WEBVTT 00:00.000 --> 00:01.500 -: Hey, and in this video you're gonna learn 00:01.500 --> 00:03.210 about some of the core features 00:03.210 --> 00:05.400 that OpenAI provides you with as a developer 00:05.400 --> 00:06.630 so that you get an awareness 00:06.630 --> 00:08.520 about all the different types of capabilities 00:08.520 --> 00:11.370 that you're gonna be able to do when using this API. 00:11.370 --> 00:14.520 For example, some of the core language capabilities 00:14.520 --> 00:15.960 are text generations, 00:15.960 --> 00:18.810 so OpenAI can easily generate text, 00:18.810 --> 00:21.270 you can use the responses API 00:21.270 --> 00:22.830 like you can see on the left here, 00:22.830 --> 00:26.190 and the models range from lightweight to advanced. 00:26.190 --> 00:27.570 You can control parameters 00:27.570 --> 00:29.640 like we were showing you earlier in the Playground, 00:29.640 --> 00:31.110 such as the temperature, 00:31.110 --> 00:33.870 the maximum number of tokens that can be generated, 00:33.870 --> 00:35.490 and also the top-p. 00:35.490 --> 00:37.110 There's also content filtering 00:37.110 --> 00:38.910 and safety measures that are built in. 00:38.910 --> 00:42.090 Secondly, you can extract structured outputs from OpenAI. 00:42.090 --> 00:45.030 This allows you to get predictable JSON response. 00:45.030 --> 00:47.580 This is perfect for doing data extraction 00:47.580 --> 00:49.740 and structured information retrieval, 00:49.740 --> 00:52.530 also enforces the output format 00:52.530 --> 00:54.930 and validates that it is in that structure. 00:54.930 --> 00:57.960 This reduces the post-processing effort that you have to do 00:57.960 --> 01:00.540 and it basically means you can get structured data 01:00.540 --> 01:02.130 from just about anything. 01:02.130 --> 01:04.770 Also, you can use ChatGPT's API 01:04.770 --> 01:08.370 or OpenAI's API to analyze images. 01:08.370 --> 01:11.730 You can provide a prompt with the image URL 01:11.730 --> 01:15.660 or attach the image as Base64 encoded string, 01:15.660 --> 01:18.060 and then you can basically get an output back 01:18.060 --> 01:21.270 directly from OpenAI about that image. 01:21.270 --> 01:22.770 You can also generate images 01:22.770 --> 01:27.090 with OpenAI's API using the dall-e-3 model at the moment, 01:27.090 --> 01:28.440 and you can put in a prompt 01:28.440 --> 01:30.840 and the number of generations you want with n 1 01:30.840 --> 01:33.120 and the size as well of the image. 01:33.120 --> 01:35.850 OpenAI also offers text-to-speech, 01:35.850 --> 01:38.070 so generating real life-like speech 01:38.070 --> 01:39.810 directly from a textual prompt 01:39.810 --> 01:42.663 and also speech-to-text, i.e. transcription. 01:43.950 --> 01:47.190 There are some more advanced features of OpenAI's API, 01:47.190 --> 01:49.680 such as function calling or tools, 01:49.680 --> 01:52.200 and you can basically provide a agent 01:52.200 --> 01:54.240 with a get_weather function, 01:54.240 --> 01:56.580 then the agent is capable of understanding 01:56.580 --> 01:58.200 when to call that tool. 01:58.200 --> 02:00.990 And you can see, for example, in the prompt on the left, 02:00.990 --> 02:02.977 we have this query from the user, 02:02.977 --> 02:05.340 "What's the weather like in Paris today?" 02:05.340 --> 02:07.710 That will likely lead the AI choosing 02:07.710 --> 02:11.580 to use a get_weather tool to answer that user's query. 02:11.580 --> 02:14.460 There are more recent models, which are reasoning models, 02:14.460 --> 02:18.000 these are very good for solving complex problems 02:18.000 --> 02:20.490 and I would recommend having a look at the newer models. 02:20.490 --> 02:22.590 So o3 is the most performant model 02:22.590 --> 02:24.210 at this time of recording. 02:24.210 --> 02:25.590 There is also o4-mini, 02:25.590 --> 02:28.050 which is slightly more cost-efficient. 02:28.050 --> 02:30.330 You have a parameter that allows you 02:30.330 --> 02:33.780 to see on the left-hand side here the ability to choose 02:33.780 --> 02:36.840 how much reasoning effort should be gone into. 02:36.840 --> 02:38.730 There is also embeddings, 02:38.730 --> 02:43.260 which is a way to represent words in a numerical format. 02:43.260 --> 02:47.340 This is useful for semantic search and similarity use cases. 02:47.340 --> 02:49.807 And you can see on the left here, we can take an input, 02:49.807 --> 02:52.740 "The quick brown fox jumps over the lazy dog," 02:52.740 --> 02:55.680 and then basically we'll produce a series of numbers, 02:55.680 --> 02:57.570 also known as an embedding. 02:57.570 --> 02:58.920 Some of the different types of ways 02:58.920 --> 03:00.270 that you might like to use this 03:00.270 --> 03:02.220 is similarity matching, 03:02.220 --> 03:04.350 dynamically bringing content back, 03:04.350 --> 03:06.000 that's similar to a query, 03:06.000 --> 03:09.030 clustering and classification, knowledge bases, 03:09.030 --> 03:12.450 and RAG, also known as retrieval augmented generation. 03:12.450 --> 03:13.950 There are lots of different things you can do 03:13.950 --> 03:16.410 inside of OpenAI beside the core features, 03:16.410 --> 03:17.880 such as fine-tuning, 03:17.880 --> 03:20.580 so training your own models on your specific data, 03:20.580 --> 03:23.700 running evaluations, or commonly known as evals, 03:23.700 --> 03:26.550 systematically evaluating your model performance, 03:26.550 --> 03:28.020 and also distillation, 03:28.020 --> 03:30.630 so creating smaller, more specialized models, 03:30.630 --> 03:34.260 which helps to reduce the inference cost and the latency. 03:34.260 --> 03:36.060 There's also a bunch of developer resources 03:36.060 --> 03:37.830 that are put into this video, 03:37.830 --> 03:39.870 so feel free to have a look at that 03:39.870 --> 03:41.460 in your journey with OpenAI. 03:41.460 --> 03:42.780 In the next video, we're gonna have a look 03:42.780 --> 03:44.310 at how you can get started. 03:44.310 --> 03:47.160 So creating an account, creating an API key, 03:47.160 --> 03:48.810 and then as well as that, you'll also need 03:48.810 --> 03:52.950 to install the SDK with this pip install openai command. 03:52.950 --> 03:54.800 Cool, I'll see you in the next video.