WEBVTT

00:00:00.000 --> 00:00:03.459
Imagine an AI. It's incredibly smart yet. Well,

00:00:03.500 --> 00:00:05.740
it's a master bluffer. Right. It gives these

00:00:05.740 --> 00:00:08.740
really confident but totally wrong answers. Why

00:00:08.740 --> 00:00:11.359
would it even do that? That's what we're diving

00:00:11.359 --> 00:00:14.919
into today, this curious world of AI honesty.

00:00:15.339 --> 00:00:19.129
Welcome to the deep dive. We are... We're really

00:00:19.129 --> 00:00:21.710
going to unpack the complex reality of artificial

00:00:21.710 --> 00:00:24.149
intelligence as it stands right now. We've got

00:00:24.149 --> 00:00:26.809
some truly fascinating sources lined up. Yeah,

00:00:26.870 --> 00:00:28.870
we're looking at OpenAI's latest research on

00:00:28.870 --> 00:00:32.710
those tricky AI hallucinations. Turns out it's

00:00:32.710 --> 00:00:34.990
not quite what we thought. Okay. And we'll also

00:00:34.990 --> 00:00:36.530
get into some cutting -edge user hacks, people

00:00:36.530 --> 00:00:38.869
doing amazing things, plus a pretty surprising

00:00:38.869 --> 00:00:42.090
report on corporate AI adoption. Might make you

00:00:42.090 --> 00:00:44.710
think twice. So our mission today, let's try

00:00:44.710 --> 00:00:48.240
to understand AI's... Current capabilities, it's

00:00:48.240 --> 00:00:51.380
unexpected quirks and how it's truly impacting

00:00:51.380 --> 00:00:53.359
our world like right now. Let's do it. Okay,

00:00:53.420 --> 00:00:55.759
let's unpack this. So first up, this idea of

00:00:55.759 --> 00:00:59.079
AI honesty or maybe dishonesty, those hallucinations.

00:00:59.079 --> 00:01:01.340
We've probably all seen them. Oh, yeah. The confident

00:01:01.340 --> 00:01:04.859
nonsense. Exactly. But OpenAI's new research

00:01:04.859 --> 00:01:08.019
suggests these aren't just like random glitches.

00:01:08.019 --> 00:01:10.239
They're saying it's more like a trained behavior.

00:01:11.340 --> 00:01:14.140
And it's not just the models themselves. It's

00:01:14.140 --> 00:01:16.879
partly how we judge them. Oh, so? Well, think

00:01:16.879 --> 00:01:19.819
of it like a multiple choice test. We, or rather

00:01:19.819 --> 00:01:23.099
the benchmarks we use, often reward a lucky guess.

00:01:23.200 --> 00:01:26.920
An AI saying, I don't know, that gets zero points,

00:01:27.000 --> 00:01:29.859
basically penalized. Oh, okay. But a polished,

00:01:29.920 --> 00:01:32.379
confident, wrong guess, sometimes that actually

00:01:32.379 --> 00:01:34.299
scores pretty well in the current systems. So

00:01:34.299 --> 00:01:36.420
we're kind of inadvertently teaching them to

00:01:36.420 --> 00:01:39.420
bluff. Wow. So the bigger models, like the really

00:01:39.420 --> 00:01:43.239
powerful ones, GPT -4, maybe GPT -5 soon, they're

00:01:43.239 --> 00:01:45.939
more likely to do this. Seems that way. They

00:01:45.939 --> 00:01:47.799
often fake confidence when they only have partial

00:01:47.799 --> 00:01:49.780
info. They bluff instead of just admitting they're

00:01:49.780 --> 00:01:53.040
not sure. That feels wrong. Risky. It is. And

00:01:53.040 --> 00:01:54.920
interestingly, the smaller models, they tend

00:01:54.920 --> 00:01:57.120
to be a bit safer. They're apparently more likely

00:01:57.120 --> 00:01:58.760
to just throw up their hands and say, I don't

00:01:58.760 --> 00:02:02.519
know. But what's really fascinating here is OpenAI's

00:02:02.519 --> 00:02:05.299
proposed fix. They're saying we need to redesign

00:02:05.299 --> 00:02:08.400
the evaluation metrics, reward calibrated responses.

00:02:09.280 --> 00:02:12.199
That means like. Give partial credit for admitting

00:02:12.199 --> 00:02:15.240
uncertainty. I'm 80 % sure about it. Okay. And

00:02:15.240 --> 00:02:18.520
penalize a confident wrong answer much, much

00:02:18.520 --> 00:02:21.539
more heavily than a simple, I'm not sure. That

00:02:21.539 --> 00:02:23.639
feels critical. I mean, imagine this in medicine

00:02:23.639 --> 00:02:28.219
or law. Exactly. A confidently wrong AI answer

00:02:28.219 --> 00:02:31.520
there could be genuinely dangerous. Silence or

00:02:31.520 --> 00:02:34.479
admitting uncertainty is actually safer. We need

00:02:34.479 --> 00:02:36.990
to know when it doesn't know. So it really changes

00:02:36.990 --> 00:02:38.889
how we should think about interacting with these

00:02:38.889 --> 00:02:41.349
things. Totally. We as users probably need to

00:02:41.349 --> 00:02:44.050
start expecting more I'm not sure answers. Yeah.

00:02:44.110 --> 00:02:46.030
And, you know, accept them. See it as a feature,

00:02:46.069 --> 00:02:47.969
not a bug. Right. So what's the core message

00:02:47.969 --> 00:02:51.449
for us, the users, from this part? Expect and

00:02:51.449 --> 00:02:54.289
accept AI saying I'm not sure. Value honesty

00:02:54.289 --> 00:02:55.849
over bluffing. Okay. That makes a lot of sense.

00:02:56.219 --> 00:02:58.900
Let's shift gears then. From the lab almost to

00:02:58.900 --> 00:03:02.219
the real world, what are people actually doing

00:03:02.219 --> 00:03:04.280
with AI? Oh, it's kind of wild out there. So

00:03:04.280 --> 00:03:06.620
much ingenuity. Like one user shared this thing

00:03:06.620 --> 00:03:08.900
they called a mega prompt. Mega prompt. Yeah.

00:03:08.979 --> 00:03:11.620
You basically feed it to chat GPT or Mistral

00:03:11.620 --> 00:03:16.219
or Gemini and it turns the AI into like a 247

00:03:16.219 --> 00:03:18.800
research agent for you. Whoa. Constantly digging,

00:03:18.939 --> 00:03:22.370
synthesizing info. And it's free. Imagine having

00:03:22.370 --> 00:03:24.189
that kind of power just running in the background.

00:03:24.530 --> 00:03:26.750
Game changer for research. That's incredible.

00:03:27.009 --> 00:03:29.409
And OpenAI itself is adding new tools too, right?

00:03:29.469 --> 00:03:32.669
Like chat branching. Yeah. In chat GPT. So you

00:03:32.669 --> 00:03:34.449
could be in a conversation, have a side question

00:03:34.449 --> 00:03:37.110
pop up. Yeah. You can just fork the chat, go

00:03:37.110 --> 00:03:39.189
down that rabbit hole, and then jump right back

00:03:39.189 --> 00:03:41.310
to where you were without losing your original

00:03:41.310 --> 00:03:44.430
thread. Oh. It's like mental bookmarks for your

00:03:44.430 --> 00:03:46.789
AI chats. Super useful. I can see that being

00:03:46.789 --> 00:03:49.189
really helpful for complex stuff. And get this,

00:03:49.310 --> 00:03:52.569
lovable's voice mode. It uses 11 labs tech. It

00:03:52.569 --> 00:03:55.069
lets you code and build applications just by

00:03:55.069 --> 00:03:57.289
talking to it. Seriously, just voice command.

00:03:57.409 --> 00:04:00.530
Yeah. Imagine scaling up, handling like a billion

00:04:00.530 --> 00:04:03.110
queries without ever touching a keyboard. Whoa.

00:04:03.330 --> 00:04:05.669
I mean, think about the accessibility implications

00:04:05.669 --> 00:04:08.250
there. That's mind blowing, actually. OK, so

00:04:08.250 --> 00:04:11.669
we have these powerful practical uses, but there's

00:04:11.669 --> 00:04:13.389
also some weirder stuff happening, right? Some

00:04:13.389 --> 00:04:16.970
speculation. Oh, yeah. The fun stuff. So there's

00:04:16.970 --> 00:04:19.620
this theory floating around about Sonoma. Maybe

00:04:19.620 --> 00:04:22.839
being a secret Grok variant. Grok, Elon Musk's

00:04:22.839 --> 00:04:26.689
AI. How so? Well, the theory goes it reads invisible

00:04:26.689 --> 00:04:29.189
Unicode characters seems to really like the number

00:04:29.189 --> 00:04:32.250
42. You know, Hitchhiker's Guide stuff. Okay.

00:04:32.290 --> 00:04:34.189
And it shows some other little quirks that remind

00:04:34.189 --> 00:04:36.990
people of Grok. Some are jokingly calling it

00:04:36.990 --> 00:04:41.050
Grok 4 .20 in disguise. Probably just a coincidence.

00:04:41.269 --> 00:04:43.709
But it shows how people are trying to find personality

00:04:43.709 --> 00:04:46.970
or secrets in these black boxes. Totally. And

00:04:46.970 --> 00:04:50.850
on the more concrete corporate side. Yeah. Atlassian.

00:04:51.339 --> 00:04:53.399
Big news there. What'd they do? They're apparently

00:04:53.399 --> 00:04:55.680
buying the browser company, the folks who make

00:04:55.680 --> 00:04:57.899
the Arc browser. Yeah. For something like $610

00:04:57.899 --> 00:05:00.620
million. Wow, that's a lot. Why? Their goal.

00:05:00.740 --> 00:05:03.439
Build an AI -first browser specifically designed

00:05:03.439 --> 00:05:05.819
for work. So think about your browser not just

00:05:05.819 --> 00:05:08.000
showing you stuff, but actively helping you do

00:05:08.000 --> 00:05:09.819
stuff. Interesting. The browser as an assistant.

00:05:10.180 --> 00:05:12.180
Exactly. And just a couple of quick hits, too.

00:05:12.339 --> 00:05:14.660
People are starting to talk about using AI to

00:05:14.660 --> 00:05:17.060
get, like, therapy recommendations. Not therapy

00:05:17.060 --> 00:05:19.379
itself, but finding the right therapist. Hmm.

00:05:19.800 --> 00:05:22.240
That's a sensitive area. Needs careful thought.

00:05:22.420 --> 00:05:25.560
Definitely. Also, rumors that Google AI mode

00:05:25.560 --> 00:05:27.920
might become the default search experience. Maybe

00:05:27.920 --> 00:05:30.720
soon. That would be a massive shift for, well,

00:05:30.759 --> 00:05:33.459
everyone. Huge. And DeepSeek is planning a big

00:05:33.459 --> 00:05:38.439
AI agent release for late 2025. So AI not just

00:05:38.439 --> 00:05:40.540
answering, but doing tasks for you. That's the

00:05:40.540 --> 00:05:42.180
next frontier. It really feels like things are

00:05:42.180 --> 00:05:45.519
accelerating, tools popping up everywhere. What's

00:05:45.519 --> 00:05:48.879
the common thread in all these user stories and

00:05:48.879 --> 00:05:51.459
developments? Users are pushing AI's boundaries,

00:05:51.740 --> 00:05:54.339
finding creative and powerful new applications

00:05:54.339 --> 00:05:57.040
daily. Right. So we've seen the potential, the

00:05:57.040 --> 00:05:59.930
innovation. It's pretty exciting. But let's look

00:05:59.930 --> 00:06:01.569
at the bigger picture. What about corporations?

00:06:01.569 --> 00:06:05.730
Is that initial like AI frenzy cooling off a

00:06:05.730 --> 00:06:07.290
bit? Well, it's interesting you ask that. There's

00:06:07.290 --> 00:06:09.430
some new U .S. census data, pretty fresh, that

00:06:09.430 --> 00:06:12.310
suggests maybe we have a slowdown anyway. What

00:06:12.310 --> 00:06:14.250
does it show? Specifically for larger firms,

00:06:14.370 --> 00:06:17.310
250 employees or more, AI adoption seems to be

00:06:17.310 --> 00:06:19.470
trending down slightly. They've been asking companies

00:06:19.470 --> 00:06:21.470
like over a million of them every two weeks.

00:06:21.709 --> 00:06:24.430
Have you used AI tools in the last two weeks?

00:06:24.779 --> 00:06:27.759
And fewer are saying yes. The trend is slightly

00:06:27.759 --> 00:06:30.879
downwards, yeah. It suggests that maybe after

00:06:30.879 --> 00:06:33.860
that initial rush to put AI in everything, companies

00:06:33.860 --> 00:06:36.680
are taking a step back. So a reality check. Maybe

00:06:36.680 --> 00:06:38.800
companies just jammed AI in everywhere without

00:06:38.800 --> 00:06:40.879
thinking. That seems to be part of it. People

00:06:40.879 --> 00:06:43.399
are talking about ripping out useless add -ons

00:06:43.399 --> 00:06:45.819
now. Yeah. Things that didn't actually provide

00:06:45.819 --> 00:06:48.620
real value. This might be the first, you know,

00:06:48.660 --> 00:06:51.620
crack in the AI hype cycle. Or maybe just...

00:06:52.029 --> 00:06:54.149
Getting smarter about it. Not abandoning it,

00:06:54.230 --> 00:06:56.430
but being more strategic. That's probably the

00:06:56.430 --> 00:06:58.829
more likely scenario, yeah. And it raises the

00:06:58.829 --> 00:07:02.209
question, what survives this kind of recalibration?

00:07:02.569 --> 00:07:04.790
What do you think? Probably only the mature tools.

00:07:05.009 --> 00:07:08.110
The ones that show clear ROI. Those kind of crash

00:07:08.110 --> 00:07:09.970
and burn experiments. The ones that were just

00:07:09.970 --> 00:07:13.449
AI for AI's sake. They likely won't last. Workplace

00:07:13.449 --> 00:07:15.800
reality is setting in. And where is it working

00:07:15.800 --> 00:07:17.980
well in the workplace based on what we're seeing?

00:07:18.139 --> 00:07:20.339
It definitely seems to excel at more mundane

00:07:20.339 --> 00:07:24.579
tasks. Things like internal search, proofreading,

00:07:24.819 --> 00:07:27.060
maybe tweaking phrasing for emails or reports,

00:07:27.279 --> 00:07:30.139
repetitive stuff. Right. Freeing up humans for

00:07:30.139 --> 00:07:32.480
other things. Exactly. But it's still, you know,

00:07:32.480 --> 00:07:34.699
pretty weak for actual writing from scratch,

00:07:34.899 --> 00:07:37.720
especially anything requiring nuance or strategy

00:07:37.720 --> 00:07:41.329
or real creativity. I mean, I still wrestle with

00:07:41.329 --> 00:07:43.889
prompt drift myself sometimes, you know, where

00:07:43.889 --> 00:07:46.990
the AI kind of wanders off topic or changes style

00:07:46.990 --> 00:07:49.149
unexpectedly when I try to get it to do something

00:07:49.149 --> 00:07:52.470
genuinely creative. It's hard. Yeah, I've seen

00:07:52.470 --> 00:07:55.290
that too. And a lot of executives, they tested

00:07:55.290 --> 00:07:58.350
AI broadly at first, put it everywhere. But it

00:07:58.350 --> 00:08:00.670
turns out not all those use cases actually added

00:08:00.670 --> 00:08:03.750
tangible value. Plus, let's be real, a lot of

00:08:03.750 --> 00:08:07.459
AI tech still sounds. Like generic LinkedIn posts.

00:08:07.759 --> 00:08:09.879
Exactly. You can often spot it. And some folks

00:08:09.879 --> 00:08:12.379
are still maybe rightly skeptical calling it

00:08:12.379 --> 00:08:14.939
a misinformation machine if the accuracy isn't

00:08:14.939 --> 00:08:16.839
rock solid. Which it often isn't yet. Right.

00:08:16.899 --> 00:08:19.339
So in a tighter market where budgets are scrutinized,

00:08:19.439 --> 00:08:23.879
that AI for AI's sake stuff is getting cut. They

00:08:23.879 --> 00:08:26.000
want results. So what's really driving this corporate

00:08:26.000 --> 00:08:28.800
recalibration then, boiling it down? Companies

00:08:28.800 --> 00:08:33.059
want clear ROI. AI for AI's sake isn't proving

00:08:33.059 --> 00:08:36.299
its value. Okay. Makes sense. It's maturing.

00:08:36.879 --> 00:08:40.580
Sponsor. So wrapping things up today, we've journeyed

00:08:40.580 --> 00:08:44.100
through the the fascinating, sometimes kind of

00:08:44.100 --> 00:08:47.360
frustrating world of AI from its built in tendency

00:08:47.360 --> 00:08:49.779
to bluff. Which OpenAI is trying to train out

00:08:49.779 --> 00:08:51.899
of it. Right. To the just incredible ways people

00:08:51.899 --> 00:08:53.899
are using these tools, pushing the boundaries.

00:08:54.039 --> 00:08:57.179
And then that dose of reality from the corporate

00:08:57.179 --> 00:08:59.259
world. Yeah, we're definitely seeing AI evolve

00:08:59.259 --> 00:09:01.940
super rapidly. The power is obvious. Right. But.

00:09:02.250 --> 00:09:04.350
you know, so are the limitations. And there's

00:09:04.350 --> 00:09:07.090
this critical need for better ways to evaluate

00:09:07.090 --> 00:09:09.850
it and for more focused, actually valuable applications.

00:09:10.230 --> 00:09:13.710
This whole conversation about trust in AI, it

00:09:13.710 --> 00:09:15.190
feels like it's really just getting started.

00:09:15.289 --> 00:09:17.049
What does trust even mean with these systems?

00:09:17.190 --> 00:09:20.129
It's complex. It really is. We hope this deep

00:09:20.129 --> 00:09:22.450
dive helps you, listening out there, navigate

00:09:22.450 --> 00:09:24.330
this landscape with a bit more clarity, more

00:09:24.330 --> 00:09:26.429
insight. As you use these tools, maybe think

00:09:26.429 --> 00:09:28.889
about how you approach AI. Yeah, like... Are

00:09:28.889 --> 00:09:31.429
you accidentally encouraging it to bluff just

00:09:31.429 --> 00:09:33.629
to get an answer? Or are you looking for that

00:09:33.629 --> 00:09:35.909
honesty, that transparency about what it doesn't

00:09:35.909 --> 00:09:37.669
know? That's a good question to ask yourself.

00:09:37.950 --> 00:09:40.850
And here's a final thought to chew on. What if

00:09:40.850 --> 00:09:44.190
AI's greatest strength down the line isn't just

00:09:44.190 --> 00:09:47.429
raw intelligence, but maybe a new kind of honesty?

00:09:48.029 --> 00:09:50.889
An ability to clearly admit its own limits. Hmm.

00:09:51.269 --> 00:09:53.009
That would certainly make it a more reliable

00:09:53.009 --> 00:09:54.549
partner. Definitely something to think about.

00:09:54.690 --> 00:09:57.029
Thank you for joining us on this deep dive. Until

00:09:57.029 --> 00:09:58.580
next time. Keep exploring.