00:00Hello everyone and welcome to this very special conversation.
00:03Remember, the AI race is accelerating and the warnings are getting louder.
00:09The world's biggest tech companies are racing to build AI that is faster, smarter and more powerful.
00:16But now, some of the very people driving that race are asking,
00:21are we moving faster than we can control?
00:24And this is where the AI battle gets complicated
00:27because Washington does not want to lose its technological edge to Beijing.
00:32Tech giants do not want to be overtaken by their rivals.
00:37And nobody wants to be the country or the company that slows down first.
00:42So the debate is no longer simply about AI versus humanity.
00:48It is about AI versus human control.
00:51Some want more time for safeguards like Sam Altman,
00:56and tropic CEO Dario.
00:58Others say slowing down the frontier could cost the world its next technological revolution
01:06and hand that wanted to China, like US President Donald Trump.
01:10He says, we must not kill the golden goose.
01:14Meta CEO Mark Zuckerberg has said that an industry-wide slowdown is unnecessary.
01:20And behind this, clash is one fundamental question.
01:25If you know AI is dangerous, then why build it?
01:29If you know AI is dangerous, then how fast should we let it evolve?
01:35Because the race may not simply be to build the most powerful AI.
01:39It may be a race to ensure that when AI becomes powerful enough, humans are still in control.
01:46And to speak on this big AI versus humanity control debate,
01:51I am being joined by Dr. Vishal Sikha,
01:55who is a leading artificial intelligence expert,
01:59former CEO of Infosys,
02:00and founder of enterprise AI companies,
02:04including Vinay systems and Hang 10 systems.
02:09Sir, thank you for speaking to India today.
02:11For years, Dr. Sikha, AI was sold as a productivity tool.
02:16Now the people building it are wanting it could end humanity.
02:21Has the industry's own story about itself changed?
02:26And why?
02:29I think that this talk about AI ending humanity and all this,
02:36this is all nonsense.
02:38This is, I think we have to,
02:41I think most importantly, we have to learn about what it is.
02:46What is this technology?
02:48What is it capable of?
02:50What can we do with it?
02:55So, you know, one of the great teachers of my life,
02:59Alan Kay, once said that the music is not in the piano.
03:05So it's not AI that is going to do anything.
03:08It is people who use AI are going to do things.
03:12And of course, people who use AI in bad ways can do bad things,
03:16and people who use AI in good ways can do good things.
03:21It's as simple as that.
03:24If you look at the hugging face incident
03:30and similar incidents that have happened recently,
03:33there is no doubt that the models are extremely powerful.
03:36They are very, very large.
03:38They are incomprehensibly large.
03:40They are running on incredible amounts of training data
03:44and post-training data that they have been trained and post-trained on.
03:48And the computers that they are running on are unbelievably powerful,
03:52just even compared to a few years ago.
03:54So all that is true.
03:56And they have been trained on huge amounts of security-related data,
04:02logs, things like that.
04:05But in the end, when a model goes out and does these things,
04:11if you look carefully at what actually happened,
04:14it was a poor collection of instructions that the models were given.
04:18The models did nothing other than what was in their training data.
04:21So I think from there to jump to this thing
04:23that there is an X percent chance of humanity ending and all that,
04:28this is, I think this is quite, I find it quite silly.
04:35At the same time, I don't know if you remember,
04:38there was an NSA leak, I think it was called Eternal Leak
04:42or something like that.
04:44Right, right.
04:44That was a very powerful, what that did was,
04:49it made it possible for regular people to do very damaging things.
04:55And I think clearly such a powerful technology
04:58in the hands of people who have not been trained on how to use it
05:02can do very damaging things.
05:04But in the end, AI itself is not doing anything.
05:07It is people who are asking it to do things.
05:09In fact, the instructions in these incidents were quite clear.
05:13They were prompting these models to actually break out of the things
05:18that then they broke out of.
05:19So, I, you know, as someone who has been practicing in this field
05:24for a long time, I find this really, it is not amusing.
05:29Okay, so...
05:30Even though I'm laughing, you know.
05:32Every technological revolution has had its skeptics
05:36and history often proves them wrong.
05:39What makes this moment different?
05:42So, Dr. Sikka, are we projecting familiar fears onto a technology
05:48we don't yet fully understand, sir?
05:53It is the latter.
05:54It is a very powerful technology.
05:56It is...
05:57You can do extremely powerful and positive things with it.
06:02Also, you can do damaging things with it.
06:05But it is the latter that you said,
06:08that projecting fears on things that we don't understand, you know.
06:13We have always...
06:14Humanity has always done that.
06:15We start to either pray to the things that we don't understand
06:19or we start to become afraid of them.
06:21And it's the same kind of a thing that is happening here.
06:26So, LLMs are extraordinarily powerful.
06:29They capture vast amounts of data in very condensed ways
06:35and can reproduce things very coherently.
06:38And then on top of it, the advances,
06:41especially in the last year or so,
06:43around what we call reasoning models,
06:45and the ability to integrate these reasoning models with tools,
06:49that has made the systems extremely powerful.
06:53Extremely powerful and able to do lots of things.
06:57So, for example, when you take a very large model
07:00and then you post-train it
07:02to do particular kinds of reasoning chains,
07:06these chains can carry out long-running complex tasks.
07:12And then, especially the integration of tools,
07:14so, for example, in software development,
07:16when you integrate tools for program verification
07:20or tests, unit tests of software
07:23to test whether a piece of functionality
07:25was accurately implemented,
07:28once those got integrated into the chain of the model itself,
07:33this makes it possible to write very powerful software
07:36very efficiently.
07:39But still, people who provide these prompts,
07:41people who provide the markdown files,
07:44who provide the MCP protocols and so forth,
07:46these people have to understand what it is that they are building.
07:50And indeed, the burden of articulating what you are building with it
07:55becomes even higher on the people who are building this.
07:58Otherwise, damaging things can happen.
08:00You not only don't get the benefit of AI,
08:03when you do a naive or an ill-informed use of it,
08:07you can end up causing significant damage.
08:11And so I see this as a problem of a lack of understanding
08:14what this technology is capable of.
08:16And also, frankly, you know, people talk about regulation.
08:20But the reality is that, like, I am sitting under this roof here,
08:24confident that this roof is not going to fall on my head.
08:27You know, we have, I got a haircut yesterday,
08:30and the person who gave me a haircut had a license
08:33to, you know, use scissors on my head.
08:36So I don't understand what the big deal around regulating AI is.
08:40AI has to be regulated.
08:42Sir, I'm going to come to regulation in just a bit.
08:45But let me talk about the makers as you highlighted.
08:48Here is Jacob Coxon, who walks out of Anthropik,
08:52saying researchers privately believe that AI could be catastrophic.
08:55And within hours, you have Amodei, Sam Altman, Elon Musk,
09:02usually at odds with each other.
09:04All of them converge on the need to pace the frontier.
09:08When rivals who rarely agree on anything
09:12suddenly speak in one voice,
09:15should we read that as a genuine safety alarm then?
09:23Maria, why the blogs were written the way they were written?
09:29Why is there a call for slowdown?
09:33These are all different things.
09:35So the researcher who left and made this statement,
09:39which got really picked up,
09:41in principle, his statement was exactly like the statement
09:44that Mrinank, who left a few months earlier,
09:48made a similar statement.
09:50And I really respect him, Mrinank Sharma.
09:54And he's a distinguished researcher.
09:56He made exactly the same statement seven or eight months ago.
09:59And then a couple of years ago,
10:01there was another researcher from OpenAI
10:03who, in fact, went to the Senate
10:05and gave it a statement testimony there.
10:08And he made the same statement.
10:11So, you know, that's a personal thing
10:15that these people who are in elite frontier labs
10:20making these statements.
10:22Who knows what the reasoning for that is?
10:25The way I see it is the dangerous situation,
10:29the damaging situation is no more than people
10:34using these very powerful models to do damaging things.
10:39And indeed, arguably, giving instructions
10:42to specifically do damaging things
10:45and then being surprised that they are doing damaging things.
10:48This is how I see it.
10:50To the point about the slowdown and all of that,
10:53I've been thinking about it.
10:55I've been talking to friends.
10:56I've been reading analysis on that.
10:59Who knows what the motivation of these people is?
11:01They are playing with extremely, you know,
11:06large stakes here.
11:08And the amount of investment is very, very large.
11:12Who knows what the motivation is?
11:14I would, I think that a good maxim for this time
11:21is don't try to judge what is happening.
11:27Just try to understand what is happening.
11:29And I think I would say that.
11:31But I have no doubt,
11:33after having worked in this field for decades
11:35and being quite close to the field
11:37since the beginning of this particular wave of AI,
11:42that there is no, you know,
11:46all this talk about catastrophic doom and so forth,
11:51I think is overblown.
11:54Jensen seems to agree with you, Dr. Sikka.
11:58Jensen Huang has in fact pushed back
12:00on this entire doomsday narrative.
12:03And he has argued that AI will create more jobs
12:06than it destroys
12:06and that the focus should be on building
12:09and scaling responsibly.
12:10Then are we seeing a split within the tech world
12:14between those selling optimism
12:16and those warning of catastrophe?
12:21I think that has always been the case.
12:24I think even in the industrial revolution
12:26and computing and all these,
12:31I remember when there was a Pentium chip
12:37that Intel made,
12:38that was the fifth generation
12:39of their x86-based processors,
12:42had a floating point error.
12:44I don't know if you remember this,
12:46about 30 years ago.
12:47And that was like,
12:49that shook the foundation on the whole industry
12:51that if these things make errors.
12:54And, you know, if you remember,
12:56there was a beautiful movie made
12:59by the NASA ladies.
13:01What was it called?
13:03Hidden Figures.
13:04Hidden Figures.
13:06Because people used to calculate.
13:08Even the astronomical calculations
13:11when the first spaceships went into the space
13:14and so on,
13:15those calculations were made,
13:17done by hand by people.
13:19And when the computers first came out
13:21and started to do these calculations,
13:23people often used to ask,
13:24how do we know
13:24that these calculations are correct?
13:27Now we just take it for granted
13:28that these things
13:29are generating correctly,
13:31that the compilers
13:32are generating correct software
13:33and all of that.
13:34And I think this kind of reliability
13:36has to get built in
13:37to the technology.
13:38It will happen over time.
13:39Right now we are not there.
13:42But the trust in the systems
13:44will build over time.
13:45And I think then this talk
13:47starts to subside.
13:48But this kind of split in camps
13:51and people who are sort of doomsayers
13:56and all that,
13:56this has always been the case.
Comments