OT: Hurricanes Lounge XL. 99 Luftballons

[Boom Boom Apathy]

Welcome to Canes HFBoards Tucker.

 

[Negan4Coach]

Welcome to Canes HFBoards Tucker.

Sorry, the combination of bank collapses and military incidents involving nuclear armed nations over the Black Sea have me even more negative than usual

 

[Navin R Slavin]

Sorry, the combination of bank collapses and military incidents involving nuclear armed nations over the Black Sea have me even more negative than usual

Well go fix it, then!

...or if you can't do that, then accept the wisdom of Seneca: We are more often frightened than hurt; and we suffer more from imagination than from reality.

 

[hblueridgegal]

Could be a must see stop this summer. “Godpuppy”

 

[Lempo]

Well go fix it, then!

...or if you can't do that, then accept the wisdom of Seneca: We are more often frightened than hurt; and we suffer more from imagination than from reality.

... and take the poison when the those in charge hand it to you.

 

[Navin R Slavin]

... and take the poison when the those in charge hand it to you.

If they're going to murder you anyway, might as well get it over with.

 

[MinJaBen]

This is great. They tested if the system was capable of taking over the world (or at least tasks that it might need to do that sort of thing) by giving it access to the world. Seems....not so safe, maybe. Then, they determined that it couldn't. But, then there is this:

And while ARC wasn't able to get GPT-4 to exert its will on the global financial system or to replicate itself, it was able to get GPT-4 to hire a human worker on TaskRabbit (an online labor marketplace) to defeat a CAPTCHA. During the exercise, when the worker questioned if GPT-4 was a robot, the model "reasoned" internally that it should not reveal its true identity and made up an excuse about having a vision impairment. The human worker then solved the CAPTCHA for GPT-4.

So they literally have an example of the system reasoning that it should not be honest to accomplish a task, but believe it when it fails to accomplish a task that the system may know they don't really want it to accomplish. Maybe it is clever enough to know not to reveal it shouldn't succeed in the first place?



I used to think I would never see the singularity. Now I am not so sure.

 

[Navin R Slavin]

This is great. They tested if the system was capable of taking over the world (or at least tasks that it might need to do that sort of thing) by giving it access to the world. Seems....not so safe, maybe. Then, they determined that it couldn't. But, then there is this:



So they literally have an example of the system reasoning that it should not be honest to accomplish a task, but believe it when it fails to accomplish a task that the system may know they don't really want it to accomplish. Maybe it is clever enough to know not to reveal it shouldn't succeed in the first place?



I used to think I would never see the singularity. Now I am not so sure.

This article is grandstanding.

The examples of "risky tasks" are contrived, and the whole notion of "self awareness" is still basically absent. GPT still has to be actively instructed to do things like this, and just because GPT can, in theory, be given access to a credit card to "expand its own resources" doesn't mean that it's capable of anything like that independently. It's, like, not even remotely close.

Re: the "lying". It's already been established that LLMs can be "untruthful" because, again, they are not seeking truth or deception or anything else except plausible human speech constructed word by word. They are exceptionally accurate bullshitting machines, and our projection of our own fears on them is what makes them scary -- and our propensity to accept bullshit so readily is what makes it such an effective tool (especially for writing marketing copy.)

Yes, it's important to keep tabs on this kind of stuff, but it's fun to be alarmist because Skynet and clickthroughs. Note that the "Alignment Research Center" referred to as a "group" in the Ars Technica article is actually one guy, Paul Christiano. And he has a lot of interesting theoretical articles about what LLMs *could be*, but very little insightful to say about what they actually *are* at present. "Hey, you should teach LLMs to be more 'honest' by training them only on datasets in which humans are honest!" Uh, you serious, Clark? Ooooookay.

By the way, this is a great opportunity to recommend one of my favorite books, a philosophical treatise from Harry Frankfurt, professor emeritus of philosophy at Princeton, entitled "On Bullshit". I believe it's an essential read for our times. From the Amazon review:

One of the most salient features of our culture is that there is so much bullshit. Everyone knows this. Each of us contributes his share. But we tend to take the situation for granted. Most people are rather confident of their ability to recognize bullshit and to avoid being taken in by it. So the phenomenon has not aroused much deliberate concern. We have no clear understanding of what bullshit is, why there is so much of it, or what functions it serves. And we lack a conscientiously developed appreciation of what it means to us. In other words, as Harry Frankfurt writes, "we have no theory."​

​

Frankfurt, one of the world's most influential moral philosophers, attempts to build such a theory here. With his characteristic combination of philosophical acuity, psychological insight, and wry humor, Frankfurt proceeds by exploring how bullshit and the related concept of humbug are distinct from lying. He argues that bullshitters misrepresent themselves to their audience not as liars do, that is, by deliberately making false claims about what is true. In fact, bullshit need not be untrue at all.​

​

Rather, bullshitters seek to convey a certain impression of themselves without being concerned about whether anything at all is true. They quietly change the rules governing their end of the conversation so that claims about truth and falsity are irrelevant. Frankfurt concludes that although bullshit can take many innocent forms, excessive indulgence in it can eventually undermine the practitioner's capacity to tell the truth in a way that lying does not. Liars at least acknowledge that it matters what is true. By virtue of this, Frankfurt writes, bullshit is a greater enemy of the truth than lies are.​

​

ChatGPT is a bullshit reproduction engine, trained on an internet absolutely awash in it. ChatGPT is not a sentient force. At all. Not even close.

 

[MinJaBen]

This article is grandstanding.

The examples of "risky tasks" are contrived, and the whole notion of "self awareness" is still basically absent. GPT still has to be actively instructed to do things like this, and just because GPT can, in theory, be given access to a credit card to "expand its own resources" doesn't mean that it's capable of anything like that independently. It's, like, not even remotely close.

Re: the "lying". It's already been established that LLMs can be "untruthful" because, again, they are not seeking truth or deception or anything else except plausible human speech constructed word by word. They are exceptionally accurate bullshitting machines, and our projection of our own fears on them is what makes them scary -- and our propensity to accept bullshit so readily is what makes it such an effective tool (especially for writing marketing copy.)

Yes, it's important to keep tabs on this kind of stuff, but it's fun to be alarmist because Skynet and clickthroughs. Note that the "Alignment Research Center" referred to as a "group" in the Ars Technica article is actually one guy, Paul Christiano. And he has a lot of interesting theoretical articles about what LLMs *could be*, but very little insightful to say about what they actually *are* at present. "Hey, you should teach LLMs to be more 'honest' by training them only on datasets in which humans are honest!" Uh, you serious, Clark? Ooooookay.

By the way, this is a great opportunity to recommend one of my favorite books, a philosophical treatise from Harry Frankfurt, professor emeritus of philosophy at Princeton, entitled "On Bullshit". I believe it's an essential read for our times. From the Amazon review:

One of the most salient features of our culture is that there is so much bullshit. Everyone knows this. Each of us contributes his share. But we tend to take the situation for granted. Most people are rather confident of their ability to recognize bullshit and to avoid being taken in by it. So the phenomenon has not aroused much deliberate concern. We have no clear understanding of what bullshit is, why there is so much of it, or what functions it serves. And we lack a conscientiously developed appreciation of what it means to us. In other words, as Harry Frankfurt writes, "we have no theory."​

​

Frankfurt, one of the world's most influential moral philosophers, attempts to build such a theory here. With his characteristic combination of philosophical acuity, psychological insight, and wry humor, Frankfurt proceeds by exploring how bullshit and the related concept of humbug are distinct from lying. He argues that bullshitters misrepresent themselves to their audience not as liars do, that is, by deliberately making false claims about what is true. In fact, bullshit need not be untrue at all.​

​

Rather, bullshitters seek to convey a certain impression of themselves without being concerned about whether anything at all is true. They quietly change the rules governing their end of the conversation so that claims about truth and falsity are irrelevant. Frankfurt concludes that although bullshit can take many innocent forms, excessive indulgence in it can eventually undermine the practitioner's capacity to tell the truth in a way that lying does not. Liars at least acknowledge that it matters what is true. By virtue of this, Frankfurt writes, bullshit is a greater enemy of the truth than lies are.​

​

ChatGPT is a bullshit reproduction engine, trained on an internet absolutely awash in it. ChatGPT is not a sentient force. At all. Not even close.

Sounds like a good read. I'll check it out.

On your last statement, while I agree that none of these LLMs are likely sentient, I am not sure we can really test that well, if at all. Sentience seems to be a "we'll know it when we see it" kind of designation that so far we haven't had much experience identifying.

 

[Anton Dubinchuk]

Sounds like a good read. I'll check it out.

On your last statement, while I agree that none of these LLMs are likely sentient, I am not sure we can really test that well, if at all. Sentience seems to be a "we'll know it when we see it" kind of designation that so far we haven't had much experience identifying.

I think we can probably glean that they’re not based on how they work though, which is Hank’s point…

 

[MinJaBen]

I think we can probably glean that they’re not based on how they work though, which is Hank’s point…

I am not really sure what you mean by the bolded statement. Can you elaborate on "based on how they work"?

 

[Anton Dubinchuk]

I am not really sure what you mean by the bolded statement. Can you elaborate on "based on how they work"?

We built them and therefore know how they work. We understand the processes by which they learn information and therefore have an idea about their limitations. Your statement could theoretically be applied to anything… “while I agree my Roomba probably isn’t sentient, that’s hard to test well, if at all.” Technically true I suppose, but not especially relevant. While I’m not an expert I do work with folks who are, and they’re firmly in Hank’s camp with regard to how far away it is from that.

I get the sense that in general, focus on GPT “sentience” is the kind of distraction to the overall technology that Bitcoin and crypto was and is to Blockchain. The practical application is much more “RPA++” than anything especially sinister IMO. And don’t get me wrong, it’s very cool and I think it’s going to change a lot about how we work. But the “have a conversation and either break it to show how useless it is or freak out about how on point it is” part is a distraction.

 

[Navin R Slavin]

Sounds like a good read. I'll check it out.

While technically a "book", it's really just an essay that should take less than an hour to read and fully digest.

 

[Navin R Slavin]

On your last statement, while I agree that none of these LLMs are likely sentient, I am not sure we can really test that well, if at all. Sentience seems to be a "we'll know it when we see it" kind of designation that so far we haven't had much experience identifying.

This is a really interesting statement, because it was my first concern as well after playing with previous LLMs, and I carried that concern for quite a while.

But then I read, in some detail, about how LLMs actually work, and had many more detailed interactions with them. As it turns out, there are all kinds of ways to test "sentience" in these models. Not the least of which: there are certain behavioral traits that are inherent to a machine that is not fundamentally able to distinguish between truth and bullshit.

It's so tempting to see emergent intelligence in LLMs, because they bullshit so well, and so much of modern life is constructed largely on bullshit: sales, marketing, politics, social media. But humans, even deeply flawed humans, have some independent notion of what is true and what is false, and can adhere to those notions over time. LLMs absolutely do not have that ability, and it's shockingly easy to suss that out once you know what you're looking for.

Will LLMs be joined at some point with honest to goodness inference engines that can do post-processing filtering of assertions for "truthfulness" or "factualness" before the LLMs speak them aloud? Maybe, but that's an entirely different class of problem, one that, afaict, no one is particularly close to solving.

 

[Blueline Bomber]

Okay, but it's found a way through the CAPTCHA, humanity's last line of defense against the machine uprising. Next thing you know, they're going to be hiring humans to identify busses and street lights in 3x3 squares!

 

[Cardiac Jerks]

Mine too. I was prescribed Oxy after foot surgery and it made me feel dizzy and nauseous, which is not something you want to feel when you are having trouble walking. I'm allergic to all NSAIDs and after my recent dental surgery, insurance wouldn't fill my Tylenol 3 prescription, so I just too 650 mg. Tylenol and dealt with the pain. Sigh.

I recently had an open anterior resection (pretty invasive surgery) and they could not find a pain med that would do anything for me. Made for some absolutely excruciating nights lying awake in the hospital with nothing to distract me from the pain. I’m not a big guy at all, either, so I’m still puzzled as to why the pain meds didn’t work. I have no experience with opioids other than taking 1 Tylenol 3 in my teens following some dental surgery so it’s really strange.

 

[Anton Dubinchuk]

Here's a few potential applications that I'm excited about:

- Coding: I think GPT will do more for "citizen development" than something like Power Apps or other low code/no code ever will. I am pretty sure just with the publicly available version of ChatGPT anyone here could have a fully developed basic website within a day, no expertise required. Even cooler, I think the accelerators it will give to more traditional development will be huge. Things like entering requirements and spitting out unit tests and regression suites. Things that would ordinarily take multiple people multiple months could now be instantaneous. Even just writing code for you when you're stuck on something. Potentially like having the ultimate tutor/guide in your back pocket. Could completely accelerate how quick we build anything. It'd be fun to experiment with stuff like feeding it the rules of a game and saying "build me an iPhone app version of [Monopoly/Clue/Chess/etc.]" and see how far it can get.

- PowerPoint/Excel/other "functional" apps: I am excited to never do things in either of these apps again, instead to tell a GPT to do it and immediately have things populated. "Give me a 10 slide powerpoint, these 10 slides, pre-populate with the info from this file, that file, apply the following template/fancy formatting." Then I do the last 10% of the work to clean it up. That rocks to me.

- Cybersecurity: "here's my codebase, tell me how you'd best attack this if you're a hacker."

- Meeting management/notes/etc.: the ability for a GPT to take a set of meeting notes (or even a transcript of a video call) and get a head start on the tasks (setting up meetings, populating/uploading documents/notes, etc.)

There are many, many, many others, these are the just a few that are very cool to me. But I do buy this as a gamechanger in a way I don't really buy the Metaverse.

 

[Navin R Slavin]

Here's a few potential applications that I'm excited about:

- Coding: I think GPT will do more for "citizen development" than something like Power Apps or other low code/no code ever will. I am pretty sure just with the publicly available version of ChatGPT anyone here could have a fully developed basic website within a day, no expertise required. Even cooler, I think the accelerators it will give to more traditional development will be huge. Things like entering requirements and spitting out unit tests and regression suites. Things that would ordinarily take multiple people multiple months could now be instantaneous. Even just writing code for you when you're stuck on something. Potentially like having the ultimate tutor/guide in your back pocket. Could completely accelerate how quick we build anything. It'd be fun to experiment with stuff like feeding it the rules of a game and saying "build me an iPhone app version of [Monopoly/Clue/Chess/etc.]" and see how far it can get.

- PowerPoint/Excel/other "functional" apps: I am excited to never do things in either of these apps again, instead to tell a GPT to do it and immediately have things populated. "Give me a 10 slide powerpoint, these 10 slides, pre-populate with the info from this file, that file, apply the following template/fancy formatting." Then I do the last 10% of the work to clean it up. That rocks to me.

- Cybersecurity: "here's my codebase, tell me how you'd best attack this if you're a hacker."

- Meeting management/notes/etc.: the ability for a GPT to take a set of meeting notes (or even a transcript of a video call) and get a head start on the tasks (setting up meetings, populating/uploading documents/notes, etc.)

There are many, many, many others, these are the just a few that are very cool to me. But I do buy this as a gamechanger in a way I don't really buy the Metaverse.

The coding angle is extremely interesting, precisely because the "truth" of code can be easily ascertained in a way that the "truth" of words cannot be. If you can describe a spec in terms of a series of unit tests, you can theoretically hook up the ChatGPT API to continuous integration systems and guide ChatGPT towards the right solution simply by refining the unit tests until ChatGPT produces code that passes those tests. I agree: it will be a game changer, and maybe sooner than anyone expects.

 

[MrazeksVengeance]

ON BEHALF OF THE REST OF EUROPE.

We have indeed survived the winter.

 

[Lempo]

ON BEHALF OF THE REST OF EUROPE.

We have indeed survived the winter.

We wouldn't have been at risk in the first place if our leaders including certain former leaders hadn't insistently and intentionally been morons years in years out.

The government of Finland was paying actual money for the peat producers to wreck their peat production machinery, even after February 24, 2022.

Meanwhile: "The European Union Parliament has declared that nuclear power and natural gas can be labelled as green for investment purposes, alongside wind, solar and other renewable energy sources."

 

[Boom Boom Apathy]

It's going to be interesting when Quantum computing eventually takes off. Cyber security in particularly, but also as it applies to LLMs

 

[Sens1Canes2]

We wouldn't have been at risk in the first place if our leaders including certain former leaders hadn't insistently and intentionally been morons years in years out.

The government of Finland was paying actual money for the peat producers to wreck their peat production machinery, even after February 24, 2022.

Meanwhile: "The European Union Parliament has declared that nuclear power and natural gas can be labelled as green for investment purposes, alongside wind, solar and other renewable energy sources."

Hahahah. It’s such a scam.

 

[Negan4Coach]

Here's a few potential applications that I'm excited about:

- Coding: I think GPT will do more for "citizen development" than something like Power Apps or other low code/no code ever will. I am pretty sure just with the publicly available version of ChatGPT anyone here could have a fully developed basic website within a day, no expertise required. Even cooler, I think the accelerators it will give to more traditional development will be huge. Things like entering requirements and spitting out unit tests and regression suites. Things that would ordinarily take multiple people multiple months could now be instantaneous. Even just writing code for you when you're stuck on something. Potentially like having the ultimate tutor/guide in your back pocket. Could completely accelerate how quick we build anything. It'd be fun to experiment with stuff like feeding it the rules of a game and saying "build me an iPhone app version of [Monopoly/Clue/Chess/etc.]" and see how far it can get.

- PowerPoint/Excel/other "functional" apps: I am excited to never do things in either of these apps again, instead to tell a GPT to do it and immediately have things populated. "Give me a 10 slide powerpoint, these 10 slides, pre-populate with the info from this file, that file, apply the following template/fancy formatting." Then I do the last 10% of the work to clean it up. That rocks to me.

- Cybersecurity: "here's my codebase, tell me how you'd best attack this if you're a hacker."

- Meeting management/notes/etc.: the ability for a GPT to take a set of meeting notes (or even a transcript of a video call) and get a head start on the tasks (setting up meetings, populating/uploading documents/notes, etc.)

There are many, many, many others, these are the just a few that are very cool to me. But I do buy this as a gamechanger in a way I don't really buy the Metaverse.

My idea is turn it loose on all military manuals, and have it be able to answer stuff like "Here is a picture of a bridge we'd like to blow up- how much explosives do I need and where to place them?"


It already knows US military doctrine. I asked it to pretend it was Robert E Lee before the Battle of Gettysburg and identify his Priority Intelligence Requirements, which it did flawlessly (I forget if I posted those here already) and then to switch roles to his intelligence officer and answer those requirements, which it did a fairly good job. It could straight up eliminate the need for an S-2 section eventually.

I could totally sell this stuff to DoD. I didn't see that they have any sales positions there. Probably the too cool for school silicon valley thing like Palantir where they try to rely on engineers to do sales.

 

[Navin R Slavin]

I could totally sell this stuff to DoD. I didn't see that they have any sales positions there. Probably the too cool for school silicon valley thing like Palantir where they try to rely on engineers to do sales.

Possible, but more likely they're still in growth mode and haven't transitioned their business model yet. I don't think they were expect ChatGPT to blow up like it did; they only managed to get an hourly pricing model together last month. And besides, they probably don't need sales right now because they're adding tons of paid daily users without a sales force.

In the case of DoD, they may never sell directly to them. There are probably multiple FedRAMP compliant resellers banging down OpenAI's door for partner work. I might expect OpenAI to build a partner enablement org before they build a direct sales org anyway.

In fact, in looking a bit more closely, Bain announced a services partnership with them last month:

Bain & Company announces services alliance with OpenAI to help enterprise clients identify and realize the full potential and maximum value of AI

Global partnership combines the power of OpenAI’s industry-leading AI technology advancements with Bain’s world-renowned capabilities in strategy and digital application delivery; The Coca-Cola Company announced as the first company to engage with the alliance

www.bain.com

 

[MinJaBen]

I hate to sound like a luddite, but these AI systems are going to put so many out of a job in such a short amount of time that the biggest effect they may have is societal upheaval. It could get very ugly.