Halls of Shrewd'm / US Policy
No. of Recommendations: 0
Gifted:
A Bursting Bubble Would Be Great for A.I.
https://www.nytimes.com/2025/12/05/opinion/ai-bubb...Large language models, for all their wonders, can only predict the next thing a human would say. Train one on texts from the late 1800s and it won’t invent airplanes or rockets. It will channel ideas from that period, when leading scientists thought human flight was impossible. If we only scale up our current approach, wasting money on fast-obsolete chips and energy-guzzling data centers, we won’t progress beyond our current technology, which still yields limited, mediocre results.
Better A.I. would remember what it learns, just as humans do, and squeeze more work from each watt. Tech companies spend billions of dollars running large language models that don’t learn while they run. A tool that does both simultaneously would come closer to approximating the human brain, allowing it to innovate more readily.Is that really how they work? If they can hold a conversation, they must remember what was said previously.
I tried using Perplexity for a programming problem I had. It faithfully reproduced what the equipment manufacturer's manuals said, and collected it in a nice, easy to read format. Yeah, not useful. I'd already read the manual before I consulted Perplexity. I guess I could have saved time by finding out a bit quicker that the manual was wrong.
No. of Recommendations: 5
Is that really how they work? If they can hold a conversation, they must remember what was said previously.
Slight overgeneralization:
They do remember, but only within their "context windows"...back (say) a few hundred words or a few thousand, depending on the product. And at the end of the session it evaporates. They all have anterograde amnesia.
One can add some "icing" on the training after a foundational model is built, as another one-time step, but they don't generally learn or remember as they run.
Jim
No. of Recommendations: 1
I spent an entire day being "assisted" by ChatGPT and Perplexity to program a new remote control (my Harmony died) finally ending up with bizarre and mediocre results (a used Harmony is on its way from EBay). It then took me about ten minutes to configure a new network router.
I used to half-joke that five years of engineering school taught me where to smack my TV to get it to work as well as how to twist a piece of aluminum foil on its rabbit ears to improve reception. I should have hired a ten-year-old to program my remote control.
Jeff
No. of Recommendations: 0
back (say) a few hundred words or a few thousand, depending on the product.
__________________________________
ChatGPT has regurgitate advice based on conversations of a few months ago that it apparently associates with me.
Jeff
No. of Recommendations: 0
ChatGPT has regurgitate advice based on conversations of a few months ago that it apparently associates with me.
Maybe something in your cache/cookies?
No. of Recommendations: 0
I hear that people are having relationships with these things, so presumably it remembers what you whispered in its ear last night, yelled at it last Thursday, and so on?
Or is that a different technology?
No. of Recommendations: 3
I had a long, complicated conversation about taxes with Google Gemini. When I referred to something Gemini had said earlier in the conversation it told me it couldn't remember what it had said earlier. I had to repeat the exact question to bring Gemini back to the subject.
Gemini does save conversations which are listed on the left-hand side of the screen. This is handy for re-thinking what Gemini said and possibly reopening the conversation.
Wendy
No. of Recommendations: 8
I had a long, complicated conversation about taxes with Google Gemini.
Are you really sure you want to do that? There are repercussions, if you get it wrong.
you win a story:
Back when the earth was young, I received my corrected homework back, in one of my calculus classes. One problem was bathed in red. I asked the teacher if it was right or wrong, because I could not tell from all the marks on it. She said she initially marked it wrong, because it didn't agree with what the answer key said. But, she continued, I showed all my work, and she couldn't find what I did wrong. So she worked it out herself. She got the same answer I had. The answer key was wrong.
There is a lesson in there, about believing what the "authority" says, without challenging it.
Steve
No. of Recommendations: 1
Maybe try different models for different things.
I have full Copilot integration at work because my employer is investing ridiculous sums of money to push AI into the workplace through licenses and training. Employees are "encouraged" to utilize AI at least 80% of workdays, I'm sure you can imagine all the creative behavior that is driving. I find Copilot incredibly good at summarizing historical email and internal data on topics to get up to speed on projects/products I haven't engaged with recently or answer questions based on previous inquiries. I was a borderline obsessive for decades about filing email in .pst/subfolders to simplify future retrieval and I now find myself drifting far beyond inbox zero with Copilot's quick retrieval and summarization. Copilot's ability to draft something in my own style using the tone and voice of my email history is borderline scary. But as many of you mention, it suffers from the same short term memory problem from recent interactions sometimes as bad as forgetting between one prompt to the next in the same conversation. I also find it's computational ability using simple prompts to be quite poor. The integration with Teams for note taking and meeting summarization is fantastic especially when you consider it will use the meeting transcriptions in any future prompts for information (Oh yeah Bill said this about risks and Sally was going to look into it).
Grok has been my primary model used for my non-work related stuff. I find its computational and analytical ability to be extremely good, far better than even 6 months ago. I currently don't want any models tied into my personal email so I can't compare it fairly against Copilot but it is much more conversational in its interactions. I find its ability to "remember" is much better than the other models even when returning to prior conversations listed under the left hand menu from weeks prior.
Gemini seems to have caught up to, if not surpassed, Grok with the recent updates. With it being in browser I find myself using it more and more just due to convenience. Maybe a positive for Google.
I like Perplexity quite a bit for research on complicated topics. I find the way it presents specific references to be convenient and useful.
Just recently Alexa+ entered my home in beta form on devices through out our home. You can't fathom the chaos in my household when Alexa began responding with a different voice one morning. Luckily it was easy enough to ask to return to the original voice. It is much more conversational and you no longer get "here's what I found on the web". It also seems to be doing more speech pattern recognition to determine who in the household it is interacting with and attempting to address that person by name.
And Apple, well at least they use google search in Safari and Gemini is easily accessible. Siri isn't even worth using. Yet my family of four still has 20+ apple devices.
The next 12-24mo should be very interesting as all of these continue to evolve.
Jeff
