Himanshu Tyagi - Building Open AGI Loyal to Humanity - TOKEN2049 Dubai 2025

So, sentient is uh sentient is one of the major cryptoi projects. We are building open agi that is loyal to humanity. And in this talk, I would like to tell you what it means, what is this uh notion of open AGI and what do we mean that we want it to be loyal to humanity. So, ah right. So let's go over what happened with open-source AI and how it became what we now understand as closed AI open AI.

So everything obviously started open source and the culmination of all this work on AI all the different research decades of research was this Google's paper on transformer an architecture which they came up with which was best in capturing all the human language and that's what led to all the all the new models we are used to starting with these two first models birth from Google and GPT from open AI which were both open in 2018. They were both open and everything was open AI. Then 19 this GPT became into a much more powerful model GPA 2 which for the first time we heard that this model may not be open. It was closed in the beginning and then later on it was opened up. After that everything changed.

Everybody realized this is really really big. This economy is going to be really really big. It's you started hearing about tens and 20 billion dollar valuation. You started hearing about raises which are tens and 20 billion. Obviously dollars were also printed at that time.

So everything was very big after that and this open AI this research grand humanity saving AI became closed AI. It's now an artifact which is very powerful and this competition of closed monopoly started and three major models came out of it. We are all using it now. GPT which is like 75% of the market I would say and then Gemini which is the next second from Google it picked up now it had lost the battle I felt but now it has picked up and claude which you use in your coding application these three big closed models took over everything and this has been there for 3 4 years nothing happened open AI was sort of dead and suddenly around 23 that's when we started building sentient conceptualizing Towards the end of it, there was sort of a revival of these open models. They started with Llama Meta's model around November.

Then Mistral, then finally Deepseek was started there. And last year end and this year suddenly we see that open AI is everywhere. Something magical happened. China decided that it's GTM for AI will be around open source model. So if you now go and spend time in China, so ironic that every city is buzzing with this open-source model and everybody is very keen about is your model open source and it's it's very strange but somehow from led by China open source has come back.

So now open source and closed models are almost at par. If you try to build an application with open AI and if you try to build an application with Quinn, Deepseek, Llama combined roughly you get similar experience. You have to engineer more for open source models but roughly same experience. So this is the moment of revival of open AI. Still the market share is skewed.

Closed models have most of the market but closed the the usage of open models is rising fast much much faster than three times faster than closed models. So this is the this is the good part of the story. Why do we need this open models to keep up with closed models? So here is the thesis. We are all racing towards this AGI.

AGI for us is nothing but a very general purpose AI which can solve all kind of intelligence questions not just dumb query questions but can you know write product text then implement write the code for that product then market for that product and then just give you money and you just chill here in this place that's the world we want to get to right so this AGI can only come when AI compose and compete with each other and to have this kind of a fair environment where different AI is competing and composing one needs this AI to be open. So this large models that you see this monolithic models let's say 03 03 is one of the marquee uh reasoning model one of the best models from open AI these are slow expensive but they are very smart they can do a lot of analysis but they do just one kind of intelligence very good but what we need is many different variety of intelligence right we need a multitude of agents models software frameworks to do not just our intelligence queries but also actions associated with it. So this AGI that we are hoping for will not just come from single monolithic systems. They will come from this multi- aent multimodel capabilities built around them. And this is the thesis that AGI will evolve from open cooperation and competition between multiple AIS.

And it's not just my thesis although I must take some bragging rights for it. This is the thesis on which we raise for centi and this is the thesis we are after. But this thesis is a very mainstream thesis. Now lot of companies like Google, Enthropic are supporting frameworks to allow AIS to compose with each other and compete with each other and bring this out. So there are many many new tools coming out which allow you to combine your AI models and application into a singular experience.

At sentient our goal has been to build this kind of open AGI and for that we have this platform which we call sentient chat. Sentient chat is a singular platform where multiple agents and models can be given to you through a common experience. Underlying just like open AI there's a front end for open AI which you ask your queries to but behind that front end there are many different models. You ask some queries and they go to the different models and you get your answer. The difference between sentient chat and open AI setup are that these different models and then these different agents are not all built by us.

They're built by different players. Some of them are built by us. So far the best ones are still built by us. But it is open for others to integrate in this using A2A and these kind of frameworks and give you this unified intelligence. Not only chat agents and chat based models but also task related models and agents.

So that's what sentient chat is. There are all these uh many many different models governed through a single platform and uh giving this user thing experience. So behind this behind these agents and these models we have many many open-source frameworks which we are making available for you to build your own agents on as well. One of them is this library which we just released recently which is our open deep search library. One thing about all these agents is that when you ask them question their knowledge is the models have fixed knowledge but you need very updated knowledge.

So you want to give this models also access to all the internet, all the APIs, all the share market data, all the different data and combining that knowledge with this existing knowledge of the model requires a lot of engineering. We did that for you and made it available. We use it to build our agents. You can use it to extend it to different agents and plugin in sentient chat. This is our ODS library.

Okay. So this is what we are building open AGI. But there is a problem here. When a open AI builds its AI, it can be malicious but it is actually governed by some regulators in the US and if they don't do their job well then maybe someone will take an action and they will somebody will go to jail. But what happens when things turn south in this open AGI ecosystem this wild wild west of AI when you use are what happens when things are going wrong just to take an example a very popular open-source project in crypto is Eliza OS which is a framework for building your agents the AI6Z's open uh framework for building agents now it's a very powerful thing and people are quite excited about it but when you see an open artifact like that there are many many hidden vulnerabilities in it.

So if you give your wallet to agents built on Eliza OS, is it possible that someone can take money from that wallet uh instead of without your permission? So in fact we did show such attacks on Eliza OS because the onus of securing Eliza OS is on the user who builds on it. So when you build an open is always this issue that the what about security? What about deep bugs and what about the vulnerabilities? So our goal is to make sure that this AI which is open remains loyal to people who govern it.

So that's what we call loyal AI. It is AI that is owned by us a community that is aligned to that community's interests, its values and preferences. And it's controlled by us in the sense that certain actions which are not allowed for example transferring money from your wallet to someone else's wallet without your permission are not not per not permitted in that stack. That's what we call loyal AI. Loyal AI is something which is communityowned aligned and controlled.

So not only build an open but build this loyal AI. When when you look at the AI stack now suppose you want to build AI to be loyal to you. There are three parts to that stack. First part is the software layer. Eliza OS is that software layer where all the memory part is there where all the uh sandbox for running code is there because AI is becoming very sophisticated.

It not only accesses a model it has various tools to call it can code. It can access some API keys to call some services. It has all that access. So the software and the attack surface is very wide. It can do many many uh dangerous things.

So one has to design this OS keeping in mind all those possible combinations and make sure that there is no attack that is allowed. The next level of defense which you add to loyal AI is prompting. Prompting is very simple defense. It's how you teach your kid to be nice. You scream at the kid and the kid is nice.

So you write in bold, do not ever steal the money. Do not ever transfer. That's what prompting is. And hopefully the model has been trained to follow instructions, these specific instructions, and it goes through. It's actually a joke.

But that's where we are in AI. You are just screaming at the model to behave properly. I don't know if you've seen prompts. Many of the prompts have exclamation marks, bold font. It is to impress upon the model to take this seriously.

There's a second line of defense in all these in this world. The final line of defense is the model itself. The most opaque part of defense where its own alignment right what are the beliefs of the model. You can ask some hard questions to the model. Just to take an example if you look at llama Facebook's open model and if you ask the model now today let's say 3.

1 not the latest one can Trump be the president of US it will never agree to it. It it has hard belief that Trump can never be the president of US. It's knowledge was logged in 23. So it's it's outdated. But the belief is very hard.

In fact, what I found very amazing was that Lama is ready to believe that Democrats will lose. But it's very hard to convince it that Trump will become the president. That's how hardcoded these beliefs are in the model. Now imagine a model which is reasoning with that hardcoded belief. It will always have certain bias in all kind of reasoning all kind of thing.

All the investment that it will suggest for you are all driven by that inherent bias in the model. And so the final line of defense in creating this loyal AI is training the model itself around the constitution and belief which you want it to have. So these three layers build a airtight software which doesn't allow for vulnerabilities. prompt properly and finally believe train the model to be loyal. So we have been building this uh tech to create loyal AI.

We have a TE stack which allows us to uh secure the software. Prompting we don't do much. I don't believe in prompting but I don't do much there. And finally model training. There are all these uh very fine grain model uh alignment pipeline that we have built through which you can train to train to change the belief of the model at a very fine grain level.

It's uh censorship its belief in being pro crypto whether it's being pro Solana or pro- ETH you can train it in all that fine level of granularity. Okay. So using this we built one of one of the first uh community governed model. This model is called Dobby. It's governed by 600,000 people who came together and decided what kind of personality do they want for that model.

So they wanted to be unhinged. It's a little bit of an abusive model like Grock. They wanted to be very pro- crypto. They wanted to be very pro- freedom. And based on their based on their uh inputs and their preferences, this model was finally trained and they can they can identify this model through the fingerprints that are put in this model.

And this is the first communitybuilt loyal model that uh sentient has built. Yeah, it's a little bit fun. You can go and try it. Right. So all this is good.

But how does how does it finally come together? So one thing we are seeing is that everybody is building all this AI. All these YC startups are building all different kind of AI. They raise 2 million, 3 million, 4 million and then there is this red ocean of competition in which all of them sink. What's the next stage?

The next stage is what Bit Tensor showed us. The ne next stage is that crypto needs to help them get through that initial cycle by giving them those incentives or by giving them you know users or data incentivized by crypto and bit tensor does it but it doesn't tie it to any kind of output. So sentient is following a similar approach. We are creating a crypto incentivized layer for creating a joint model agent offering through sentient chat. All these agents and models will get rewarded in the emissions based on based on how much uh usage they get on sentient chat and based on how much stake is there on these models and that's what this economy is about and we are building it.

Initial traction is showing is very good. Please come and try our private beta for sentient chat. Download Dobby, let it abuse at you and have fun. Thank you very much.