#485: Creating memories

00:00 Hello and welcome to Python Bytes, where we deliver Python news and headlines directly to

00:05 your earbuds. This is episode 485, recorded on Tuesday, June 23rd. I'm Calvin Hendryx-Parker.

00:12 And I'm Michael Kennedy.

00:13 Thanks, Michael, for being here with me today. I'm excited to be your co-host for Python Bytes.

00:19 We've got a lot going on. We're going to dive right in. But first, make sure you go check out

00:26 our sponsors. It is sponsored by us in our work.

00:29 Michael's great work over at the Talk Python courses.

00:32 So check out Talk Python's coursework over there.

00:35 And if you are in the Midwest tomorrow, since this does go live today,

00:41 AWS Community Day Midwest is tomorrow, June 24th in downtown Indianapolis.

00:46 Six Feet Up is sponsoring and there'll be two Sixies presenting.

00:49 So you've got, that's actually what Six Feet Up people refer to themselves as.

00:54 So if you want to come check out two of us presenting at the conference tomorrow, check that out.

00:58 If you want to connect with the hosts, you can find Michael and myself at all the various social spots on Mastodon, Bluesky.

01:06 Do you still use X?

01:08 I do still use X.

01:09 And I updated our show notes template so that we have our X links and LinkedIn.

01:14 Perfect.

01:15 Yeah, because I'm probably a little more active on LinkedIn.

01:17 But maybe this is a nudge or a motivation that I should go be active over on the other platforms too.

01:22 And you can check out the show on Mastodon and Bluesky at those various links.

01:26 And X there as well.

01:27 So join us on YouTube for the live version of the show at pythonbytes.fm/live to be part of the audience.

01:35 Usually we record Tuesday at 7 a.m. Pacific, 10 a.m. Eastern for those of us on the East Coast.

01:41 I guess we cover both the coasts, Michael.

01:44 And then older video versions are available there as well.

01:47 Finally, if you want a bonus digest of every week's show notes from an email form, just add your name

01:54 and email to our friends of the show list. And we'll never share that email address with anyone

01:59 else. You'll just get great sets of links and show notes from the team over here at Python Bytes.

02:04 All right, Michael, let's kick it off. You've got the first topic up here.

02:08 And let me tell you, I think this one is going to resonate with you.

02:12 Yeah, I already saw why I peeked at some of the headlines. I was like, this is interesting.

02:16 Backing up Docker volumes.

02:18 So maybe we should do a quick, what the heck?

02:21 What are Docker volumes?

02:23 So if you do Docker a lot, you know exactly what Docker volumes are.

02:26 But Docker's a little bit one of those things like, oh, it seems so complicated, so weird.

02:31 And it's just really different than what I'm used to doing.

02:33 And then you work on the Go.

02:34 That's kind of like Linux, but just in a file.

02:36 You know what I mean?

02:38 But one of the things that is different about Docker is it's very transient.

02:42 It's not quite Nix, that it can't be changed once it's fired up.

02:46 but it's generally ephemeral, right?

02:50 I start up a Docker container, my app or whatever app is running there might write to it,

02:54 but that's right inside the Docker container.

02:56 And then when it goes away, if I restart it, rebuild it, especially,

03:01 then whatever it wrote is just gone.

03:03 Sometimes you want that.

03:04 But say if you want your web server logs to stick around,

03:07 you don't want it.

03:08 If you want your database data to stick around, you really don't want it.

03:11 So there's a lot of times that you want a permanent storage,

03:16 but an ephemeral operated system and code and so on for Docker.

03:20 So what you can create is a Docker volume, which is a managed persistent virtual hard drive type thing

03:28 associated with that particular Docker container.

03:31 And you can move around theoretically afterwards, right?

03:33 But typically you say, okay, my Docker setup, this database server gets a volume for its data, something like that.

03:39 Yeah, that's one of my favorite features of Docker

03:41 is the immutability of the containers.

03:43 And if people respect that, things go a lot smoother generally.

03:47 Yes, exactly.

03:47 Like, well, what I typically do, Calvin, when my Docker doesn't work, I just log into the container.

03:53 And then I like get pulled a new version and just like read.

03:56 No, just kidding.

03:57 This is this is the modern equivalent of change mod 777 across all your files.

04:03 Exactly.

04:03 It's not working.

04:04 That always fixes it.

04:05 Just fix it.

04:06 I just open all the ports.

04:07 It says it can't get to it.

04:08 What's the problem?

04:08 Passwords.

04:09 Exactly.

04:10 So what do you do with those volumes, Michael?

04:14 So you've got to back that data up.

04:16 I mean, unless you just don't like yourself.

04:18 You've got to back up your data because if something goes wrong,

04:21 from a whole host of reasons, things can go wrong.

04:24 You need a backup of your data.

04:25 That's just living in servers, right?

04:27 Yeah.

04:28 Or just computers.

04:29 So a lot of times what you can do is you can run some kind of command,

04:33 like I use MongoDB.

04:34 A lot of these people, I also use Postgres.

04:36 For either of those, there's a command you can run that says backup this

04:39 and like back it up to the host or somewhere like that and then send it off to a encrypted storage

04:45 or something. But if it's just the hard drive, the volume that you want to back up specifically,

04:52 that's way more challenging. So Brian Weber, thanks Brian, sent over this idea called this Rust

05:00 utility called Docker Volume Backup. And the idea is you can back up Docker volumes to in the,

05:07 all the ways I can imagine I would possibly want to back them up.

05:10 So you can back them up locally.

05:12 You can back them up to an S3 bucket.

05:14 And when you say S3, that means not just AWS that you can learn about tomorrow,

05:19 but many of the cloud storage providers have adopted some kind of S3 compatible API.

05:25 So you could back it up to, I don't know, Digitalization or something if you want as well.

05:29 WebDev, Azure Blob Storage, which you probably could just treat as S3 directly.

05:34 Dropbox, Google Drive, some SSH compatible storage, I'm guessing SCP or something like that.

05:39 All those things.

05:40 And how sweet is that?

05:42 So the idea is when you've got a Docker Compose type of setup,

05:47 you can go and add a separate service, right?

05:50 In your Compose.

05:52 And it just says the image is this project, this Docker volume backup,

05:56 and you give it environment files and then it can just back up.

05:59 That's really easy.

06:01 Really, if it's easy to do, you will do it.

06:03 Yes, it's nice, right?

06:04 So you basically just add it as a separate service.

06:08 And then in your ENV, you give it a schedule and a destination.

06:12 And then you're on the ENV settings, right?

06:14 The backup settings.

06:15 You just give it those kinds of things.

06:16 And if you've got a home lab and you've got a NAS on there,

06:18 most NASs have S3 compatible storage.

06:20 So you can keep it all, I'm kind of touting the local first operation again,

06:25 keeping all your data, your data.

06:27 Yeah, absolutely.

06:28 So this thing is super small, written in Go, 25 megs.

06:31 So there was also a Docker volume backup Docker shell based image, and this is 1/20th the size.

06:38 So that's cool.

06:39 let's see, you can set up a cron job in your composed setup, or you can just CLI run it.

06:45 It, this is, this is part of the stuff that I thought was really cool.

06:47 It rotates away your old backups.

06:49 You don't have to worry about that kind of stuff.

06:50 You say like, just keep me, give me a backup every day or for a month, you know, or, or

06:55 something like that.

06:55 And then you move it on.

06:57 local GPT encryption, like at rest encryption of your backups, which I think is super important

07:03 because the last thing you want to do is like, well, I want to make sure I back it up.

07:06 So it doesn't, if things get hacked, it doesn't go wrong.

07:08 And then that actually being the thing that gets stole.

07:11 Get notification on finished infill build, filled runs.

07:14 you can have it stop the container before it does the snapshot and start it back up.

07:20 So if you're doing like, I don't know, a SQL light sort of thing as your storage

07:25 and you don't need that much uptime, you want to make sure it's not in some like half-committed journaling state

07:30 or whatever, you know.

07:32 You can run custom commands during the backup lifecycle.

07:34 That was my next question.

07:35 Can you run a custom command?

07:36 And here's the one that maybe go start going down this list.

07:40 It has Docker Swarm support plus ARM64 and ARMv7, which means it runs on Raspberry Pi Home Labs.

07:47 Oh, that's nice.

07:49 It's hitting all the check marks.

07:51 Yeah, it's pretty calm, right?

07:52 It's all right.

07:54 That's very cool.

07:54 This also gives me a chance to shout out Idlembyte's search.

07:59 So people come over and search for like Docker backup.

08:02 And Brian pointed out like, oh, if you search for volume until today, it came back with no results.

08:08 And he thought it was broken.

08:09 But actually, he said, it's so fast, I thought it didn't work.

08:12 It felt kind of like a rough thing.

08:14 But I actually put a ton of energy into this custom search engine that I built for this.

08:19 If you're interested in what we cover some of them, please check out our search engine.

08:22 And you can even add it.

08:24 Now, this is like next level right here.

08:26 You can add it as a custom search engine to your browser.

08:28 So you just say PB space and search our site.

08:31 So you can have Python bites all the things.

08:34 Actually, I was looking for that this morning.

08:35 Because I was like, did we cover this?

08:36 Have we talked about this in the past?

08:38 So that's super cool.

08:40 Now, that's awesome that they went and looked at your site to see if this thing existed

08:45 and didn't find it and scratched that itch and made that thing happen.

08:49 So very cool.

08:50 Well, speaking of making things happen.

08:52 Before we move on, I just want to give a shout out to Chris out there.

08:56 It says, backing up a volume is surprisingly hard when you think about application state,

09:01 stateful data, restore points.

09:03 Yes.

09:03 So absolutely.

09:04 And this is a really nice option there.

09:06 Yeah, that is very true.

09:08 Like I said, if you make a thing easy to do, it will get done.

09:10 If it is hard to do, it will go into that queue of things still to do.

09:16 All right.

09:16 Well, thanks for that tip on backing up my Docker volumes

09:20 into S3 compatible storage.

09:22 But the next one up here actually is kind of exciting

09:25 because it's not just a release of Pyodide 314,

09:29 which you'll notice a new numbering scheme there going on.

09:32 The real news here is actually the fact that PEP783

09:37 has made this possible.

09:39 The Pyodide maintainers were basically hand building

09:43 300 plus packages for your browser to use Python in the browser.

09:47 So this is all about using Python in the browser.

09:49 If you've ever thought about why do I have to use JavaScript

09:52 to make fun web apps, you can actually use Python

09:55 to do that with tools like Piodide.

09:57 And now that the 783 has actually been accepted,

10:00 you can use cibuildwheel and put wheels, like compiled binary packages onto PyPy

10:06 so that if you write an app for the browser, someone can quickly download and get the latest versions

10:11 of the dependencies you may be including in that application itself.

10:15 So again, I don't know, they may be seeing hopefully

10:18 a surge of browser-based Pyodide packages onto PyPI or PyPI coming up here in the future,

10:26 but that's actually really cool to see that ease of use now.

10:30 And so there's a whole tool chain available to build these packages and make them available.

10:34 A couple other changes that have happened in this 314 release, you'll notice the numbering change.

10:39 There's a version jump from 0.29 to 314.

10:44 that's intentional. Basically, you're going to have the Pyodide version track the Python version

10:49 that it closely matches with. So if you're looking for the 3.12 version of Pyodide, that's the older

10:56 0.29. But moving forward, you'll quickly spot and know which version of Python you're getting when

11:02 you install and use Pyodide in your projects based on the version number of Pyodide itself. So you'll

11:07 get 314 tracks to 3.14.

11:11 So for binary compatibility, it's locked within the Python cycle,

11:15 meaning the packages you build today won't break

11:17 with the next Pyodide release.

11:18 So another big benefit to this piece here.

11:21 They've also brought back in some packages that were taken out of the standard library

11:27 for Pyodide to make the packaging smaller.

11:30 Things like SQLite3, SSL, and LZMA are back in the default standard library,

11:36 which is nice, so you don't have to basically install

11:38 something like SQLite if you're using that in your browser-based application.

11:43 There's some bug fixes and other things in there.

11:46 There's also experimental TCP socket support in Node.js,

11:50 and you can now connect Pyodide to a real database.

11:54 So that's kind of a wordy, if you've got Python running in the browser,

11:57 but you can't connect to a database or to sockets,

12:00 you were kind of limited in what you could do with local storage, for example,

12:03 but now that opens up a wider spectrum of things you can do.

12:07 It looks like they've tested it on MySQL, Postgres, and Redis

12:11 when running on server side.

12:13 So it definitely blurs the lines between Python in the browser

12:16 and Python runtime anywhere WASM runs, which I think is, again, super cool.

12:21 So I don't know, what do you think?

12:22 Are you building some web-based Python apps, Michael?

12:25 I would love a Vue.js-like equivalent or something along those lines built on top of PyScript,

12:33 which can run Pyodide or even MicroPython, which I think would be super neat.

12:37 I built some examples actually when I built a PWA based on PyScript when it first came out

12:43 and that was based on Pyodide.

12:45 And it worked really, really well, but I have not.

12:48 I've not been doing it lately.

12:49 You guys see a lot more variations with all your projects that you work on.

12:52 How about you?

12:53 I mean, we've got some interesting stuff going on

12:55 in the appliance world, but this might actually open the door to,

12:59 again, more of those at-edge applications where connectivity may be limited to,

13:05 I just see a lot of special use cases that open up when you can run Python in the browser.

13:09 I've always wanted Python in the browser.

13:11 And when PyScript came out, what was that back?

13:13 Oh, it's been a few years back now.

13:15 That was exciting and kind of big news, but obviously it was slow in the early days.

13:19 There was limitations, but this, I think this is a big move

13:23 that's going to open up a lot of options for folks to build either embedded applications

13:27 or things that can kind of stand on their own and run at the edge.

13:31 i'm excited i think it's going to offer a lot of interesting uh moves there yeah uh i'm very

13:36 excited about um about wasm i think it's going to be i think it's going to be cool i think it's

13:40 going to be super cool uh now let's continue on you've been talking a lot about tools uh lately

13:46 you've got another one here i never stop we don't stop talking about the tools so i won't talk about

13:51 won't stop exactly i want to talk about the nbcli so this is a command line interface or automation

14:00 people but the primary reason for bullying is ai agents and notebooks specifically the envy

14:05 obviously is for notebooks so this is a jupyter ai contrib project and i'm sure you've noticed

14:12 this calvin but so many projects now have some form of cli as kind of the final integration

14:19 pointer the ultimate integration should be the ultimate integration point yeah i mean the

14:24 motivation allows like well we our agents can't get to our data and it doesn't know the authenticated

14:29 API is too complicated. So what can we just do to make it simple? But it opens up so much more,

14:34 right, to just have a CLI for whatever you're working on. And so, so here's one for Jupyter

14:41 notebooks. And if you do Marimo, like hang tight, I got something there as well. But the idea here

14:46 is that instead of having the agents try to parse the IPYNB files, which are just gnarly JSON,

14:54 especially if like lots of output is in there, you know, it's like there's your actual code and your

15:01 cell definitions, but then there's maybe 10,000 lines of output to the next cell. That's not a

15:06 great place for agents to be working. Right. Right. Context windows are a precious space.

15:12 Exactly. And you're, you're messing with that. And also it's hard for it to actually kind of see

15:18 the ambient variables and stuff like that that are in the kernel. So the idea here is with this CLI,

15:23 you get more direct access to those types of things.

15:27 So let me over here.

15:30 First of all, experimental.

15:31 So heads up.

15:32 It is Rust-based, so it's likely you can see VC funding.

15:35 That's super good.

15:36 But the idea is that it can read, write, execute, and search Jupyter Notebooks and just skip that whole thing.

15:41 It doesn't necessarily need to parse the IPYMB files,

15:45 and it even works without a Jupyter server.

15:48 So it can just talk directly.

15:51 So it works with files directly, And it talks to the kernels using zero MQ, which is super cool.

15:57 And instead of using that token heavy JSON or just markdown, it uses at at cell and at

16:03 at output sentinels with inline metadata.

16:06 So as I said, less wasted context and you compose these things and it even ships with

16:13 an agent skill.

16:14 This is another thing I'm starting to see a lot of and have been working on.

16:17 I should have prepared an announcement on this that I'm not really talking about, but I got

16:20 some stuff over.

16:21 Some skills.

16:22 I got skills.

16:23 So if you want to install this, you can just npx skills install jupyter-a-dash-contrib-nb-cli

16:32 to install it with Claude Code.

16:33 Yeah, I think that's smart because MCP is not the answer to everything.

16:38 I think there's a lot of value in skills and CLI combined together because, again, preserving

16:43 that context window in a tool like this opens up a lot for folks who are doing data science

16:48 and they don't know how to write or want to write an MCP,

16:51 but they're definitely a whole in on AI and agentic and LLMs.

16:54 Yeah.

16:55 And MCPs basically say, hey, AI, here are some tools you can use,

16:59 but they don't set the context of, here's why you need to use the tools.

17:03 Here's the ways in which they work together.

17:05 You know, like you need an agent or a skill.

17:07 You need a skill to kind of set the stage, even if it's an MCP,

17:10 like, hey, you can use all this stuff in the MCP, by the way.

17:13 Makes me think I should improve my skills.

17:15 Okay.

17:15 So I said I got skills, but we'll see.

17:18 there's an article about it by so you check that out

17:23 and if you're a Marimo person I am going to have

17:27 Trevor on from the Marimo team to talk about Marimo pair

17:31 which is a similar not the same idea but a pretty similar idea

17:34 of basically how you get the agents inside the kernel instead

17:39 of just having them work with the files and so on

17:42 so look forward to that over to you

17:45 I like that I like that. The Marimo pair of aspects. I've become a Marimo, like I mentioned in the last

17:50 episode of Marimo fan. Okay. So pairing well with something like that last tool you just mentioned

17:56 is that a lot of agents and a lot of these agentic workflows are lacking the ability to recall memory

18:03 in an efficient way. So this next piece here is actually a tool you can add into your agentic

18:09 workflow to do AI memory. So, and more than just memory, memory that learns. So thinking about this

18:17 memory problem is actually kind of hard. AIs forget everything between sessions, obviously.

18:21 So you're constantly writing project specs, implementation detail plans. You know, those are

18:27 good practices, but wouldn't it be nice if it could remember like last Tuesday when you use this tool,

18:32 this kind of output happened. And maybe in this context, it's actually better to use this other

18:36 tool, but that kind of, that temporal, like using time and associating facts and details together

18:44 isn't something that just a markdown memory.md file is going to help you out with. So there's

18:49 an open source tool called Hindsight from the folks at Vectorize, and they offer a cloud version

18:54 of this, but you can run it locally. You can install it with Python. It's got, I think,

18:58 adapters for all the major like Python, JavaScript, Node, et cetera. But basically what they're saying

19:05 is simple vector search is not enough.

19:07 Even if you had a ChromaDB locally with your documents all loaded into it,

19:11 that isn't enough to solve the problem about what did you do last spring, for example,

19:16 like the example they give here.

19:18 That requires temporal reasoning and not just semantic similarity.

19:21 So hindsight attempts to add in this extra layer of temporal retrieval,

19:28 which runs semantic keyword graph and temporal search in parallel

19:33 to store, retrieve, and reason over those memories.

19:36 So you actually get kind of three main functions.

19:39 There's three, it's a simple API.

19:41 There's just retain, recall, and reflect.

19:44 And then it uses this temper technique to do the search across all in parallel.

19:50 And then it brings it all together and it automatically consolidates the related facts

19:55 into durable observations instead of piling up duplicates.

19:58 So again, preserving that context window, putting in what is most valuable,

20:03 when it means the most to what you're doing, it's probably the next step in memory.

20:08 I think this is, again, a next front that has to be conquered

20:12 is giving these agents better recall and memory.

20:16 I think a lot of people got a little taste of this

20:17 with OpenClaw, but OpenClaw's default memory, I don't think gives you this level out of the box,

20:22 but you could integrate this into your existing OpenClaws.

20:26 It is on GitHub, it's open source.

20:28 You can pip install hindsight all, And that actually brings down and installs the whole server and process includes like a line chain,

20:35 llama index, Pydantic AI and crew AI and more all kind of bundled together into a nice package that

20:42 you can just start using with your Claude Code or cursor or whatever agent you want. You can now give

20:48 it better memory that can actually learn from what you're doing along the way. So I don't know if

20:53 Have you run into memory issues, Michael?

20:56 Especially as we age.

20:57 Oh, yes.

20:58 100%.

20:59 I have my command book app that I mentioned last week.

21:02 Oh, yeah.

21:03 You're all about this.

21:04 It doesn't work as well as Python in terms of agentic stuff,

21:08 but obviously you can do stuff like that.

21:10 But just to give you an example, I had to add a feature to the CLI.

21:14 I've been doing a lot of search work to the CLI for the command book UI GUI app,

21:19 which is weird, but will make sense the next time I talk about it.

21:22 But the way that those things are built is the Xcode project has them linked in the project

21:28 and you don't have to build them separately.

21:30 But the agent, even though it's worked on it a ton of times, is like, oh, I got to figure

21:35 out how to build the CLI.

21:37 I can't find it.

21:37 It doesn't seem to be a CLI project.

21:39 What is going on?

21:40 And it spent like five minutes on this.

21:41 And what are you doing?

21:42 Save a memory that you just build the project this way.

21:45 And he goes, oh, yes, that would be great for next time.

21:48 Right?

21:48 Right.

21:48 It's like having a fresh intern.

21:50 Every hour you start a new thing.

21:52 Every day they show up, they're like, hey, I'm here for my first day of work.

21:55 Like, no, you're not.

21:56 What are you doing?

21:56 You've been here for weeks.

21:58 And so I think that would be great.

21:59 One of the things that I think would be cool would be, and I don't know if you could do

22:03 this, is have it sync across computers because I work on my MacBook and I work on my mini

22:09 and I'm like, I told it to save on my mini.

22:12 But, you know, why is that not part of your somehow part of your account rather than part

22:17 of your, you know, like on my account, it knows that this project is like this, right?

22:21 so this must have files right maybe you could save them to your dropbox yeah well i was thinking that

22:29 after our tools episode from last week i'm really focusing on building out my framework desktop

22:35 machine to be that central place i go um i mentioned tmux last week is my favorite go-to

22:40 tool for basically using mosh and logging into that machine but i think i'm going to switch over

22:44 to zelige if you've tried z-e-l-l-i-j i don't know how you pronounce it these are zelige or zelige

22:50 but having this mental model, this hindsight running over there

22:55 where no matter where I'm out and about or if I'm at my desk or sitting on the couch,

22:59 I'm always actually working off that one machine

23:01 where I've got great local model availability.

23:03 You can talk to it over the network?

23:05 Yeah, talk to it over the network.

23:05 Exactly.

23:06 Well, combining that with the tail scale or I've got the Ubiquiti,

23:12 what do you call it?

23:13 Teleport here.

23:14 Yeah.

23:14 Okay, this is cool.

23:15 I'm standing up on my mini.

23:16 Yeah.

23:17 Yeah.

23:17 I think you put it on one spot.

23:18 So you can see here, the memory bank consists of not just like

23:21 a single markdown file of memories, but it's like middle models, observations,

23:24 memories and entities, chunks and documents, which is important because whole documents

23:29 won't do it for you.

23:30 Sometimes it needs to be able to do that reasoning

23:32 and reflecting and pull those chunks into specific spots.

23:37 And based on, again, I think the key bit here is time,

23:40 a clever name, hindsight.

23:42 In hindsight, time would matter for memory.

23:45 And so I think they've definitely solved this problem,

23:46 which I think is pretty cool.

23:48 So check it out.

23:49 Yep.

23:49 Check it out.

23:50 I'm going to, I'm going to install it on my framework desktop and start using this for

23:53 memory inside of my projects.

23:55 I think it's going to be pretty cool.

23:57 All right.

23:57 Let's head into, let's go to the extra part of the show.

24:02 And I think both of us have quite a few extras here.

24:05 When you want to kick it off, Michael.

24:06 You got your screen.

24:07 Okay.

24:07 Yeah.

24:07 Okay.

24:08 Okay.

24:08 Okay.

24:08 I can do that.

24:09 Yeah.

24:09 Let me, let's see here.

24:10 The first one I was going to talk about, oh my gosh, I've got too many, I've got too many

24:14 things on my screen now.

24:16 That's basically what this, I guess it's about.

24:18 How do you refer to your AI agents when you're talking to them?

24:22 Do you have a term of endearment for them, Michael?

24:26 Not in a personalized way, but I refer to Claude as Claude.

24:30 And I refer to...

24:32 With their name, their branding, their branding.

24:34 Yeah, I refer to ChatGPT as Chat.

24:36 Like, hey, Chat, let's talk about this.

24:37 Or like, hey, Claude, good job on that.

24:39 I kind of like Armin's approach of calling it clanker.

24:42 I think it puts it in...

24:43 Clanker.

24:44 Clanker.

24:44 It kind of puts it in the right spot.

24:45 And when I think about it, I think about this.

24:48 I assume this is the Star Wars reference because in Star Wars, they called all the Trade Federation robot bots clankers as a derogatory term in their case.

24:57 But I feel like it doesn't make it as human.

24:59 And you can yell at a clanker a lot easier than you could if you're yelling at Claude.

25:03 People might get upset about you yelling at something named Claude.

25:06 So I thought clanker was kind of interesting.

25:08 Check that out.

25:08 There's a blog post from Armin about it from a couple weeks old, but I think it's still very relevant to check out.

25:15 So that was my first one. And the next one is everyone has one of these in their office.

25:20 Someone decided to make it more official. Do I not have the, let me grab the link for that one.

25:26 There's a project called Ponytail. And now I've got it open over here.

25:30 I read this article. I know I watched a video evaluating this on YouTube just last night.

25:35 There it is. Yes. So make your AI agent think like the laziest senior dev in the room.

25:40 The best code is the code you never wrote.

25:43 So instead of the, it's kind of the not invented here,

25:48 like our Yagni, like I'm not gonna rewrite my own requests library.

25:53 Make sure you always use the libraries that are available.

25:56 I kind of like, I haven't tried it, but I thought it was kind of clever.

25:59 Less code, less tokens, obviously.

26:02 And you basically can just insert into your code,

26:05 hey, ponytail, browser has this, and put in there like a placeholder of like what you want,

26:10 if you want to date that for example in this case is a date picker so it's like input type date and

26:14 then let your agent figure it out with this uh skill installed to figure out what what um uh what

26:21 dependency it should go install instead of trying to write its own so a lot of times these ai agents

26:24 are going to try and write their own versions of things that probably already exist why would you

26:28 they're very happy to go oh yeah enterprise on you when it's like i just i don't need this is not in

26:34 need of that. It is a very simple project. I just need like a dialogue. So maybe you could just use,

26:40 hey, guess what? HTML is recently has a built-in dialogue thing. You don't need to install React

26:45 just so I can have a dialogue. You know, that's kind of the vibe. All right. The next one is,

26:49 we've been talking a lot about agents, AI, and giving them control over our computers.

26:55 How can we do that in a safe way? And how can we do it for multi-user? A good friend of mine,

26:59 Chris McDonough from the Zope Python clone communities. He has wrote a agentic wrapper

27:07 around all this. So it kind of does containerized sandbox workspaces where you can actually do it

27:11 multiplayer. So check out Dave. I don't believe he's done much fanfare around this project. It's

27:17 very new. Just started a couple of weeks back, but I think it's already made a ton of progress.

27:20 There's already like well over a thousand commits in there from him and a couple other contributors

27:24 into this, but basically use it with Pi or Claude Code, but intentionally given wide permissions.

27:32 So you want to be able to have them read, write, and execute code, but do it in a safe space,

27:37 in its own safely isolated workspace. And then for Teams, you basically, I believe it incorporates

27:44 Tmux in for allowing kind of the multi-user, multiplayer. If you've ever done Tmux with

27:49 multi-user, it's actually kind of magical. It feels like Google Docs where multiple people can edit.

27:54 simultaneously, well, multiple people can be in the terminal

27:58 simultaneously working together.

28:00 And so when you wanna collaborate and wanna be using an agent

28:03 and actually collaborate with other humans, this actually may be a really good option for you.

28:06 It is called Clang, kind of maybe a play on the word clankers again.

28:12 So Clang with the K at the end, how do you pronounce that?

28:15 These projects need to get better naming.

28:17 Clang I don't know.

28:20 Yeah.

28:20 I honestly think when projects are like this and they could be pronounced many ways,

28:23 There should be a little MP3, just like, here's how it sounds.

28:26 How do I pronounce this?

28:27 All right.

28:28 Another one.

28:29 Who's been having struggles with GitHub and their outages recently?

28:33 Well, Cursor has announced Origin.

28:36 It's currently on a public wait list, but it might be something worth watching.

28:39 But it is basically a source control optimized for the agentic world.

28:43 I think they are claiming something like over 22 commits per second support out of the box.

28:48 So if you've got a lot of users going at it with their agents

28:52 and be able to commit at that level of scale, I think this is the future

28:56 and where GitHub better get their act together because it sounds like Cursor's coming for their lunch.

29:01 And the last extra I wanted to cover was performative UI.

29:06 Michael, have you ever wanted to launch your own AI startup

29:11 and you needed to have the right AI components onto the page?

29:14 Oh yeah.

29:14 This is AI native React components for the next you name it.

29:18 So any fancy sparkles, any fancy pills with, you know, status lights in them.

29:25 You can just npm install performative UI.

29:27 I just added a pill to my, one of my websites and I didn't even think that I could use the framework for it.

29:32 Oh yeah.

29:33 Well, so there's, for example, golden eye, the golden eye effect.

29:36 If you'd like to have that, you can, you can get that right out of here.

29:39 I kind of, I kind of don't want to like this, what I do kind of like it.

29:41 I know.

29:41 I was, when I saw it, I was like, oh my gosh, the things, I mean, you, you scroll down through here, they have thought of, well, or

29:48 every possible trope that most of the ai startups are doing uh if you want to have your chat bubbles

29:53 look like they are coming straight out of the existing chat client uh if you want to have an ide

29:59 they've got a mock ide component that you can land in your page to show developers that you're very

30:03 developer friendly so that was a fun one uh that i thought i would share with the group here uh

30:09 michael what do you got in the way extras all right i got it i got a couple of things um by the way

30:13 before I jump over there, it's just, you talked about Armin's article on clinker, clinkers in that

30:18 word. I have gone the opposite direction. And I found it surprisingly makes me feel better working

30:24 with these things is I've gotten Claude to address me in a way more personal way. Okay. So instead of

30:31 going, the user has requested or the project, you know, it refers to me as Michael and like uses me.

30:38 So it'll be like, and Michael already committed this.

30:42 And so we're going to do this.

30:43 And it's because it sounds so silly, but here's the thing.

30:47 I find some, you know, without those kinds of things, like doing a lot of this agentic

30:50 stuff, it's super sterile.

30:51 It's still like working with other people.

30:53 I'm just now working with this thing that like refers to me as user.

30:55 I'm like, I am not a number.

30:58 I am a man, you know, like that's what I feel like.

31:00 I'm going to be worried, Michael, when you start referring to yourself in the third person.

31:03 I know.

31:05 Well, well, well, Michael wants to cover.

31:08 No, just kidding.

31:09 So a really cool Talk Python episode that I released last week, and it will still be the

31:15 live one for a day or two before I release.

31:17 The next one is I interviewed Charlie Marsh.

31:20 Now that Astral has been part of OpenAI for a little while.

31:23 What a journey.

31:24 Yeah, basically it talks about the whole journey he's been on.

31:27 It talks about what does this acquisition by OpenAI mean for the Python community, for

31:33 uv ty rough all those kinds of things so uh i think people if you care about like what is happening to

31:40 uv check this out and you mentioned cursor they're gonna have a little extra money to you know build

31:46 build origin because they just got acquired by spacex or are going to be i mean what do you do

31:53 what do you do with that extra money other than build more stuff exactly you could run a lot of

31:59 agents for 60 billion dollars because that's what they're apparently getting acquired for 60 billion

32:04 dollars a vs code clone oh my gosh wow wow wow wow wow um i was a huge lover of of cursor for a long

32:13 time and never been a cursor user it was not my not my normal toolkit not your vibe no it's i'm

32:19 cli's i know i know see this is why it's good you're here we have different perspectives i love

32:24 of cursor but it's been moving more and more towards just like just fire off a bunch of agents

32:28 that don't even look at what it does sort of that's been it's what it's embracing and the more it does

32:33 that the less like i kind of want to still see what's going on a little bit anyway um also astral

32:39 intersect open ai but this is not the same thing charlie marsh posted i'm proud to announce that

32:44 we're renewing and expanding our support for maintainers of open source projects used across

32:50 Astral and Codex tool chains at OpenAI.

32:53 Today, that starts with committing more than $160,000

32:56 in direct sponsorship through GitHub sponsors.

32:59 Nice.

32:59 So there's the Astral open source fund and so on, and it shows you how to apply.

33:05 So if you want to get, also the Codex open source, a million dollar fund

33:09 that includes six months of ChatGP Pro, API credits, and so on and so on.

33:14 Which have you noticed in Codex, if you're using ChatGPT in Codex,

33:18 it now gives you an option to reset your quota twice.

33:20 Like you get two resets of, oh,

33:23 I messed up and I went over.

33:25 Okay.

33:25 Interesting.

33:26 Watch, keep an eye on tokenomics.

33:28 I think this is going to be the next three months,

33:31 two months are going to be very interesting in the token space.

33:34 I agree.

33:35 And things like, not necessarily endorsement,

33:37 but things like ponytail.

33:39 Yeah.

33:39 In memories and other stuff that keep it more focused are going to become more and more

33:43 important.

33:43 I imagine.

33:44 All right.

33:46 What am I?

33:47 Oh yes.

33:47 I know what I'm talking about here.

33:48 If I go to any of these courses on Talk Python, what is really cool is I can now, I guess I had to log in to show you, which I'm not.

33:57 I won't bother.

33:58 But now every single course at Talk Python training, all 283.9 hours of them have Spanish subtitles, German subtitles, which they always have been there for a little while.

34:14 But the announcement is now Portuguese.

34:16 Every course has full high quality Portuguese transcripts.

34:21 And boy, boy, was that a lot of words to generate those correctly.

34:24 So who QAs those?

34:26 Do you have some folks in the community who are QAing these translations?

34:29 No, I have a whole series of AI agents and Python software that like re-evaluates them.

34:37 And it's just a whole back and forth.

34:40 I was able to QA them somewhat in German because I can speak some German, but my technical German is not great.

34:47 So, you know, somewhat.

34:48 But it worked well for there.

34:49 So I'm just trying to apply the same process.

34:51 We'll see.

34:51 People let me know how they're turning out.

34:53 But I think they're pretty good.

34:54 Well, I'm glad you're making the effort to make it accessible.

34:57 Yeah, exactly.

34:58 I'm thinking, even if they're not perfect, you know what?

35:00 The English subtitles aren't perfect either.

35:02 They were done mostly by humans directly, but non-technical humans.

35:07 So they still need a little bit of fixes here and there.

35:09 That's the challenge.

35:10 right like even if you get a human being to do it then you got to find a human being who both is

35:15 patient diligent and does python which is like really and wants to do transgress this is a very

35:21 rare person i'll tell you what okay a couple more quick announcements uh this one um we've got the

35:27 jango developer survey is out but the analysis in the search for a dsf executive director so

35:35 So boosted by our friend Jeff Triplett out there.

35:39 So they're looking to hire somebody for this.

35:42 And I believe you deserve a bit of a shout out here.

35:46 Hey, look at that.

35:47 Shout out to Six Feet Up, who is one of the key sponsors making this happen.

35:52 Do you want to say one quick word about that?

35:54 You know more than I do.

35:55 Yeah.

35:55 No, I think we're definitely very involved in wanting to be engaged in getting the Django

36:01 Software Foundation and executive director.

36:03 it's gonna be a big move for that group to have that kind of dedicated leadership.

36:07 Right now it's very all volunteer run.

36:09 So things like the conferences are done through another volunteer group called Defna

36:13 and the Software Foundation managed some pieces of it.

36:15 There are some paid positions like the fellows who work on the code,

36:19 but this is working on growing the community.

36:22 And there's a huge tale of people using Django out there

36:26 in the world and they don't know how to give back.

36:29 And I think that's part of this executive director's mission

36:32 is going to be how do they go and find the folks

36:36 who are getting a lot of value out of Django and making sure that that value kind of gets put back

36:41 into the community in the form of sponsorships for the software project itself.

36:44 So we can have more fellows, we can have bigger sponsors at the conferences,

36:49 people can get greater adoption of Django into more places.

36:52 I think there's a lot of behind the firewall usage here

36:55 that this is gonna help bring to light and get more sponsorships.

36:58 So yeah, 6Whip is super excited to sponsor and help find the right person.

37:02 That's awesome, Calvin.

37:04 One of the, I think, key secret sauces of Django is the community.

37:08 So this is right in that space, yeah.

37:09 Yeah, yeah.

37:10 Especially now in this agetic world, right?

37:13 Human contact is even more important than ever.

37:16 I agree, 100%.

37:18 All right, I've got a joke.

37:19 So you got a joke for us?

37:20 You got a joke?

37:21 It's called, oh, babe.

37:22 Oh, babe.

37:23 Oh, this is rough, but it's right on point for what we've been talking about, some of the things.

37:29 All right, hit me with it.

37:30 You talked about tokenomics.

37:33 So this is a screenshot of an iMessage.

37:37 It says, babe, heart, heart.

37:38 $15,000 out of our ticket to go.

37:40 Oh, my God.

37:41 Oh, is this what I think it is?

37:43 Ring emoji, ring emoji, ring emoji.

37:45 Hands over.

37:46 Anthropic PC.

37:47 Invoice from Anthropic.

37:49 $15,000.

37:51 Keep an eye on that token bill.

37:53 I don't think it's going to go down.

37:55 I'm so sorry, but it's not a ring.

37:57 It's just a monthly Claude bill.

37:59 How rough is that?

38:00 It's very real.

38:02 I feel seen.

38:05 All right, Michael, thank you so much for joining us for another episode of Python Bytes.

38:11 We will be back next week with more Python news and headlines right to your earbuds.

38:17 So we'll see you all later.

38:18 Thanks, everyone.

38:19 Thanks, Calvin.