00:00
At Real Python, we’re always excited when new models come out in this world of AI. And just a few days ago, OpenAI released their Astra model. So we thought let’s jump on a call, talk about it.
00:11
And joining me today is Dan Bader. Hello, Dan. Hey, how’s it going?
00:17
And Martin Breuss. Hi, Martin. Basically, we don’t have any big agenda. Just talking a bit about our first impressions of the model. So let’s make a quick round.
00:28
Dan, have you tried the new model? What are your first impressions? How are you feeling about this? Martin and I stayed up in the middle of the night, you know, after our kids were asleep to try out the model.
00:39
It sounds like we share a family, but that’s actually…
00:44
Any parent knows that’s a big sacrifice. We’re sacrificing sleep to play with new computer stuff. For me, at least, you know, that’s like a little sacrifice there health-wise, and
00:55
I’ll make it up with coffee intake the next day. But I started playing with it as soon as we got access to it. It was kind of a weird launch, really, where there was that big outage where all kinds of services went down.
01:06
We experienced it, too, at Real Python, where our video hosting went down for a short amount of time. There was some blip on, who knows, Amazon or whatever, AWS.
01:15
And I think it kind of interfered with OpenAI’s launch for this model a little bit. And so there was a blog post out, and then it disappeared. And it was like, it just felt like very like, oh my God, you know, this is the big one.
01:27
This is like the super important release. And then what is happening? Like, is it already taking over and hacking servers?
01:35
You know, the internet’s going to go out soon. But then, yeah, it was just some, you know, infrastructure issue. And we ended up getting access on Friday.
01:43
And that’s when it showed up on OpenRouter and stuff. Definitely my favorite model from OpenAI, like from their offering there. It feels to me comparable to Claude Fable.
01:55
And I guess at some point we’re going to talk about like, you know, Fable 5.1 and all of these other ones. It doesn’t feel like I’m taking a step back in like model capabilities when I’m working in Codex compared to Claude Code or when I have access to Fable.
02:09
Because just Fable to me was like such a jump in kind of capabilities. And now it feels like Astra is basically there.
02:16
Since then, so this was, we got access late Friday night. And so now it’s Monday afternoon. So I used it for a bunch of things. Man, it seems very, very capable.
02:28
I had it rewrite just for fun. I had it rewrite a code base from Python into Rust.
02:36
Don’t say this on the record. Yeah, I know. It’s just amazing. I mean, it’s kind of an easy task for these models these days, but it just did such a brilliant job.
02:44
And I’m not much of a Rust programmer at all. I was like, damn, this is identical to what I had before, but now it’s three times faster. Like, what does that mean?
02:53
Right? So that was very impressive.
02:56
It feels very economical in its output. Like it’s not as verbose. And it’s sort of more of a proactive partner, kind of like what working with Fable felt like.
03:06
And it’s always hard to say, is that just innate to the model? Is that the harness or whatever they ship around the model? Right. Is that improvements they made to Codex or like different system prompts and what have you?
03:16
But it just felt like it was a noticeable improvement in that regard. And so, yeah, I don’t know. I had a lot of fun with it. I found actually on the lower effort settings, that’s all I need really.
03:28
I found myself using the fast mode too. And you can kind of do it even though it’s this big sort of expensive model. If you use it on the lower settings, it’s very fun to use that way.
03:39
Very capable.
03:40
I had it screen about a thousand emails today. And it did a great job emulating sort of my voice, like pre-writing some drafts and whatnot.
03:49
I’m very happy with it. I mean, definitely feels like a big step up and very important release for OpenAI.
03:55
Martin, do you share that sentiment? How was your first impression? Yeah, I mean, I guess I didn’t get to play with it that much. I had a lot of childcare this weekend
04:05
because my wife isn’t here at the moment, but I did get to work with it a bit. And I feel like especially the things I’ve been using it primarily for were voice-related, so text, text-based stuff, and also like the interactions that you get with the model when co-working with it, essentially.
04:22
I currently prefer it over the Anthropic… I’ve been mostly working with Anthropic models, but like just the way that the interactions feel when working with it is an improvement, in my, for my feeling.
04:35
And then also, like the way, so what Dan mentioned, that it actually got his voice right in terms of writing those emails. I also had some experiences with that when experimenting with updating content that is like quite outdated, and like figuring out how can I get a kickstart on having an updated content piece that still feels like the original, essentially, right?
04:56
But it’s technically accurate and up to date. And that’s been working really well with the Astra model as well. My main experiments and experience, that’s what I was trying to say, text and language related.
05:08
It’s a really nice feeling. And like, even maybe you’ve seen these launch videos where they’re like, okay, so, you know, I have this couple of people in the room and they’re like, do this now, do that now, can we talk about that for a second?
05:20
Like, what does everybody feel? Well, I’m gonna say I have something to say about it. Yeah, yeah. Which is like, I actually felt like that, you know. Yeah, I didn’t feel it was like super overblown.
05:29
I mean, I didn’t like print out a little rocket ship or whatever, or make a game out of it. But the actual interactions, I felt myself similarly, where I could imagine being in a room.
05:41
And then the voice interactions, which is like how I interact with these models a lot. Like, I mean, I switch between typing and speaking. But yeah, the things I was working with Astra didn’t feel so far off from that very, you know, hyped ad, obviously.
05:57
So that was kind of an interesting experience, I guess. It does feel like the Star Trek computer or something, right? Especially with the voice mode stuff, which, by the way, I always liked OpenAI’s integration.
06:10
They’re better for some reason. I think Anthropic, they have now caught up with Claude Code and their mobile app. They’re using a better transcription model.
06:17
But it feels like this is the computer from Star Trek and you just ask it to do stuff and things are happening. So I thought the video they did was quite cool, with this sort of retro aesthetic and stuff, nice in their marketing.
06:30
Although I feel like a lot of the use cases were kind of, you know, these models are so powerful, and it’s like what we’re showing there in the marketing is like, oh, I’ll make like a little like 3D game.
06:40
And that’s all cool stuff, but that’s not really what’s going to have the big impact, really, right? Like the impact’s going to be like, oh, this thing is going to write all the code, like this thing is going to do all the accounting and all of these things.
06:52
So they’re kind of, I think they’re pulling their punches in the marketing, but it was a cool video. Yeah, and it’s like, it’s the consumer orientation too, right?
07:00
Like OpenAI has been doing that a lot. And if you compare it to how Anthropic’s been handling this kind of stuff, where they’re much more focused on different business, you know.
07:10
But then OpenAI has always been focused on the consumer market, essentially. Whether or not that’s working out is a different question, I guess. But yeah, I agree.
07:19
It feels like a… which maybe also makes it, you know, easier to identify yourself as being that person. So maybe it’s just, it’s good marketing that, like in that sense, I remembered it.
07:29
You know, I was thinking about it, I was like, yeah, look, that’s me.
07:34
Yeah, but I feel it’s like, what kind of like my little issue with releases and also the time after that, like you always see these examples where it’s like this big 3D rendering and my game that I one-shot prompted, and everything.
07:55
And I feel it’s, new models and the capabilities will especially show in your day-to-day and how it kind of like feels interacting with it, and if it’s actually like supporting your work and you don’t have to work against it that much.
08:09
Which personally I felt a bit with Opus 5, where the outputs felt not that great to me. And I think with Astra it got more on the, okay, like we’re more thinking in the same ways.
08:23
I’m wondering, Martin, you were saying like that it was also for you, like with the content that it created and the text output that you had, did you have to adjust something with your prompting that you did with a model before?
08:37
Or it just kind of like moved on with all the skills and prompts you had before? No, I’m, I mean, honestly, I’m just giving these models my thoughts as directly as possible these days.
08:49
There’s not much prompt engineering that I do in my day to day. I mean, I feel like it’s more about how you use the harness rather than like how you write individual prompts.
08:58
Yeah, it’s kind of interesting, you know, when you speak to other people, how do they interact with AI? I’ve always had the feeling that I do it kind of like automatically differently or something.
09:07
But that’s, I mean, that’s a selection of people that you speak to, I’m sure. But, you know, I hear and read about people arguing with AI, for example. That’s just never been a thing for me.
09:17
I’ve never gotten into this, you know. So it’s, okay, this conversation isn’t leading me anywhere. I cut the context, I go back, I go somewhere else, or I just start a new conversation when I do something.
09:27
I haven’t changed anything. That’s been, I guess, a practice that I’ve established working with these tools over the past, what’s it been? A long year and a half or what?
09:36
I don’t know. Like dog years, right? It feels like decades. AI years. I don’t engineer my prompts. I speak my mind as much as I can, basically. That’s my prompting.
09:49
Regarding the prompting, one thing I noticed is that the Fable-class model and the Astra, they seem to be size-wise similar. And a step up, nobody knows for sure, but the scuttlebutt rumors have a lot more parameters, so a bigger model, meaning more innate knowledge baked into the model.
10:10
And what’s interesting is that both Anthropic for Fable and OpenAI for Astra release kind of new prompting guidance on how you’re supposed to prompt it. And they even have automations that allow you to sort of review your old prompts.
10:24
And so I did that today on a couple of models. Like, I think I posted on our Basecamp for the company. It was pretty interesting, because both companies now, they tell you to basically don’t prompt engineer, right?
10:38
And keep your skills and sort of prompts generally, I guess, shorter. Don’t do stuff like, “You must blah blah blah,” and like, “You are like a very experienced, highly intelligent, like galaxy brain.”
10:51
Don’t do that stuff. “That depends on the health of the future of my family.
10:57
You will be severely punished.” Don’t do that stuff, right? Just sort of almost be more minimal, to not, like the way I understood or interpreted it, is like you don’t want to immediately put the model into these rails, but you wanted to allow it kind of more freedom in how it’s going to get to the goal, because it might actually have a better capability, and sort of, you know, a new model might have a better idea on how to do that.
11:24
And I thought that was really interesting, because I noticed that too, that a couple of months ago, we would have to use, for coding, we would use like plan mode and all of these things.
11:31
Because if you just told the model like, “Hey, I just want to talk about this, like, don’t go and start editing all my files,” it wouldn’t really work. And it would forget after a couple of turns.
11:40
And now with Astra, same as Fable, you can just ask it stuff and say like, “Hey, I just want to discuss it or brainstorm with me,” and it’ll… like plan mode is just a capability of the model now, it feels like.
11:51
And so to me, like that is almost, that was like the biggest shift in like moving from something like Opus or 5.6 Sol to these like larger classes of models.
12:01
And now both providers have arrived there with their respective models, which is, it’s kind of cool. Like it’s just so much capability. I had a similar experience where first when I read the press announcement, like those blog posts, which I always find, it’s like a mixed bag because it’s like all these benchmarks and stuff like that.
12:22
But one thing that jumped out to me was that OpenAI was talking about that this model is more aligned. And I was a bit confused about like, what do they mean with aligned?
12:34
And the more I worked with Astra over the last few days, that’s actually the word that jumped to me. It’s kind of like it really is aligned to my way of thinking, and not kind of like in a “it can read my mind” sense.
12:48
But for example, when I was pointing something out about like a typo I made myself in a prompt, it just went on with this typo without kind of like updating it on the way.
13:01
And in the past, like every model I used, it was just like, “Oh, there is a typo. I will fix that for you.”
13:07
And this was like an interesting thing. I don’t know if I liked it or not, because after that I was also pointing out, like, “Hey, didn’t you see I had a typo in there?”
13:14
It was like, yeah, there was a typo. The correct spelling would be this. Obviously. Like no question after. So I was like, yeah, please fix that. And maybe that was kind of like something where I also have to relearn a bit.
13:29
Be like, yeah, if I really want it to go ahead and fix it already, it’s something that I also have to tell. And that’s what you were just saying, like in the past you had to be like very protective that it doesn’t run away with things.
13:43
And now it feels more like it sticks more to what you were asking, and if you didn’t ask for it, it also won’t do it. That’s really interesting, because I noticed a different, like kind of similar, you know, in this kind of usage or UX, like what does it feel like working with the model?
13:59
Like definitely feeling very strongly that it’s a better collaborator, right? Where it’s like, it’s just more, it’s just easier to work with. There’s less kind of rough edges, right?
14:08
So it’s a better tool, better collaborator to work with. I did notice though, I don’t know, maybe it’s just because I’ve been using these Anthropic models so much recently, because they were just generally stronger.
14:19
I did have a couple of moments where it’s like this classic sort of OpenAI model moment, of like, there is like a very obvious thing that I was asking, that I’m trying to come up with like a concrete… This was just like doing some random research or like summarizing an article, just for, you know, I wanted to churn through some stuff that had piled up.
14:36
And I want to say I asked it like a really obvious question, of like, well, what does that mean or something? And it was like, what are you referring to? Like, it just didn’t… It was like some glitch or like, I had a couple of moments like that with Astra, and that hasn’t really happened with Fable for me, where it’s just, I can be as like unclear as I want, almost.
14:57
It’s generally kind of, you know, reeling me back in and sort of, I end up where I want to go. And I felt like just Astra was, whether that’s prompting or whether that’s the harness or whatever, it just felt like a little bit weaker in that regard.
15:09
But it could totally be, you know, I’m just so used to the Claude models. Kind of, they have a certain, you know, personality is maybe not the right word, but they have sort of a certain way of you interacting with them and steering them.
15:23
And maybe that doesn’t translate as well to the OpenAI models. But I was kind of surprised when I hit that moment, because like, how did you just, you know, write all of this code and like told me you fixed a bunch of bugs in my original implementation, which was also something that I loved and that I don’t really, you know, didn’t really experience with smaller models where, you know, just, okay, well, I translate everything and the bugs are still in there.
15:44
And then this thing was like, “Oh, by the way, there were like three bugs in this. I fixed them in the Rust implementation. Do you want me to go back and also fix it in your Python original?”
15:52
And I was like, yes, of course. And it was good. So there it is: bugs, but not typos. Yeah. It’s like the headlines. It feels to me the same way, that going from Opus to Fable, now going from 5.6 Sol to Astra, it’s like, okay, OpenAI has like their equivalent.
16:11
But it also doesn’t feel like stronger than Fable, as far as I can tell. Like I would say they’re sort of on par, and it’s like personal preference which one you like better.
16:19
Do you guys feel the same? I guess I probably haven’t used Fable for the same tasks that I’ve tried using Astra for now. So I don’t think I have a good answer for that at the moment.
16:33
I’ve used Fable quite a bit when I, you know, I had the feeling I needed more intelligence for some tasks. But it’s like, it’s a different billing, so I wouldn’t just go run Fable for everything.
16:43
I’m still using Opus for most of the things that I do, and then I would selectively jump to using Fable because I was like, this is a crucial, like, planning step, or I want this, this is really important, so I want to get it right, so I’m gonna up the intelligence for that.
16:57
And those were not, you know, text, those were not tasks that rely on producing text or anything like that. I don’t really have that comparison, but I really like using Fable for those other tasks, and I thought it was doing a really good job.
17:10
I mean, maybe I’m just tired of some of these AI phrasings that it uses to communicate with me, that are also happening on Fable for me, like similar to Opus.
17:19
Just, you know, just the communication, just the back and forth, basically. It just keeps having, I don’t know, just, and maybe it’s Anthropic models, I don’t know that.
17:28
Yeah, keep using the same phrases where I’m like, okay, yeah. And I just, you know, I just read over it. I’ve seen it so much that I just jump over it. Yeah, and it’s been refreshing for me that with Astra I don’t have to do that, because it’s just a different back and forth that doesn’t feel like very, like, “But here’s the thing,” or “Here’s the…” or, you know, yeah, I’m sure you know all those phrases.
17:49
The voice is very, I guess I’ve seen people call it like Claude-isms or Claudies, which definitely, I felt like the Opus 5 model kind of took that to the next level, and that was just crazy.
18:02
So yeah, I agree. I guess I see it a lot when I do Opus, but I do a lot of Opus work, right? So it just jumps out now when you go back to… And I totally agree, Astra is very, I mean, for me, it felt just really kind of terse and to the point, which was refreshing actually coming from, you know, like using Opus 5 and then to a certain degree Fable or like Fable 5 versus 5.1.
18:22
There’s also a bit of a difference there for me. But yeah, it felt refreshing. That was so economical. And one thing, I’m curious if you had the same experience with it, but I just felt that Astra especially, I just use at the lowest effort setting and I feel like I’m actually getting better results out of it.
18:40
Like it’s certainly when I’m working with it interactively, like it just doesn’t… If I put the effort to high, it just wants to go and like do all the things.
18:48
With low or medium effort, it just felt kind of the sweet spot, which was interesting to me, because with the previous models I would be like, yeah, okay, I can’t really, you know, you got to crank this thing up to the max.
18:59
And now with low effort, it’s even faster because it’s spending less time generating thinking tokens, but it’s still getting to a very good outcome. So have you noticed that too?
19:08
I think it’s interesting that you’re saying this, because for me, I feel as well, same for Fable, like that with the low effort, it’s actually like it produces good output.
19:18
And I was wondering, like, because you have like the pricing, and especially when new models come out, they’re more on the expensive side, so you want to be a bit more economic with your tokens, which makes me switch to the lower end of things faster than in the past.
19:33
And the other thought I had was maybe it is like what you were saying, and I have to check, like, when didn’t it fix my typo? So maybe it was when it was in low mode or something like that.
19:45
That the more you use the model, the more you can also anticipate how it does things. So maybe cranking it up will happen over time again, because you know what to expect.
19:55
With the bigger jump, so maybe that is also this time at the beginning where you kind of need to get a feeling for it. And additionally to that, what you were saying, Martin, about like how Opus was phrasing things, I’m wondering if we might talk in three weeks about Astra and being like, oh gosh, like these phrases that Astra is using, because we’re more used to it and then we start seeing the pattern.
20:20
So maybe currently it feels new and feels very special and like… Refreshing. Just a different person to speak to. Yeah. And then it’s like this pattern again.
20:31
So I’m wondering how much is there at the honeymoon phase of this new model as well. But I agree. I feel like these effort level settings really make a difference.
20:43
OpenAI hasn’t said that publicly, but apparently it uses a different architecture for the model itself. So it’s called a recurrent transformer, where basically multiple transformer passes are happening,
20:54
which means the model is thinking in that sense, or again, that puts it in very human language, like it’s sort of processing stuff.
21:03
And it doesn’t generate output tokens. Where the previous models, like 5.6 Sol, when it’s thinking, it’s generating thinking tokens that we don’t, for all the security reasons and stuff, currently see just plain in the API.
21:17
But if you run a local model, you can see like a perfect sort of thinking token output, and then you get the true answer. We only get a summary with the OpenAI API.
21:24
But apparently, this new class of model, Astra, there is more thinking that happens within the model, which it does not generate visible thinking tokens, which is actually a problem for this kind of alignment and safety.
21:38
Well, previously, I think how this worked is you just had a judge model look at the reasoning tokens and be like, “Hey, I think there’s something bad, right?
21:47
Throw up a flag.” And now you’ll have to re-engineer that, where it’s directly looking at the activations and the weights, or, I don’t actually know how that works or how feasible that is.
21:58
But it is interesting how, yeah, it seems to be very economical in its token output. Like when you look at what is happening behind the scenes, it’s like, it feels very different from 5.6 Sol, which probably means new architecture and maybe new exciting developments there over the next couple of months.
22:15
It’s interesting that they, you know, what Philipp mentioned that they called it the most aligned model, when, like, I wonder what they have in terms of checking up on alignment.
22:24
Visibility into its thinking process is so much harder to do. I mean, maybe it’s not harder to do, maybe they have a different way of doing it. But I mean, that’s interesting for me, especially in the light of like the recent news, right?
22:38
With all these like hacks and the Hugging Face incident and all of that stuff. Yeah, I think that was one reason why it was so much in the marketing, because they also then shared like an example of like the Sol versus Astra, and how much like Sol went off the rails with certain tests, and how much or how little Astra was doing that, and therefore being like more aligned with the task at hand.
23:04
I want to come back to one thing you were saying then about this Python to Rust conversion. So maybe to get a bit more practical towards the end of the conversation, it was an experiment that you were running with Astra to kind of like see how it performs.
23:19
And do you feel like with one of the former OpenAI models or with the Anthropic models, like before Fable 5.1, that wouldn’t have worked as pleasantly? I don’t think so.
23:33
To be fair, I haven’t tried this, right? So this is more like a vibes gut feeling check. Fable, absolutely, yeah. I’ve totally, I tried this with other code bases and similar results.
23:45
So I feel like they probably end up in the same place. I don’t think it would have worked as well with 5.6 Sol or Opus. I’m sure I would have gotten something out of it.
23:54
But if it had been, you know, this, I don’t know, it just felt like it was really, a really like well-designed project too, where, you know, it’s like, it’s just writing up what it did.
24:03
And it’s like the core was like well-architected. Like, and we’re talking about like a text, little text editor, right? So this is not like some crazy amount of complexity or whatever.
24:14
It’s not a lot of states in this thing. But it just did a really good job of extracting out a core, cleaning that up a little bit. And I don’t, I mean, again, this is, I’m just saying this.
24:25
I haven’t actually validated it. Maybe 5.6 Sol would have done as good a job of it. Maybe it would have taken much longer or whatever. It would have actually been more expensive.
24:34
I don’t know. But I was certainly very happy with what Astra did there.
24:39
Yeah. Yeah, it feels like this model launch, both Fable 5.1 and Astra, feels a bit different. And I think that’s also one reason why we wanted to jump on a call, because there were model releases in the last, I mean, last AI years, that were sometimes more exciting and sometimes less exciting.
24:58
And it feels right now that it’s on the more exciting point. And I’m also curious how all of this will play out in, I don’t know, two to three weeks or something, if we’re still excited about Astra or if it’s kind of like, yeah, it lost its thunder a bit.
25:15
Yeah, there’s definitely that effect. I think you said that earlier, Philipp, where it’s so exciting. There’s a new model and it does things a little bit differently.
25:23
And it’s like, oh, this is the best thing. And then working with it day in and out or a couple of weeks, it’s like, oh, no, it has its sort of attractor states where you end up in, like, It just keeps on using the same phrases and stuff, like we now recognize because we work with Claude Opus a lot in our company infrastructure.
25:39
And so I wonder if I’ll get to that point, too. One thing I like a lot about how OpenAI does stuff, just their developer relations.
25:49
Right now, they feel like they’re more on our side. And this is totally going to change when it’s opportune for them or whatever. But they’re very generous with their usage resets.
25:58
You can actually use like your Codex subscription and stuff with other harnesses, which generally, my understanding is, Anthropic does not allow. So I have like, I used to have 5.6 Sol running in the Hermes harness.
26:15
Being able to now just go and upgrade that to Astra was amazing. And generally, Anthropic doesn’t allow that kind of thing. It’s like, oh, Fable is like in its own class and it’s like separate billing.
26:26
Right. I think you said that earlier, Martin. And you can’t use your own harness. And there’s just so many more restrictions. Also on the security and safety side.
26:36
I mean, for better or for worse, right? Fable seems to refuse security-related work a lot earlier. And then you’re falling back. It actually puts you back to Opus 4.8 for whatever reason.
26:47
And it’s like all of a sudden half your brain disappeared, right? And with Astra, I haven’t really encountered that yet. And so I had to do a bug sweep and some stuff.
26:57
I’ve actually found like some, nothing crazy, but like valid bugs that definitely I hadn’t noticed before. Like, well, I haven’t again, I haven’t repeated this with Fable exactly.
27:06
But generally what happens with Fable, it’s like, okay, safety classifier kicks in after a couple of turns, and then you get whatever Opus 4.8 is going to find, which is not that useful a comparison to Fable.
27:18
And so I like that a lot. Like, I like how OpenAI is handling sort of that developer relations piece. I mean, we’re all benefiting from it as customers. I agree with that.
27:29
But I also find it interesting, like the personal feeling, like towards these weird companies, like how it also shifts. Like one month Anthropic is like the good guy and then OpenAI catches up.
27:42
Same thing with the models.
27:44
It’s nice that we are independent and we can switch and mix and change depending on how we’re feeling about it. And I feel like, yeah, currently it feels a bit more like an OpenAI Astra moment, but it’s also, to me at least, reassuring.
28:00
Like, it’s nothing like now everything is OpenAI only, but it’s like, yeah, let’s see how Fable 5.1 catches up with things or how it does in the day-to-day.
28:09
And we can be very opportunistic about that. I mean, totally. And it does feel like these model releases are, I mean, they’re just happening more frequently and they’re more impactful.
28:21
Like I used to be, you know, I used to get so excited about like Apple, like the developer conferences and all this stuff, you know, because it was like every year it’s like game-changing new hardware.
28:31
It’s like, oh my God, like this laptop is like the best, I have to get it, and like try and, you know, make enough money to afford it. Or like this phone, it’s so amazing.
28:39
It’s like every year there was like some massive exciting drop. And that hasn’t really happened in recent, I mean, maybe it’s also, you know, I just turned 40, so you kind of grow out of it a little bit, I guess, too.
28:50
But I haven’t really had that experience as much recently with other technology, and then with these AI models, It’s like I’m getting that same feeling. I mean, quarterly or almost like monthly or potentially weekly at this point where it’s like, oh, wow, like this is some new thing that has like really exciting new capabilities.
29:09
Like, what am I going to do with this? Right. It’s just the sense of like something new has been unlocked. And it’s like the same price as it was before. And I can like immediately use it.
29:19
That to me is like super exciting. So it’ll be really interesting to see, you know, what’s going to happen with like Astra 6.1 and like what is like Anthropic’s next move.
29:27
So yeah, to wrap things up, maybe you have some take on that. Like what would you recommend for our viewership? Do you feel then they should like try it out or like, is it kind of like, it’s not for you, it’s kind of like it is a big leap.
29:42
At the very least, if coding is your career, I think you want to stay in touch with these jumps in capability. Because then you understand, okay, this class of model, you know, working with that, here’s what it can do in general.
29:53
And then the differences between the models are maybe not as big, you know. Also pricing-wise, there’s lots of competition, so they’re going to be similar.
30:01
And so I, yeah, I would encourage everyone, you know. And it’s, for some people it’s fun. Like for me, it’s a lot of fun, like trying out the new models, and I get very excited about it.
30:10
And so for that reason, I would also recommend it. But at the very least, it’s like, yeah, every time there’s a big release, it’s worth trying it out and sort of just getting a sense of like, you know, what is happening there.
30:20
With the point releases, you know, usually it’s not that big of a difference. Yeah, a lot of us are probably used to, you know, like some other release cadence of things, where, you know, software or whatever, you know, I don’t know, Python improves a little bit with each release.
30:37
And that’s maybe more like the speed that you’re used to. And it’s like, it’s okay to not try Python 3.14 for a year and then check in or something. You know, you’ll still get most of the same things.
30:50
But that’s just not true with a lot of this AI model releases. Like, a new release can make a gigantic difference in what ends up and what’s actually even possible to do with this.
31:01
And we all, because we’re all humans, wired up with some sort of a, you know, this is how quick progress happens, you know, that we’ve experienced throughout our lifetimes.
31:12
It’s like very easy to fall into this trap that you’re like, okay, well, maybe I’ll try in a couple of months or yeah, it’s probably going to be a little bit better.
31:20
But then this development is so quick and I don’t know, exponential maybe, I don’t know, but like they certainly feel like it. So it’s like any sort of beliefs you hold on to for like, oh yeah, this AI model is just going to produce garbage code for me or whatever.
31:37
And it’s just going to be very much outdated in a short amount of time.
31:42
Totally, very short amount of time, right? I think every time I hear from someone like, oh, you know, I tried like AI, but it made like extra fingers in the photos that it made for me or something.
31:52
It’s like this stuff changes so quickly. You really, at least on a like weekly or monthly cadence, if you want to keep up with it, which again, I would recommend to anyone who’s, I mean, anyone period, I would say.
32:07
But we’re talking specifically, I think, with regards to developers, Python developers, you know, technical people.
32:16
It’s like it is changing so fast and developing so fast. It’s, yeah, really the time, the timelines there for significant changes are on the order of like month, if not weeks now.
32:29
And it is, I mean, it’s the flip side is like it’s overwhelming. Kind of the cool, interesting side is that, or fascinating side is that, I mean, when have we had that where on a monthly basis you have technology that unlocks like new use cases and all of this cool stuff that all of a sudden we can do.
32:46
And, you know, we’re gonna release some cool new product features there soon on Real Python that weren’t really possible or economically possible for us with previous models.
32:56
And so, you know, we’ll do a dedicated video on that and stuff too, but it’s a very exciting time. And, well, the downside is just RAM and everything gets more and more expensive.
33:06
So goodbye gaming PCs. Goodbye Steam Frame and whatnot. But, you know, the models are pretty cool. So it’s been fun playing with those. I agree with both of you.
33:16
I feel if you haven’t played around with AI in a while, I think now is a good moment again to give it a try. If all of what we were talking about really didn’t make much sense for you, check out realpython.com/plan.
33:30
There you can create your own learning plan, and we will pick you up where you are. And if you want to learn about AI, we have the right tutorials and video courses for you.
33:39
And if you have watched so far in this video, I will also give you a quick hint that we have something very special that we will share with you in one of our next videos.
33:49
So make sure to subscribe. If you have any questions, let us know in the comments and see you next time. Bye-bye.
