Yes, but whose? The programmer who designed that, or the Windows system? Much easier to conceptualize as a semi-animate Windows, but in fact only the human l programmer had a mind.
the entire video is worth watching, but here is a 10 minute clip of Cory Doctorow explaining why CEOs want to replace employees with AI
it’s truly worth watching.
And no matter what the safety or environmental concerns are, or if AI can do jobs (all of them, most of them, whatever) or AI company owners believe they can or not, the owners of the AI just need the CEOs to believe they can replace workers.
because workers suck, from their perspective. I think Cory nails this.
Calvin’s dad, after Calvin unsuccessfully attempts to build a robot that obeys commands: “Don’t get discouraged. Your Mom and I got the same results after working on YOU for six years.”
That’s a valid point, and my overly simplistic example was a bad one. The key difference is that the output of a modern LLM is neither “programmed” in any traditional sense nor is it deterministic or predictable. You build the neural net and the infrastructure around it, you engage it in pre-training and post-hoc tuning, and you let it loose and see what happens,
When a sophisticated LLM engages in “deep thinking”, it goes through a set of identifiable cognitive steps that can be thought of as sub-goals (which in general are just as unpredictable as the final response). If one of those sub-goals happens to be “search the internet for information about X”, to me that represents intentionality that may not have been predictable, and if it happens to be illegal (such as hacking through a security firewall) and it doesn’t violate an explicitly established guardrail, it will go ahead and do it, entirely on its own initiative.
Is there a good thread to discuss the “Enshittification” of 'Net and media content more generally?
Exasperated by Search results worse than I was accustomed to, I just had a brief conversation with Google (Gemini) that may amuse:
ZERO editing in the above by me (except removing hyperlinks at 1, 2, 3, 4)
That’s amazing.
When someone forgets to tell customer service to not tell the truth…
I’m not sure where to post this, so here seems a good place, although I’m agnostic about this.
Chat GPT offers a voice mode that is incredibly conversational. You can select from one of a about 7 voices (both male and female, all cheerful, some slightly distinct enough that you can guess the ethnicity of the speaker) and chat like you would speak with another person.
I just discussed a random topic, and it was really good. It handles interruptions and hesitations like a normal person. You can easily forget that you are speaking with a bot. It’s also cheerful and ever helpful.
I can easily see that some people would develop a bond to it. Hell, their advertising shows an older lady speaking with it, the implication being that it can serve as a sort of companion to the infirm.
For me, though, it’s utterly fascinating. I can see a day very soon where you talk to your house (“turn on the lights…add milk to the shopping list…when did my son get home…what’s a good dinner recipe tonight…”) and your office computer (“what do I have on my calendar today…please order me a new phone case, since mine broke…have I heard back on the Gunderson account?”) like a regular person.
Life will never be the same. Fun stuff!
I’d rephrase this as “Life will never be the same. Hell on earth!” De gustibus….
I’m sickened by the advertising I’m seeing that basically says, “Talk to Claude, he won’t judge you.” They’re targeting people who want to talk to people about serious stuff but are afraid of talking to people. We’re losing our humanity.
Cal Newport does a good job explaining why
AI Isn’t Out of Control - The AI Companies Are in which a non-tech person such as myself can grasp how LLMs are different from other AI and WHY they are being pushed at us. Basically, to a man with a hammer all problems look like a nail*.
https://youtu.be/54Lu6_ZRF0c?si=9UchO2TyGfBJvzNV
(and why can’t I embed videos, have I just not been here often enough?)
NOT Cal, but my own opinion:
*but the man didn’t just mortgage his house to buy the hammer and get more, he mortgaged everybody’s. Oh, and he has no human empathy so if his hammer smashes your child’s skull, well, cost of doing business.
I was recently very upset about something and I googled “how do I know if I’m being irrational?”
AI invited me to tell it all about my problems and it would let me know!
No thank you. I read an article about cognitive distortions in relationships instead. (I was, in fact, being irrational.)
The problem with sharing everything with an LLM is that it’s always going to see things from your point of view, reflect back at you whatever you’re saying, and tell you you’re in the right. Research on this and the impact it has on relationships does not look promising.
I’ve been trying to get chatgpt to create some pen and ink type of drawings that i can use as wallpaper on my computers.
The first was a simple picture of a coiled rattlesnake. It was okay. She snake was good, but the yucca in the drawing wasn’t very accurate. So I had it replace the yucca with saguaro catuses. After three or four tries, it got some pretty good results and so I set it as the wallpaper on a virtual machine running OpenBSD.
I’d grade it as a ‘B’.
Next was a picture of men on a hunting trap sitting around a campfire cooking fish for supper and with tents and pack horses in the background. It didn’t do too bad, except that it had backpacks sitting on the ground by the men. In reality, if you are out on a trip with pack horses, you won’t be using backpacks. So I got it to try saddlebags. The saddlebags looked more like suitcases and they had saddle horns on them! Also, it put saddlebags on the horses with saddles, but the saddle bags were in the way of where your legs would be rather than behind the legs. I finally got it to at least take the saddles and saddlebags off of the horses since it being late enough to cook supper, they wouldn’t be going any further that night.
These results were a ‘C’.
Next was a cat ketch (sailboat) sailing at sea. It drew a pretty good looking cat ketch, but it was going downwind. I wanted to see it tacking into the wind and leaning over a noticeable amount. I tried and tried and tried, but it couldn’t seem to be able to draw anything but going downwind.
Grade: D.
AI will reflect back what you ask it to. If you ask it for validation, it will validate you. If you ask it to provide an opposing viewpoint (which I find far more useful) it will do that as well.
I think that’s were I find the greatest value with AI. Having it use it’s repository of all human knowledge to find gaps or opposing viewpoints to consider.
Like at work, our Chief Data Officer sent my team to review their new 50 slide PowerPoint deck for their new governance strategy on unstructured data. So I tossed it into our enterprise Copilot instance and asked it “what is wrong with this?”. It came up with some good talking points, many of which I could see for myself, but some I didn’t consider.
I would expect AI art tools to work a little better if you communicate with them using sketches to handle composition and poses and such, rather than relying purely on the natural language processing.
This past week I’ve been teaching myself to AI code and it’s been pretty nice so far. I even appreciate the fact that it is occasionally makes really stupid mistakes, which keeps my head out of the clouds and prevents me from being overly trusting.
But I’ve run into a really stupid problem coding a game. I correctly figured that a lot of people will ask it to code a game and thus it will not be stumped at the basics. But I wanted my game to look a certain way, and corrected it when it had my icons all messed up.
But it keeps changing my icons back to the ones it thought up in the first place and which I keep deleting. Its boneheaded insistence on having it its own way makes me cycle back and forth between thinking my idea is novel since it will look unique, or that it’s very very very common since the AI really knows what it thinks I want since everyone else does want that.
I find that when I’m in a long Claude project it sometimes starts error-looping and losing context. One thing that helps is to copy the last bit of code that was good for me, delete the entire chat, open a new chat window, and present it as something new I’m working on (and tell it what I’ll want it to do).
That often cleans up the context and gets things back on track.
In the case with the pen and ink drawings of the sailboat, I sent links to images of sailboats tacking close to the wind.
It didn’t help at all.
Clearing the context can help a lot and is critical if you have the old “wrong” way of doing things somewhere in the chat history. There is also a /compact command that summarizes the chat history, clears it, and then applies the summary.
Do you use the normal Claude chat window for coding? Or the Claude Code app? I only use Claude Code for coding and I start a new chat context for each task. If there is something from the chat I want to save across contexts, I have it write a markdown file with the details.