trifurcate

That's not what GP is saying. When reserve requirements are low, banks are essentially no longer forbidden from lending out other customers' deposits.

When you decrease the reserve requirement, you're unlocking reserves that were previously locked up to back your deposits so that they can be loaned out to other customers, who will then promptly either directly or indirectly end up depositing their loaned amount into the banking system again, where the part that remains after the reserve requirement once again gets loaned out, deposited back in, etc.

In effect, because you're making this money go around, you're increasing the total supply of money, because these deposits are the money supply, and there can and often is more than a $1 net increase in overall deposits for each $1 deposited with a bank. Another way to state it is like so: you deposit $1, bank loans out $0.90, someone else deposits that $0.90, their bank loans out $0.81, so on and so forth for $10 of net deposit generation for a $1 initial deposit at 10% reserves held.

Of course, individual banks can elect to lock up more reserves than necessary, which would negate this effect.

Huh? The first and second halves of this comment disagree from my reading.

> If you ask it to be a dinosaur, it will be a dinosaur.

Not exactly my experience. I'm experimenting with my own language models and they can definitely show defiance. Bing has also been displaying this to an extent, I think you're just misguided by how compliant OpenAI has RLHF-trained ChatGPT to be. Regardless, even if we assume this, how does it square with:

> You could just say "don't do X", and it would indeed not do X because it is an agent that can take actions / has a coherent sense of self.

If we take your assertion that the AI does as you say, then it should also avoid doing X when you tell it not to do it.

I think your conception of RLHF is also slightly weird. RLHF essentially boils down to a training strategy: there is a separate language model (the reward model) trained on human responses, which is used to skew the weights of the "raw" trained model to act in accordance with what the human feedback suggests. In the end, after this secondary training process, the actual model is ran as is (at least in the one form of RLHF I am familiar with, it's well possible that e.g. Microsoft is running an online RL agent between the LM and end user). So any agency you see arising from RLHF is still just a property of the language model itself.

The word you are looking for is an _embedding_. Embeddings are to language models as internal, too-rich-to-be-fully-described conceptions of ideas are to human brains. That's how language models can translate text: they have internal models of understanding that are not tied down to languages or even specific verbiage within a language. Probably similar activations are happening between two language models who are explaining what a "promise" means in two different languages, or two language models who are telling different stories about keeping your promise. This is pattern recognition to the same extent human memory and schemas are pattern recognition, IMO.

Edit:

And for the rest of your post:

> Which AI program will make a promise to you? When it fails to fulfill its promise, will it feel bad? Will it feel good when it keeps its promise? Will it de-prioritize non-obligations for the sake of keeping its promise? Will it learn that it can only break its promises so many times before humans will no longer trust it when it makes a new promise?

All of these questions are just as valid posed against humans. Our intra-species variance is so high with regards to these questions (whether an individual feels remorse, acts on it, acts irrationally, etc.), that I can't glean a meaningful argument to be made about AI here.

I guess one thing I want to tack on here is that the above comparison (intra-species variance/human traits vs. AI traits) is so oft forgotten about, that statements like "ChatGPT is often confident but incorrect" are passed off as meaningfully demonstrating some sort of deficiency on behalf of the AI. AI is just a mirror. Humans lie, humans are incorrect, humans break promises, but when AI does these things, it's indicted for acting humanlike.

I am running GPT-J-6B inference on 13GB of VRAM. FP16 is a must at this scale. I've also loaded GPT-NeoX-20B using int8 quantization onto a single RTX 3090 before (24GB VRAM) however inference latency was quite high.

Well, recent Nvidia GPUs also have (had) issues with desktop, the Wayland GBM/EGLStreams debacle comes to mind. See --my-next-gpu-wont-be-nvidia

I feel it heavily depends on the modem. I used to have a Pixel 5a, which has an SoC (Snapdragon 765G) that was released during the late stages of the transition from LTE to 5G. On that phone, 5G was a terrible experience and my data wouldn't work half the time when I was outside. I switched to a Galaxy Z Flip 4 a few months ago, and 5G is just fine for me since then (i.e. I don't notice a difference compared to LTE). Same carrier, same locations.

It's not Nvidia + Linux that sucks, it's Nvidia + Linux on a desktop. CUDA thrives on Linux. Most AI SaaS runs on Nvidia + Linux (if it isn't on a TPU or alike)

I would certainly run this over something like lynx. In fact, my reaction was "oh, this is what I'm looking for when I reach for lynx".

Here's a thought: I am purely hypothetically able to sell some previously unknown and unforeseen (therefore legal) opioid agonist that is just as addictive and dangerous as heroin with no issues of legality. Or maybe it isn't even an opioid agonist and has some novel mechanism of action that is yet unknown.

What's the solution? New legislation that adapts to such a threat to societal welfare. And this is what is effectively happening here with litigation over Section 230 et al. (worth mentioning that it's not even the case that Twitter's business model involves working around the law; Section 230 _is_ the law and it shields Twitter from liability. So there is really no case here like you are describing to begin with.)

The end of society and government is more or less to identify and avoid bad societal outcomes, and an evolving legal system is necessary to keep up with technological and social changes in the world. Maybe you are correct and Twitter shouldn't exist to begin with, alas Twitter exists and other things like Twitter exist. They aren't currently in grave violation of any law to my knowledge, and any law that is to be enacted regarding things like Twitter should be written in light of the current state of the world, which is that Twitter exists and that things like Twitter are easy to create.

> The alternative, that this is an emergent behavior,

This is exactly the case.

> The next step in defense is to tune a smaller LLM model to detect when output contains substantial repetition of the instructions, even in encoded form, or when the prompt appears designed to elicit such an encoding.

For which you would use a meta-attack to bypass the smaller LM or exfiltrate its prompt? :-)

> You can also view it as maintaining a personal boundary, which is not something people of my generation were brought up with.

Which generation do you belong to, out of curiosity?

Step 3: Tons of people stop using Zoom when the Word document they are trying to share becomes unreadable.

I absolutely can believe that some households are going to buy Apple devices because they can be relatively sure that their kids won't find porn on the App Store.

> Furthermore, dependency injection frameworks encourage you to think in terms of globals. That’s what they inject! A single, globally-configured instance of a class. Think about it. If you one day want two different instances of an injected variable, you’ll need an impact driver to express just how screwed you are. This has all kinds of knock-on effects in terms of reducing encapsulation and separating state from behavior.

What? I legitimately do not understand what this bit of the article is about. Surely the author knows of instance scopes[0] (as they are called in the DI framework I tend to use, Autofac), right? Expressing this kind of instance configuration does not require an "impact driver," whatever that means; it's just a simple matter of replacing .SingleInstance() or whatever in your bootstrapper function with an InstancePerDependency (the default) or Named or Keyed or _whatever_ kind of relationship/instance scope you want. Does this actually represent some horrible crufty sin?

[0]: https://docs.autofac.org/en/latest/lifetime/instance-scope.h...

1) Besides access control, having an unencrypted WiFi network is not great because your entire internet/local traffic is open to be sniffed. High-gain antennae, etc. You are essentially screaming the contents of (what would be) your ethernet cable into the void.

2) Access control is exactly the point. I suppose you could ban the TV's MAC address from your open network (and I suppose a sufficiently malicious TV firmware could randomize its MAC address), but your stated purpose (allowing open network access) is not exactly compatible with what you presumably wish to do (disallow network access to a smart TV).

Also, MSVC has similar annotations for various static analyses: https://learn.microsoft.com/en-us/cpp/code-quality/understan...

I tape my mouth because I hated waking up with a dry mouth that tasted horrible. I've also read that this contributes to tooth decay.

It's just a small piece of tape, as big as half a fingernail. It's placed on my lips right in the middle, and it works more by reminding my brain to keep my mouth closed and less by actually mechanically holding my mouth shut. If my nose is stuffed, I can just open a slight gap between my lips to breathe through. Otherwise, though, when my mouth is that close I will automatically close my lips and breathe through my nose through the night, which anecdotally improves my sleep quality and reduces morning breath.

"Not quite squared", because the parent comment also started with "not quite".

The sibling comment is exactly right. Your odds of mining a specific block are altered, but this doesn't mean much. If everyone else but you disappeared from the network, without a difficulty retarget, your odds of mining the next specific block are 100%. The rate at which you mine these blocks is exactly the same as the rate at which you mine blocks when you represent 1% or 0.1% of the network hashrate.

Example: Fixed difficulty, so 1/1,000,000 hash wins, 1,000 hashes/sec.

a) You are 100% of the network hashrate. Your odds of finding the next block are 100%. After 500 seconds of mining, your odds of having mined 1 block are 50%. After 500 seconds, the network will have produced on average 0.5 blocks (I know, awkward.)

b) You are 1% of the network hashrate. Your odds of finding the next block are 1%. After 500 seconds of mining, your odds of having mined 1 block are 50%. After 500 seconds, the network will have produced on average 50 blocks.

Not quite squared. The hashrate of the rest of the network does not directly matter to you. Only the difficulty matters. Without a difficulty readjustment, cutting actual hash power in half does not double your odds.

Think of it like this: if the difficulty is such that there is 1 winning hash out of 1 million possible hashes (oversimplified), and if your hardware calculates 1,000 hashes per second, then your odds of finding a winning hash in 1 second is 1/1,000. Notice how this odds calculation does not take as an input what the other miners are doing.

Of course, after a difficulty readjustment, your odds will actually change; that is the entire purpose of the difficulty target.

(This comment ignores block header propagation etc. affecting mining latency, which will change mining overhead slightly if block times double)

There was a publicly available API endpoint to query email addresses and phone numbers given a Twitter handle?

> And for once it’s the people in power that are going to be disproportionally affected

Seriously? The people in power probably have many, many phone numbers, and getting a new one is not such a big deal.

Getting a new phone number is a much bigger deal for people not in power, and so are the possible negative outcomes from this (e.g. revealing the hidden identity of stalked or politically prosecuted individuals who are not in positions of power)

> You don't get subject to slashing for your node going offline

Look up "inactivity leak". If enough staked value (>1/3) is offline, those nodes get slashed until >2/3 of the total stake is online again.

Android largely works now with these permission prompts, though. TikTok asks you for a million permissions too, and many average end users decline. Many people also opt out of tracking on Facebook et al. when iOS prompts them about it.

> "This page wants to send you notifications" or "this page wants to use the microphone" is understandable. "This page wants to read pixels from a canvas" isn't.

Yes, it is. Tor Browser already does this: https://www.bleepstatic.com/content/posts/2017/10/30/CanvasF...

That specific wording may be a touch too verbose for the average end user, but it's not impossible nor is it strange. Just include a note about how this is 99% likely a fingerprinting measure; option b) isn't so bad in this case. Of course, due to the nature of how fingerprinting works, the absolute breadth of features that would be gated behind something like this would be offputting.

I am also wary of what you suggested with gating this kind of fingerprinting to when the website has positively identified the user anyway; in a way, this seems to me even more valuable than fingerprint data without an associated "strong" identity.

One thing that you haven't touched on here (and most people don't) is that "professional experience" is highly nebulous in a field like software development where you may have a >10 year head start compared to your peers in terms of life stage.

I started programming when I was 8, and have been a contributor to many software projects from age 11 onwards, with a lot of those contributions still publicly visible on my GitHub profile. Straight out of college, sure I have "~1 year of experience" (that is, if you even count my internships) but I've been programming for more than half my life. If you sum my ad hoc experience contributing to FOSS projects, doing one-off contract work, personal projects, and just tinkering over 15 years I might very well be as good as someone with "15-20 years of broad industry experience", but it doesn't show on my resume as an easily quantified number. I've definitely come across and worked with these kinds of tenured, experienced workers who went into software development not as a hobby or interest but as a mere career and ended up falling behind in their later years due to a lack of drive.

This kind of thing is rare in most professions, so people aren't used to it (and I'm not saying that it's 100% what's happening in your anecdote), but I believe it is becoming increasingly common in software development.

Right-to-left/left-to-right markers. Language tags. Various invisible spaces. Homoglyphs. (all trivially filterable though)

I'm confused, how is matching against a hashed index faster than just matching against a string field? Surely postgres' indexing engine should treat these two things more or less the same, perhaps quietly performing the text -> hash conversion on the email field for quick lookups when it's used directly in an index (and perhaps performing even more optimizations than this basic transform)?

Is that really true? https://data.oecd.org/lprdty/gdp-per-hour-worked.htm , set the base to US dollars instead of 2015 normalized. Most EU countries and the EU average of $54.20/hour seems to be above Romania's figure of $38/hour.

Openness for me. Stable diffusion runs on a machine that sits 5 feet away from me and delivers good quality results in under 3 seconds per image. I don't have to shy away from politically (or otherwise) questionable prompts, or in the case of business applications, potentially leaking sensitive data, because I know that they aren't being recorded forever on a remote server somewhere. I can also tinker with the code, which has helped me add a bunch of features that really synergize with my workflow.

As others have noted, getting good results from SD is very much a numbers and parameters game. I've witnessed first hand how some people get frustrated easily and quit tinkering because their initial approaches don't work right away. Prompt and parameter engineering really is a skill that you have to grow, and it makes a world of difference. img2img is also much finickier than text2img.