WEBVTT

00:00:00.000 --> 00:00:05.600
Hello! A new model has sparked widespread discussion over the last couple of days, named JF.

00:00:05.600 --> 00:00:14.240
A former OpenAI researcher spent two years developing it after leaving, creating a new model trained in a way that differs from the past.

00:00:14.240 --> 00:00:23.600
I think its features are very distinct and powerful. For specific tasks, it can run at a fraction of the cost and hundreds of times the speed of traditional models,

00:00:23.600 --> 00:00:29.440
with very low hallucination rates, generating the answers you need, which is highly significant.

00:00:29.440 --> 00:00:34.320
它是重要的,因為它選了完全不同的一路由我們現代的AI,

00:00:34.320 --> 00:00:36.320
它是在原來的N-Token Prediction。

00:00:36.320 --> 00:00:39.840
當大家是比拼了強硬的解決,

00:00:39.840 --> 00:00:44.480
硬的解決,和能力的解決方法,

00:00:44.480 --> 00:00:47.360
它其實是很小的地步。

00:00:47.360 --> 00:00:49.600
我們不需要等待太多的信息,

00:00:49.600 --> 00:00:54.960
或多多的解決,我們只需要一個比較的測試,

00:00:54.960 --> 00:00:57.040
或是個比例的不同的結果。

00:00:57.040 --> 00:00:59.600
我给你一个例子

00:00:59.600 --> 00:01:03.180
例如,在一个公司的行业,

00:01:03.180 --> 00:01:05.480
如果我们使用了AI的相机,

00:01:05.480 --> 00:01:07.140
我们必须要进入到的模特,

00:01:07.140 --> 00:01:08.880
等待在模特上的模特,

00:01:08.880 --> 00:01:10.980
这些时间需要几秒,对吗?

00:01:10.980 --> 00:01:12.100
这一终的效果会给你,

00:01:12.100 --> 00:01:14.100
你会在哪个部份接触,

00:01:14.100 --> 00:01:15.200
谁要跟他在谈论,

00:01:15.200 --> 00:01:18.100
或者是指出的优质是低的,

00:01:18.100 --> 00:01:19.380
需要更加注意力。

00:01:19.380 --> 00:01:20.880
它给你一个样的效果。

00:01:20.880 --> 00:01:23.100
在实际上,用这个新模特,

00:01:23.100 --> 00:01:24.900
它可以非常快速,

00:01:24.900 --> 00:01:28.900
这几个百分钟的几个百分钟的几个百分钟

00:01:28.900 --> 00:01:30.900
这几个百分钟的几个百分钟

00:01:30.900 --> 00:01:32.900
这几个百分钟的几个百分钟

00:01:32.900 --> 00:01:38.180
the billing department,estimatesthe customer churn rate,and assessesthe likelihoodofneeding

00:01:38.180 --> 00:01:41.580
human intervention.

00:01:41.580 --> 00:01:47.900
Once theseassessmentsare made,I canassignthe tasktoaspecificdepartmentorouteittoahumanagent,

00:01:47.900 --> 00:01:54.520
allowingme tomakethecorrespondingadjustmentsanddecisions.So itsefficiency andcosteffectivenessareincrediblyimpressive.It

00:01:54.520 --> 00:02:02.740
isjustextremely,extremelyfast.Even initslivedemonstrationofstructuredoutput,youcanclearlyseethatitgeneratesresults

00:02:02.740 --> 00:02:05.080
但是传统的 models

00:02:05.080 --> 00:02:06.320
也会一直在用评论

00:02:06.320 --> 00:02:08.180
词为文字的评论

00:02:08.180 --> 00:02:10.340
所以这两个技术的 approaches

00:02:10.340 --> 00:02:12.220
其实完全不同

00:02:12.220 --> 00:02:13.960
当 the founder released this video

00:02:13.960 --> 00:02:15.140
他说了一篇,

00:02:15.140 --> 00:02:17.920
他说了一篇,

00:02:17.920 --> 00:02:20.060
我们的认识,

00:02:20.060 --> 00:02:22.060
其实是一篇的答案

00:02:22.060 --> 00:02:23.860
合适的答案

00:02:23.860 --> 00:02:25.820
不太不太确定,

00:02:25.820 --> 00:02:27.060
我们喜欢他们,

00:02:27.060 --> 00:02:27.640
对吗?

00:02:27.640 --> 00:02:28.540
例如,

00:02:28.540 --> 00:02:30.200
他们可能会更多 empathetic

00:02:30.200 --> 00:02:32.100
或是更多理解我们的

00:02:32.100 --> 00:02:35.300
但是,他們的問題是他們的解答問題

00:02:35.300 --> 00:02:36.800
人們不可能會不喜歡,

00:02:36.800 --> 00:02:39.700
甚至是有個方法的方法。

00:02:39.700 --> 00:02:44.100
這帶來更加盡量的信息和 consistency。

00:02:44.100 --> 00:02:47.400
但是,他們的解答案是叫做RLCD,

00:02:47.400 --> 00:02:50.800
名為 Reinforcement Learning from Calibrated Decisions。

00:02:50.800 --> 00:02:52.800
它指的是,在 Reinforcement Learning,

00:02:52.800 --> 00:02:54.600
我實在是在調整的解答案是

00:02:54.600 --> 00:02:56.600
的解答案是一個解答案。

00:02:56.600 --> 00:02:59.800
解答案是每次它提供答案,

00:02:59.800 --> 00:03:01.600
它提供一個 probability。

00:03:01.600 --> 00:03:02.760
A confidence score

00:03:02.760 --> 00:03:05.160
For instance, with this billing inquiry

00:03:05.160 --> 00:03:09.760
the probability that I should route it to the billing department is about 90%

00:03:09.760 --> 00:03:12.360
But if the customer also mentioned other issues

00:03:12.360 --> 00:03:16.360
I believe there's a certain probability it should be prioritized for other departments

00:03:16.360 --> 00:03:20.160
So I might assign a 10% or 20% probability to those as well

00:03:20.160 --> 00:03:22.960
At this point, you can set a threshold

00:03:22.960 --> 00:03:26.360
For example, if the churn risk is very high

00:03:26.360 --> 00:03:28.760
I can have a human intervene directly

00:03:28.760 --> 00:03:32.760
如果它是低,我可以让它 AI model handle the response.

00:03:32.760 --> 00:03:36.760
所以,它帮助我们去解决哪个性能量的权量使用

00:03:36.760 --> 00:03:39.520
来解决于不同的情况下。

00:03:39.520 --> 00:03:44.240
所以,我相信这个模型会将将其进行我们的生活生活

00:03:44.240 --> 00:03:46.360
和很多不同的专业。

00:03:46.360 --> 00:03:50.200
以前我们曾经用了一枪创造的创造,

00:03:50.200 --> 00:03:53.000
或者我们不需要太多的时间。

00:03:53.000 --> 00:03:57.480
但这些随意的随意的选择是我们的选择。

00:03:57.480 --> 00:04:02.600
但是在这个新的模式中,我们的选择和认证能更加更加更加适合。

00:04:02.600 --> 00:04:06.680
如果我们说我们的选择这个进入我们的页务,

00:04:06.680 --> 00:04:11.880
我们肯定能够用这个模式,因为现在,我们的页务员工程,

00:04:11.880 --> 00:04:15.640
他们其实会有很多认证和评价的评价。

00:04:15.640 --> 00:04:19.480
例如,如果我们的收入的收入我们的预期,

00:04:19.480 --> 00:04:21.720
我们还是需要这些认证,

00:04:21.720 --> 00:04:24.680
这会常常需要一个人的路程式的方法。

00:04:24.680 --> 00:04:26.600
但是如果我们有一个模式的模式,

00:04:26.600 --> 00:04:29.800
我们可以定义它的质量从开始

00:04:29.800 --> 00:04:34.080
即使是80%,60%,40%的距离,

00:04:34.080 --> 00:04:38.520
我们只需要在一个很低的40%的距离。

00:04:38.520 --> 00:04:41.080
如果它基本上是在结构的质量,

00:04:41.080 --> 00:04:43.480
我们不需要保持人的质量,

00:04:43.480 --> 00:04:46.600
我们不需要保持人的质量。

00:04:46.600 --> 00:04:50.040
所以,在这个将来的更多的模式,

00:04:50.040 --> 00:04:52.280
它可以更加大幅并且在这些模式上

00:04:52.280 --> 00:04:55.280
这个模式其实只发了一个新版本

00:04:55.280 --> 00:04:59.040
没有提供的技术和计程度

00:04:59.040 --> 00:05:01.960
我觉得这个方法非常值得考虑

00:05:01.960 --> 00:05:04.160
它使用一个系统一式的方法

00:05:04.160 --> 00:05:05.640
是快思考

00:05:05.640 --> 00:05:06.880
Daniel Kahneman的作词

00:05:06.880 --> 00:05:09.280
说一大部分人的认识

00:05:09.280 --> 00:05:11.360
不在实际上的认识

00:05:11.360 --> 00:05:12.720
我们立刻的认识

00:05:12.720 --> 00:05:14.640
我们立刻的认识

00:05:14.640 --> 00:05:17.440
例如,选择哪个衣服

00:05:17.440 --> 00:05:18.880
很自然的衣服

00:05:18.880 --> 00:05:21.440
我只穿着衣服

00:05:21.440 --> 00:05:24.200
and when I'm going to work I might wear a different pair.

00:05:24.200 --> 00:05:26.940
So this judgment is actually very intuitive.

00:05:26.940 --> 00:05:29.280
It's the same for AI models.

00:05:29.280 --> 00:05:34.040
You don't actually need to generate or think through every single judgment

00:05:34.040 --> 00:05:36.200
or use chain of thought reasoning.

00:05:36.200 --> 00:05:40.200
I can use a smaller model or a model designed for fast thinking

00:05:40.200 --> 00:05:41.680
with parallel computing

00:05:41.680 --> 00:05:44.680
to accomplish our tasks and goals very efficiently.

00:05:44.680 --> 00:05:48.600
So I think it might really make its way into our society soon

00:05:48.600 --> 00:05:50.700
because it's cheap, right?

00:05:50.700 --> 00:05:53.780
而且是最高的,因為它是很高的,

00:05:53.780 --> 00:05:57.100
在0.042萬元的價值,

00:05:57.100 --> 00:05:59.780
比 GPT-4的價值,可能是10萬元。

00:05:59.780 --> 00:06:03.300
價值是這麼高,他們現在不在賣現在的價值。

00:06:03.300 --> 00:06:07.820
所以它可以使用在一大多數的評測 scenarios。

00:06:07.820 --> 00:06:10.740
我們看到每個小型的手法

00:06:10.740 --> 00:06:13.420
可能有一個決定的模式背後。

00:06:13.420 --> 00:06:18.580
以前,一名 classifier would have required someone to spend ages training it, right?

00:06:18.580 --> 00:06:22.740
我們以前要有的要求,但現在是沒有必要的

00:06:22.740 --> 00:06:25.780
現在,這 model可以幫我們做了一切的評判,

00:06:25.780 --> 00:06:28.660
決定定 specific outcomes based on the context

00:06:28.660 --> 00:06:34.460
這代表,這代表,已經融化了更多的技術和考驗

00:06:34.460 --> 00:06:38.860
我們會看到這個功能伸展開更多的使用

00:06:38.860 --> 00:06:42.620
是的,我认为这个是非常有趣的新方式,

00:06:42.620 --> 00:06:46.380
一遍完全不同的方法使用力

00:06:46.380 --> 00:06:47.980
我們之前說明,

00:06:47.980 --> 00:06:50.700
未來可能會有小型的模式

00:06:50.700 --> 00:06:52.420
我們只需要有小型的模式

00:06:52.420 --> 00:06:54.500
提供的模式

00:06:54.500 --> 00:06:56.500
而是用販的模式

00:06:56.500 --> 00:06:57.380
每個都市場的模式

00:06:57.380 --> 00:06:59.620
我們最近的書

00:06:59.620 --> 00:07:00.940
我們最近的書

00:07:00.940 --> 00:07:02.540
也有同一個同意

00:07:02.540 --> 00:07:05.060
成為我們目標的最適合的方法

00:07:05.060 --> 00:07:06.500
是完全不平的

00:07:06.500 --> 00:07:11.180
所以我們想分享這些 JB的模式

00:07:11.180 --> 00:07:15.360
They've opened their waitlist, so if you'd like to try them out, feel free to join.

00:07:15.820 --> 00:07:17.680
Have a great weekend. Bye!