lee/hpr-knowledge-base
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial commit: HPR Knowledge Base MCP Server

Browse Source 
- MCP server with stdio transport for local use
- Search episodes, transcripts, hosts, and series
- 4,511 episodes with metadata and transcripts
- Data loader with in-memory JSON storage

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
Lee Hanken
2025-10-26 10:54:13 +00:00
commit 7c8efd2228
4494 changed files with 1705541 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

120
hpr_transcripts/hpr4064.txt Normal file
View File

@@ -0,0 +1,120 @@
Episode: 4064
Title: HPR4064: Large Language Models
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr4064/hpr4064.mp3
Transcribed: 2025-10-25 19:09:48
---
This is Hacker Public Radio episode 4,064 for Thursday, the 29th of February, 2024.
Today's show is entitled, Large Language Models.
It is hosted by Daniel Persson, and is about 13 minutes long.
It carries a clean flag.
The summary is, what are they good for?
Hello Hacker's, and welcome to another episode.
Today I'm going to talk about machine learning and LLM, so large language models.
To start with, when it comes to application of machine learning,
I think that there is a bunch of really interesting and good things to do with machine learning.
We have machine learning that trying to find proteins and how to fold proteins,
and so on, that is using in the medical field, which is really great.
You have machine learning that is trying to look at images and find particular faults in people's
eyes, for instance, to find sickness, or you could have machine learning that finds different
objects in traffic in order to help cameras in life and so on, which could really benefit society.
There is a lot of these kind of creepy things that is going on there as well,
so everything can be used for good and bad, of course.
Then you have these kind of machine learning AIs that looks at images and
finds, for instance, characters, so you can actually read things, translate things,
which is also helping people to live their lives.
Really great applications.
When it comes to working and creating new material, there is AIs that could create sound,
like music, and could that create voices, these kind of TTS engines, that sound really good,
and for people that is either blind or have reading disabilities,
then that is a really good application to give them the possibility to actually listen to media
and that kind of thing. So there is a great application in that as well.
When it comes to creating images and videos, there is a bunch of creepy applications,
but for a professional, these kind of tools can be really helpful as well.
So if you take, for instance, an image and you need to remove something in the image,
we have already algorithms for that, but AI improves those algorithms.
If you have, for instance, square images, which we had at our company,
and we needed rectangle images as product images, that would have cost a lot of money
for the customer to hire its staff and reshoot all those product images,
just because we needed a different resolution or a different crop of the image.
So having the tools in Photoshop and other programs in order to create things that weren't there,
in order to create a different aspect ratio of an image,
or adding material that isn't really there, where the thing is not that important.
You are not trying to fool anyone. You want something to look natural and what is actually the
real thing is still there. So in our case, we had, I think it was somebody selling rings.
So it was a ring on a finger where you saw the product really well,
but we needed a little bit bigger image. So the AI generated an arm that stretched a little bit
further and a door frame that stretched a little bit further. So you could actually see
a little bit more of the image and have a little bit of a different aspect ratio.
So in those applications, I find very helpful and very interesting ways of using AI
or machine learning pretty much in order to do something better. I don't really like the term
AI, but everybody uses it. When it comes to large language models, I still have not found any
viable applications for those. So there is a bunch of applications that people use them for,
but I haven't really found the key thing that is helpful for everyone.
So when it comes to different applications, you could, for instance, create text.
And there is a bunch of different text things that you could create. You could, for instance,
create these kind of blurbs you see on web pages where people are talking about their company.
And those are usually these kinds of we are a consultancy that is trying to help our company to
reach newer heights and work with blah blah blah and so on, which doesn't really say anything
about that company. We also have these kind of extra information that you need to put on your
products in order to sell them, which if they don't contain any real facts like this product is
this large, this product is this fast, this product has this particular color or whatever it might
might be that is objectively correct information that you need to have in your description.
If those kind of things aren't really what you're looking for, you just want to create some text
that describes this product. You could do that with a large language model and get a reasonable
result. These kind of things people don't usually read anyway, so it doesn't really matter.
Another application which I think is more sinister is that you could create really good spam.
So I have a lot of people creating a bunch of very interesting spam. I've gotten a bunch of them.
People that are trying to sell me steel, trying to sell me all kinds of different
strange products that I would never be interesting in, but they are starting it with
creating an information that is seems reasonable as an email to me and then when you have read
two paragraphs or something, you realize that this is a product I'm not interested in. Before
you could see these emails because they were so badly produced that you realize directly that
this is not something I'm interested in, but now that is so cheap and so easy to create these
kind of spam emails, you could do that much better. So we will see a lot more spam.
Another thing that is really easy to create with machine large language models is of course
content. So now Google results are much worse than last year this time because there is so much
content pumped out that doesn't really say anything. So if you're searching for a solution to a problem,
the other day I was searching for how to train a large language model. So I could figure that
one thing out. I had to go to the fourth page of the result in Google in order to find somebody
that actually was talking about training a model. The other ones was generated by a model talking
about the different steps, but didn't have any extra knowledge or anything that actually gave me
any information that could help me in the process of training a model. It was talking about
how the process was looking. What kind of steps you could have like in order to train a model,
you need to find data, refine data, and then train with the data, verify the data, and then
use the data. Those five steps in 10,000 words doesn't help me at all. Yes, for somebody that
has never seen the topic before, that is very helpful, but for me that actually trying to figure
out how can I do this myself, it's totally useless. And I can see much more of this kind of
wasteful content being produced in order to get clicks, in order to get search, ranking, and in order
to just put content out there to earn money. But I haven't really found what kind of use case could
benefit society using an LLM. Of course, you could create your own emails and send to your colleagues
that is 2,000 words where you're trying to say one thing, or you could just write that thing that
you want to say in order to writing it in 2,000 words and really waste their time. So that is my point
of view of the LLM so far, or else we will might end up in a world where we try to write down five
facts, send them through an email, we have an LLM, then generates an email with 10,000 words,
send it over to someone, and they have an LLM in the other end, or an AI in their machine learning
model in the other end, which summarized that 10,000 words email into these five points again.
So we get a actual message across. That might be what we are doing in the future,
and all this translation back and forth will of course create hilarious side effects and
misunderstanding between people. So that might be the world that we are living in in two years,
but hopefully not. Hopefully we will have found what we should use large language model for by then.
So I am looking forward to see what we should use these things for, still today I have no clue
because asking an LLM for an answer, it will not give you the answer, it will
figure out what is the most commonly used opinion about this, but probably not the correct answer.
It's probably something like, yeah, I think it's this, and then you said, yeah, but that doesn't
make any sense. So I have actually talked to a chat GPT over and over again, where it has said,
okay, try this, and I said, but that is not correct. I know about this subject, I want to do know
how to do this thing, but that what you send me is not correct. And I have done that over and over
again, and when you have told somebody that they are wrong ten times, then you are wasting time
when you can just look it up instead. So yeah, hopefully we will find something that these things
are useful for, or they might be much better so they can actually generate truth in the future,
or actually generating value. So this is my way of thinking about LLMs, do you have a different
opinion? Have you found the use case for this yet? Please leave a comment about that, or record
your own episode so I can hear about what you have used LLMs for. I hope to see you in the next episode.
You have been listening to Hacker Public Radio, as Hacker Public Radio does work. Today's show was
contributed by a HBR listener like yourself. If you ever thought of recording podcasts,
you click on our contribute link to find out how easy it really is. Hosting for HBR has been
kindly provided by an honesthost.com, the internet archive, and our syncs.net. On the Sadois
stages, today's show is released on their creative commons, attribution, 4.0 International
License.