No matter how many times I get ChatGPT to write my rules to long-term memory (I checked, and multiple rules exist in LTM multiple times), it inevitably forgets some or all of the rules because after a while, it can only see what's right in front of it, and not (what should be) the defining schema that you might provide.
I haven't used ChatGPT in a while. I used to run into a problem that sounds similar. If you're talking about:
1. Rules that get prefixed in front of your prompt as part of the real prompt ChatGPT gets. Like what they do with the system prompt.
And
2. Some content makes your prompt too big for the context windows where the rules get cut off.
Then, it might help to measure the tokens in the overall prompt, have a max number, and warn if it goes over it. I had a custom, chat app that used their API's with this feature built in.
Another possibility is, when this is detected, it asks you if you want to use one with a larger, context window. Those cost more. So, it would be presented as an option. My app let me select any of their models to do that manually.
I use Sunshine and Moonlight when I'm out away from my Windows laptop in my home office and back near the router in the front room over my Macbook air, it's great when you can get directly on the LAN with it. But then, RDP is also better over the LAN, and casual disconnection is easier.
But if I want to do graphics-intensive work on the headless tower, I'll use Sunshine and Moonlight.
I used this all the time twenty years ago. Tried it out again for some reason recently, I think at the suggestion of ChatGPT (!), for some archiving, and it actually did some damage.
I do wish there was a modern version of this that could embed the videos in some of my old blog posts so I could save them entire locally as something other than an HTML mystery blob. None of the archive sites preserve video, and neither do extensions like SingleFile. If you're lucky, they'll embed a link to the original file, but that won't help later when the original posts go offline.
ArchiveBox is your friend. (When it doesn't hate you :)
It's pretty good at archiving most web pages - it relies on SingleFile and other tools to get the job done. Depends on how you saved the video, but in general it works decently well.
> Working one-on-one with an expert personal tutor is generally regarded as the most efficient form of education [14]
> [14] B. S. Bloom, "The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring," Educational researcher 13, no. 6, 4-16 (1984).
> You know what was published 40 years ago (okay in 1986, not 1985)? Benjamin Bloom's "2-Sigma Problem," a paper that many ed-tech and education reform folks still love to cite, even though its findings on the incredible effect of one-on-one tutoring have not been replicated.
> An experimental intervention in the 1980s raised certain test scores by two standard deviations. It wasn't just tutoring, and it's never been replicated, but it continues to inspire. ...
> As the computing and telecommunication revolutions advanced, visionaries repeatedly highlighted the potential of technology to answer Bloom's challenge. Starting in the 1980s, researchers and technologists developed and eventually brought to market "cognitive computer tutors," ... Sal Khan, founder of Khan Academy, highlighted this promise in a May 2023 TedX talk, "The Two Sigma Solution," which promoted the launch of his AI-driven Khanmigo tutoring software. ...
> It all sounds great, but if it also sounds a little farfetched to you, you're not alone. In 2020, Matthew Kraft at Brown University suggested that Bloom's claim "helped to anchor education researchers' expectations for unrealistically large effect sizes." Kraft's review found that most educational interventions produce effects of 0.1 standard deviations or less.
Huh, and 'Matt Kraft reported that effects of educational interventions generally-not just tutoring-are about twice as large when they are evaluated based on narrow as opposed to broad tests.' as well as 'Burke's and Anania's two-sigma intervention did involve tutoring, but it also had other features. Perhaps the most important was that tutored students received extra testing and feedback'.
That's a pretty interesting read about the limitations to tutoring! Now back to the Harvard paper.
My quick read of this paper says the researchers, who are from physics and engineering, studied one class over two weeks, computed a bunch of numbers, and decided it was significant.
The intro points out "Despite this recent excitement, previous studies show mixed results on the effectiveness of learning, even with the most advanced AI models2,3." so they know the possible signal is not that strong.
Give the history, I think it's unwise to make such a strong claim without a longer baseline. I don't see how the test can eliminate possible other factors.
I would also have liked to see the test pre-registered.
And there's selection bias as there are likely a lot of experiments like this being run where the lack of success is deemed not interesting enough to publish.
If you hard-code effective learned distributions from a trained model, I suppose that could be described as an 'AI algorithm', even though the final output is a flat algorithm.
I moved to FastMail three years ago, and, for a contrasting experience, found that spam filtering was almost on a par with Gmail. I had feared it would be otherwise.
