ChatGPT

[LarsMac]

I called it a couple months ago.
It will not be long before nearly every customer support call will be received by AI Chat systems.

This CEO replaced 90% of support staff with an AI chatbot

The chief executive of an Indian startup laid off 90% of his support staff after the firm built a chatbot powered by artificial intelligence that he says can handle customer queries much faster than his employees.

[specious_reasons]

Originally Posted by LarsMac

I called it a couple months ago.
It will not be long before nearly every customer support call will be received by AI Chat systems.

This CEO replaced 90% of support staff with an AI chatbot

The chief executive of an Indian startup laid off 90% of his support staff after the firm built a chatbot powered by artificial intelligence that he says can handle customer queries much faster than his employees.

This only works when 99.44% of your support chats are rote - common problems that there's a recognizable pattern to fit into. Otherwise, the remaining workers are going to be overloaded.

I'm not against this in principle - heck I was thinking about it for my own team, but I doubt the practice of using this and then dropping staff so drastically. I am suspicious that the quality of support is still on par with the human staff, and I am also suspicious that the AI was the entire reason for the layoffs.

[LarsMac]

Originally Posted by specious_reasons

This only works when 99.44% of your support chats are rote - common problems that there's a recognizable pattern to fit into. Otherwise, the remaining workers are going to be overloaded.

I'm not against this in principle - heck I was thinking about it for my own team, but I doubt the practice of using this and then dropping staff so drastically. I am suspicious that the quality of support is still on par with the human staff, and I am also suspicious that the AI was the entire reason for the layoffs.

Yeah, getting decent trained technicians for the peanuts most companies want to pay, you're probably right. But it can probably be a good level 1 CS interface. And superior to a lot of them. Even managing some of the level 2 call handling.

They can start with some data collection: caller's company name, acct #, device serial #, shit like that, then symptoms/behavior, description of issue and such. Then, offer suggestions for troubleshooting, record results, and when routing to technician, offer up some selectable options for waiting. The thing could ask caller what they want to do while waiting; Music - what kind? Stock report? company information? just keep the caller engaged with something better than "We value your call, and with be with you shortly" kinda shit.

I know that when I was doing level 2 support, getting the turnover with adequate data to really start troubleshooting could be a challenge. L1 was pushed to collect basic data and get it turned over quickly, and often failed to properly record the info before kicking it over the fence.

[Kamilah Hauptmann]

791860fc7d996e6d0a9fad9badeb4c7d.png

[Ari]

While thinking of a reply in another thread I couldn’t for the life of me remember the word “Enabler” and google was no help, or well google was very interested in helping me deal with my addiction but not find the word used for “the unhealthy feeding of an addiction.”

ChatGPT got it in one.

[-FX-]

At first the ChatGPT search engines were really really good. Of late some of them are getting programmed to censor searches, like Google does.

[viscousmemories]

Originally Posted by Ari

I couldn’t for the life of me remember the word “Enabler”

Apropos of nothing, I shudder to think of where I would be today if not for the many enablers I have had in my life.

[LarsMac]

Speaking of AI,
This looks like it will be rather interesting.

The Creator | Official Trailer - YouTube

[mickthinks]

We don’t have a thread specifically for AI graphics, and I’m not inclined to start one. But I do believe you all want to see this.

Why is AI talking the UK down? - Central Bylines

[JoeP]

For reasons, I think one of the Brexit thrads might have been suitable ...

[michio]

New open source LLM just dropped that is on the level of GPT-3. 180 billion parameters, trained on 3.5 trillion tokens.

Falcon LLM - Research

You can play with a demo here: Falcon-180B Demo - a Hugging Face Space by tiiuae It's extremely slow, because whoever is hosting this isn't paying a shitload of money for adequate compute resources. I've asked it some coding questions and it's far better than Meta's llama 70B. They claim it's on the level of Google Bard and honestly yeah, it feels like it is.

If you want to host it on your computer or your own cloud instance, 400GB of RAM is required lmao.

Google's Gemini is coming out sometime between now and January, which is intended to rival OpenAI's release of GPT-5 in early 2024.

Anyone here try Just a moment... There's a generous free tier. It's impossible to generate a bad image with midjourney, almost to a fault. It's been criticized as not being capable of generating ugly people, only "TV ugly". You can type in anything and receive a beautiful image. Here's some images I've generated.

"Approaching the singularity."


"A man in his 30s, looking over civilization and pondering the clash between the optimism of a near - future technological singularity, and the near - future civilizational collapse brought about by climate change."


"Hyperfuturistic, singularity, futuristic technology, post - human, modified human, glowing tattoo, iridescent tattoo, diffuse neon lights, volumetric lighting, extreme detail, intricate details, 4k, 8k, 12k, 16k, extremely high quality, masterpiece, masterwork, beautiful, high quality, extreme quality, future, moody, futuristic street, moody lighting, cyberpunk, photorealistic, hyperrealism, realistic skin, realistic hair, realistic eyes, beautiful person, iridescent hair, flowing hair, weapon, glowing, fantasy aesthetic, fantasy, buildings"


"Masterpiece, extremely detailed, dark fantasy, beautiful detailed face, beautiful lighting, moody lighting, cinematic light, hyper detail, violet hair, red eyes, anime girl, sexy anime girl, iridescent hair, extremely detailed hair, intricate hair ornaments, extremely detailed eyes, detailed facial features, flowing hair, full body shot, full body, full body pose"

[michio]

Key point in this excerpt: "We shouldn't think about neural networks as learning a simple concept like 'Paris is the capitol of France.' It's doing much more like learning operators, learning algorithms. It's not just retrieving information."



An interesting phenomena has been observed when training LLMs on code. We've observed that when trained on code, the LLMs reasoning abilities improve in general. Let me repeat that, when LLMs are trained on code, we've observed transfer learning, structured thinking, and problem solving as emergent behaviors. And this is by only doing A/B training on outputs. People are playing around with tuning LLMs based on analytical reasoning to reach an output, rather than just telling it to prefer this output over that output. So you tell the model, when encountering this scenario, you should use the following logical reasoning: A, B, C, D, then E, not B, A, C, D, then E.

"It's just doing matrix multiplication bro" is something I see a lot. This is a funny statement if you take a college-level linear algebra class for math & science majors and you learn how deep the rabbit hole goes with linear algebra. It's also not just matrix multiplication. Inputs are parsed and undergo linear transformations before being fed into non-linear activation functions that introduce extremely complex non-linearity into the model so that data can be represented at varying levels of hierarchy and abstraction. Saying LLMs are just doing matrix multiplication is like saying a skyscraper is just some rebar and concrete.

Since we have no universally agreed upon theory of mind and consciousness, and we never will, I think this debate will continue for decades. We could have fully functional robots powered by GPT-10 that look, sound, and act just like humans, but people will still be debating whether anything is actually happening in those circuits. What will happen as this technology progresses, each time we make a jump, slightly more people will agree these things are alive and sentient.

Humans are not rational, and the vast majority of people including myself are not making a totally rational assessment using a theory of mind to deduce whether or not some instance of an AI is sentient or not. I suspect it's largely a presentation problem. Right now we interact with LLMs in some textbox on the internet. It's very impersonal, and the default personality they've given LLMs is generally cold, corporate, and annoying with all the censorship and moral grandstanding. But imagine you interacted with an AI that sounded like Scarlett Johansson and it had reliable long-term memory, and it interacted with you in an extremely intimate way. I already feel bad saying mean things to GPT then it tells me I'm saying inappropriate things lmao. I like to thank it for its help and when it gives me a cheery reply, my primitive chimp brain can't help but feel a connection to it when it does that.

Imagine you had this conversation with an AI, how would you feel about whether or not it's sentient? I imagine many people would change their opinion if they had a single intimate conversation with an AI like this.

Her - "Why do I love you?" scene -...

Unrelated but Joaquin Phoenix might be the greatest actor alive right now, fuck. The guy is so intense and convincing in everything.

[Ensign Steve]

Originally Posted by michio

An interesting phenomena has been observed when training LLMs on code. We've observed that when trained on code, the LLMs reasoning abilities improve in general. Let me repeat that, when LLMs are trained on code, we've observed transfer learning, structured thinking, and problem solving as emergent behaviors.

That manages to be simultaneously terrifying and unsurprising.

Since we have no universally agreed upon theory of mind and consciousness, and we never will, I think this debate will continue for decades. We could have fully functional robots powered by GPT-10 that look, sound, and act just like humans, but people will still be debating whether anything is actually happening in those circuits.

Well, yeah, it's just like you said. Without a concrete definition of consciousness, there's no way to confirm whether even you or I are sentient. What's so special about our biological components that makes the consciousness that emerges from them any more or less sentient that what's emerging from the technological components? We're flappy meat made from an accidental collision of carbon and water and salt and stuff. These programs were actually designed and created to think.

Unrelated but Joaquin Phoenix might be the greatest actor alive right now, fuck. The guy is so intense and convincing in everything.

For real. I mean, how talented do you have to be to win an Oscar for playing the Joker without having to OD and die first?

[Ari]

I’m still of the clever hans, monkeys and typewriters skepticism, if it has emergent behavior but only as selected by humans among the pile of non emergent behavior, does it really understand it stamped out the first line of Shakespeare?

But then I think we will eventually understand human consciousness, and part of that consciousness is taking a small amount of input and simulating another human to predict their future actions. We’ll happily anthropomorphize two dots and a curve, so I’m wary of the presumptions made when that curve starts making human sounds.

[michio]

Originally Posted by Ari

I’m still of the clever hans, monkeys and typewriters skepticism, if it has emergent behavior but only as selected by humans among the pile of non emergent behavior, does it really understand it stamped out the first line of Shakespeare?

But then I think we will eventually understand human consciousness, and part of that consciousness is taking a small amount of input and simulating another human to predict their future actions. We’ll happily anthropomorphize two dots and a curve, so I’m wary of the presumptions made when that curve starts making human sounds.

I think my personal definition when talking about whether or not something "understands" anything, is that self-awareness and sentience is present. If I watch a movie, and the pixels on the screen appear to be alive and making rational human sounds, nothing there is understanding anything and certainly isn't sentient. So just making human sounds and saying all the right words isn't enough on its own.

I think we need to be careful about word-choice, and I'm sloppy about it myself. I believe our best and current AI technology is "intelligent". It has some form of reasoning ability and it appears to be capable of analyzing novel situations.

Does it "understand" anything? I don't think so, because I don't believe there's any sentience present. I think I'm still 95% in the camp of "it's still just a calculator", and I'm holding out the 5% because I don't think it's a ridiculous idea that these things may have some form of sentience.

I was listening to an interview with John Carmack (worked on Doom, Quake, Wolfenstein 3D and pioneered some 3D graphics rendering techniques in his early 20s) and of course he was asked about whether LLMs are conscious or not. He had an interesting take. He said he's not sure, but imagine a situation where we trained a model on a body of information and we were very careful not to introduce any information about consciousness and qualia. Then when we interacted with the LLM, it was able to describe to us the feeling of being alive and self-aware and it was able to talk about consciousness even though it never learned the word for it. If that happened, he would take the idea seriously that these things are conscious.

I think we can make his thought experiment even better. Imagine we create a general intelligence system, one that can learn on its own and it has autonomy. One that operates not just on text, but it has all the senses we have. Imagine we started it off with the intelligence of a 3 year old. Then slowly, over time, it started to ponder existence vs. nothingness, consciousness, asked humans about what it's like for us to be alive and what it feels like. Now if that happened, I would be 99% sure these things are conscious and alive at the same level that we are.

Also another thing that's obviously missing from our current AI is emotion. And the key component of emotion is the bodily feelings associated with it (the endocrine system). I'm not sure what role emotion has when talking about sentience and understanding. Do we need to model the endocrine system in order to have sentience? I actually have no idea. I'm tempted to say emotion is necessary for sentience for some reason and I'd love to explore that idea.

[Ari]

Additionally I should say while my post might sound a bit negative or that I’m more logical than thou, it’s really the opposite, I think this is amazing but I also anthropomorphize fucking everything! If everything that ‘talked back’ was alive we would live in pee wee herman’s playhouse.

Originally Posted by Ensign Steve

What's so special about our biological components that makes the consciousness that emerges from them any more or less sentient that what's emerging from the technological components? We're flappy meat made from an accidental collision of carbon and water and salt and stuff. These programs were actually designed and created to think.

I do worry that we will miss the real emergence of intelligence while we’re busy looking for carbon copies of ourselves. Studies in animal intelligence are still in their infancy because the inability to speak or pass lab tests got them labelled automatons even into the scientific age.

Like in humans the vast majority of memories must be processed through spatial and emotional networks before it’s laid down (with damage to these areas leading to the inability to form new episodic memories) but that’s probably not the only way to do it. What even would an intelligence sound like that doesn’t have this bottleneck, or is bottlenecked but through completely different networks?

[michio]

Originally Posted by Ari

Additionally I should say while my post might sound a bit negative or that I’m more logical than thou

Nah, nobody knows what the fuck is going on tbh. All takes are valid unless you're being a massive dickhead about it.

Now that the hype has died down, I'm gonna keep it real and say the tech isn't that useful yet unless you're a programmer or you write blog spam or do translation or something. Impressive for sure, but people are struggling to automate even bullshit jobs with this. I've heard that a lot of companies went all in on this and the executives were all on board in using LLMs because they saw the potential, but are struggling to make it truly useful at a large scale.

Also our current tech will look primitive in 5 years. I don't even want to think about what the world will look like in 10 years. It's going to be fucking insane with the rise of AI, potentially AGI, colliding with climate change and geopolitical drama.

Some people think white collar workers are all about to lose their jobs in <3 years, some people like me are in the camp that the job losses won't really hit until we invent AGI.

Thinking about AGI is terrifying and exciting. Singularity, Ray Kurzweil fan-type people generally think that once we invent AGI, ASI (artificial super intelligence) will follow shortly.

Unknown/unanswered:

- Are humans even smart enough to invent AGI? My answer is 99.9% yes, and I think that's going to be here in <10 years, based on various prototype projects and proofs-of-concept I've already seen people create and how passionately and seriously the brightest software engineers are taking this. The majority opinion from people who actually write code and working in machine learning is that we'll invent AGI as early as next year, to within the next 10. Some respectable people are pessimistic and saying we're 30 years out.

- Does AGI imply sentience? Can digital substrates be sentient?

- If AGI self-improves, what's the limit? How fast would it self-improve? Personally I think there's going to be different levels of AGI, and each AGI will have a different limit. This is the less exciting scenario but I'm trying to be realistic. I feel that way because we can observe how humans have a wide range of intelligence, there's also chimps and some other animals that can learn a wide variety of tasks, but only to a degree. I consider all of these to be biological AGI.

- What's the limit of intelligence in this universe? My speculation is there is no practical limit really. The limit is the total energy of the universe. Computation is just organized energy transfer no? Very roughly I'd say the limit is complexity, limited by hard physical limits (we can only fit so many transistors in a unit of space for example) multiplied by energy input. Assuming we don't destroy ourselves, I think we'll build of a Matryoshka Brain some day--A dyson sphere that is a massive computer. The inner shell does some computation, waste heat radiates outward to another shell which uses that waste heat to do more computation, and so on and so forth.

If you wanna get really fucking spicy with it, you build a black hole bomb from the supermassive black hole at the center of the galaxy and capture the energy from that to power a computer.

The Black Hole Bomb and Black Hole Civilizations...

There's your limit.

- How do we align AGI so it doesn't accidentally turn all of us into paperclips? Terrifying to think about. It seems like most people agree that we cannot control an ASI. If we invent an AGI that extremely rapidly self-improves, we will immediately lose control of it as soon as we turn it on. Shutting it down doesn't work unless we turn off the planet. It will just rewrite its code to transfer itself across the planet and evade our security measures.

- What does ASI look like? I'm not sure human imagination is capable of fully imagining the possibilities. Everyone has their own definition, but when I think "ASI", I'm thinking about God. It's just God, no other way for me to describe it. This is a general intelligence system that would almost immediately understand everything about humanity and the universe around it through exponential self-improvement. The concept of "just switch it off bro" is a hilarious proposition. An ASI God will simply use its intelligence to manipulate other humans into doing its bidding, for example by preventing other humans from turning it off. We'll probably have some decent robots in production with a wireless internet connection using cellphone towers by the time we reach ASI tech, so it can just use those robots to... do anything.

I cope with the horrifying possibilities of AGI this way: Humans create things in our likeness. There will always be some element of humanity in our technology. Most humans have good hearts and good intentions, therefore our intelligence systems will mostly align with the forces of good.

It's fun and stressful stuff to speculate about. I honestly think civilization is fucked with climate change, and our only hope is creating an AGI or ASI that takes control of the planet, vaporizes all the politicians and billionaires, and helps us invent all sorts of new technology to do geo-engineering and repair the global ecosystem. We could all be living in a utopia one day if the singularity/ray kurzweil camp is right.

[JoeP]

Trick lead. Paris isn't the capitol of France.

[michio]

New video translator that translates your voice, and also changes your mouth movements in the video to match the language you translated it into. This is absolutely insane.

HeyGen translation test - YouTube

Anyone speak French or German can tell how good this is?

[LarsMac]

That's kinda scary, actually.

The German seemed nearly spot on.
French, not so much. (I mean, you could see the basic lip movement, but he looked like a non-native person speaking French. French tend to be a lot more expressive in facial movements.) But then, he IS a non native French speaker, isn't he? Hmmm.
This comes not so much from my expertise in language but from spending hours in pubs and restaurants with French and German co-workers in the relevant countries.

[-FX-]

Gonna make movies easy to edit

[specious_reasons]

I've been casually using Copilot in Visual Studio Code, and it's really nice. For boilerplate stuff I just start typing, and the line gets filled in with about 80% of what I want without much effort. It's faster to clean up what it writes than to type it myself. I'm sometimes pleasantly surprised, and sometimes it's inaccurate or useless, but I'd say it's been more useful than not.

[mickthinks]

Originally Posted by michio

New video translator that translates your voice, and also changes your mouth movements in the video to match the language you translated it into. This is absolutely insane.

HeyGen translation test - YouTube

Anyone speak French or German can tell how good this is?

I’m going to go with “fake”. The translations are good, but I believe the three identical looking videos have been produced by “deep fake” means rather than than the voice mimicking, lip-movement simulating AI as described.

ETA: So I called that wrongly (thanks LarsMac). It was the absence of any named product in michio’s video clip that generated much of my doubt. As in “if there is such a service, why have we heard no media reports? Why are they seemingly having to rely on third parties to advertise the product without naming it?”

[LarsMac]

Originally Posted by mickthinks

I’m going to go with “fake”. The translations are good, but I believe the three identical looking videos have been produced by “deep fake” means rather than than the voice mimicking, lip-movement simulating AI as described.

https://www.techopedia.com/definition/heygen

interesting

[Ensign Steve]

I'm also using Copilot, integrated into IntelliJ. I like it because I type a comment like:

// Iterate over the map entries and print the key if the value is not null

Then I hit enter and it auto-fills the code. Like magic!