100:00:06,480-->00:00:08,400
Good morning. We have a banger for you200:00:08,400-->00:00:09,840
today. We're going to launch chatbt300:00:09,840-->00:00:11,519
agent. But before jumping into that, I'd400:00:11,519-->00:00:12,559
like to ask the team to introduce500:00:12,559-->00:00:14,080
themselves. Starting with Yosh.600:00:14,080-->00:00:17,840
Hi, I'm Yash. I work on agent team and700:00:17,840-->00:00:20,080
before that I used to work on operator.800:00:20,080-->00:00:22,560
Hi, I'm Jing. I work on agents research900:00:22,560-->00:00:24,400
previously on deep research.1000:00:24,400-->00:00:26,000
Hi, I'm Casey. I'm a researcher on1100:00:26,000-->00:00:27,920
agents formerly operator.1200:00:27,920-->00:00:30,560
Hi, I'm Issa. I'm a researcher on agent1300:00:30,560-->00:00:32,640
formerly on deep research.1400:00:32,640-->00:00:34,880
So we we started launching agents1500:00:34,880-->00:00:36,800
earlier this year. Uh we launched deep1600:00:36,800-->00:00:38,879
research, we launched operator and1700:00:38,879-->00:00:40,160
people were very excited about this.1800:00:40,160-->00:00:42,480
People could see that now uh AI was1900:00:42,480-->00:00:44,640
going off to do complex tasks for them.2000:00:44,640-->00:00:46,079
But it became clear to us that what2100:00:46,079-->00:00:48,000
people really wanted was for us to bring2200:00:48,000-->00:00:49,760
those capabilities and more together.2300:00:49,760-->00:00:51,920
People wanted a unified agent that could2400:00:51,920-->00:00:55,039
go off, use its own computer and do real2500:00:55,039-->00:00:57,360
complex tasks for them, that could uh2600:00:57,360-->00:00:59,359
seamlessly transition from thinking2700:00:59,359-->00:01:01,520
about something to taking actions to2800:01:01,520-->00:01:03,359
using lots of tools using the terminal,2900:01:03,359-->00:01:05,360
clicking around the web, even producing3000:01:05,360-->00:01:06,880
things like spreadsheets and slides and3100:01:06,880-->00:01:08,960
and much more. And wanted people want to3200:01:08,960-->00:01:10,159
be able to do this over a long time3300:01:10,159-->00:01:12,159
horizon and a sort of for universal3400:01:12,159-->00:01:13,840
tasks. So the team has been working3500:01:13,840-->00:01:16,400
super hard to bring that together. And3600:01:16,400-->00:01:18,080
today we have chat with the agent. Um,3700:01:18,080-->00:01:19,680
it's probably easier to show it to you3800:01:19,680-->00:01:21,439
than to keep talking about it. It is one3900:01:21,439-->00:01:23,360
of the feel the aon moments for me to4000:01:23,360-->00:01:25,280
watch it work. So, let's take a look.4100:01:25,280-->00:01:27,840
Awesome. Thanks, Sam. Hello, everyone.4200:01:27,840-->00:01:29,920
Very excited to share chat GBD agent4300:01:29,920-->00:01:31,600
with everybody. And as Sam said, let's4400:01:31,600-->00:01:33,759
just dive right into the demo. Okay, so4500:01:33,759-->00:01:36,159
we are on Chad GBD as we all know and4600:01:36,159-->00:01:39,119
love. And to turn on the agent mode, you4700:01:39,119-->00:01:40,880
just click the tools menu and select4800:01:40,880-->00:01:43,280
agent. You can also just type agent in4900:01:43,280-->00:01:45,040
the composer bar and it'll take you to5000:01:45,040-->00:01:47,520
agent mode. Um, Edward and I have a5100:01:47,520-->00:01:49,360
wedding to go to later this year. Uh,5200:01:49,360-->00:01:51,119
it's for one of our mutual friends.5300:01:51,119-->00:01:52,560
Should we should we have the Asian5400:01:52,560-->00:01:53,280
planet?5500:01:53,280-->00:01:55,680
Yeah, let's do it. I need an outfit. And5600:01:55,680-->00:01:56,799
don't forget the gift.5700:01:56,799-->00:01:58,719
Okay, great. We won't forget the gift.5800:01:58,719-->00:02:00,240
Um, it's a little bit of a longer5900:02:00,240-->00:02:01,680
prompt, so I have it copied in my6000:02:01,680-->00:02:02,799
buffer, so I'm just going to go ahead6100:02:02,799-->00:02:05,759
and paste it. Um, okay. So, let's see.6200:02:05,759-->00:02:07,360
Let's see what it says. Our friends are6300:02:07,360-->00:02:08,640
getting married later this year, as I6400:02:08,640-->00:02:10,720
said, Minia and Sarah. And we want the6500:02:10,720-->00:02:12,879
agent to help us find an outfit that6600:02:12,879-->00:02:15,520
matches the dress code. uh propose a few6700:02:15,520-->00:02:17,840
options. Nice mid luxury taking into6800:02:17,840-->00:02:21,040
account venue and weather. We also want6900:02:21,040-->00:02:23,280
to find us some hotels and as Edward7000:02:23,280-->00:02:25,760
said, don't forget the gift. Um so let's7100:02:25,760-->00:02:27,840
see and7200:02:27,840-->00:02:30,319
send the prompt away. As Sam said, agent7300:02:30,319-->00:02:32,640
uses a computer. Uh so in the beginning7400:02:32,640-->00:02:34,959
it sets up its environment. It it you7500:02:34,959-->00:02:38,000
know it'll take a minute or two or not7600:02:38,000-->00:02:39,680
really 5 seconds to set up its7700:02:39,680-->00:02:41,440
environment. And in this case, as you7800:02:41,440-->00:02:43,840
see, it understands the prompt. It's7900:02:43,840-->00:02:46,319
asking for me for a clarification. I'm8000:02:46,319-->00:02:48,000
just going to let it just continue and8100:02:48,000-->00:02:51,120
work. Anyway, um I think it got confused8200:02:51,120-->00:02:54,239
by saying,"Oh, where's the um what8300:02:54,239-->00:02:55,680
exactly is the time of the date of the8400:02:55,680-->00:02:57,200
wedding?" I think it'll figure out using8500:02:57,200-->00:02:59,840
the website. Okay, cool. So, now it's8600:02:59,840-->00:03:01,760
kicked off. It's starting the process,8700:03:01,760-->00:03:03,920
the prompt, and it's open up a browser.8800:03:03,920-->00:03:04,959
And to walk you through what's8900:03:04,959-->00:03:06,800
happening, here's9000:03:06,800-->00:03:09,040
Yeah. So, as mentioned, we gave the9100:03:09,040-->00:03:10,879
agent access to its own virtual9200:03:10,879-->00:03:13,280
computer, and the computer has many9300:03:13,280-->00:03:14,720
different tools installed, and it can9400:03:14,720-->00:03:16,239
choose which to use as it's working9500:03:16,239-->00:03:18,640
through the task. So, in chat GPT, you9600:03:18,640-->00:03:21,360
can see a visualization of the agent's9700:03:21,360-->00:03:23,680
computer screen, and you can see9800:03:23,680-->00:03:25,519
overlaid its chain of thought in text,9900:03:25,519-->00:03:27,200
and that's what it's thinking as it's10000:03:27,200-->00:03:28,480
working through the task and deciding10100:03:28,480-->00:03:30,799
what to do next. We gave the agent10200:03:30,799-->00:03:32,400
access to two different ways to browse10300:03:32,400-->00:03:34,560
the internet. First, we gave it a text10400:03:34,560-->00:03:36,159
browser, and this is similar to the deep10500:03:36,159-->00:03:38,000
research tool. And this is what lets it10600:03:38,000-->00:03:40,159
really efficiently and quickly read many10700:03:40,159-->00:03:43,440
web pages um um and search for them. And10800:03:43,440-->00:03:45,040
we also gave it access to a visual10900:03:45,040-->00:03:46,319
browser. And this is similar to the11000:03:46,319-->00:03:48,239
operator tool. And this is what lets it11100:03:48,239-->00:03:50,159
actually interact with the UI of a web11200:03:50,159-->00:03:52,720
page. So it can um drag things. It can11300:03:52,720-->00:03:54,879
use the cursor to click around. It can11400:03:54,879-->00:03:57,280
open UI components. It can fill out11500:03:57,280-->00:03:59,920
forms and enter text and text areas.11600:03:59,920-->00:04:02,560
It's very flexible. So those two tools11700:04:02,560-->00:04:04,720
are very complimentary. And then we also11800:04:04,720-->00:04:06,720
gave it access to its own terminal so11900:04:06,720-->00:04:08,720
that it can run code and it can also12000:04:08,720-->00:04:10,640
generate and analyze files like slide12100:04:10,640-->00:04:12,879
decks and spreadsheets. And then through12200:04:12,879-->00:04:14,560
the terminal it's also able to call12300:04:14,560-->00:04:17,840
APIs. So both public APIs and APIs to12400:04:17,840-->00:04:19,840
access your private data sources like12500:04:19,840-->00:04:22,479
Google Drive, Google Calendar, GitHub,12600:04:22,479-->00:04:25,360
SharePoint and many others um and only12700:04:25,360-->00:04:26,960if you explicitly connect them similar12800:04:26,960-->00:04:28,960
to deep research connectors. And then it12900:04:28,960-->00:04:31,680
also has access to the image gen API so13000:04:31,680-->00:04:34,240
it can create nice visuals for um slide13100:04:34,240-->00:04:36,080
decks and other things as it's working13200:04:36,080-->00:04:38,240
through its tasks.13300:04:38,240-->00:04:40,800
How is deciding which tools to use here?13400:04:40,800-->00:04:42,560
Yes, we train the model to move between13500:04:42,560-->00:04:44,160
these capabilities with reinforcement13600:04:44,160-->00:04:46,080
learning. This is the first model we13700:04:46,080-->00:04:48,880
trained that has access to this unified13800:04:48,880-->00:04:52,000
tool box. A text browser, a GUI browser13900:04:52,000-->00:04:53,840
and a terminal all in one virtual14000:04:53,840-->00:04:57,120
machine. To guide its learning, we14100:04:57,120-->00:04:59,360
created hard tasks that require using14200:04:59,360-->00:05:01,919
all these tools. This allows the model14300:05:01,919-->00:05:04,000
not only to learn how to use these14400:05:04,000-->00:05:06,160
tools, but also when to use which tool14500:05:06,160-->00:05:08,400
depending on the task at hand. At the14600:05:08,400-->00:05:10,400
beginning of the training, the model14700:05:10,400-->00:05:12,880
might attempt to use all these tools to14800:05:12,880-->00:05:15,600
solve a relatively simple problem. Over14900:05:15,600-->00:05:17,840
time, as we reward the model for solving15000:05:17,840-->00:05:20,560
problems correctly and efficiently, the15100:05:20,560-->00:05:24,080
model will have smarter tool choice.15200:05:24,080-->00:05:27,360
For example,if you ask a model to uh15300:05:27,360-->00:05:29,039
find a restaurant with specific15400:05:29,039-->00:05:31,919
requirements and make a reservation, the15500:05:31,919-->00:05:34,479
model may typically just start a deep15600:05:34,479-->00:05:36,160
research in the text browser to find15700:05:36,160-->00:05:39,039
some candidates, then switch to the GUI15800:05:39,039-->00:05:42,160
browser to view photos of food, uh check15900:05:42,160-->00:05:45,600
availability, and complete the booking.16000:05:45,600-->00:05:48,000
Similarly,for creative task like16100:05:48,000-->00:05:50,160
creating an artifact, the model will16200:05:50,160-->00:05:51,680
first search online for public16300:05:51,680-->00:05:54,479
resources, then switch to the terminal16400:05:54,479-->00:05:57,039
to do some code editing to compile the16500:05:57,039-->00:05:59,919
artifact and finally verify the final16600:05:59,919-->00:06:02,960
outputs in the GUI browser. With this,16700:06:02,960-->00:06:05,600
we truly feel like we brought together16800:06:05,600-->00:06:08,240
the best of deep research and operator16900:06:08,240-->00:06:11,759
and added some extra sparkle.17000:06:11,759-->00:06:14,000
That's right. Yeah. So to put this17100:06:14,000-->00:06:15,520
project in context, I want to give a bit17200:06:15,520-->00:06:18,000
of history. So a few months ago, we17300:06:18,000-->00:06:20,960
shipped operator in January and this was17400:06:20,960-->00:06:23,120
our agent that lets you do online tasks17500:06:23,120-->00:06:25,759
like book reservations and um send17600:06:25,759-->00:06:27,840
emails and then two weeks later we17700:06:27,840-->00:06:29,919
shipped deep research and deep research17800:06:29,919-->00:06:31,919
is a tool that lets you do in-depth17900:06:31,919-->00:06:35,759
internet research and output highquality18000:06:35,759-->00:06:39,280
um um research reports. And after launch18100:06:39,280-->00:06:41,039
we realized that actually these two18200:06:41,039-->00:06:42,319
approaches are actually deeply18300:06:42,319-->00:06:44,160
complimentary.18400:06:44,160-->00:06:46,400
Um for example operator has some trouble18500:06:46,400-->00:06:48,720
reading super long articles. Um it has18600:06:48,720-->00:06:50,400
to scroll. It takes a long time. But18700:06:50,400-->00:06:51,759
that's something that deep research is18800:06:51,759-->00:06:56,240
good at. Conversely operator uh uh deep18900:06:56,240-->00:06:58,240
research isn't as good at interacting19000:06:58,240-->00:07:00,319
with web pages interactive elements19100:07:00,319-->00:07:03,199
visual uh highly visual web pages but19200:07:03,199-->00:07:04,800
that's something that operator excels19300:07:04,800-->00:07:08,639
at. So uh yeah we felt these approaches19400:07:08,639-->00:07:11,120
were complimentary and then we we were19500:07:11,120-->00:07:13,120
also looking at some customer feedback.19600:07:13,120-->00:07:14,880
So for example one of our most highly19700:07:14,880-->00:07:17,120
requested features for deep research was19800:07:17,120-->00:07:18,960
the ability to log into websites and19900:07:18,960-->00:07:20,960
access authenticated sources. That's20000:07:20,960-->00:07:22,880
something that operator can do.20100:07:22,880-->00:07:24,000
I've been waiting for that for a long20200:07:24,000-->00:07:24,560
time.20300:07:24,560-->00:07:26,160
Yeah.20400:07:26,160-->00:07:28,479
Um another thing is that we were looking20500:07:28,479-->00:07:29,840
at the prompts that people were trying20600:07:29,840-->00:07:31,520for operator and we saw that they were20700:07:31,520-->00:07:32,880
actually more deep research type20800:07:32,880-->00:07:35,199
prompts.for example, plan a trip and20900:07:35,199-->00:07:38,240
then book it. And so, yeah, we we really21000:07:38,240-->00:07:39,360
feel like we're bringing the best of21100:07:39,360-->00:07:41,440
both worlds here. And on a personal21200:07:41,440-->00:07:42,800
note, we've all been friends for a21300:07:42,800-->00:07:44,160while, and it's really exciting to be21400:07:44,160-->00:07:46,479
working together. So, speaking of21500:07:46,479-->00:07:48,960
matches made in heaven, how is the21600:07:48,960-->00:07:50,319
wedding planning going?21700:07:50,319-->00:07:51,759
It's amazing to watch. This is an21800:07:51,759-->00:07:53,599
example of a task I hate doing. This can21900:07:53,599-->00:07:55,520
like ruin like, you know, multiple hours22000:07:55,520-->00:07:56,960for me as I get sucked into these rabbit22100:07:56,960-->00:07:58,160
holes. So, just watching this as you22200:07:58,160-->00:07:59,520
guys have been talking click through22300:07:59,520-->00:08:01,199
this and just like do the whole thing is22400:08:01,199-->00:08:03,360
really quite remarkable. Yeah, totally.22500:08:03,360-->00:08:06,560
Um, looks like it started off by22600:08:06,560-->00:08:08,560
figuring out the weather. One of the22700:08:08,560-->00:08:11,280
cool features, um, is that, you know, as22800:08:11,280-->00:08:12,560
some of these tasks may take a little22900:08:12,560-->00:08:14,160
bit longer, you can just go back and see23000:08:14,160-->00:08:15,759
what it was doing. So, that's what we're23100:08:15,759-->00:08:17,199
exactly going to do. Looks like it went23200:08:17,199-->00:08:18,720
through the website to use the text23300:08:18,720-->00:08:21,039
browser. Interestingly,for that, now23400:08:21,039-->00:08:22,400
it's looking through the suits for23500:08:22,400-->00:08:23,919
Edward. I think it'll find something23600:08:23,919-->00:08:25,360
good. Here you can see it switched over23700:08:25,360-->00:08:27,199
to actually a visual browser to make23800:08:27,199-->00:08:28,960
sure suit will look really good on23900:08:28,960-->00:08:31,280
Edward.24000:08:31,280-->00:08:34,560
And now looks like yeah, it's got24100:08:34,560-->00:08:36,880
chugging along, figuring out what to do.24200:08:36,880-->00:08:39,599
Um, and still on suits and now probably24300:08:39,599-->00:08:41,919
getting to the gifts section. Um, okay,24400:08:41,919-->00:08:43,279
cool. So, this is going to take a while.24500:08:43,279-->00:08:44,959
As Sam said, these tasks sometimes can24600:08:44,959-->00:08:46,160
take a long time. So, it's going to24700:08:46,160-->00:08:47,680continue doing hopefully much faster24800:08:47,680-->00:08:49,760
than we will do. Um, should we do24900:08:49,760-->00:08:51,600
something elsewhile it's doing it? I25000:08:51,600-->00:08:53,519
think the team really wanted the um25100:08:53,519-->00:08:55,279
stickers, some stickers for the for the25200:08:55,279-->00:08:56,480
launch. Should we do that?25300:08:56,480-->00:08:57,279
Yeah, cool.25400:08:57,279-->00:08:59,040
All right. So, we have a team mascot,25500:08:59,040-->00:09:00,320
which is one of our colleagues, Bunny25600:09:00,320-->00:09:03,279
Doodle. really really cute tell you. Um25700:09:03,279-->00:09:06,080
and we're going to try and bring um get25800:09:06,080-->00:09:08,480
some laptop stickers for everybody. Uh25900:09:08,480-->00:09:10,480
one of the favorite features for agent26000:09:10,480-->00:09:13,120
is given that trajectories can take 1526100:09:13,120-->00:09:15,040
minutes,20 minutes,30 minutes26200:09:15,040-->00:09:17,120
depending on the complexity of the task.26300:09:17,120-->00:09:19,120
Um a lot of times the you might need to26400:09:19,120-->00:09:20,560
help the agent. Agent might need to ask26500:09:20,560-->00:09:22,480
you clarifications, confirmations and26600:09:22,480-->00:09:25,040
things like that. Um so I love to use it26700:09:25,040-->00:09:26,640
on the go. So I'm going to use my mobile26800:09:26,640-->00:09:28,160
phone to actually send the query this26900:09:28,160-->00:09:30,240
time and then see how it goes.27000:09:30,240-->00:09:32,880
Okay, so let's see. Okay, so we are on27100:09:32,880-->00:09:35,519
Chad Gibbdi. Uh I have already selected27200:09:35,519-->00:09:38,560
the agent mode. I've also inputed our uh27300:09:38,560-->00:09:40,560
cute mascot and I'm going to quickly27400:09:40,560-->00:09:43,040
paste a query. So query says make some27500:09:43,040-->00:09:45,279
swag for the team one by one laptop27600:09:45,279-->00:09:47,920
stickers and order 500 of them. I'll27700:09:47,920-->00:09:52,959
also say I like sticker mule27800:09:52,959-->00:09:55,279
which we have used in the past and send27900:09:55,279-->00:09:57,200
it off.28000:09:57,200-->00:10:00,080
Okay. So, just like it was doing on the28100:10:00,080-->00:10:02,080
web, it's going to take some time, think28200:10:02,080-->00:10:04,080
about like what's it doing, and it'll28300:10:04,080-->00:10:07,120
kick off kick off the query. And as it's28400:10:07,120-->00:10:08,880
going, it'll take some time to kick it28500:10:08,880-->00:10:11,200
off. Is it Oh, there we go. So, it'll28600:10:11,200-->00:10:12,480
start working on it. Looks like it's28700:10:12,480-->00:10:14,720
starting to create the anime art. It'll28800:10:14,720-->00:10:16,640
probably use image that Isa referred28900:10:16,640-->00:10:18,399
earlier on to create hopefully an anime29000:10:18,399-->00:10:20,240
art. We'll see how it comes out. While29100:10:20,240-->00:10:21,760
that's going, anything else we want to29200:10:21,760-->00:10:22,399do?29300:10:22,399-->00:10:24,720
Oh, yeah. I also need a pair of shoes29400:10:24,720-->00:10:26,320
because my shoes got damaged.29500:10:26,320-->00:10:27,360
How did they get damaged?29600:10:27,360-->00:10:28,560
Uh, by the rain29700:10:28,560-->00:10:30,000
in SF.29800:10:30,000-->00:10:30,800
Yes.29900:10:30,800-->00:10:32,160
Cool. All right. Uh, well, let's get30000:10:32,160-->00:10:34,240
Edward a pair of shoes as well. So, oh,30100:10:34,240-->00:10:40,320
can you also find us um pair of men's30200:10:40,320-->00:10:43,519
dress black shoes in size30300:10:43,519-->00:10:44,2409.5?30400:10:44,240-->00:10:46,0009.5.30500:10:46,000-->00:10:47,920
So, one of the key capabilities of the30600:10:47,920-->00:10:49,920
model is being able to interrupt. I30700:10:49,920-->00:10:51,920
think you know as trajectories take long30800:10:51,920-->00:10:53,760
time or whatever time it's really30900:10:53,760-->00:10:56,720
important for us to for it to feel very31000:10:56,720-->00:10:59,120
multi-turn so the users can interject31100:10:59,120-->00:11:01,120
user can direct it user can give it more31200:11:01,120-->00:11:02,640
guidance less guidance whatever we want31300:11:02,640-->00:11:04,320
to do and that's what we're doing here31400:11:04,320-->00:11:07,040
we essentially the the model was31500:11:07,040-->00:11:08,720
chugging along figuring out all the31600:11:08,720-->00:11:10,240
things that we had asked before and in31700:11:10,240-->00:11:12,320
this case we essentially said hey can31800:11:12,320-->00:11:16,000
you also uh get us a pair of men's black31900:11:16,000-->00:11:18,160
shoes and now it's thinking and soon32000:11:18,160-->00:11:19,839
enough hopefully it'll take that into32100:11:19,839-->00:11:22,000
account and keep going uh into its32200:11:22,000-->00:11:23,600
trajectory. There we go. So, it said32300:11:23,600-->00:11:25,120
acknowledge the interruption. It said,32400:11:25,120-->00:11:26,880"Okay, cool. I'll also research men's32500:11:26,880-->00:11:29,600
black shoes in size 9.5." Um, and then32600:11:29,600-->00:11:31,680
it'll probably get on its way. Um, but32700:11:31,680-->00:11:33,120
maybe Issa can tell us a little bit more32800:11:33,120-->00:11:34,240
about how that works.32900:11:34,240-->00:11:36,320
Yeah, sure. So, as you can see, the33000:11:36,320-->00:11:38,079
agent is very collaborative, and this33100:11:38,079-->00:11:39,920
was really important to us when we were33200:11:39,920-->00:11:41,200
training the model and building the33300:11:41,200-->00:11:42,880
product. If you were asking another33400:11:42,880-->00:11:44,399
person to do a task for you that would33500:11:44,399-->00:11:45,519
take them a really long time to33600:11:45,519-->00:11:46,959
complete, you'd probably give them some33700:11:46,959-->00:11:48,800
instructions to start and then they33800:11:48,800-->00:11:50,640
might ask you some clarifying questions33900:11:50,640-->00:11:52,320
and then they'd start the task and maybe34000:11:52,320-->00:11:53,600
realize, oh, they need more34100:11:53,600-->00:11:55,440
clarification from you or they need your34200:11:55,440-->00:11:56,880
permission to sign into something or do34300:11:56,880-->00:11:58,560
something on your behalf and then you34400:11:58,560-->00:12:00,240
might realize, oh, I forgot to mention34500:12:00,240-->00:12:02,640
this thing or um what's your status? How34600:12:02,640-->00:12:04,240
are you doing? Can I help redirect you34700:12:04,240-->00:12:05,760if you're getting along the wrong path34800:12:05,760-->00:12:07,760
or something? And so similarly for these34900:12:07,760-->00:12:09,680
really longrunning agentic tasks, it's35000:12:09,680-->00:12:11,519
very important that both the user and35100:12:11,519-->00:12:13,600
the agent are able to initiate35200:12:13,600-->00:12:15,519
communication with each other so that um35300:12:15,519-->00:12:17,200
the agent is able to most effectively35400:12:17,200-->00:12:19,360
help you with your tasks. And so this is35500:12:19,360-->00:12:20,560
something that we actually trained into35600:12:20,560-->00:12:22,320
the model. We trained it to be able to35700:12:22,320-->00:12:24,160
ask clarifying questions, not every35800:12:24,160-->00:12:26,240
single time like deep research. Um we35900:12:26,240-->00:12:28,800
also asked it we also trained it to be36000:12:28,800-->00:12:30,560
interruptible as Yash just showed. And36100:12:30,560-->00:12:32,000
also sometimes it will ask you for36200:12:32,000-->00:12:33,519
clarification and confirmation36300:12:33,519-->00:12:35,680
mid-trajectory.36400:12:35,680-->00:12:38,079
Yeah. And part of working with agent is36500:12:38,079-->00:12:40,480
that well sometimes it'll make mistakes.36600:12:40,480-->00:12:42,079
And that's why we felt it was important36700:12:42,079-->00:12:44,079
to train the model to ask you for36800:12:44,079-->00:12:45,920
confirmation at the last step of36900:12:45,920-->00:12:49,279
important steps. Um so for example maybe37000:12:49,279-->00:12:51,519
before it's going to send the email um37100:12:51,519-->00:12:53,440
it'll ask you to take a look at the37200:12:53,440-->00:12:54,720
draft and whether it makes sense and37300:12:54,720-->00:12:56,079
whether there are any embarrassing37400:12:56,079-->00:12:59,200
typos. Um, and if there are, then you37500:12:59,200-->00:13:01,360
can either ask it to fix it or you can37600:13:01,360-->00:13:03,440
directly take over the browser and jump37700:13:03,440-->00:13:06,079
right into the um, agents environment37800:13:06,079-->00:13:09,040
and correct it yourself. And that way it37900:13:09,040-->00:13:10,720
feels collaborative and you can um,38000:13:10,720-->00:13:13,680
really work with the agent.38100:13:13,680-->00:13:15,120
Should we look at maybe one more demo?38200:13:15,120-->00:13:17,279
We've got this uh, sort of fun tradition38300:13:17,279-->00:13:19,600
in live streams of using uh, using our38400:13:19,600-->00:13:21,120
newest models to sort of evaluate38500:13:21,120-->00:13:23,040
themselves or do something kind of meta.38600:13:23,040-->00:13:24,240
Anything like that we could do?38700:13:24,240-->00:13:27,440
Yeah, let's do it.38800:13:27,440-->00:13:28,320
So um38900:13:28,320-->00:13:29,440
I think people would love to know how39000:13:29,440-->00:13:30,320
good the model is.39100:13:30,320-->00:13:33,920
Yes. So this is a prompt we previously39200:13:33,920-->00:13:36,880
gave the a agent yesterday. So basically39300:13:36,880-->00:13:38,959
it asks the model to pull its own39400:13:38,959-->00:13:40,959
evalution number from our Google job39500:13:40,959-->00:13:43,440
connector and make some slides. So we39600:13:43,440-->00:13:44,959
want to keep it simple like no39700:13:44,959-->00:13:47,360
introduction no conclusion just present39800:13:47,360-->00:13:50,000
the results with in the charts. As you39900:13:50,000-->00:13:52,160
can see now the model is connecting to40000:13:52,160-->00:13:55,120
the Google Drive API and uh then search40100:13:55,120-->00:13:57,600
within API it right now it looks like40200:13:57,600-->00:13:59,920
the first result is very relevant. So40300:13:59,920-->00:14:02,720
it's reading the first result.40400:14:02,720-->00:14:04,959
Now it's reading the first result uh in40500:14:04,959-->00:14:07,920
details. Uh let's accelerate this uh40600:14:07,920-->00:14:12,800
replay. So then the model might read40700:14:12,800-->00:14:15,279
from the result again and write some40800:14:15,279-->00:14:16,959
code.40900:14:16,959-->00:14:19,519
So here you can see that the model is41000:14:19,519-->00:14:21,920
using the image generation model called41100:14:21,920-->00:14:24,480
image generation tool to generate some41200:14:24,480-->00:14:28,079
decorations for the slides.41300:14:28,079-->00:14:30,160
And let's see what's the first slide the41400:14:30,160-->00:14:33,399
model made.41500:14:33,920-->00:14:35,920
So here the model is writing some code41600:14:35,920-->00:14:38,399
that will be compiled to be the final41700:14:38,399-->00:14:41,120
slides. So this is the first slide the41800:14:41,120-->00:14:44,160
model make in this demo which looks okay41900:14:44,160-->00:14:46,240
but it's not polished enough.42000:14:46,240-->00:14:48,240
One of the key feature in reinforcement42100:14:48,240-->00:14:50,160
learning is that the model will re42200:14:50,160-->00:14:52,240
review its own results and refine the42300:14:52,240-->00:14:55,120
results to to deliver a good final42400:14:55,120-->00:14:57,839
results. Let's see what's the finally42500:14:57,839-->00:15:00,320
what the model give us.42600:15:00,320-->00:15:04,000
We can click skip and then the model42700:15:04,000-->00:15:07,519
give us a good uh PowerPoint file. So42800:15:07,519-->00:15:09,040
it's a real PowerPoint that you can42900:15:09,040-->00:15:14,040
download and open it in any software.43000:15:14,639-->00:15:19,279
Let's open it in uh in the office. So43100:15:19,279-->00:15:22,160
let's present the slides the model just43200:15:22,160-->00:15:23,839
generated.43300:15:23,839-->00:15:27,120
First are two intelligence benchmarks.43400:15:27,120-->00:15:30,480
Humanities last exam is a benchmark that43500:15:30,480-->00:15:33,519
measures AI's ability to solve a broad43600:15:33,519-->00:15:37,120
range of subjects on hard problems. We43700:15:37,120-->00:15:40,320
evaluate the models with two settings43800:15:40,320-->00:15:43,440
with and without tool use.43900:15:43,440-->00:15:45,920
We can see that the agent modes the raw44000:15:45,920-->00:15:48,720
intelligence is already pretty nice and44100:15:48,720-->00:15:50,880
with access to all tools nearly double44200:15:50,880-->00:15:54,720
the performance to 42%.44300:15:54,720-->00:15:56,720
When evaluating models on humanity's44400:15:56,720-->00:15:59,360
last exam, especially with the browsing44500:15:59,360-->00:16:01,759
ability, we have a two-layer44600:16:01,759-->00:16:04,399
decontamination that ensure that the44700:16:04,399-->00:16:07,680
model doesn't cheat on this benchmark.44800:16:07,680-->00:16:10,079
Front TMS is a benchmark that measures44900:16:10,079-->00:16:11,839
advanced mathematical reasoning ability45000:16:11,839-->00:16:13,680
of models.45100:16:13,680-->00:16:16,000
Different from our baseline of mini and45200:16:16,000-->00:16:18,56003 which use Python with function45300:16:18,560-->00:16:21,440
coding. We give the agent model all45400:16:21,440-->00:16:23,440
available tools like a browser, a45500:16:23,440-->00:16:26,320
computer and a terminal. The agent45600:16:26,320-->00:16:29,360
achieves new state art of 27% on this45700:16:29,360-->00:16:31,440
benchmark with the help of all these45800:16:31,440-->00:16:34,440
tools.45900:16:34,639-->00:16:36,880
Next, we evaluated the model on two46000:16:36,880-->00:16:39,519
agentic benchmarks. Web arena is a46100:16:39,519-->00:16:41,519
benchmark that measures web agents46200:16:41,519-->00:16:43,600
ability so to solve real world web46300:16:43,600-->00:16:47,279
tasks. The agent model improves over46400:16:47,279-->00:16:51,360
previous O3 model that powers the core.46500:16:51,360-->00:16:54,399
Browse comp is a benchmark we introduced46600:16:54,399-->00:16:56,240
earlier this year that measures the46700:16:56,240-->00:16:58,880
browsing agents ability to search and46800:16:58,880-->00:17:02,320
find uh how to locate information.46900:17:02,320-->00:17:03,839
The agent model significantly47000:17:03,839-->00:17:06,160
outperforms 03 and deep research on this47100:17:06,160-->00:17:11,679
benchmark achieving 69% pass rate.47200:17:11,679-->00:17:14,559
Finally, we care about how the users47300:17:14,559-->00:17:16,959
will benefit from our model in the real47400:17:16,959-->00:17:19,919
world. Spreadsheet bench is a benchmark47500:17:19,919-->00:17:21,919
that measures the model's ability to47600:17:21,919-->00:17:24,400
edit spreadsheets derived from the real47700:17:24,400-->00:17:28,079
world use case. Here the agent model47800:17:28,079-->00:17:30,480
with the liberal office and the computer47900:17:30,480-->00:17:34,000
tool can already solve 30% of the task48000:17:34,000-->00:17:36,480
when we give the model the access to the48100:17:36,480-->00:17:39,840
raw Excel file in the terminal which48200:17:39,840-->00:17:44,000
further boost the performance to 45%.48300:17:44,000-->00:17:46,000
Finally we evated the model on an48400:17:46,000-->00:17:48,000
internal banking benchmark. The bench48500:17:48,000-->00:17:49,760
this benchmark evaluated the model's48600:17:49,760-->00:17:52,559
ability to to conduct first to third48700:17:52,559-->00:17:55,679
year investment bank uh banking analyst48800:17:55,679-->00:17:58,799
tasks such as like putting together a48900:17:58,799-->00:18:00,559
three statement financial model for49000:18:00,559-->00:18:04,000
Fortune uh 500 company in this49100:18:04,000-->00:18:06,160
benchmark. The agent model significantly49200:18:06,160-->00:18:08,080
outperforms the previous deep research49300:18:08,080-->00:18:11,760
and all three models. As you can see49400:18:11,760-->00:18:13,919
this model is one of the most powerful49500:18:13,919-->00:18:16,080
model we've ever trained.49600:18:16,080-->00:18:18,960
It's not only good on benchmarks, it's49700:18:18,960-->00:18:22,480
also capable of reasoning, browsing, and49800:18:22,480-->00:18:24,720
tackling real world tasks at a level49900:18:24,720-->00:18:28,480
that we cannot imagine three months ago.50000:18:28,480-->00:18:31,600
That's right. Um, as Edward said, um, we50100:18:31,600-->00:18:32,799
think we've trained a very powerful50200:18:32,799-->00:18:35,280
model and a lot of the power comes from50300:18:35,280-->00:18:38,240
its ability to browse the internet. And50400:18:38,240-->00:18:40,240
as we know, the internet can be a scary50500:18:40,240-->00:18:42,400
place. There are all sorts of hackers50600:18:42,400-->00:18:45,120
trying to steal your information, scams,50700:18:45,120-->00:18:48,480
uh fishing attempts. Um and agent isn't50800:18:48,480-->00:18:51,120
immune to all these things. Um one50900:18:51,120-->00:18:53,360
particular thing we're worried about is51000:18:53,360-->00:18:55,520
a new uh attack called prompt51100:18:55,520-->00:18:57,120
injections.51200:18:57,120-->00:18:59,840
This is where let's say you ask agent to51300:18:59,840-->00:19:02,080
buy you a book and you give it your51400:19:02,080-->00:19:04,400
credit card information to do that.51500:19:04,400-->00:19:06,240
Agent might stumble upon a malicious51600:19:06,240-->00:19:08,559
website that asks it,"Oh, enter your51700:19:08,559-->00:19:10,400
credit card information here. it'll help51800:19:10,400-->00:19:12,799
you with your task. An agent, which is51900:19:12,799-->00:19:15,200
trained to be helpful, might decide52000:19:15,200-->00:19:18,080
that's a good idea.52100:19:18,080-->00:19:19,760
We've done a lot of work to try to52200:19:19,760-->00:19:22,320
ensure that this doesn't happen. We've52300:19:22,320-->00:19:24,240
trained our model to ignore suspicious52400:19:24,240-->00:19:27,120
instructions on on suspicious websites.52500:19:27,120-->00:19:29,039
We've also have uh we also have layers52600:19:29,039-->00:19:32,000
of monitors that kind of peer over the52700:19:32,000-->00:19:33,760
agent's shoulder and watch it as it's52800:19:33,760-->00:19:36,480
going um and stop the trajectory if52900:19:36,480-->00:19:38,799
anything looks suspicious. We can even53000:19:38,799-->00:19:41,919
update these in real time if new attacks53100:19:41,919-->00:19:44,160
are found in the wild.53200:19:44,160-->00:19:45,919
That said though, you know, this is a53300:19:45,919-->00:19:47,760
cutting edge product. This is a new53400:19:47,760-->00:19:50,000
surface and we can't stop everything.53500:19:50,000-->00:19:51,280
And so that's why I feel it's very53600:19:51,280-->00:19:52,559
important for the audience to be aware53700:19:52,559-->00:19:55,360
of the risks involved in using agent.53800:19:55,360-->00:19:57,440
And um we encourage users to be53900:19:57,440-->00:19:59,520
proactive in kind of thinking about how54000:19:59,520-->00:20:01,120
they share their information. You know,54100:20:01,120-->00:20:02,880if it's highly sensitive information,54200:20:02,880-->00:20:06,799
maybe don't share that. um maybe um uh54300:20:06,799-->00:20:08,799
use our features like takeover mode to54400:20:08,799-->00:20:10,799
directly input your credit credit card54500:20:10,799-->00:20:12,880
information into the browser instead of54600:20:12,880-->00:20:15,679
um giving it to agent. Um we feel like54700:20:15,679-->00:20:18,640
we've built a very powerful product but54800:20:18,640-->00:20:20,480
again it's important for our users to54900:20:20,480-->00:20:21,760
understand the risk involved.55000:20:21,760-->00:20:23,280
Yeah, I really want to emphasize that I55100:20:23,280-->00:20:25,520
think this is a new level of capability55200:20:25,520-->00:20:27,120
in AI. It's a new way to use AI, but55300:20:27,120-->00:20:28,799
there will be a new set of attacks that55400:20:28,799-->00:20:30,799
come with that. And society and the55500:20:30,799-->00:20:33,120
technology will have to evolve and learn55600:20:33,120-->00:20:34,320
how we're going to mitigate things that55700:20:34,320-->00:20:36,159
we can't even really imagine yet. Uh, as55800:20:36,159-->00:20:37,360
people start doing more and more work55900:20:37,360-->00:20:39,679
this way. Before I wrap up, should we56000:20:39,679-->00:20:41,840
check in on some of the tasks you kicked56100:20:41,840-->00:20:42,080
off?56200:20:42,080-->00:20:46,159
Yeah, let's do it. Um, okay. So, I am56300:20:46,159-->00:20:48,240
going to open a new tab and make sure56400:20:48,240-->00:20:51,840
that we can see the progress of our um,56500:20:51,840-->00:20:55,679
stickers as well. Okay. Let's see. All56600:20:55,679-->00:20:58,159
right. So, sounds like stickers are56700:20:58,159-->00:21:00,880
ready. Let me see what it actually Okay.56800:21:00,880-->00:21:03,200
So, cool thing. This is sort of the end56900:21:03,200-->00:21:06,720
end result of the took about 7 minutes.57000:21:06,720-->00:21:08,480
Highly likely figured out everything.57100:21:08,480-->00:21:09,840
We'll go back and look at the trajectory57200:21:09,840-->00:21:11,679
and see how it did. But at the end57300:21:11,679-->00:21:13,679
result, it looks like it's added to the57400:21:13,679-->00:21:15,360
cart. This is the subtotal. I can just57500:21:15,360-->00:21:17,360
go ahead and look at it and then figure57600:21:17,360-->00:21:20,000
out uh I can just take over at this57700:21:20,000-->00:21:21,600
point as Casey said to enter my credit57800:21:21,600-->00:21:23,039
card information and then place the57900:21:23,039-->00:21:25,200
order really quickly. model is asking58000:21:25,200-->00:21:27,120for confirmations, etc. as it's supposed58100:21:27,120-->00:21:29,280
to do. Let's just quickly browse through58200:21:29,280-->00:21:31,039
the trajectory and see what it actually58300:21:31,039-->00:21:33,280
did. Oh, it looks like it generated some58400:21:33,280-->00:21:35,840
stickers. Oh, look at that. That's what58500:21:35,840-->00:21:38,880
it generated sticker. Cool. So, yeah,58600:21:38,880-->00:21:40,640
that's the task. I think I can at this58700:21:40,640-->00:21:42,559
point finish up by myself or I can ask58800:21:42,559-->00:21:43,919
the model to actually go ahead and do it58900:21:43,919-->00:21:46,720for me as well. Let's check on the59000:21:46,720-->00:21:49,840
wedding. Okay, great. Looks like it just59100:21:49,840-->00:21:52,720
finished in the nick of time. Uh, okay,59200:21:52,720-->00:21:55,520
cool. So in this case, as as we said, we59300:21:55,520-->00:21:57,840
were looking for hotel, stress, uh59400:21:57,840-->00:22:01,919
suits, and also shoes. So it's come out59500:22:01,919-->00:22:03,520
with a pretty comprehensive report. It59600:22:03,520-->00:22:05,840
looks like wedding venue, date, when it59700:22:05,840-->00:22:10,240
is with the Zilla links, dress codes. It59800:22:10,240-->00:22:11,600
figured out like what the suit59900:22:11,600-->00:22:12,960
recommendation should be, where you can60000:22:12,960-->00:22:14,799
buy. Now I can go ahead and buy myself60100:22:14,799-->00:22:17,120
or I can ask the agent to go and buy for60200:22:17,120-->00:22:20,960
me. Um also figured out footwear hurdle60300:22:20,960-->00:22:23,360
options. It actually looked through all60400:22:23,360-->00:22:27,120
the oop sorry it looked through all the60500:22:27,120-->00:22:29,360
availability. You can see actually it60600:22:29,360-->00:22:31,440
gives screenshots of what it checked. In60700:22:31,440-->00:22:33,120
this case we use booking.com and it's60800:22:33,120-->00:22:35,280
able to do that. Also has gift60900:22:35,280-->00:22:37,360
suggestions etc. And next step I can ask61000:22:37,360-->00:22:39,760
it as you said the agent says hey if you61100:22:39,760-->00:22:41,520
need assistance purchasing any item or61200:22:41,520-->00:22:42,960
have any further adjustments let me know61300:22:42,960-->00:22:44,880
so we can do that. Um, and I want to61400:22:44,880-->00:22:46,320
show one last demo which we didn't61500:22:46,320-->00:22:48,640
really run live but I think it's really61600:22:48,640-->00:22:51,280
cool and especially because the folks61700:22:51,280-->00:22:52,880
who are getting married are really into61800:22:52,880-->00:22:57,679
MLB. U so we asked the agent uh to go61900:22:57,679-->00:22:59,679
and build an optimal itinary for62000:22:59,679-->00:23:02,640
visiting all 30 MLB stadiums in just in62100:23:02,640-->00:23:05,200case you're thinking of a satical uh and62200:23:05,200-->00:23:08,159
then design the optimal route prioritize62300:23:08,159-->00:23:10,960
Hello Kitty nights and whatnot and62400:23:10,960-->00:23:12,400
present a final plan as a detailed62500:23:12,400-->00:23:13,520
spreadsheet. I'll really quickly run62600:23:13,520-->00:23:15,440
through this. Um I think it's just so62700:23:15,440-->00:23:18,240
fun to see. So again like as we have62800:23:18,240-->00:23:20,720
thrown shown throughout the the live62900:23:20,720-->00:23:23,919
stream it uses a multitude of tools uses63000:23:23,919-->00:23:26,240
container the terminal use using the63100:23:26,240-->00:23:28,799
browser working through all the details.63200:23:28,799-->00:23:30,400
It'll probably use again back to the63300:23:30,400-->00:23:33,200
browser figuring out Hello Kitty nights63400:23:33,200-->00:23:36,559
and then sports stadium and whatnot. Oh63500:23:36,559-->00:23:39,520
let's see did I miss the Oh go map.63600:23:39,520-->00:23:42,080
building a map using code to actually63700:23:42,080-->00:23:43,919
build it out and then overall we get63800:23:43,919-->00:23:46,159
like a pretty solid result I think at63900:23:46,159-->00:23:48,880
the end takes 25 minutes to work where64000:23:48,880-->00:23:50,400
does the season start and what not you64100:23:50,400-->00:23:51,919
have a spreadsheet that you can quickly64200:23:51,919-->00:23:55,760
view inside just right inside Chad GBD64300:23:55,760-->00:23:57,919
you can map the journey cool looking map64400:23:57,919-->00:24:00,400
I guess and that's it so this is Chad64500:24:00,400-->00:24:02,240
GBD agent we hope you really like it and64600:24:02,240-->00:24:04,000
over to Sam64700:24:04,000-->00:24:05,919
amazing work all of you and and to your64800:24:05,919-->00:24:07,440
teams this is I think uh really64900:24:07,440-->00:24:08,720
something that's going to help people65000:24:08,720-->00:24:10,720
get worked done uh and have more time to65100:24:10,720-->00:24:12,240do the things they want to do. Um I65200:24:12,240-->00:24:13,520
think it's it's really amazing how much65300:24:13,520-->00:24:15,360
you've brought together to deliver this65400:24:15,360-->00:24:17,760
experience and watching the agent sort65500:24:17,760-->00:24:19,120
of use the internet, make these65600:24:19,120-->00:24:20,640
spreadsheets, make PowerPoints, whatever65700:24:20,640-->00:24:22,960else uh and do all this work is is quite65800:24:22,960-->00:24:26,000
amazing. We're going live today for pro65900:24:26,000-->00:24:28,880
plus and team users. Pro users will get66000:24:28,880-->00:24:30,720
uh 400 queries a month plus some team66100:24:30,720-->00:24:32,720
users will get 40 a month. Uh the66200:24:32,720-->00:24:34,000
rollout should be finished by the end of66300:24:34,000-->00:24:36,159
the day for pro and very soon for plus66400:24:36,159-->00:24:38,400
and team users. will try to be live for66500:24:38,400-->00:24:40,799
enterprise and edu by the end of this66600:24:40,799-->00:24:43,360
month. As Casey mentioned, although this66700:24:43,360-->00:24:45,360
is an extremely exciting new technology,66800:24:45,360-->00:24:48,080
there are new risks. Uh people learned66900:24:48,080-->00:24:49,520
how to use the internet generally pretty67000:24:49,520-->00:24:50,880
safely, although of course there are67100:24:50,880-->00:24:52,880
still scammers and other attacks. People67200:24:52,880-->00:24:54,559
are going to need to learn to use AI67300:24:54,559-->00:24:56,080
agents. Uh and societyy's going to need67400:24:56,080-->00:24:57,919
to learn to build up defenses against67500:24:57,919-->00:25:00,080
attacks on AI agents as well. So we're67600:25:00,080-->00:25:02,080
starting with a very robust system, lots67700:25:02,080-->00:25:04,240
of warnings. We will relax that over67800:25:04,240-->00:25:05,679
time as people get more comfortable with67900:25:05,679-->00:25:07,600
it. But we do want people to treat this68000:25:07,600-->00:25:09,919
as a new technology and a new risk68100:25:09,919-->00:25:12,080
surface and use all of the caution that68200:25:12,080-->00:25:14,799
Casey talked about. Um, but that said,68300:25:14,799-->00:25:16,720
we hope you'll love it. Uh, this is68400:25:16,720-->00:25:18,159
still very early. We will improve it68500:25:18,159-->00:25:20,640
rapidly and we're excited to see where68600:25:20,640-->00:25:22,640
it all goes. So, congrats again. Thank68700:25:22,640-->00:25:26,440
you very much. Hope you enjoy.
文章目录递归什么是递归为什么会用到递归如何理解递归如何写好一个递归搜索 vs 深度优先遍历 vs 深度优先搜索 vs 宽度（广度）优先遍历 vs 宽度（广度）优先搜索 vs 暴搜深度优先遍历 vs 深度优先搜索（dfs）宽度…
一、进程概念
在 Linux 系统中，进程（Process） 是程序执行的动态实例，是操作系统进行资源分配和调度的基本单位。
1. 程序 vs 进程
程序（Program）：是静态的代码集合&…
SpringBoot 是面试中后端开发岗位的高频考点，以下是核心考点整理：1. SpringBoot 基础概念- 定义：SpringBoot 是 Spring 框架的简化版，通过“自动配置”“起步依赖”等特性，简化 Spring 应用的搭建和开发，减…
在容器中使用 CGroups)
)