that we think about computers today. I prefer to think about them as "math machines"."
>"But subsequent machines were able to do math. From what I'm seeing, it sounds like a lot of it was military use. A ton of code-breaking efforts during World War II. Also a bunch of projectile calculations for artillery fire."
3. Poor description of the advent of electronic computers.
>"Then in the 1940s, there was a breakthrough.[10] The vacuum tube took computers from being mechanical to being electric. In doing so, they made computers cheaper, quieter, more reliable, more energy efficient, about 10-100x smaller, and about 100x faster. They enabled computations to be done in seconds rather than hours or days. It was big."
It was certainly a breakthrough, but the idea that computers immediately became quieter, cheaper, and more reliable is false. They were much larger, initially, compared to analog computers of the era. By almost any measure, they were also much less efficient with energy, though this may depend on what sort of calculations you are doing - I'm less sure of this.
4. Incomplete and incorrect descriptions of programming languages and the history of digital logic. No mention of information theory and Claude Shannon, digital circuits.
This is a poor analogy that misleads a reader who is unfamiliar with programming languages, it obscures the abstraction:
>"Think of it like this. It's translating between two languages. Assembly is one language and looks like this: LOAD R1, #10. Machine code is another language and looks like this: 10010010110101010011110101000010101000100101. Just like how English and Spanish are two different languages."
5. Lack of understanding of digital hardware.
The author never describes why or how vacuum tubes and then transistors allowed computers to use logic that is both digital and electronic.
The author jumbles a lot of ideas into one and does not seem to understand the relationship and distinction between the evolution of transistor technology (point-contact -> BJT -> FET -> MOSFET) and the creation of integrated circuits.
>"Before 1966, transistors were a thing, but they weren't the transistors that we imagine today. Today we think of transistors as tiny little things on computer chips that are so small you can't even see them. But before 1966, transistors were much larger. Macroscopic. Millimeters long. I don't really understand the scientific or engineering breakthroughs that allowed this to happen, but something called photolithography allowed them to actually manufacture the transistors directly on the computer chips."
6. Lack of historical context. No mention of the motivations for creating the vacuum tube or transistor: amplification and switching for use in telegraph and phone networks. No mention of the role the US government played beyond the 1860 Census, no mention of continued investments motivated by the Cold War, Apollo Program, ICBMs, etc. They briefly cover artillery firing solutions and mention code-breaking.
7. Over reliance on LLMs to research and write this.
Hard to take a history which includes this seriously:
>"And from what ChatGPT tells me, it's likely that this would have been an investment with a positive ROI. It'd make the construction of mathematical tables significantly faster and more reliable, and there was a big demand for such tables. It makes sense to me that it'd be a worthwhile investment. After all, they were already employing similar numbers of people to construct the tables by hand."
>"Anyway, all of this goodness lead to things really picking up pace. I'm allowed to quote Claude, right?"
Author implies (in the quote above which occurs after the discussion of the invention of personal computers) that the early personal computers from the 1970s did not use microprocessors. This of course is false: All the early "personal computers" used microprocessors. For example the IMSAI used an 8080, the Apple II used a 6502, and the TRS-80 used a Z80. Microprocessors -- which were never intended to be the basis for entire general-purpose computers -- were repurposed for exactly that application by visionaries like Woz. Microprocessors made personal computers possible.
It would be more correct to state that in the 1980s microprocessors began to replace integrated circuits in mini and larger computers.
A subtle related point is that it would be even better to point out that by "integrated circuit" above the author really means discrete small-scale integrated circuit. All microprocessors are integrated circuits, but not all integrated circuits are microprocessors. Microprocessors are large-scale integrated (LSI) circuits or nowadays very large-scale integrated (VLSI) circuits.
Not only against the analog computers of the era. Early vacuum tube computers were significantly less reliable and less energy-efficient than electromechanical digital computers, like the Harward Mark 1.
https://www.theregister.com/2012/12/22/verity_stob_8086_and_...