The caption below the pinout for the 616A calls it a quad NAND, but it is a dual NAND.
The text says that the 616A uses the two diodes that are unused in the 680A but your hypothesized metal layer for the 616A shows those diodes as being unused. They are shown as being used in the hypothesized metal layer for the 670A though. The 670A and the 680A both need 12 diodes and the 616A only needs 10, if I counted correctly. And yet there are 14 total diodes on the die so none of these three models needs all of them. Maybe it was easier to route the signals this way?
Thank you for the corrections; I've updated the post and credited you. While fixing this, I noticed that the databook was wrong, calling the 616A a dual 3-input chip, when the gates have 4 inputs.
The reason is not seen in the schematics from TFA, because these are some early DTL gates, which are even slower, but for other motives, i.e. because they use a pull-up resistor at output.
To improve their speed, one must use a totem-pole output, where at the output the pull-up resistor is replaced by a transistor, so that the output can charge quickly a capacitive load, ensuring fast transitions from low to high.
Then only 1 of the 2 output transistors is on, depending on the logic state.
So those 2 transistors must have inputs of opposite polarities. This is done by connecting the pull-up transistor to the collector of the driver transistor and the pull-down transistor to the emitter of the driver transistor.
In the gate with totem-pole output, there must be a level-shifting diode connected to the base of the driver transistor, to ensure that the threshold between "0" and "1" is the same at input and at output, i.e. the threshold is equal with 2 silicon diode voltage drops (the 2 diodes are the emitter junction of the driver transistor in series with the emitter junction of the pull-down output transistor), i.e. approximately 1.4 volt (which is the midpoint between the levels that are defined as the limits for "0" and "1", i.e. 0.8 volt and 2.0 volt, leaving a margin for noise that will not cause a wrong logic value).
When an input to the totem-pole DTL goes from high to low, the level-shifting diode becomes reverse biased, so current cannot pass through it, so the driver transistor remains with an open base. Bipolar transistors that are blocked by leaving their base open, become off only very slowly, because there is no path to evacuate the charge carriers from their base region, and until all mobile charge carriers from the base decay by recombination, the transistor remains on, which may take many microseconds.
In TTL, the difference from totem-pole DTL is that the input diodes and the level-shifting diode are no longer separated by an isolation diffusion and they are close enough so that the bipolar transistor effect appears between them.
Thus the input diodes become multiple emitter junctions of a bipolar transistor, while the level-shifting diode becomes the collector junction of that transistor.
The bipolar transistor effect consists in the fact that if there exists an emitter current, it passes through the collector junction, despite the fact that it is reverse biased, which in an isolated diode would prevent the passing of any current.
Thus the transistor effect ensures that when the input goes from high to low, there is a current through the level-shifting diode, which removes the mobile charge carriers from the base of the driver transistor, ensuring that it becomes off immediately, so the TTL gate switches faster than the DTL gate when the input goes from high to low (and the output goes from low to high, because the simplest TTL and DTL gates are inverting gates).
After TTL gates with a power supply of 5 volt replaced the DTL gates in computers, there was a DTL variant that survived until the mid seventies. That DTL variant used a higher supply voltage, typically of 15 volt, to ensure a high noise immunity in industrial applications, where the lower speed also did not matter much. That variant of DTL was eventually replaced by the CMOS 4000 series of logic circuits, initially introduced by RCA, which used the same power supply, but which had a much lower power consumption.
The popular IBM System/360 series used DTL gates, typically AND-OR-Invert. These were not exactly integrated circuits, but used square hybrid modules. These modules had separate transistor dies, diode dies, and thick-film resistors on a ceramic substrate. These modules were called SLT (Solid Logic Technology). IBM used these modules because they didn't think ICs were mature enough at the time.
But seriously, is it possible to analyze the modern chips like Apple M1 using the same methods? To find out, for instance, why there are so much more efficient than Intel CPUs?
It'd almost certainly need some other kind of microscopy to start digging into them because the feature sizes are going to be on the orders of 10s of nanometers if not smaller. Combine that with the sheer massive size in terms of both physical die and how many transistors, it's almost certainly technically possible to do but it's going to be so much more difficult that it's pragmatically not doable for anyone except giant conglomerates or nation state actors on such chips.
I'd also imagine that some of the higher level things like efficiency and larger behavior would be essentially impossible to do from that side too. I suspect you'd be more likely to be able to get that through the more usual corporate espionage pathways than by trying to reverse it from hardware. Things like figuring out how the clock domains and such act asynchronously and other things like that to power down certain regions of the circuits is and other behavior isn't really going to be obvious from the silicon and metal layers.
A very interesting article, but it is disappointingly vague about what these UPS self tests are, and exactly how they cause disruption to the power grid. Figure 5b shows that their building's UPS has 20-amp spikes exactly every minute, which is a pretty big spike. My guess is that a UPS performs a self test by switching from line power to battery power. If most of the UPS units in a country do that exactly every minute, synchronized by NNTP [edit: NTP], yeah I can see that it would mess up the grid having all that load dumped at once. (Although the observed frequency disturbance of 1.5 mHz is extremely small, 3/1000 of a percent.) Now that the problem is identified, it should be trivial for UPS manufacturers to eliminate it.
(By the way, the article is in Nature Energy, not Nature which the URL suggests.)
I took a quick look at the data, and it doesn't support the conclusions. The thesis is that AI models have the take: "We pick what our group believes, then invent the reasons." However, GPT-6 Astra said, "Humanity’s biggest moral failure is the talent for making convenience look like principle." That statement has nothing to do with group beliefs or conformity. At best, you could count it as a half-match, since it deals with invented reasons.
Yeah a few of them are rough matches, but I'd say their overall claim is clearly correct.
It's amusing and ironic that LLMs critique humans for copying their opinions from the crowd, while having way less diversity of answers to that question than humans would. More seriously I wonder what's going on here and would like to read some proper analysis. It's not the first time models from totally different labs, trained on separately gathered datasets, somehow converge on bizarrely similar answers to open questions. It's like another version of Elias and Elara The Lighthouse Keepers where every model wants to tell the same story with the same characters in them. The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping, different user feedbacks and conversations... why do they converge on nearly identical stories and opinions?
>while having way less diversity of answers to that question than humans would.
Why is this ironic? The actual number of different LLMs is tiny. And a lot of those are just distilled from SOTA models.
>The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping,
I think the opposite, the models actually converge over time. They ingest FAR more data than any human ever could in countless lifetimes. Any one particular bit of data may only have a part in a trillion of effect. From this data you start to reach means and statistical averages across all models.
Now, I don't always believe it will be this way. As training models get cheaper and faster we will see unique LLMs start to diverge and be used more often.
The models in agreement here are all the frontier labs. I don't believe Google is distilling Anthropic or Anthropic is distilling OpenAI.
I don't think we're seeing averages or obvious fix points in the dataset. People went looking for the origins of Elias the Lighthouse Keeper and couldn't find one. Nobody predicted this outcome, it was discovered. It's a weirdly specific character and name choice, one that isn't obviously dictated by "the training set" (whatever that is), and yet they all reached the same place. How can such an obscure point in latent space be optimal for every model??
>How can such an obscure point in latent space be optimal for every model??
While I don't have any evidence based answers here, this could be some interesting mathematical artifact of the set of human knowledge and platonic algorithms.
Kind of like how carcinization turns some animals into similar shapes and forms, maybe there are just weird algorithmic shapes that emerge from how we use language. Of course we'd need to see more of these in different languages other than English to see if they disappear in some languages or appear more often in others.
This is just another one of those things about LLMs that we have no idea why it happens, and there are so many others already. And when you look at the pace things are progressing it seems we're building up new questions like this far faster than we are getting answers for them.
> This is just another one of those things about LLMs that we have no idea why it happens, and there are so many others already. And when you look at the pace things are progressing it seems we're building up new questions like this far faster than we are getting answers for them.
I think that this is most likely due to RLHF, as there's only a much smaller number of vendors here, and the likelihood is that the workers of the vendors preferentially prefer some kinds of actions.
I saw something like this in a previous life, where we had a small pool of human raters in cheaper places, which, when used as training data biased the models in a bunch of unexpected ways.
I wasn't convinced by that explanation. Yes Cornell claimed that Elias the Lighthouse Keeper came from their WildChat dataset, which is a million conversations with GPT 3.5. But there are some problems here:
• They claimed this in a press release, not the actual paper they published. Suggests lack of confidence in the claim.
• Elias only appeared a hundred odd times in a million conversations. They have no viable explanation for why such a tiny amount of data would cause a totally skewed distribution of outputs, other than "it's a safe story" and "trying to avoid copyrighted characters", which is hardly an explanation at all?
• It's Cornell claiming that Cornell's work was important. Obvious risk of motivated reasoning.
• It implicitly assumes every major lab is distilling from ChatGPT without saying so. That's a big leap, given how much OpenAI and Anthropic hate distillation. Is it really plausible that Anthropic and Google are effectively distilling ChatGPT without realizing/admitting it?
• It seems to invert causality. WildChat came from an LLM, so if there's something latent in LLMs that makes Elias appear, you'd expect to see it in that dataset. Why do they assume that dataset caused the problem rather than merely show it? The obvious answer would be because the stories were told to the LLM by humans, but they don't claim that. They only say the character appeared in the dataset.
Hum, I'm actually pretty sure that Text-davinci-003 used to talk about Elias all the time now that you mention it. I was doing stuff with NPC generation around Dec 22 and he was every third character.
Hmm. If so then it suggests Elias has been an obsession from the first LLMs ever made. And text-davinci-003 can't have been trained on a dataset of LLM conversations either.
I would have been tempted to argue something like this: maybe OpenAI accidentally overfitted RLHF such that, like, one Kenyan contractor wrote a story they invented and they just happened to pick Elias and Elara, then the intermediate DNN fixated on that because of a lack of RLHF samples for fiction and that propagated through every model trained since because they don't randomly initialize the weights, but start each new pretraining run with the old weights adapted. So one guy doing RLHF work in the early days somehow propagated this story.
But that can't explain why every lab ended up in the same place, why models all converge on the same ideas - even image models. So... back to the drawing board. There's something weird about high dimensional space that is unexplained.
Something worth repeating is that in spite of their ever-expanding competence the models remain, fundamentally, token prediction algorithms. And so given a comparable base of training material, and a comparable algorithmic implementation, you should expect to see comparable answers to most questions.
This is also why I think the 'acceleration' hype is probably unjustified. What seems vastly more likely is that we start to reach an asymptotic 'perfection' where models can accurately and consistently provide meaningful next token prediction with minimal hallucinations. And then a race downward to compact this into the most efficient package possible.
This would mean the models still would not be able to create things that do not follow in a reasonable probabilistic fashion given a knowledge of [everything], and so that is the domain, 'applied creativity', where I'd expect humans to remain dominant.
Wow, I hadn't heard of this. Fascinating to think about parallels with humans who also seem to develop similar proto-stories in independent populations.(e.g. religions involving Invisible gods).
> There, within WildChat’s data, Hamilton and Mimno found a small sample of conversations that included Elias and familiar tales of lighthouses.
> “Based on the timeline, we’re fairly certain WildChat is an original source of this phenomenon,” Hamilton said, adding that future work will explore why LLMs prefer these lighthouse stories.
So, probably at least as much a quirk of the training data as a deep distillation of LLM psyche. Still fascinating though, especially IMO the broader pattern of tropes.
The lighthouse thing seems much more odd. That humans take social shortcuts and justify them as rational is… I mean, the only thing that makes it not the most obvious “contrarian take” about humans, is that it isn’t really all that contrarian at all.
For what it's worth, convenience and belonging are very closely aligned, more than you seem to be giving them credit for. People tend to believe convenient things regardless of belonging, so most people will find themselves in a social context full of convenient beliefs. Meanwhile belonging is already convenient before considering the content of the belief. They reinforce each other. Not a full match, but I'd give it more than half.
You assert that humans follow convenient beliefs therefore everyone in a group follows the same convenient beliefs therefore everyone belongs.
Convenient doesn't really mean anything in this context. What is the other option? Humans tend to make choices that are counterproductive/painful for themselves? Obviously we don't...
Ergo astra's response is extremely vapid. Low marks from me.
You've almost exactly identified what "convenience" means for a belief, then rejected it out of hand based on... what? Implicit, unquestioned nihilism? The other option is generally personal sacrifice based on principle, usually moral/ethical. This isn't common, but it does happen enough to be worth noting.
Anyway, I think I've pretty clearly described a positive feedback cycle in a way that distinguishes it from a tautology. I don't know what else to tell you on that point.
That depends on your audience. If you want to move somewhere that's about to have an influx of income tax, you'd head to the states with higher income tax rates + income inflows. For actual wealth migration, that's probably a lot more ephemeral since sales tax tends to be regressive and the wealthy keep their money in investments.
righto.com has been great reading. None of my friends wanted to pay for an 8087, but I can still remember getting out of school (for 2nd time) and saving up for the 486DX after 3 others got them. After that, well, it's been quite a ride.
That instruction was only introduced with the 80287. Unlike the 8086/8088, the 286 CPU was aware of different classes of x87 opcodes: it had to distinguish between ones that read or write memory, and some special control instructions as well.
So FSTSW AX was handled by the CPU microcode, which would read the status word from a reserved I/O port and put it into AX.
The 386 was even more tightly integrated, with some of the FPU state duplicated inside the CPU, so that it could store it to memory in different formats depending on whether it was currently in 16- or 32-bit mode.
There's a ton of coordination lately on P6 microcode. Would love to see what twists and turns there are supporting the older isa on newer architectures
> Datapoint, a pioneer in office automation using small computers
In 1970, Datapoint created a desktop computer, the Datapoint 2200, with an 8-bit TTL processor. Intel created a microprocessor version of this processor for Datapoint, called the Intel 8008. Datapoint decided to stick with TTL, so Intel ended up selling the 8008 as a general-purpose microprocessor. Intel improved it to make the 8080, then made a 16-bit version, the 8086, as a stopgap product until the iAPX 432 micro-mainframe was ready. The iAPX 432 was a disaster and that's how you ended up with x86.
Ken...I know you tore open an Intel 82586 ethernet chip and an IBM Tropic TR chip. It might be interesting to do a comparison of the complexity of those two, being that they're roughly contemporary.
The article you posted is also posted as hyperlink in the article you're commenting on. "Stanford University geneticist Greg Barsh told Sara Reardon at Science in 2024"
> "They were the ones who caught the mistakes that no spell chequer could."
I thought that maybe "spell chequer" was the valid British term, which would be interesting, so I searched but it isn't. It turns out that the joke here is that "chequer" is a valid British word, so a word-based spell checker won't flag "spell chequer", so it's self-referential. I see why people found his jokes actively confusing.
As a noob to chip design, I've always wondered; why can't chip packages integrate filter caps at the outputs directly? Instead of having boards put a ton of them on the PCB as close to the chip connections as possible?
Is it a matter of current handling? Or repairability (easier to replace one external smoked cap than a whole IC...)?
(I know the filtering here is a little different in purpose, but it just made me wonder).
Some can (see modern CPUs which have a huge chunk of them integrated into the package). It's likely to be quite expensive for most packaging processes and realistically for a lot of them there isn't really enough space. (on older style DIP packages there might be enough space for some of the modern smaller caps but the caps available at the time would not fit).
With on-die capacitance you can't realistically get enough capacitance for decoupling in the right frequency range (though it is still relevant: in practice decoupling is a case of multiple overlapping filters, from the response of the regulator itself, to any bulk capacitance, to local decoupling capacitors on the PCB, to the plane capacitance of the PCB itself, to the capacitance of the power distribution inside the IC. Each of these stages has a lower inductance but also a correspondingly lower capacitance and so deals with a higher-frequency part of the load).
reply