WEBVTT

00:00:00.140 --> 00:00:18.660
Speaker 0: So last week I was on vacation, so i only watched the launch of the GPT-Six Astra model from afar with some delays since on X or in the internet, but of course the launch followed a typical pattern.

00:00:18.760 --> 00:00:40.220
Speaker 0: We had posts by people hyping up the model and then after a couple days later we had more sober or maybe disappointed posts by people that shared.

00:00:40.960 --> 00:00:44.860
Speaker 0: the model seemed overhyped and isn't amazing.

00:00:45.240 --> 00:00:51.500
Speaker 0: Now this week I also got my hands on it, where do i start?

00:00:51.660 --> 00:00:53.360
Speaker 0: The model is really great at.

00:00:55.480 --> 00:01:00.240
Speaker 0: I don't know about you, but my timeline has been drowning in blender.

00:01:00.440 --> 00:01:04.920
Speaker 0: Three D art and games created by Astra.

00:01:05.160 --> 00:01:22.100
Speaker 0: now it's hard to verify if everything is truly generated by astra But i got some hints because i also played around with it And at least for the games there are Some recurring elements You can spot in these different wiped up games.

00:01:22.280 --> 00:01:31.480
Speaker 0: nonetheless Clearly, the three-D capabilities of Astra are really amazing especially compared to what models were able do before.

00:01:31.560 --> 00:01:36.780
Speaker 0: So clearly it has been heavily fine tuned and optimized for better at that kind.

00:01:39.300 --> 00:01:45.300
Speaker 0: three D stuff in the post training data and a fine tuning process, I would guess.

00:01:45.660 --> 00:01:47.080
Speaker 0: But what about coding?

00:01:47.180 --> 00:02:02.640
Speaker 0: That is the part i personally care most apart About an in this launch blog post we of course also have Coding related benchmarks And here Astra Also does really Really well.

00:02:03.140 --> 00:02:09.080
Speaker 0: now i will say that benchmarks are one thing you know, we had in all these launches.

00:02:09.180 --> 00:02:12.320
Speaker 0: it is possible to optimize models.

00:02:12.400 --> 00:02:21.500
Speaker 0: To do well in benchmarks and of course your heavily incentivized as a lab or an AI lab make your model do well on benchmarks.

00:02:21.560 --> 00:02:25.040
Speaker 0: because imagine releasing the model which does not do well with benchmarks.

00:02:25.260 --> 00:02:30.720
Speaker 0: It wouldn't count if it did well in actual usage Because nobody would start using this in first place.

00:02:31.740 --> 00:02:35.280
Speaker 0: I don't blame AI labs for benchmarks.

00:02:35.320 --> 00:02:38.420
Speaker 0: You basically have to do it, but benchmarks they offer also.

00:02:39.440 --> 00:02:40.640
Speaker 0: Don't tell us that much.

00:02:41.020 --> 00:02:43.760
Speaker 0: It's a nice indicator But yeah you all know it.

00:02:44.020 --> 00:02:45.660
Speaker 0: Benchmarks are not everything.

00:02:46.740 --> 00:02:54.280
Speaker 0: It's more the vibe of using it and how it behaves when you actually start Using in your day-to-day work.

00:02:54.360 --> 00:03:02.020
Speaker 0: And here GPT-Six Astra For me at least feels a lot like GPT-Five point six sol.

00:03:02.420 --> 00:03:09.860
Speaker 0: It maybe has a different set of quirks and features, but in the end it's still a good coding model!

00:03:10.360 --> 00:03:18.820
Speaker 0: It is not horrible at least for me... But its also not ground breakingly better.

00:03:19.220 --> 00:03:23.040
Speaker 0: You still have to know what you're doing and what's your asking for.

00:03:23.120 --> 00:03:31.180
Speaker 0: And of course, you can also wipe code stuff that was possible before in the can get you quite far on four many tools or things.

00:03:31.240 --> 00:03:32.860
Speaker 0: you need it maybe all unique.

00:03:33.020 --> 00:03:49.120
Speaker 0: You don't need to have a perfect code base for every little project or internal tool unit but if you do care about your codebase architecture building something plan maintaining up one It's the same as with previous models.

00:03:49.460 --> 00:03:51.680
Speaker 0: You need... to have proper constraints.

00:03:51.760 --> 00:03:54.840
Speaker 0: You need to know what you're doing, What's your asking for?

00:03:55.020 --> 00:04:01.400
Speaker 0: You need To Know how to work with such a model and then you can get good results out of Astra.

00:04:01.520 --> 00:04:02.940
Speaker 0: that has been my experience.

00:04:03.300 --> 00:04:05.640
Speaker 0: it still Can lead to ugly code.

00:04:05.680 --> 00:04:11.060
Speaker 0: It Still Can Lead to unnecessarily complex code though at least in That respect.

00:04:11.140 --> 00:04:22.780
Speaker 0: For me it tended A bit less to add layer over Layer Of complexity And fallback Code here and the legacy compatible code here, which GPT-Five point six sold it a lot.

00:04:23.440 --> 00:04:34.800
Speaker 0: It always treated every piece of software as if it were big well established enterprise software that needs to be maintained and that existed for ten years but... That's the different story.

00:04:35.340 --> 00:04:45.700
Speaker 0: Astra is better there for me But still you need to steer and control And I don't know when will change.

00:04:46.320 --> 00:04:49.100
Speaker 0: Clearly these models are getting better and clearly.

00:04:49.200 --> 00:04:52.420
Speaker 0: These models are good at following instructions.

00:04:52.560 --> 00:04:58.040
Speaker 0: I mean, it's easy to forget this but a year ago we had models that stopped early.

00:04:58.500 --> 00:04:59.820
Speaker 0: That didn't do that.

00:04:59.980 --> 00:05:00.200
Speaker 0: well.

00:05:00.320 --> 00:05:04.860
Speaker 0: when it came too follow instructions that often went astray.

00:05:05.000 --> 00:05:09.080
Speaker 0: and whilst this still happens today, they're of course much better.

00:05:09.220 --> 00:05:13.360
Speaker 0: They can work on their own or not task for a very long time frame.

00:05:13.460 --> 00:05:23.540
Speaker 0: you just want to make sure that you define the tasks such as have guidelines in place so it doesn't wander off but are much better there!

00:05:23.600 --> 00:05:34.200
Speaker 0: And especially if you give them something it can evaluate It can really dig into a task and keep on going.

00:05:34.300 --> 00:05:35.900
Speaker 0: And ultimately solve it.

00:05:36.400 --> 00:05:40.420
Speaker 0: maybe not in the most beautiful way, but that again doesn't count for every task.

00:05:40.600 --> 00:05:44.520
Speaker 0: And depending on your setup ,it can also solve of course in good ways.

00:05:44.600 --> 00:05:47.160
Speaker 0: so GPT-Six Astra is a good model .

00:05:47.440 --> 00:05:51.020
Speaker 0: But it's also not The last model will ever need an.

00:05:51.120 --> 00:05:53.780
Speaker 0: I mean That shouldn't come as a surprise.

00:05:53.820 --> 00:06:04.060
Speaker 0: On the other hand i think its just also the current meta in which we are, where all these labs or especially OpenAI Anthropic.

00:06:04.160 --> 00:06:11.700
Speaker 0: in the end they hype up their models so much when they're released and probably have to please investors.

00:06:11.760 --> 00:06:13.940
Speaker 0: And get the traction that you want.

00:06:14.300 --> 00:06:29.660
Speaker 0: for them to keep people subscribed They have to hype up everything so much, that you just... You HAVE TO be disappointed at some point because nothing can live up to the type.

00:06:30.020 --> 00:06:32.380
Speaker 0: I'm a gamer and you know this same from video games!

00:06:32.480 --> 00:06:35.900
Speaker 0: Some games are just hyped-up THAT MUCH, GTA VI maybe?

00:06:36.420 --> 00:06:45.980
Speaker 0: Well i think it will be good game but we have games that are hyped up SO MUCH That when they're actually released It's underwhelming or it just doesn't live up to the hype.

00:06:46.040 --> 00:06:48.680
Speaker 0: And I think we're constantly seeing that with these models.

00:06:49.000 --> 00:07:00.920
Speaker 0: now, of course combined With The fact That We Get New Models Almost Every Week It's Really Exhausting Especially Also Combined With The Fact That We're Talking About Our Future Here!

00:07:01.020 --> 00:07:05.720
Speaker 0: WE ALL DON'T KNOW WHAT WE'RE GOING TO DO IN A FEW YEARS!

00:07:05.940 --> 00:07:08.020
Speaker 0: WE DONT KNOW WHERE OUR PROPHESION IS HEADING.

00:07:08.080 --> 00:07:11.740
Speaker 0: AND THAT'S NOT JUST TRUE FOR PROGRAMMING IT'S TRUE for ALL WHITE COLOR WORK.

00:07:12.640 --> 00:07:14.160
Speaker 0: We don't know where we're heading.

00:07:37.800 --> 00:07:40.460
Speaker 0: These are my two cents after being back from vacation.

00:07:40.520 --> 00:07:44.740
Speaker 0: It's a good model, it is much better at three D work for coding.

00:07:45.420 --> 00:07:46.140
Speaker 0: just the good model.

00:07:46.780 --> 00:07:46.900
Speaker 0: use.

00:07:47.100 --> 00:07:52.140
Speaker 0: learn how to work with AI properly and I will soon have more share on that.