see there is this aspect of yeah Bad actors taking advantage of some of the Innovations in open source so clearly there is a long-term strategic Advantage for meta and that's why they're clearly motivated at this point of time so those are aspects that uh definitely will Foster more and more of innovation and collaboration in the space hello and welcome to AI Connect episode 12 on this episode we are going to talk about two very important and relevant topics with respect to AI the first topic is related to open source driven generative Ai and the development

and Innovation that is happening in that space and the second topic is going to be focused on AI powered automation both are important topics but let's first start with the first topic um let me welcome first Shiva Shiva welcome to AI Connect episode 12 I'm Hing that you I'm hoping that you are all geared up for today's discussion yes yes all right okay so let's dive in Shiva so talking about the open source specifically in the context of La meta launched llama 3 in July 2024 and uh I think with passing months we see that

meta is taking an aggressive stance in the open source driven by their their llm llama I want to touch upon today about the impact of Open Source in AI basically AI driven Innovation right how do you see open source AI for example meta's Lama influ influencing the pace of innovation across different Industries so if you look at traditionally the open source movement has really democratized the use and access of software across right from the days of Unix Linux and other open source Technologies and my feeling is we can expect the same with llm and generative

AI Technologies as well uh with generative AI allowing developers to inspect the inner workings of the model like the model weight is one of the key for future development of this technology I would say because you can look at Innovative applications innovative ways of utilizing infrastructure with all this open source coming in and by lowering entry barriers to coders around the world open- source llms can foster Innovation and improve the models by reducing biases increasing accuracy and the overall performance as well okay so one of the inherent I think proposition or promise that open source

brings is collaboration right do you think because collaboration is an important factor in to drive Innovation that's going to Foster or faster Innovation with the adoption of more open source based llms correct correct absolutely like organizations which are very keen on sort of using some of the smaller language models and then open- Source Foundation models like Lama can be downsized through various techniques so those are aspects that uh definitely will Foster more and more of innovation and collaboration in the space one of the first thing that comes in mind that how do you balance the

transparency and the security concerns when it comes to open source and some of it could be our bias or some old experiences but I wanted to highlight that meta in in this case particularly argues that open source AI is uh much safer due to this transparency and and wider scrutiny because they're more exposed to public so what are your thoughts um on the balance between transparency and the potential risk of misuse of an open source AI environment look that is still early days I would say to really comment on that because there is a lot

of innovation lot of things happening on the ground and see there is this aspect of yeah Bad actors taking advantage of some of the Innovations in open source and then utilizing them for nefarious purposes as well so there is one Chain of Thought around that but in general yes what MAA is saying has been true for all the software transparent open source has fostered transparency and more security but then there are also examples of more closed ecosystems like the one touted by Apple right which again places a lot of emphasis on security as well I

would say still early days but what I would I'm also seeing is organization which are keen on protecting data within their firewalls or within the Enterprise are reluctant to adopt some of the close Source Technologies like GPD or cloud and more gravitating towards deploying a model like llama in their private cloud or on Prem systems to utilize the power of LMS as well in general I would say it is still early days to pass a verdict on this but there there are pros and cons of each of these approaches with respect to security I think

uh one valid point that you brought up here is obviously I think we'll have to look at the Enterprise segment a little more and it is still early days right as you mentioned the other two aspects obviously there there's a promise of lower cost and therefore there is an economic benefit that automatically comes with open source and then there is an ethical implications as well so how do you foresee this affecting the economic landscape in general with open source uh and even some of the other ethical considerations especially in developing countries or um smaller Enterprises

from a economic point of view obviously since the open- source software is free the uper cost for let's say a model is not there but at the same time there is a cost which is going to be incurred from an infrastructure standpoint you will need developers or data scientists with little more sort of experience in utilizing some of these models to really develop a business application for you so those costs do not go away whether you're using open source or a close Source model the costs have to be looked at holistically and what we have

seen is whenever the volumes are not very high purely from an economic standpoint it makes sense to you use a pay as you go model which is probably close Source rather than trying to deploy an opal Source model on your infrastructure and train it Etc uh so it's going to be scale issue from a cost perspective and from ethical point of view see all these models have been trained on data that is publicly available and then there are aspects around copyright infringements Etc that organizations are battl out I guess the verdict is not out yet

there from an Ethics standpoint um it's really uh very hard to distinguish from an ethical point of view whether you're open source or closed Source whichever is better right I think this debate will evolve and will take shape in various forms before it gets set got it so clearly right now the promise is obviously the cost and secondly I think it is the wider access um right basically will probably drive or Foster Innovation bringing in on board a lot more diverse sort of talent to work on this space right correct now switching gear from from

a strategic advantages perspective for open source especially for Enterprises right corre meta particularly emphasizes that the advantage of open source for avoiding dependency on closed ecosystem how important do you think this is for Enterprises in today's rapidly evolving Tech environment and what challenges might they face in adopting open- Source AI solutions from an Enterprise standpoint definitely Enterprises which are looking to develop products for their customers for wider use have generally expressed interest in owning the entire stack uh of the solution which includes some of these large language models definitely Enterprises will see it as a

strategic Advantage when they are leveraging an open source model than a close Source model but at the same time because this field is evolving rapidly and of course a company like meta coming from Models coming from companies like meta are fairly trustworthy and then also can be relied upon from a product road map point of view in terms of further enhancements coming in but then it is also important for prises to uh take a view of the evolving nature of this whole space and place the bets accordingly in terms of which vendor or which development

partner uh to choose as well so because that also like for example on the open source site lama lama from meta and mistl as the two leading models however on the same length there are close Source models like open Ai and Cloud which are again equally impressive from a performance point of view or more and are also coming from Enterprises which are fairly well trusted providers of software in the Enterprise space right one has to the CIO has to take a very balanced view of this from a strategic point of view I would say and

I can also see where meta is coming from because because the space is new they want to open source and then have skilled teams find different uses of the model and there will be optimizations to the way the models are fine tuned which will feed back into the meta EOS system right these are aspects which are again meta I believe is considering while they are trying to open source a lot of the stuff yeah so clearly there is a long-term strategic Advantage for meta and that's why they're clearly motivated at this point of time while

they are ignoring any immediate short-term return from this right correct is is my now one question I I I was dabbling in my mind is that Lama 3 I think is based on some 70 billion parameters right and I know that chat gity 4 uh o version um they haven't disclosed really what what is their parameters in terms of numbers I think they all started highlighting the number of parameters that are are important aspect of a particular version but slowly and gradually I'm seeing that all these llms are actually going in different direction for example

Germany is focusing more on efficiency and speed I think chat is focusing more on the accuracy and so on right while Lama is actually focusing on bringing the ca down making it more accessible and so on so I think there a space is clearly heating up still early days and we will hopefully continue to see what is bringing in uh in the future in the coming days so at this point point of time Shiva maybe we'll switch GE and and move to the other interesting area that we going to discuss today and which is about

AI power automation I wanted to quickly jump onto that side let's start with the fundamental of automation right we have seen that uh traditionally at least still few years ago automation was predominantly driven by RPA as we see and talk about aiod automation there's a shift that is happening obviously towards AI how do you see this transition from traditional RPA to AI power automation impacting industries that have heavily relied on the rule-based Automation in the past and what challenges or opportunities do you think would arise when moving from automating simple task to handling complex decision-based

processes as we see the case today um when we looked at traditional it systems automating a business process PA we were really looking at automation of simple tasks especially with RPA would be like filling up a form reading specific data elements from a email Etc and then really leaving it up to the traditional it systems to perform the rest of the automation or a human who is basically reviewing all the work and then completing the workflow with AI powered automation we are more and more looking at more complex tasks for example could read email subjects

date sender Etc and then dump into a database but then now anyi agent can read an email analyze the sentiment and then even autogenerate a response which is very much in line with what the human response would be right so we are looking at a little more sort of complexity getting introduced especially with generative AI coming in as well one is the aspect of making it more inent and then the second aspect is also the complexity of the task that can be handled these are advantages however at the same time we need to guard against

in inaccurate responses inaccurate sort of process flows which is where I think we are more and more relying on uh very special purpose agents right which can do a complex tasks which are broken down into series of activities and then are connected through the chain so that's the way AI power automation or agentic automation is evolving right now okay yeah talking about Evolution I think AI agents are the driving force today behind complex automation tasks right um can you share some insights on into how these AI agents are designed to make decisions autonomously and what

are the key consideration when balancing the autonomy of AI agents a need for human oversight in critical processes why agents we think or AI agents would be better um human interventions yeah one one aspect is AI agents can analyze large volumes of data for example like fraud detection agent for let's say banking or credit card transactions right definitely the there has been much improvement with the AI and machine learning coming into this space and large volumes of data can be analyzed and then flagged and typically how these agents work is we sort of configure or

certain thresholds within which autonomous decisions can be made and then rest of the decisions can be flagged to or pointed out to a human uh who can intervene and then actually take into account a lot many lot more sort of uh aspects while the AI agent can um provide the data for decision making the decision making ultimately can reide with with a human being who is trained in that particular area right so that's way it is evolving uh as and will continue to evolve as we go along yeah so can you maybe give a couple

of examples of such agents maybe our viewers can identify um easily with that maybe are already yeah one example could be an agent which is while you are paying a glame right they agent can the gaming agent will play along with you play learn and adapt to your behavior and making the game World feel more immersive and responsive so that's one example and then there are agents in business processes for example and let's say invoice matching process where you need to three-way matching of an invoice right you need to match the invoice with the goods

received and then the PO that has been issued and a lot of the process currently is manual because somebody has to go and read the email which is from where the invoice is coming in and then look at the PO system and then look at what are the goods received the inward register all the process can be automated using an AI agent and then the agent can determine whether the invoice needs to be partially paid or fully paid or needs to be put on hold right uh and then if the agent is not able to

make decisions then it can flag it off to a human invoice processor to look at it and then make that decision so um in a way it answers my question in terms of how do you ensure that these specialized AI agents communicate and coordinate effectively to deler Accurate right right in terms of I was speaking to other day someone and they said that okay we still do automation but human in the loop so it's human aided automation for a certain use case of language because of accuracy spe specifically in the translation domain um while AI

agents can handle complex tasks as we discussed there is mention of the importance of having human in the in in what scenarios do you believe human involvement in most critical in AI driven workflows and how do you foresee the role of humans evolving as the AI automation becomes more sophisticated and pervasive so answer this with an example of an agent or an agentic flow that we recently created this is to help key account managers manage their who are managing about hundreds of vendors for their company and the key account managers have to answer questions around

okay where is my payment St where is my payment stuck or has the I have sent the material or Goods uh a couple of days back have you received it is there anything else that is pending what is the sort of pipeline looking for you so that the vendors can plan for the the material shipment or manufacturing right so these are common questions that the key account managers receive on a day-to-day basis and then the way we try to solve is create mult multiple special purpose agents which talk to each other for example there is

an agent which looks at Goods received and matches against the invoice and then there is another agent which looks at what were the what was the total material that was intended what was it how much was received and then whether the posos are in place uh and then also look at another agent looks at the demand side what has been entered as a forecast and then the demand for each of these vendors and then creates for from a vendor perspective so now uh some of these processes are entirely automated by the agent like for example

when is the when when is the expected date of payment right because you just need matching of the invoice against the received and then the PO and then you can actually uh uh tell the respond automatically saying this is when your payment would be paid out or the amount will be paid out but then there are aspects around forecasting uh when the vendor is asking the question how much do I need to plan for next month or next quarter then those are queries that will require a little bit of a human intervention that's where the

query gets rooted through answer gets rooted through the key account manager who is managing that set of vendors that's that's the balance that we need to maintain where there are still some complex more ambiguous tasks which agents cannot perform automatically and needs to be routed through human intervention yeah I think said because in real life the context can be very Dynamic and the situations can be changing quite dramatically um and not very linear right probably that's why the human Loop will still be relevant at least for now uh looks like right great I think we

discussed two important topics today the influence and the penetration of open source in AI Innovation AI Le Innovation and I think the second point that we discussed today is basically AI powered automation both are very relevant and important topic as as for today's time at least and in terms of learning more about some of these topics do visit our blog and also look out for some of the other videos in the AI connect series we continue to cover some of these areas and we look for to your comments and suggestions but for now today sha

thank you very much for your time and we will again meet in another episode till then bye from me thank you thank you J