0:00.000–0:00.500
enHello!
0:00.500–0:01.818
enA new model has sparked
0:01.818–0:03.273
enwidespread discussion
0:03.273–0:05.018
enover the last couple of days,
0:05.018–0:05.600
ennamed JF.
0:05.600–0:06.585
enA former OpenAI
0:06.585–0:07.949
enresearcher spent two
0:07.949–0:10.223
enyears developing it after leaving,
0:10.223–0:11.512
encreating a new model
0:11.512–0:12.800
entrained in a way that
0:12.800–0:14.240
endiffers from the past.
0:14.240–0:15.627
enI think its features are
0:15.627–0:17.291
envery distinct and powerful.
0:17.291–0:18.469
enFor specific tasks,
0:18.469–0:19.787
enit can run at a fraction
0:19.787–0:21.173
enof the cost and hundreds
0:21.173–0:22.352
enof times the speed of
0:22.352–0:23.600
entraditional models,
0:23.600–0:25.711
enwith very low hallucination rates,
0:25.711–0:27.681
engenerating the answers you need,
0:27.681–0:29.440
enwhich is highly significant.
0:29.440–0:34.320
zh它是重要的,因為它選了完全不同的一路由我們現代的AI,
0:34.320–0:36.320
zh它是在原來的N-Token Prediction。
0:36.320–0:39.840
zh當大家是比拼了強硬的解決,
0:39.840–0:44.480
zh硬的解決,和能力的解決方法,
0:44.480–0:47.360
zh它其實是很小的地步。
0:47.360–0:49.600
zh我們不需要等待太多的信息,
0:49.600–0:54.960
zh或多多的解決,我們只需要一個比較的測試,
0:54.960–0:57.040
zh或是個比例的不同的結果。
0:57.040–0:59.600
zh我给你一个例子
0:59.600–1:03.180
zh例如,在一个公司的行业,
1:03.180–1:05.480
zh如果我们使用了AI的相机,
1:05.480–1:07.140
zh我们必须要进入到的模特,
1:07.140–1:08.880
zh等待在模特上的模特,
1:08.880–1:10.980
zh这些时间需要几秒,对吗?
1:10.980–1:12.100
zh这一终的效果会给你,
1:12.100–1:14.100
zh你会在哪个部份接触,
1:14.100–1:15.200
zh谁要跟他在谈论,
1:15.200–1:18.100
zh或者是指出的优质是低的,
1:18.100–1:19.380
zh需要更加注意力。
1:19.380–1:20.880
zh它给你一个样的效果。
1:20.880–1:23.100
zh在实际上,用这个新模特,
1:23.100–1:24.900
zh它可以非常快速,
1:24.900–1:28.900
zh这几个百分钟的几个百分钟的几个百分钟
1:28.900–1:30.900
zh这几个百分钟的几个百分钟
1:30.900–1:32.900
zh这几个百分钟的几个百分钟
1:32.900–1:34.300
enthe billing department,
1:34.300–1:36.300
enestimatesthe customer churn rate,
1:36.300–1:38.500
enand assessesthe likelihoodofneeding
1:38.500–1:41.580
enhuman intervention.
1:41.580–1:43.591
enOnce theseassessmentsare made,
1:43.591–1:44.525
enI canassignthe
1:44.525–1:47.900
entasktoaspecificdepartmentorouteittoahumanagent,
1:47.900–1:48.567
enallowingme
1:48.567–1:51.633
entomakethecorrespondingadjustmentsanddecisions.
1:51.633–1:52.633
enSo itsefficiency
1:52.633–1:55.700
enandcosteffectivenessareincrediblyimpressive.It
1:55.700–1:56.767
enisjustextremely,
1:56.767–1:57.700
enextremelyfast.
1:57.700–1:58.200
enEven
1:58.200–2:00.933
eninitslivedemonstrationofstructuredoutput,
2:00.933–2:03.467
enyoucanclearlyseethatitgeneratesresults
2:03.467–2:05.080
zh但是传统的 models
2:05.080–2:06.320
zh也会一直在用评论
2:06.320–2:08.180
zh词为文字的评论
2:08.180–2:10.340
zh所以这两个技术的 approaches
2:10.340–2:12.220
zh其实完全不同
2:12.220–2:14.087
zh当 the founder released this video
2:14.087–2:15.140
zh他说了一篇,
2:15.140–2:17.920
zh他说了一篇,
2:17.920–2:20.060
zh我们的认识,
2:20.060–2:22.060
zh其实是一篇的答案
2:22.060–2:23.860
zh合适的答案
2:23.860–2:25.820
zh不太不太确定,
2:25.820–2:27.060
zh我们喜欢他们,
2:27.060–2:27.640
zh对吗?
2:27.640–2:28.540
zh例如,
2:28.540–2:30.200
zh他们可能会更多 empathetic
2:30.200–2:32.100
zh或是更多理解我们的
2:32.100–2:35.300
zh但是,他們的問題是他們的解答問題
2:35.300–2:36.800
zh人們不可能會不喜歡,
2:36.800–2:39.700
zh甚至是有個方法的方法。
2:39.700–2:44.100
zh這帶來更加盡量的信息和 consistency。
2:44.100–2:47.400
zh但是,他們的解答案是叫做RLCD,
2:47.400–2:49.064
zh名為 Reinforcement Learning
2:49.064–2:50.800
enfrom Calibrated Decisions。
2:50.800–2:52.800
zh它指的是,在 Reinforcement Learning,
2:52.800–2:54.600
zh我實在是在調整的解答案是
2:54.600–2:56.600
zh的解答案是一個解答案。
2:56.600–2:59.800
zh解答案是每次它提供答案,
2:59.800–3:01.600
zh它提供一個 probability。
3:01.600–3:02.760
enA confidence score
3:02.760–3:03.607
enFor instance,
3:03.607–3:05.160
enwith this billing inquiry
3:05.160–3:06.526
enthe probability that I
3:06.526–3:07.819
enshould route it to the
3:07.819–3:09.760
enbilling department is about 90%
3:09.760–3:11.093
enBut if the customer also
3:11.093–3:12.427
enmentioned other issues
3:12.427–3:13.493
enI believe there's a
3:13.493–3:14.827
encertain probability it
3:14.827–3:16.093
enshould be prioritized
3:16.093–3:17.360
enfor other departments
3:17.360–3:18.560
enSo I might assign a 10%
3:18.560–3:20.493
enor 20% probability to those as well
3:20.493–3:21.293
enAt this point,
3:21.293–3:22.960
enyou can set a threshold
3:22.960–3:24.029
enFor example,
3:24.029–3:26.360
enif the churn risk is very high
3:26.360–3:28.760
enI can have a human intervene directly
3:28.760–3:29.427
zh如果它是低,
3:29.427–3:32.760
zh我可以让它 AI model handle the response.
3:32.760–3:36.760
zh所以,它帮助我们去解决哪个性能量的权量使用
3:36.760–3:39.520
zh来解决于不同的情况下。
3:39.520–3:44.240
zh所以,我相信这个模型会将将其进行我们的生活生活
3:44.240–3:46.360
zh和很多不同的专业。
3:46.360–3:50.200
zh以前我们曾经用了一枪创造的创造,
3:50.200–3:53.000
zh或者我们不需要太多的时间。
3:53.000–3:57.480
zh但这些随意的随意的选择是我们的选择。
3:57.480–4:02.600
zh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。
4:02.600–4:06.680
zh如果我们说我们的选择这个进入我们的页务,
4:06.680–4:11.880
zh我们肯定能够用这个模式,因为现在,我们的页务员工程,
4:11.880–4:15.640
zh他们其实会有很多认证和评价的评价。
4:15.640–4:19.480
zh例如,如果我们的收入的收入我们的预期,
4:19.480–4:21.720
zh我们还是需要这些认证,
4:21.720–4:24.680
zh这会常常需要一个人的路程式的方法。
4:24.680–4:26.600
zh但是如果我们有一个模式的模式,
4:26.600–4:29.800
zh我们可以定义它的质量从开始
4:29.800–4:34.080
zh即使是80%,60%,40%的距离,
4:34.080–4:38.520
zh我们只需要在一个很低的40%的距离。
4:38.520–4:41.080
zh如果它基本上是在结构的质量,
4:41.080–4:43.480
zh我们不需要保持人的质量,
4:43.480–4:46.600
zh我们不需要保持人的质量。
4:46.600–4:50.040
zh所以,在这个将来的更多的模式,
4:50.040–4:52.280
zh它可以更加大幅并且在这些模式上
4:52.280–4:55.280
zh这个模式其实只发了一个新版本
4:55.280–4:59.040
zh没有提供的技术和计程度
4:59.040–5:01.960
zh我觉得这个方法非常值得考虑
5:01.960–5:04.160
zh它使用一个系统一式的方法
5:04.160–5:05.640
zh是快思考
5:05.640–5:06.880
zhDaniel Kahneman的作词
5:06.880–5:09.280
zh说一大部分人的认识
5:09.280–5:11.360
zh不在实际上的认识
5:11.360–5:12.720
zh我们立刻的认识
5:12.720–5:14.640
zh我们立刻的认识
5:14.640–5:17.440
zh例如,选择哪个衣服
5:17.440–5:18.880
zh很自然的衣服
5:18.880–5:21.440
zh我只穿着衣服
5:21.440–5:22.573
enand when I'm going to
5:22.573–5:24.507
enwork I might wear a different pair.
5:24.507–5:27.040
enSo this judgment is actually very intuitive.
5:27.040–5:29.280
enIt's the same for AI models.
5:29.280–5:30.791
enYou don't actually need
5:30.791–5:32.076
ento generate or think
5:32.076–5:34.040
enthrough every single judgment
5:34.040–5:36.200
enor use chain of thought reasoning.
5:36.200–5:38.396
enI can use a smaller model or a model
5:38.396–5:40.200
endesigned for fast thinking
5:40.200–5:41.680
enwith parallel computing
5:41.680–5:43.044
ento accomplish our tasks
5:43.044–5:44.680
enand goals very efficiently.
5:44.680–5:45.880
enSo I think it might
5:45.880–5:47.480
enreally make its way into
5:47.480–5:48.600
enour society soon
5:48.600–5:50.700
enbecause it's cheap, right?
5:50.700–5:53.780
zh而且是最高的,因為它是很高的,
5:53.780–5:57.100
zh在0.042萬元的價值,
5:57.100–5:59.780
zh比 GPT-4的價值,可能是10萬元。
5:59.780–6:03.300
zh價值是這麼高,他們現在不在賣現在的價值。
6:03.300–6:07.820
zh所以它可以使用在一大多數的評測 scenarios。
6:07.820–6:10.740
zh我們看到每個小型的手法
6:10.740–6:13.420
zh可能有一個決定的模式背後。
6:13.420–6:15.268
zh以前,一名 classifier would have
6:15.268–6:16.578
enrequired someone to
6:16.578–6:18.118
enspend ages training it,
6:18.118–6:18.618
enright?
6:18.618–6:22.740
zh我們以前要有的要求,但現在是沒有必要的
6:22.740–6:25.780
zh現在,這 model可以幫我們做了一切的評判,
6:25.780–6:28.660
zh決定定 specific outcomes based on the context
6:28.660–6:34.460
zh這代表,這代表,已經融化了更多的技術和考驗
6:34.460–6:38.860
zh我們會看到這個功能伸展開更多的使用
6:38.860–6:42.620
zh是的,我认为这个是非常有趣的新方式,
6:42.620–6:46.380
zh一遍完全不同的方法使用力
6:46.380–6:47.980
zh我們之前說明,
6:47.980–6:50.700
zh未來可能會有小型的模式
6:50.700–6:52.420
zh我們只需要有小型的模式
6:52.420–6:54.500
zh提供的模式
6:54.500–6:56.500
zh而是用販的模式
6:56.500–6:57.380
zh每個都市場的模式
6:57.380–6:59.620
zh我們最近的書
6:59.620–7:00.940
zh我們最近的書
7:00.940–7:02.540
zh也有同一個同意
7:02.540–7:05.060
zh成為我們目標的最適合的方法
7:05.060–7:06.500
zh是完全不平的
7:06.500–7:11.180
zh所以我們想分享這些 JB的模式
7:11.180–7:12.980
enThey've opened their waitlist,
7:12.980–7:14.713
enso if you'd like to try them out,
7:14.713–7:15.713
enfeel free to join.
7:15.820–7:17.680
enHave a great weekend. Bye!
0:00.000–0:00.500
Hello!
0:00.500–0:01.818
A new model has sparked
0:01.818–0:03.273
widespread discussion
0:03.273–0:05.018
over the last couple of days,
0:05.018–0:05.600
named JF.
0:05.600–0:06.585
A former OpenAI
0:06.585–0:07.949
researcher spent two
0:07.949–0:10.223
years developing it after leaving,
0:10.223–0:11.512
creating a new model
0:11.512–0:12.800
trained in a way that
0:12.800–0:14.240
differs from the past.
0:14.240–0:15.627
I think its features are
0:15.627–0:17.291
very distinct and powerful.
0:17.291–0:18.469
For specific tasks,
0:18.469–0:19.787
it can run at a fraction
0:19.787–0:21.173
of the cost and hundreds
0:21.173–0:22.352
of times the speed of
0:22.352–0:23.600
traditional models,
0:23.600–0:25.711
with very low hallucination rates,
0:25.711–0:27.681
generating the answers you need,
0:27.681–0:29.440
which is highly significant.
0:29.440–0:34.320
它是重要的,因為它選了完全不同的一路由我們現代的AI,
0:34.320–0:36.320
它是在原來的N-Token Prediction。
0:36.320–0:39.840
當大家是比拼了強硬的解決,
0:39.840–0:44.480
硬的解決,和能力的解決方法,
0:44.480–0:47.360
它其實是很小的地步。
0:47.360–0:49.600
我們不需要等待太多的信息,
0:49.600–0:54.960
或多多的解決,我們只需要一個比較的測試,
0:54.960–0:57.040
或是個比例的不同的結果。
0:57.040–0:59.600
我给你一个例子
0:59.600–1:03.180
例如,在一个公司的行业,
1:03.180–1:05.480
如果我们使用了AI的相机,
1:05.480–1:07.140
我们必须要进入到的模特,
1:07.140–1:08.880
等待在模特上的模特,
1:08.880–1:10.980
这些时间需要几秒,对吗?
1:10.980–1:12.100
这一终的效果会给你,
1:12.100–1:14.100
你会在哪个部份接触,
1:14.100–1:15.200
谁要跟他在谈论,
1:15.200–1:18.100
或者是指出的优质是低的,
1:18.100–1:19.380
需要更加注意力。
1:19.380–1:20.880
它给你一个样的效果。
1:20.880–1:23.100
在实际上,用这个新模特,
1:23.100–1:24.900
它可以非常快速,
1:24.900–1:28.900
这几个百分钟的几个百分钟的几个百分钟
1:28.900–1:30.900
这几个百分钟的几个百分钟
1:30.900–1:32.900
这几个百分钟的几个百分钟
1:32.900–1:34.300
the billing department,
1:34.300–1:36.300
estimatesthe customer churn rate,
1:36.300–1:38.500
and assessesthe likelihoodofneeding
1:38.500–1:41.580
human intervention.
1:41.580–1:43.591
Once theseassessmentsare made,
1:43.591–1:44.525
I canassignthe
1:44.525–1:47.900
tasktoaspecificdepartmentorouteittoahumanagent,
1:47.900–1:48.567
allowingme
1:48.567–1:51.633
tomakethecorrespondingadjustmentsanddecisions.
1:51.633–1:52.633
So itsefficiency
1:52.633–1:55.700
andcosteffectivenessareincrediblyimpressive.It
1:55.700–1:56.767
isjustextremely,
1:56.767–1:57.700
extremelyfast.
1:57.700–1:58.200
Even
1:58.200–2:00.933
initslivedemonstrationofstructuredoutput,
2:00.933–2:03.467
youcanclearlyseethatitgeneratesresults
2:03.467–2:05.080
但是传统的 models
2:05.080–2:06.320
也会一直在用评论
2:06.320–2:08.180
词为文字的评论
2:08.180–2:10.340
所以这两个技术的 approaches
2:10.340–2:12.220
其实完全不同
2:12.220–2:14.087
当 the founder released this video
2:14.087–2:15.140
他说了一篇,
2:15.140–2:17.920
他说了一篇,
2:17.920–2:20.060
我们的认识,
2:20.060–2:22.060
其实是一篇的答案
2:22.060–2:23.860
合适的答案
2:23.860–2:25.820
不太不太确定,
2:25.820–2:27.060
我们喜欢他们,
2:27.060–2:27.640
对吗?
2:27.640–2:28.540
例如,
2:28.540–2:30.200
他们可能会更多 empathetic
2:30.200–2:32.100
或是更多理解我们的
2:32.100–2:35.300
但是,他們的問題是他們的解答問題
2:35.300–2:36.800
人們不可能會不喜歡,
2:36.800–2:39.700
甚至是有個方法的方法。
2:39.700–2:44.100
這帶來更加盡量的信息和 consistency。
2:44.100–2:47.400
但是,他們的解答案是叫做RLCD,
2:47.400–2:49.064
名為 Reinforcement Learning
2:49.064–2:50.800
from Calibrated Decisions。
2:50.800–2:52.800
它指的是,在 Reinforcement Learning,
2:52.800–2:54.600
我實在是在調整的解答案是
2:54.600–2:56.600
的解答案是一個解答案。
2:56.600–2:59.800
解答案是每次它提供答案,
2:59.800–3:01.600
它提供一個 probability。
3:01.600–3:02.760
A confidence score
3:02.760–3:03.607
For instance,
3:03.607–3:05.160
with this billing inquiry
3:05.160–3:06.526
the probability that I
3:06.526–3:07.819
should route it to the
3:07.819–3:09.760
billing department is about 90%
3:09.760–3:11.093
But if the customer also
3:11.093–3:12.427
mentioned other issues
3:12.427–3:13.493
I believe there's a
3:13.493–3:14.827
certain probability it
3:14.827–3:16.093
should be prioritized
3:16.093–3:17.360
for other departments
3:17.360–3:18.560
So I might assign a 10%
3:18.560–3:20.493
or 20% probability to those as well
3:20.493–3:21.293
At this point,
3:21.293–3:22.960
you can set a threshold
3:22.960–3:24.029
For example,
3:24.029–3:26.360
if the churn risk is very high
3:26.360–3:28.760
I can have a human intervene directly
3:28.760–3:29.427
如果它是低,
3:29.427–3:32.760
我可以让它 AI model handle the response.
3:32.760–3:36.760
所以,它帮助我们去解决哪个性能量的权量使用
3:36.760–3:39.520
来解决于不同的情况下。
3:39.520–3:44.240
所以,我相信这个模型会将将其进行我们的生活生活
3:44.240–3:46.360
和很多不同的专业。
3:46.360–3:50.200
以前我们曾经用了一枪创造的创造,
3:50.200–3:53.000
或者我们不需要太多的时间。
3:53.000–3:57.480
但这些随意的随意的选择是我们的选择。
3:57.480–4:02.600
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。
4:02.600–4:06.680
如果我们说我们的选择这个进入我们的页务,
4:06.680–4:11.880
我们肯定能够用这个模式,因为现在,我们的页务员工程,
4:11.880–4:15.640
他们其实会有很多认证和评价的评价。
4:15.640–4:19.480
例如,如果我们的收入的收入我们的预期,
4:19.480–4:21.720
我们还是需要这些认证,
4:21.720–4:24.680
这会常常需要一个人的路程式的方法。
4:24.680–4:26.600
但是如果我们有一个模式的模式,
4:26.600–4:29.800
我们可以定义它的质量从开始
4:29.800–4:34.080
即使是80%,60%,40%的距离,
4:34.080–4:38.520
我们只需要在一个很低的40%的距离。
4:38.520–4:41.080
如果它基本上是在结构的质量,
4:41.080–4:43.480
我们不需要保持人的质量,
4:43.480–4:46.600
我们不需要保持人的质量。
4:46.600–4:50.040
所以,在这个将来的更多的模式,
4:50.040–4:52.280
它可以更加大幅并且在这些模式上
4:52.280–4:55.280
这个模式其实只发了一个新版本
4:55.280–4:59.040
没有提供的技术和计程度
4:59.040–5:01.960
我觉得这个方法非常值得考虑
5:01.960–5:04.160
它使用一个系统一式的方法
5:04.160–5:05.640
是快思考
5:05.640–5:06.880
Daniel Kahneman的作词
5:06.880–5:09.280
说一大部分人的认识
5:09.280–5:11.360
不在实际上的认识
5:11.360–5:12.720
我们立刻的认识
5:12.720–5:14.640
我们立刻的认识
5:14.640–5:17.440
例如,选择哪个衣服
5:17.440–5:18.880
很自然的衣服
5:18.880–5:21.440
我只穿着衣服
5:21.440–5:22.573
and when I'm going to
5:22.573–5:24.507
work I might wear a different pair.
5:24.507–5:27.040
So this judgment is actually very intuitive.
5:27.040–5:29.280
It's the same for AI models.
5:29.280–5:30.791
You don't actually need
5:30.791–5:32.076
to generate or think
5:32.076–5:34.040
through every single judgment
5:34.040–5:36.200
or use chain of thought reasoning.
5:36.200–5:38.396
I can use a smaller model or a model
5:38.396–5:40.200
designed for fast thinking
5:40.200–5:41.680
with parallel computing
5:41.680–5:43.044
to accomplish our tasks
5:43.044–5:44.680
and goals very efficiently.
5:44.680–5:45.880
So I think it might
5:45.880–5:47.480
really make its way into
5:47.480–5:48.600
our society soon
5:48.600–5:50.700
because it's cheap, right?
5:50.700–5:53.780
而且是最高的,因為它是很高的,
5:53.780–5:57.100
在0.042萬元的價值,
5:57.100–5:59.780
比 GPT-4的價值,可能是10萬元。
5:59.780–6:03.300
價值是這麼高,他們現在不在賣現在的價值。
6:03.300–6:07.820
所以它可以使用在一大多數的評測 scenarios。
6:07.820–6:10.740
我們看到每個小型的手法
6:10.740–6:13.420
可能有一個決定的模式背後。
6:13.420–6:15.268
以前,一名 classifier would have
6:15.268–6:16.578
required someone to
6:16.578–6:18.118
spend ages training it,
6:18.118–6:18.618
right?
6:18.618–6:22.740
我們以前要有的要求,但現在是沒有必要的
6:22.740–6:25.780
現在,這 model可以幫我們做了一切的評判,
6:25.780–6:28.660
決定定 specific outcomes based on the context
6:28.660–6:34.460
這代表,這代表,已經融化了更多的技術和考驗
6:34.460–6:38.860
我們會看到這個功能伸展開更多的使用
6:38.860–6:42.620
是的,我认为这个是非常有趣的新方式,
6:42.620–6:46.380
一遍完全不同的方法使用力
6:46.380–6:47.980
我們之前說明,
6:47.980–6:50.700
未來可能會有小型的模式
6:50.700–6:52.420
我們只需要有小型的模式
6:52.420–6:54.500
提供的模式
6:54.500–6:56.500
而是用販的模式
6:56.500–6:57.380
每個都市場的模式
6:57.380–6:59.620
我們最近的書
6:59.620–7:00.940
我們最近的書
7:00.940–7:02.540
也有同一個同意
7:02.540–7:05.060
成為我們目標的最適合的方法
7:05.060–7:06.500
是完全不平的
7:06.500–7:11.180
所以我們想分享這些 JB的模式
7:11.180–7:12.980
They've opened their waitlist,
7:12.980–7:14.713
so if you'd like to try them out,
7:14.713–7:15.713
feel free to join.
7:15.820–7:17.680
Have a great weekend. Bye!
0:00.000–0:00.500
enHello!
Hello!
0:00.500–0:01.818
enA new model has sparked
A new model has sparked
0:01.818–0:03.273
enwidespread discussion
widespread discussion
0:03.273–0:05.018
enover the last couple of days,
over the last couple of days,
0:05.018–0:05.600
ennamed JF.
named JF.
0:05.600–0:06.585
enA former OpenAI
A former OpenAI
0:06.585–0:07.949
enresearcher spent two
researcher spent two
0:07.949–0:10.223
enyears developing it after leaving,
years developing it after leaving,
0:10.223–0:11.512
encreating a new model
creating a new model
0:11.512–0:12.800
entrained in a way that
trained in a way that
0:12.800–0:14.240
endiffers from the past.
differs from the past.
0:14.240–0:15.627
enI think its features are
I think its features are
0:15.627–0:17.291
envery distinct and powerful.
very distinct and powerful.
0:17.291–0:18.469
enFor specific tasks,
For specific tasks,
0:18.469–0:19.787
enit can run at a fraction
it can run at a fraction
0:19.787–0:21.173
enof the cost and hundreds
of the cost and hundreds
0:21.173–0:22.352
enof times the speed of
of times the speed of
0:22.352–0:23.600
entraditional models,
traditional models,
0:23.600–0:25.711
enwith very low hallucination rates,
with very low hallucination rates,
0:25.711–0:27.681
engenerating the answers you need,
generating the answers you need,
0:27.681–0:29.440
enwhich is highly significant.
which is highly significant.
0:29.440–0:34.320
zh它是重要的,因為它選了完全不同的一路由我們現代的AI,
它是重要的,因為它選了完全不同的一路由我們現代的AI,
0:34.320–0:36.320
zh它是在原來的N-Token Prediction。
它是在原來的N-Token Prediction。
0:36.320–0:39.840
zh當大家是比拼了強硬的解決,
當大家是比拼了強硬的解決,
0:39.840–0:44.480
zh硬的解決,和能力的解決方法,
硬的解決,和能力的解決方法,
0:44.480–0:47.360
zh它其實是很小的地步。
它其實是很小的地步。
0:47.360–0:49.600
zh我們不需要等待太多的信息,
我們不需要等待太多的信息,
0:49.600–0:54.960
zh或多多的解決,我們只需要一個比較的測試,
或多多的解決,我們只需要一個比較的測試,
0:54.960–0:57.040
zh或是個比例的不同的結果。
或是個比例的不同的結果。
0:57.040–0:59.600
zh我给你一个例子
我给你一个例子
0:59.600–1:03.180
zh例如,在一个公司的行业,
例如,在一个公司的行业,
1:03.180–1:05.480
zh如果我们使用了AI的相机,
如果我们使用了AI的相机,
1:05.480–1:07.140
zh我们必须要进入到的模特,
我们必须要进入到的模特,
1:07.140–1:08.880
zh等待在模特上的模特,
等待在模特上的模特,
1:08.880–1:10.980
zh这些时间需要几秒,对吗?
这些时间需要几秒,对吗?
1:10.980–1:12.100
zh这一终的效果会给你,
这一终的效果会给你,
1:12.100–1:14.100
zh你会在哪个部份接触,
你会在哪个部份接触,
1:14.100–1:15.200
zh谁要跟他在谈论,
谁要跟他在谈论,
1:15.200–1:18.100
zh或者是指出的优质是低的,
或者是指出的优质是低的,
1:18.100–1:19.380
zh需要更加注意力。
需要更加注意力。
1:19.380–1:20.880
zh它给你一个样的效果。
它给你一个样的效果。
1:20.880–1:23.100
zh在实际上,用这个新模特,
在实际上,用这个新模特,
1:23.100–1:24.900
zh它可以非常快速,
它可以非常快速,
1:24.900–1:28.900
zh这几个百分钟的几个百分钟的几个百分钟
这几个百分钟的几个百分钟的几个百分钟
1:28.900–1:30.900
zh这几个百分钟的几个百分钟
这几个百分钟的几个百分钟
1:30.900–1:32.900
zh这几个百分钟的几个百分钟
这几个百分钟的几个百分钟
1:32.900–1:34.300
enthe billing department,
the billing department,
1:34.300–1:36.300
enestimatesthe customer churn rate,
estimatesthe customer churn rate,
1:36.300–1:38.500
enand assessesthe likelihoodofneeding
and assessesthe likelihoodofneeding
1:38.500–1:41.580
enhuman intervention.
human intervention.
1:41.580–1:43.591
enOnce theseassessmentsare made,
Once theseassessmentsare made,
1:43.591–1:44.525
enI canassignthe
I canassignthe
1:44.525–1:47.900
entasktoaspecificdepartmentorouteittoahumanagent,
tasktoaspecificdepartmentorouteittoahumanagent,
1:47.900–1:48.567
enallowingme
allowingme
1:48.567–1:51.633
entomakethecorrespondingadjustmentsanddecisions.
tomakethecorrespondingadjustmentsanddecisions.
1:51.633–1:52.633
enSo itsefficiency
So itsefficiency
1:52.633–1:55.700
enandcosteffectivenessareincrediblyimpressive.It
andcosteffectivenessareincrediblyimpressive.It
1:55.700–1:56.767
enisjustextremely,
isjustextremely,
1:56.767–1:57.700
enextremelyfast.
extremelyfast.
1:57.700–1:58.200
enEven
Even
1:58.200–2:00.933
eninitslivedemonstrationofstructuredoutput,
initslivedemonstrationofstructuredoutput,
2:00.933–2:03.467
enyoucanclearlyseethatitgeneratesresults
youcanclearlyseethatitgeneratesresults
2:03.467–2:05.080
zh但是传统的 models
但是传统的 models
2:05.080–2:06.320
zh也会一直在用评论
也会一直在用评论
2:06.320–2:08.180
zh词为文字的评论
词为文字的评论
2:08.180–2:10.340
zh所以这两个技术的 approaches
所以这两个技术的 approaches
2:10.340–2:12.220
zh其实完全不同
其实完全不同
2:12.220–2:14.087
zh当 the founder released this video
当 the founder released this video
2:14.087–2:15.140
zh他说了一篇,
他说了一篇,
2:15.140–2:17.920
zh他说了一篇,
他说了一篇,
2:17.920–2:20.060
zh我们的认识,
我们的认识,
2:20.060–2:22.060
zh其实是一篇的答案
其实是一篇的答案
2:22.060–2:23.860
zh合适的答案
合适的答案
2:23.860–2:25.820
zh不太不太确定,
不太不太确定,
2:25.820–2:27.060
zh我们喜欢他们,
我们喜欢他们,
2:27.060–2:27.640
zh对吗?
对吗?
2:27.640–2:28.540
zh例如,
例如,
2:28.540–2:30.200
zh他们可能会更多 empathetic
他们可能会更多 empathetic
2:30.200–2:32.100
zh或是更多理解我们的
或是更多理解我们的
2:32.100–2:35.300
zh但是,他們的問題是他們的解答問題
但是,他們的問題是他們的解答問題
2:35.300–2:36.800
zh人們不可能會不喜歡,
人們不可能會不喜歡,
2:36.800–2:39.700
zh甚至是有個方法的方法。
甚至是有個方法的方法。
2:39.700–2:44.100
zh這帶來更加盡量的信息和 consistency。
這帶來更加盡量的信息和 consistency。
2:44.100–2:47.400
zh但是,他們的解答案是叫做RLCD,
但是,他們的解答案是叫做RLCD,
2:47.400–2:49.064
zh名為 Reinforcement Learning
名為 Reinforcement Learning
2:49.064–2:50.800
enfrom Calibrated Decisions。
from Calibrated Decisions。
2:50.800–2:52.800
zh它指的是,在 Reinforcement Learning,
它指的是,在 Reinforcement Learning,
2:52.800–2:54.600
zh我實在是在調整的解答案是
我實在是在調整的解答案是
2:54.600–2:56.600
zh的解答案是一個解答案。
的解答案是一個解答案。
2:56.600–2:59.800
zh解答案是每次它提供答案,
解答案是每次它提供答案,
2:59.800–3:01.600
zh它提供一個 probability。
它提供一個 probability。
3:01.600–3:02.760
enA confidence score
A confidence score
3:02.760–3:03.607
enFor instance,
For instance,
3:03.607–3:05.160
enwith this billing inquiry
with this billing inquiry
3:05.160–3:06.526
enthe probability that I
the probability that I
3:06.526–3:07.819
enshould route it to the
should route it to the
3:07.819–3:09.760
enbilling department is about 90%
billing department is about 90%
3:09.760–3:11.093
enBut if the customer also
But if the customer also
3:11.093–3:12.427
enmentioned other issues
mentioned other issues
3:12.427–3:13.493
enI believe there's a
I believe there's a
3:13.493–3:14.827
encertain probability it
certain probability it
3:14.827–3:16.093
enshould be prioritized
should be prioritized
3:16.093–3:17.360
enfor other departments
for other departments
3:17.360–3:18.560
enSo I might assign a 10%
So I might assign a 10%
3:18.560–3:20.493
enor 20% probability to those as well
or 20% probability to those as well
3:20.493–3:21.293
enAt this point,
At this point,
3:21.293–3:22.960
enyou can set a threshold
you can set a threshold
3:22.960–3:24.029
enFor example,
For example,
3:24.029–3:26.360
enif the churn risk is very high
if the churn risk is very high
3:26.360–3:28.760
enI can have a human intervene directly
I can have a human intervene directly
3:28.760–3:29.427
zh如果它是低,
如果它是低,
3:29.427–3:32.760
zh我可以让它 AI model handle the response.
我可以让它 AI model handle the response.
3:32.760–3:36.760
zh所以,它帮助我们去解决哪个性能量的权量使用
所以,它帮助我们去解决哪个性能量的权量使用
3:36.760–3:39.520
zh来解决于不同的情况下。
来解决于不同的情况下。
3:39.520–3:44.240
zh所以,我相信这个模型会将将其进行我们的生活生活
所以,我相信这个模型会将将其进行我们的生活生活
3:44.240–3:46.360
zh和很多不同的专业。
和很多不同的专业。
3:46.360–3:50.200
zh以前我们曾经用了一枪创造的创造,
以前我们曾经用了一枪创造的创造,
3:50.200–3:53.000
zh或者我们不需要太多的时间。
或者我们不需要太多的时间。
3:53.000–3:57.480
zh但这些随意的随意的选择是我们的选择。
但这些随意的随意的选择是我们的选择。
3:57.480–4:02.600
zh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。
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。
4:02.600–4:06.680
zh如果我们说我们的选择这个进入我们的页务,
如果我们说我们的选择这个进入我们的页务,
4:06.680–4:11.880
zh我们肯定能够用这个模式,因为现在,我们的页务员工程,
我们肯定能够用这个模式,因为现在,我们的页务员工程,
4:11.880–4:15.640
zh他们其实会有很多认证和评价的评价。
他们其实会有很多认证和评价的评价。
4:15.640–4:19.480
zh例如,如果我们的收入的收入我们的预期,
例如,如果我们的收入的收入我们的预期,
4:19.480–4:21.720
zh我们还是需要这些认证,
我们还是需要这些认证,
4:21.720–4:24.680
zh这会常常需要一个人的路程式的方法。
这会常常需要一个人的路程式的方法。
4:24.680–4:26.600
zh但是如果我们有一个模式的模式,
但是如果我们有一个模式的模式,
4:26.600–4:29.800
zh我们可以定义它的质量从开始
我们可以定义它的质量从开始
4:29.800–4:34.080
zh即使是80%,60%,40%的距离,
即使是80%,60%,40%的距离,
4:34.080–4:38.520
zh我们只需要在一个很低的40%的距离。
我们只需要在一个很低的40%的距离。
4:38.520–4:41.080
zh如果它基本上是在结构的质量,
如果它基本上是在结构的质量,
4:41.080–4:43.480
zh我们不需要保持人的质量,
我们不需要保持人的质量,
4:43.480–4:46.600
zh我们不需要保持人的质量。
我们不需要保持人的质量。
4:46.600–4:50.040
zh所以,在这个将来的更多的模式,
所以,在这个将来的更多的模式,
4:50.040–4:52.280
zh它可以更加大幅并且在这些模式上
它可以更加大幅并且在这些模式上
4:52.280–4:55.280
zh这个模式其实只发了一个新版本
这个模式其实只发了一个新版本
4:55.280–4:59.040
zh没有提供的技术和计程度
没有提供的技术和计程度
4:59.040–5:01.960
zh我觉得这个方法非常值得考虑
我觉得这个方法非常值得考虑
5:01.960–5:04.160
zh它使用一个系统一式的方法
它使用一个系统一式的方法
5:04.160–5:05.640
zh是快思考
是快思考
5:05.640–5:06.880
zhDaniel Kahneman的作词
Daniel Kahneman的作词
5:06.880–5:09.280
zh说一大部分人的认识
说一大部分人的认识
5:09.280–5:11.360
zh不在实际上的认识
不在实际上的认识
5:11.360–5:12.720
zh我们立刻的认识
我们立刻的认识
5:12.720–5:14.640
zh我们立刻的认识
我们立刻的认识
5:14.640–5:17.440
zh例如,选择哪个衣服
例如,选择哪个衣服
5:17.440–5:18.880
zh很自然的衣服
很自然的衣服
5:18.880–5:21.440
zh我只穿着衣服
我只穿着衣服
5:21.440–5:22.573
enand when I'm going to
and when I'm going to
5:22.573–5:24.507
enwork I might wear a different pair.
work I might wear a different pair.
5:24.507–5:27.040
enSo this judgment is actually very intuitive.
So this judgment is actually very intuitive.
5:27.040–5:29.280
enIt's the same for AI models.
It's the same for AI models.
5:29.280–5:30.791
enYou don't actually need
You don't actually need
5:30.791–5:32.076
ento generate or think
to generate or think
5:32.076–5:34.040
enthrough every single judgment
through every single judgment
5:34.040–5:36.200
enor use chain of thought reasoning.
or use chain of thought reasoning.
5:36.200–5:38.396
enI can use a smaller model or a model
I can use a smaller model or a model
5:38.396–5:40.200
endesigned for fast thinking
designed for fast thinking
5:40.200–5:41.680
enwith parallel computing
with parallel computing
5:41.680–5:43.044
ento accomplish our tasks
to accomplish our tasks
5:43.044–5:44.680
enand goals very efficiently.
and goals very efficiently.
5:44.680–5:45.880
enSo I think it might
So I think it might
5:45.880–5:47.480
enreally make its way into
really make its way into
5:47.480–5:48.600
enour society soon
our society soon
5:48.600–5:50.700
enbecause it's cheap, right?
because it's cheap, right?
5:50.700–5:53.780
zh而且是最高的,因為它是很高的,
而且是最高的,因為它是很高的,
5:53.780–5:57.100
zh在0.042萬元的價值,
在0.042萬元的價值,
5:57.100–5:59.780
zh比 GPT-4的價值,可能是10萬元。
比 GPT-4的價值,可能是10萬元。
5:59.780–6:03.300
zh價值是這麼高,他們現在不在賣現在的價值。
價值是這麼高,他們現在不在賣現在的價值。
6:03.300–6:07.820
zh所以它可以使用在一大多數的評測 scenarios。
所以它可以使用在一大多數的評測 scenarios。
6:07.820–6:10.740
zh我們看到每個小型的手法
我們看到每個小型的手法
6:10.740–6:13.420
zh可能有一個決定的模式背後。
可能有一個決定的模式背後。
6:13.420–6:15.268
zh以前,一名 classifier would have
以前,一名 classifier would have
6:15.268–6:16.578
enrequired someone to
required someone to
6:16.578–6:18.118
enspend ages training it,
spend ages training it,
6:18.118–6:18.618
enright?
right?
6:18.618–6:22.740
zh我們以前要有的要求,但現在是沒有必要的
我們以前要有的要求,但現在是沒有必要的
6:22.740–6:25.780
zh現在,這 model可以幫我們做了一切的評判,
現在,這 model可以幫我們做了一切的評判,
6:25.780–6:28.660
zh決定定 specific outcomes based on the context
決定定 specific outcomes based on the context
6:28.660–6:34.460
zh這代表,這代表,已經融化了更多的技術和考驗
這代表,這代表,已經融化了更多的技術和考驗
6:34.460–6:38.860
zh我們會看到這個功能伸展開更多的使用
我們會看到這個功能伸展開更多的使用
6:38.860–6:42.620
zh是的,我认为这个是非常有趣的新方式,
是的,我认为这个是非常有趣的新方式,
6:42.620–6:46.380
zh一遍完全不同的方法使用力
一遍完全不同的方法使用力
6:46.380–6:47.980
zh我們之前說明,
我們之前說明,
6:47.980–6:50.700
zh未來可能會有小型的模式
未來可能會有小型的模式
6:50.700–6:52.420
zh我們只需要有小型的模式
我們只需要有小型的模式
6:52.420–6:54.500
zh提供的模式
提供的模式
6:54.500–6:56.500
zh而是用販的模式
而是用販的模式
6:56.500–6:57.380
zh每個都市場的模式
每個都市場的模式
6:57.380–6:59.620
zh我們最近的書
我們最近的書
6:59.620–7:00.940
zh我們最近的書
我們最近的書
7:00.940–7:02.540
zh也有同一個同意
也有同一個同意
7:02.540–7:05.060
zh成為我們目標的最適合的方法
成為我們目標的最適合的方法
7:05.060–7:06.500
zh是完全不平的
是完全不平的
7:06.500–7:11.180
zh所以我們想分享這些 JB的模式
所以我們想分享這些 JB的模式
7:11.180–7:12.980
enThey've opened their waitlist,
They've opened their waitlist,
7:12.980–7:14.713
enso if you'd like to try them out,
so if you'd like to try them out,
7:14.713–7:15.713
enfeel free to join.
feel free to join.
7:15.820–7:17.680
enHave a great weekend. Bye!
Have a great weekend. Bye!