WEBVTT
00:00:00.479 -->
00:00:09.599Hello and welcome back to AvanTest Talk Semi, where we dive deep into the technologies and ideas that are shaping the world of semiconductors.
00:00:09.839 -->
00:00:13.119Today we're exploring a question that goes beyond silicon.
00:00:13.279 -->
00:00:17.839What if AI could truly understand the language of semiconductors?
00:00:18.079 -->
00:00:29.440To help us unpack all this, we're joined by Dow, the CEO and co-founder from Contexture AI, the pioneer behind the unified context layer for enterprise AI.
00:00:29.839 -->
00:00:46.560Unlike generic large language models that often stumble over technical jargons and high-stakes data, Contextual's AI approach is designed to fuse structured and unstructured information, ground its reasoning in facts, and minimize costly errors.
00:00:47.119 -->
00:00:54.159In an industry where a single wrong decision could scrap a wafer worth millions of dollars, this is a game changer.
00:00:54.399 -->
00:01:04.640Looking ahead to LightSouth Factory and knowledge-driven digital tweens, contextual intelligence might just be the missing piece that makes all this possible.
00:01:05.359 -->
00:01:31.680So in this episode, we'll explore how specialized AI can number one speak the language of Fab testers and EDA's tools, how it fuse wafer maps, test logs, and design documents into meaningful insights, prevent error snowballing in high-stakes production, and also enable a secured, trusted path towards fully automated semiconductor factories.
00:01:31.840 -->
00:01:33.040So let's welcome Dow.
00:01:33.439 -->
00:01:34.079Hi everyone.
00:01:36.560 -->
00:01:37.120Alright, Dow.
00:01:37.280 -->
00:01:42.480So you know, when we think about semiconductors, the language itself can be a real barrier.
00:01:42.719 -->
00:01:49.439Acronyms, beaning codes, wafer slot locks, it's almost a dialect that only insider speaks.
00:01:49.680 -->
00:01:54.480But generic AI models, they often misinterpret or even hallucinate.
00:01:54.719 -->
00:01:59.599Imagine if we had AI that truly spoke the language of semiconductors.
00:01:59.840 -->
00:02:06.560So, Dow, why do generic LLMs struggle so much with semiconductor jargons and acronyms?
00:02:06.640 -->
00:02:10.879Can you share an example where this gap becomes critical in practice?
00:02:11.120 -->
00:02:13.439Yeah, it's a good observation, right?
00:02:13.919 -->
00:02:15.199Not every field is the same.
00:02:15.360 -->
00:02:23.039I I think uh semiconductors is is is a pretty unique domain in that uh there's a lot of sensitivity around intellectual property.
00:02:23.199 -->
00:02:30.240Uh so that means that a lot of that that data hasn't been seen necessarily by language models, which which makes it harder.
00:02:30.400 -->
00:02:35.360Uh, and like you said, there's there's just a lot of unique language uh to it.
00:02:35.439 -->
00:02:43.759And and uh you know, as a as a generally trained, non-specialized language model, it's very easy to get confused by very simple terms.
00:02:43.840 -->
00:02:54.400So if you think about like what's the meaning of EDA or SLT or sort of basic terms that everybody in the field knows, um for an AI model that can mean many different things.
00:02:54.719 -->
00:03:18.719Um so um yeah, it's uh I I think a really nice example of a very specific domain where you care a lot about accuracy, you care a lot about having things work on very large-scale, very complex data and solving very complex, highly valuable problems, and then doing that in a way where you can protect um sort of the privacy and and the underlying IP and make sure that uh that doesn't leak.
00:03:19.039 -->
00:03:27.120What are the real-world consequences if an AI misinterpreted terminology in our industry, for example, in geoanalysis or test engineering?
00:03:27.439 -->
00:03:31.280Yeah, uh a small mistake can be very expensive, right?
00:03:31.360 -->
00:03:42.159If you uh I don't know, if you uh give a wrong answer for the voltage on uh pin seven and uh you know uh something goes wrong and and fries, uh that could have have massive consequences.
00:03:42.319 -->
00:03:52.319So um hallucination is a very big problem in general uh with language models, and and that's one of the things that we are really focused on removing from these AI systems.
00:03:52.560 -->
00:04:01.039Um but yeah, I I I think in in semis, it's just a a mistake can be incredibly costly and you have a very, very low tolerance for hallucination.
00:04:01.199 -->
00:04:13.280Um, and that's why you need to be extra careful with the AI systems that you that you build yourself or or that you build with others and make sure that you have the right guardrails in in place uh so that you can avoid those costly mistakes.
00:04:13.599 -->
00:04:14.240Yeah, exactly.
00:04:14.400 -->
00:04:22.639So it if we get a false positive down the semiconductor value chain, we end up with a very expensive product that doesn't work.
00:04:22.720 -->
00:04:24.160So that's something we want to avoid.
00:04:24.399 -->
00:04:27.839So let me move a little bit into the technology that you guys have.
00:04:28.000 -->
00:04:36.720So, how does your context engineering approach actually teach AI to understand semiconductor specific language rather than just memorizing terms?
00:04:37.040 -->
00:04:46.800Yeah, so so one of the core concepts in in AI that maybe if if people have been looking into AI that they might have heard of is something called RAG, retrieval augmented generation.
00:04:47.040 -->
00:04:47.279Right?
00:04:47.439 -->
00:04:54.160So my claim to fame originally as an AI researcher uh is that I'm one of the co-authors of the RAG paper.
00:04:54.319 -->
00:04:59.600So I I worked with my team at at Facebook AI Research at the time on making this possible.
00:04:59.759 -->
00:05:02.000And the idea of RAG is actually very simple.
00:05:02.160 -->
00:05:12.879It's how do I connect a language model, a generator, to data in a way where I can use information from that data to uh give the right answer, right?
00:05:13.120 -->
00:05:20.800So um over time, I I think RAG has emerged into something much more powerful than than when we originally introduced the idea.
00:05:20.959 -->
00:05:30.480And it's really about having so much data like you do in in semiconductors or or mostly uh in any large enterprise, you have way too much data, right?
00:05:30.560 -->
00:05:34.720And it's all over the place, it's very noisy, it it um it's a mess, right?
00:05:34.800 -->
00:05:37.040And everybody will will agree that it's a mess.
00:05:37.199 -->
00:05:51.759So then the real problem that we have right now when it comes to getting value from Gen AI is how can I get the right context from all of that data to give to a language model or an agent so that it can solve a valuable problem for me.
00:05:51.920 -->
00:05:54.639And that could be a very simple thing, answer a technical question.
00:05:54.800 -->
00:06:05.600It could be a very difficult problem like generate a test program or um you know um uh analyze a log file and give me the root cause of what went wrong with this kind of device here.
00:06:05.759 -->
00:06:08.480So so there's all kinds of things you can do.
00:06:08.639 -->
00:06:14.959The the common problem that we have is how get how do you get the right context from the data to solve the problem.
00:06:15.279 -->
00:06:20.319Do you see semiconductor manufacturing as a proving ground for domain-specific AI?
00:06:20.560 -->
00:06:28.000Could the lesson learned here be applied to other complex technical industries like aerospace or biotech?
00:06:28.240 -->
00:06:28.879Yeah, totally.
00:06:28.959 -->
00:06:29.600I believe so.
00:06:29.759 -->
00:06:34.319I I do think that semi semiconductors is uniquely suited for this.
00:06:34.560 -->
00:06:44.480Um, but but I I think uh in general there is this uh problem of making general systems work on specialized problems, right?
00:06:44.639 -->
00:06:50.879And if these general systems are really meant to work for everybody, like a standard like ChatGPT, it's amazing, right?
00:06:50.959 -->
00:07:02.800It's an amazing piece of software, but it is meant to serve everybody, uh, which means that you know it needs to be able to brainstorm creative stories with the marketing department when you're doing creative writing.
00:07:03.040 -->
00:07:12.560And so that type of hallucination, maybe if you want to call it that, is really something you don't want to have when you have a sensitive domain like like semiconductors or aerospace, like you said.
00:07:12.720 -->
00:07:19.279So different domains have different kinds of trade-offs, and that's why you don't want to have a generic model that works for everybody.
00:07:19.360 -->
00:07:32.800You want to have a specialized model that understands your jargon, that works on your data, that um understands things slightly more deeply, just in the same way that you know, if you want to get into semiconductors, you need to go to school for that.
00:07:32.959 -->
00:07:38.399You can't just go and and just you know walk into a place and and expect to understand the conversation.
00:07:38.560 -->
00:07:41.439Um so yeah, specialization is is pretty important.
00:07:41.759 -->
00:07:43.519Totally, totally, totally agree.
00:07:43.759 -->
00:07:50.560And uh I'd like to dive a little bit deeper into the structured and structured and unstructured data that you just mentioned, right?
00:07:50.800 -->
00:07:54.480So in semiconductor, data is not just big, it's messy.
00:07:54.560 -->
00:07:58.879We have a lot of different players in throughout the entire semiconductor value chain.
00:07:59.040 -->
00:08:07.600So we've got structured data like NES logs, uh the SPC charts and yield dashboards, and that's structured.
00:08:07.680 -->
00:08:12.480But for unstructured, we have like engineering notes, PDFs, design documents.
00:08:12.720 -->
00:08:18.879So the challenge for us has never been storing it, but it's about connecting the dots across all these silos.
00:08:19.360 -->
00:08:29.680So, Dow, why is it so difficult to connect structured data like NES logs and with unstructured sources such as engineering reports or design nodes?
00:08:30.000 -->
00:08:32.159Everything is about context, right?
00:08:32.320 -->
00:08:50.720So if you want to understand unstructured documents, then you probably have to understand them with relation to maybe some things that are in the database somewhere, or or like you said, device logs, or maybe you can understand the logs better if you understand the manual of the the chip that you're sort of analyzing, right?
00:08:50.799 -->
00:08:58.159So there's there's all of these interdependencies across different data sources, and different data sources have different data formats.
00:08:58.720 -->
00:09:05.279So uh the way to resolve those dependencies is by by having very good contextual understanding.
00:09:05.440 -->
00:09:10.080Um and the hard part when you think about how do you analyze a big log.
00:09:10.159 -->
00:09:24.240Like let's say you have gigabytes of logs uh somewhere, and um, as a human, what you would have to do is basically kind of scroll through it, or you would maybe control F for like fail or error or something like that, and then you would read the logs, right?
00:09:24.399 -->
00:09:27.679So um an AI system can do a very, very similar thing.
00:09:27.840 -->
00:09:32.879And what used to be very hard was that we had what we call text as equal, right?
00:09:32.960 -->
00:09:44.399Or so or or sort of like basic like lexical search over log files, and then we have embedding-based search, which means that you can get the semantic information out of unstructured data.
00:09:44.559 -->
00:09:51.440And it was very hard to figure out which do I use when and how do I combine the results out of all of these different data sources.
00:09:51.679 -->
00:09:55.679So in modern RAG pipelines, we have things like re-rankers.
00:09:55.840 -->
00:09:56.879Maybe you've heard of that.
00:09:56.960 -->
00:10:10.480So you do an initial retrieval step from all the different data sources, and then you use a re-ranker on top of the initial retrieval results to find the final pieces of information that you want to give as context to the language model.
00:10:10.639 -->
00:10:18.000And when you have a good re-ranker, like we do, uh, we have probably the best re-ranker in the world that can follow instructions as well.
00:10:18.080 -->
00:10:28.799So you can give it very specific uh contextual knowledge that it uses to find the right pieces of information, uh, then you can really uh solve very, very hard problems.
00:10:29.120 -->
00:10:36.559When you talk about re-ranker across structured and unstructured data, so how does that differs or or does that matter at all?
00:10:36.799 -->
00:10:40.559Yeah, so so uh every piece of data is different, um, right?
00:10:40.960 -->
00:10:50.799So so um um that that's why you need to have a re-ranker that can make sense of all of these different data types and also understand uh what we call the metadata basically.
00:10:50.879 -->
00:10:55.440So the source of the underlying data, does this come from a device log?
00:10:55.600 -->
00:10:56.960Does this come from a database?
00:10:57.120 -->
00:10:58.799Which database does it come from?
00:10:59.039 -->
00:11:04.399Like all of that information is is important when you have to figure out what the right context is.
00:11:04.639 -->
00:11:09.360Um and and so with this new world of, I'm sure everybody's heard about agents, right?
00:11:09.440 -->
00:11:18.320Like agents are are really the future, and and so what agents can do is is use tools to do retrieval steps.
00:11:18.480 -->
00:11:30.399So you can actually when you do an initial search and you find some promising uh results, but maybe you want to get some more information as an agent, it can just go and look for for other agent elsewhere.
00:11:30.559 -->
00:11:39.679So it's really a very similar process to how you would tackle this as a human, but then much, much faster, much more efficient, and you don't have to do it, right?
00:11:39.840 -->
00:11:41.039AI does it for you.
00:11:41.279 -->
00:11:53.600So so like device log analysis, I think, is a great example where in some cases companies we work with they can take days to analyze a log file just because of of its sheer size, and they need to like cross-reference.
00:11:53.679 -->
00:11:56.159Maybe they need to call somebody to get some more information.
00:11:56.399 -->
00:11:59.120It's like what if you could just solve that in five minutes?
00:11:59.279 -->
00:12:01.759Uh that's massive time saved.
00:12:01.840 -->
00:12:06.080Um, so there's a lot of potential for for AI uh across the sector.
00:12:06.320 -->
00:12:07.759All right, totally agree.
00:12:07.840 -->
00:12:15.279We we need to get all capture and use all this unstructured data today, where which we're not using it today, right?
00:12:15.519 -->
00:12:17.840So um I'd like to move a little bit forward.
00:12:17.919 -->
00:12:29.440So if we take wafer level failure and want to trace it to the assembly logs, line logs, so how does contacts AI approach help engineer discover that connection faster?
00:12:29.759 -->
00:12:32.799Yeah, I mean, uh, so that's really a reasoning problem, right?
00:12:32.879 -->
00:12:34.879So so this is what agents are amazing at.
00:12:35.039 -->
00:12:42.559And and obviously, like a just an out-of-the-box agent is probably not going to be great at at solving a problem as specific as that one.
00:12:42.720 -->
00:12:56.000But if you specialize the agent and if you inject all of this contextual understanding and if you make it sort of see the broader context of of kind of the the process that that you're trying to analyze, then you can reason backwards.
00:12:56.159 -->
00:13:07.200Um, and and so uh this will help you uh optimize uh processes like this and really identify the boot cause that might be causing a failure or whatever you're trying to analyze.
00:13:07.519 -->
00:13:14.559Um so so what it's really about kind of um compounding performance in the end with these systems, right?
00:13:14.639 -->
00:13:24.639It's like you know you have all of these different kind of like small decisions that you need to make, and they're they're like sub-agents kind of working together, analyzing different parts maybe of a supply chain or something.
00:13:24.799 -->
00:13:32.559And and if they all work together, then you can really just um optimize the entire chain, and and that leads to to massive value.
00:13:32.799 -->
00:13:40.960So it's interesting you talk about agents as well, because as we implement more AI agents along the way, people start to think about does do you have a job?
00:13:41.039 -->
00:13:44.159And then we start talking about human in the loop automations.
00:13:44.480 -->
00:13:51.600So our engineers often jump between dashboards, PDFs, emails just to make sense of, you know, if there's an issue.
00:13:51.840 -->
00:13:59.279But how do you see AI reducing that friction, but keeping human in control of critical decisions?
00:13:59.600 -->
00:14:08.480Yeah, it's extremely important to keep the human in the loop, especially in in uh in a high-stakes settings, the setting like like the one you described.
00:14:08.639 -->
00:14:20.080I I think in in in your sector, it's really important that you are not just going to hand over the keys to AI and say go run with it and like you know, design something cool and we'll we'll just manufacture it at scale.
00:14:20.159 -->
00:14:22.879Like that would be very uh expensive mistakes, I think.
00:14:23.120 -->
00:14:28.480So um, so you need to have humans in the loop and you need to figure out when you when you want to put them in the loop.
00:14:28.559 -->
00:14:35.120So the promise of AI, I think in some sectors it might replace people's jobs.
00:14:35.279 -->
00:14:41.679I think in more specialized sectors, it is going to augment the people uh in those sectors already, right?
00:14:41.759 -->
00:14:46.639So think of think of like a customer engineer at like a large semiconductor company.
00:14:46.879 -->
00:14:56.559So all the work that they're doing, there's actually a lot of work that doesn't necessarily require a lot of like deep skill or expertise or intelligence.
00:14:56.720 -->
00:14:58.480It's kind of just mundane, right?
00:14:58.559 -->
00:15:08.240Like you said, basically, like checking some emails and reading some files and maybe scrolling through a log and like all of that is kind of like boring, boring stuff that you have to do to solve the problem, right?
00:15:08.320 -->
00:15:13.919So if AI can do that for you, think of all the the productivity gains that you get from that, right?
00:15:14.000 -->
00:15:23.600You can just deliver a lot more value as an expert in the field by having kind of an AI sidekick that augments you and and does all the boring stuff and all the plumbing for you.
00:15:23.679 -->
00:15:25.600And that that sounds pretty wonderful, I think.
00:15:25.840 -->
00:15:28.320Yeah, that's why we call operational efficiency, right?
00:15:28.399 -->
00:15:34.240We get more, our employees get more and more efficient, and then we get more and more productivity.
00:15:34.559 -->
00:15:48.080So as we move forward on this, uh, the key trend right now is talking about lights out factory, where there's no humans in the factory floor, and also digital tween, some people call it knowledge tween.
00:15:48.240 -->
00:16:02.799How do you see this uh fusion of structured and unstructured data as a step towards creating a true digital tween of the fab, you know, something that could eventually drive real-time decision making in LightSouth Factory?
00:16:03.120 -->
00:16:10.000Yeah, so so so I I think um with these types of complex processes, which that obviously is, right?
00:16:10.080 -->
00:16:11.679That is not a trivial task.
00:16:11.840 -->
00:16:18.080Uh, that just requires a very, very broad set of data sources that you need to reason over.
00:16:18.240 -->
00:16:21.279And a lot of these are structured, a lot of these are unstructured.
00:16:21.440 -->
00:16:37.679Um and and so um I I do do believe that that we are going in this direction where you can just automate a lot of these processes uh in very clean ways, as long as as as we discussed, you you identify the spots where you want to have humans in the loop.
00:16:37.919 -->
00:16:48.240Um, but yeah, it's it is really about understanding the contextual pieces of all these different uh uh data sources that you need to work over, right?
00:16:48.399 -->
00:17:02.480It's like understanding sort of the physical world that you're in, and then understanding kind of the the sort of um uh the problems you're trying to solve, understanding the processes that you have, the unique intellectual property that goes into uh this those processes.
00:17:02.559 -->
00:17:05.279So all of that that context needs to come together.
00:17:05.359 -->
00:17:07.279Um, and that's uh that's a hard problem.
00:17:07.519 -->
00:17:10.640And and I really like the way how you structure this.
00:17:10.720 -->
00:17:16.640So we could have human interloop in where we design where we want human interloop, kind of like a control point, a control tower.
00:17:16.799 -->
00:17:29.680And then the rest of the stuff uh that can be automated will have AI robots, AI agents with REG that talks to each other and understand each other from a context perspective and do a lot of automation, right?
00:17:30.000 -->
00:17:31.759I I have a nice example there, actually.
00:17:31.839 -->
00:17:34.880Maybe so we work with a lot of like finance customers too.
00:17:35.039 -->
00:17:37.519It's actually surprisingly similar in a way, right?
00:17:37.599 -->
00:17:46.000This complex domain, sensitive about IP, lots of data that is sort of all over the place, and you want to solve like complex problems on top.
00:17:46.160 -->
00:17:55.519And so one of the big problems they have is that they have very noisy, large documents that they want to understand uh in a kind of structured way.
00:17:55.680 -->
00:18:02.000So they basically want to extract, let's say, 10 different fields from a large report.
00:18:02.079 -->
00:18:06.960It's like what's the company's name, what is their revenue, what is their projected revenue, or something like that.
00:18:07.279 -->
00:18:18.880So one of the things that we do in those use cases is we extract the information, we're very good at kind of document understanding, but then as a part of the extraction, we also assign a confidence score.
00:18:19.039 -->
00:18:24.960Um, so that means that if the system isn't really sure, you probably want to have a human in the loop.
00:18:25.119 -->
00:18:30.880If the system is like 100% sure, like the revenue was X, you probably don't have to double check it, right?
00:18:30.960 -->
00:18:38.160But maybe if it's like, well, I'm kind of 50-50 because there are like two different tables and they sort of report different numbers for different time periods.
00:18:38.319 -->
00:18:41.440Is that this is something that as a human you want to double check?
00:18:41.519 -->
00:18:52.000So then you can build that into your your flows as well to make sure that whenever AI sort of self-identifies that it's not quite sure about the answer, then that's where you call in the humans.
00:18:52.160 -->
00:18:57.359Um, and I think that yeah, those those kind of processes are are becoming pretty well established at this point.
00:18:57.599 -->
00:19:01.759Let's talk about the cause of errors and how it propagates.
00:19:01.920 -->
00:19:07.920In semiconductor and you know, in cheap design, precision isn't optional, it's survival.
00:19:08.160 -->
00:19:16.319A single wrong insight could mean scrapping wafer worth millions of dollars or misdirecting our engineers to the wrong root causes.
00:19:16.480 -->
00:19:21.119And unlike other industries, error here doesn't just add up, they snowball.
00:19:21.279 -->
00:19:28.799Uh, that's why zero defect expectations and adaptive testing put such high demand on AI systems.
00:19:29.119 -->
00:19:37.200So, so Dao, why is error propagation so uniquely costly in the world of semiconductor compared to other industries?
00:19:37.519 -->
00:19:40.079I mean, I honestly that's a question for you to answer.
00:19:40.160 -->
00:19:44.079Like you're you're the semiconductor expert, I'm the AI expert, right?
00:19:44.160 -->
00:19:47.519But I I think it's it's a very, very uh good observation.
00:19:47.599 -->
00:19:58.880I I think uh especially when you have sort of very, very complex processes with lots of interdependent steps, like one small mistake in the beginning at the end can have huge repercussions.
00:19:59.119 -->
00:20:06.240Um and and so for for us, actually, in AI, um it is is is similar in a way.
00:20:06.400 -->
00:20:14.960So the the accuracy of the overall system is dependent on all the individual AI model components that exist in that system.
00:20:15.119 -->
00:20:34.000So if you think about RAG, retrieval augmented generation, you usually have some sort of document extraction that you do first, so that you can store the information in a vector database, so you encode it with embeddings, and then you do retrieval, and then you do re-ranking, and then you do generation, and then maybe post-processing or guardrails.
00:20:34.240 -->
00:20:36.000So you have lots of different steps.
00:20:36.240 -->
00:20:52.240And so if you're in a domain like semiconductors and 1% of accuracy in the final answer can have huge repercussions down the line, then you really want to make sure that all of these parts are optimized so that you get the max possible performance.
00:20:52.400 -->
00:20:58.880And and just like in the semis and and sort of the supply chain there, in AI, all of these dependencies compact.
00:20:59.039 -->
00:21:04.240So if you have a very bad extraction system, your retrieval is not going to be very good.
00:21:04.400 -->
00:21:18.400If your retrieval is not very good, then you can re-rank the results, but it's not going to give you the right result, which means that you give the wrong context to a language model so you can have an amazing language model, but it will still not be able to get the right answer because it doesn't have the right context.
00:21:18.640 -->
00:21:24.480Um, so you see the dissimilar kind of snowballing uh happening in AI systems kind of across the board.
00:21:24.799 -->
00:21:25.680Exactly, exactly.
00:21:25.759 -->
00:21:32.799And and that's why in semiconductor is so important for us to get proper testing done and get accurate results.
00:21:32.960 -->
00:21:40.400So we prevent false positive de facto going into the next step, which just costs us a lot of money just to scrap it later.
00:21:40.640 -->
00:21:44.079But um, so in AI, as you mentioned, which is the same, right?
00:21:44.160 -->
00:21:57.359You when you have an error, it's just if you let an error go into the AI system, the LM system, that's just going to snowball because it understands things differently and it's going to make different suggestions, uh, maybe a wrong suggestion in that sense.
00:21:57.519 -->
00:22:02.640But those so how does contextual How does AI prevent this snowball effect?
00:22:02.799 -->
00:22:09.839How do you catch the error when you're trying to digest a document or you digest an NES log?
00:22:10.079 -->
00:22:13.200How do you prevent this error from snowballing?
00:22:13.279 -->
00:22:18.799And especially when you know our engineers really rely on outputs for adapted test strategies.
00:22:19.119 -->
00:22:23.279I feel like it's in a way pretty similar to your industry.
00:22:23.440 -->
00:22:25.039So we do a lot of testing.
00:22:25.200 -->
00:22:28.559I mean, uh, I'm sure you uh agree with me that testing is important.
00:22:29.440 -->
00:22:35.759You really want to understand the performance of each of the individual components and you want to think about it holistically.
00:22:35.920 -->
00:22:36.240Right?
00:22:36.319 -->
00:22:52.079So one of the things that we often see a lot of our competitors do wrong, or that we see some of our uh you know, folks in in the field before they come to us, the mistakes they make, they're they're very often about thinking about it from a single model perspective.
00:22:52.240 -->
00:22:57.200It's like, oh, I have like a language model and I have like a retrieval model or something, right?
00:22:57.680 -->
00:23:07.359Rather than thinking about it holistically, saying I have this problem I want to solve, and I need to have a whole system that optimizes against solving that problem.
00:23:07.599 -->
00:23:17.759Uh, and so when you take this systems approach, um then then you can just have have much better performance because you account for that kind of compounding over time, right?
00:23:17.839 -->
00:23:21.200Or the cascading of errors, uh, if you think about it that way.
00:23:21.359 -->
00:23:30.880So that's something that we take a lot of pride in, is really systems over models, as we call it, and then specializing those systems so that you can solve hard problems.
00:23:31.359 -->
00:23:32.160Very true.
00:23:32.400 -->
00:23:40.160So we talk about zero de facto expectation, and and for semiconductor, that is very important for us.
00:23:40.480 -->
00:23:49.759So, in in in your mind, how realistic is it for AI to support that kind of mindset for a zero de facto without over compromising?
00:23:50.000 -->
00:23:59.119Yeah, so for AI on its own, I would say there is a zero percent chance of that being uh a realistic uh goal.
00:23:59.279 -->
00:24:08.799And and so so that's why humans in the loop are important and why you need to have specialists like double check results and and just verify uh that things are correct.
00:24:08.880 -->
00:24:21.279So AI can handle a lot, but one of the the the big things about AI is that you know it's a it's it can hallucinate and it makes mistakes, just like normal humans sometimes make mistakes, right?
00:24:21.440 -->
00:24:45.680So if I go back again to like finance as an example where it's a different sector but sort of similar problems, there are very good guardrails in place when it comes to the behavior of the bank that come from regulators like the SEC or folks like that who say, like, this is your process that you have to follow, where humans need to kind of sign off on different things to make sure that something is actually correct, right?
00:24:45.759 -->
00:24:49.839Because otherwise, like our global financial system is at risk or something like that.
00:24:50.000 -->
00:24:57.119So, in those types of of kind of decision uh trees, there have to be double checks and humans.
00:24:57.200 -->
00:25:07.680So some of the parts can be AI, but you probably want to keep humans in the loop there and then make sure that you leverage AI in a way where if the AI is not sure, then you call in the human, right?
00:25:07.759 -->
00:25:10.559Or sometimes you can have AI double check the human.
00:25:10.799 -->
00:25:26.000So it it's very symbiotic, I think, in in in the future, uh, when it comes to um solving high-stakes use cases, and we will be working together with AI, but letting AI on its own do it, I think is is still a bit of a pipe dream.
00:25:26.319 -->
00:25:39.839So I want to touch on a point that you make slightly earlier to talk about, you know, uh AI is going to take over some jobs and uh that is maybe lower skilled and not so much for jobs that are higher skilled.
00:25:40.000 -->
00:25:47.440In the semiconductor world, as we do testing, especially for advanced tests, which do a uh we sell task equipments, so we do a lot of testings.
00:25:47.599 -->
00:25:53.119Do you see contextual AI as an enabler for a new form of adaptive testing?
00:25:53.200 -->
00:26:09.359So where AI not only analyzes results and helps engineers make decisions, but it also helps dynamically optimize our task programs in real time so you you get your task programs get smarter and shorter and more efficient.
00:26:09.680 -->
00:26:10.160Absolutely.
00:26:10.319 -->
00:26:13.839I I think that that's sort of the the the promise of AI here, right?
00:26:13.920 -->
00:26:21.839It's like like we already talked about like context matters, and and the more of that contextual understanding you have, the more you can really tailor things.
00:26:21.920 -->
00:26:27.759So if you have to design a very specific test program for what you care about, AI can help you with that.
00:26:28.000 -->
00:26:43.599As long as you can provide the right context and the right instructions for what you're looking for, um, and maybe even set up a feedback loop, right, where over time it actually gets better and really like adjusts to you, maybe even like individually as a developer, like what you care about.
00:26:43.759 -->
00:26:47.519You can really start to do that when you have that contextual understanding.
00:26:47.680 -->
00:26:53.920So I think there's a ton of promise for um yeah, taking uh kind of advanced testing to the the next level.
00:26:54.160 -->
00:26:57.279So I like to move forward into lights out factory.
00:26:57.359 -->
00:27:02.559I think a lot of our our audience and and for us, we think that's the future as well.
00:27:02.799 -->
00:27:05.759So a lot of our customers are working on this.
00:27:06.000 -->
00:27:19.759Um, think about in a lights out factory when there is no humans, but there's a ton of a lot of different machines in in the lights out factory, and they need to talk to each other to understand each other so they can work together.
00:27:20.000 -->
00:27:28.160So the dream of a fully automated lab or a true lights out factory, you know, it's been on the horizon for years.
00:27:28.240 -->
00:27:31.519We talk about that, and there are some people who started working on that.
00:27:31.680 -->
00:27:40.960So we need robots, we need sensors, we need MES systems that really power parts of this vision, but something is still missing.
00:27:41.440 -->
00:27:43.920It's a brain that can orchestrate it all.
00:27:44.079 -->
00:27:54.720I think this is where context or where AI comes in, is not just retrieving data, but understanding, reasoning, acting as a trusted digital tween of factory knowledge.
00:27:55.039 -->
00:28:15.200So, Dow, when people talk about lights out factory, what do you think is the hype and what's possible, actually possible today in in the in the semiconductor fabs and task floor if we use contexture and contact can contexture AI be that brain for our likes out factory?
00:28:15.519 -->
00:28:19.039I would like to say yes, but I I guess we have to show that in practice.
00:28:19.839 -->
00:28:20.160Right?
00:28:20.240 -->
00:28:35.440But uh so yeah, I I I think what's so interesting about that is is just that it's a very complicated um uh process with so many different interdependencies, and like we've been talking about, right, there are lots of opportunities for cascading errors there.
00:28:35.680 -->
00:28:42.079So um I think it's a really good test bed for thinking about this future of multi-agent systems, right?
00:28:42.160 -->
00:28:52.720So that's really one of the big things that a lot of us are in AI are are really excited about, is it's not going to be just one agent, one Chat GPT or whatever that solves your problem.
00:28:52.799 -->
00:28:58.079It will be lots of different agents, and they're all specialized in in very, very different ways.
00:28:58.319 -->
00:29:02.319They have their own context and their own sort of personalization, each of them.
00:29:02.400 -->
00:29:12.799And then to your point, it's really a question of how do you combine all of that and how do you make that work together and how do you jointly optimize that as sort of a system of systems, maybe.
00:29:13.039 -->
00:29:20.000Um and and um, yeah, I I I think um it's a matter of time before we we can do that.
00:29:20.160 -->
00:29:22.960I think from the AI perspective, we're basically there, right?
00:29:23.039 -->
00:29:35.200I think one of the harder problems is actually just like physical uh intelligence, so robotics that actually have have like fine-grained enough control to do what they need to do uh in a robust way.
00:29:35.279 -->
00:29:46.000And so I I I think um 10 years ago, probably a lot of folks would have expected robotics to be much easier and language to be much harder, uh, but it turns out to be uh the opposite.
00:29:46.240 -->
00:29:47.440Totally agree.
00:29:47.680 -->
00:29:49.920So as you mentioned, we talk about, right?
00:29:50.079 -->
00:29:57.119Um, it's about having multiple AI agents working together, and then now we're just waiting for kind of waiting for the robots to catch up.
00:29:57.279 -->
00:30:03.359But how does contextual AI function as a trusted digital tween of factory knowledge base?
00:30:03.680 -->
00:30:16.480You know, it's not about working with all these AI agents, multiple AI agents, to get this knowledge or to get this data from them and then understand that, combine that together, having and then having them talk to each other.
00:30:16.640 -->
00:30:21.359Um, is is contextual AI uh helping in that sense?
00:30:21.680 -->
00:30:27.279So the the talking to each other, I think is a very open research problem for a lot of folks in AI, right?
00:30:27.519 -->
00:30:41.759It's like um I think in many cases uh it makes sense to talk to each other in natural language, because then you know a human can always jump in and double check whether whether what the agents are talking to each other about makes sense.
00:30:42.079 -->
00:30:45.599But for machines, language is very inefficient, right?
00:30:45.839 -->
00:30:59.359It's like or at least natural language that we speak, whether it's English or Chinese or whatever, is extremely inefficient when you think about it from a computer's perspective, where maybe you just want to send code or like raw data, right?
00:30:59.440 -->
00:31:01.920Um, but then you sacrifice interpretability.
00:31:02.000 -->
00:31:09.839And I think one of the big um things that we care about in in these very complex multi-agent systems is interpretability.
00:31:10.000 -->
00:31:16.960Uh so we need to make sure that we can observe what's happening in the process and always kind of interject uh human oversight.
00:31:17.119 -->
00:31:23.359Um so so uh you know, that's why it will probably be some some version of some language that we can understand.
00:31:23.680 -->
00:31:26.000I like to pick your brain a little bit.
00:31:26.160 -->
00:31:29.279So let's move forward five to ten years from now.
00:31:29.519 -->
00:31:42.400If if you can imagine that we have a semiconductor fab or OSART uh task that is fully automated and contextual AI is embedded as this decision-making layer, how do you envision that?
00:31:42.559 -->
00:31:45.759How do you see that happening and and what else is missing?
00:31:45.920 -->
00:31:52.480What else is required to make this happen, to make contextual AI as part of that decision-making layer?
00:31:52.799 -->
00:32:11.599I I honestly think that that in terms of pure technology, everything is there uh outside of the sort of the robotics, physical intelligence side of things, and and maybe sort of some of the processes around it, or even like regulatory processes and and sort of um maybe some of the like change management and and you know a lot of the vaguer things.
00:32:11.680 -->
00:32:25.759I I think one of the the exciting uh things about where we are right now in time is that we had this core capability of of intelligence, and then we are starting to really solve this problem of context.
00:32:25.920 -->
00:32:34.640So now we can give the intelligence the context that it needs, um, and and you know, so that these intelligences or agents can also work together.
00:32:34.880 -->
00:32:42.720Um, and then um the the real real question now, and that goes goes, I think, much much further than than the fab.
00:32:42.960 -->
00:32:47.440It basically extends to the whole world, is how how does this manifest itself?
00:32:47.519 -->
00:32:49.839Like what how does this become a product, right?
00:32:49.920 -->
00:32:52.240It's like how do we productize AI?
00:32:52.480 -->
00:33:13.279Um, and and I think um we're going to find out in the next uh uh five years what it will really look like and and so um how much human involvement there will be, um, you know, how we are going to make sure we can rely on these systems uh and uh how can we make sure that they don't kill us and and that kind of kind of stuff.
00:33:13.440 -->
00:33:20.480Uh so so some very, very deep, profound questions uh that we need to uh solve as a society in the next few years.
00:33:20.720 -->
00:33:21.519Yeah, yeah.
00:33:21.759 -->
00:33:46.640And and you know, also personally, uh as we're implementing AI as well, I think having um more AI in our processes, in our workflow, in our factories, uh, that also gives us a chance to redesign our processes and also redesign the workflows, redesign a lot of things that we can make it easier for and and it's not possible because uh we don't have that technology.
00:33:46.720 -->
00:33:53.039Now AI comes in with a lot of automation, a lot of understanding, and just enabling that for us to change.
00:33:53.200 -->
00:34:00.559And so I think besides just technology change, a lot of processes, SOP workflows, and how we design our machines is gonna change.
00:34:01.200 -->
00:34:02.240What do you think about that?
00:34:02.960 -->
00:34:08.559The change is coming, managing the change, so change management, that's really the big challenge, I think.
00:34:08.719 -->
00:34:25.119Um and and you know, like you mentioned SOPs and and sort of like basic things that feel basic to humans, like a like a machine could just like mix up the order of the SOP and like you know, in a way they're like incredibly smart, but at the same time, they're incredibly dumb.
00:34:25.360 -->
00:34:30.639Um and and where they are very dumb is often when it comes to this this kind of context problem.
00:34:31.360 -->
00:34:37.760Um so so that's why um when we solve that, uh it's going to unlock a lot of very exciting things.
00:34:38.000 -->
00:34:42.480In semiconductor data isn't just valuable, it's extremely important.
00:34:42.639 -->
00:34:50.559So design files, Tesla, guild data are crowned jewels in our industries, and protecting them is non-negotiable.
00:34:50.719 -->
00:34:56.960Uh, that's why any AI system that we use must deliver insights without ever putting any IP at risk.
00:34:57.119 -->
00:35:07.280So, Dow, why is data security, IP protection such a unique concern, you know, uh when we come when it comes to adopting AI systems?
00:35:07.599 -->
00:35:09.760Yeah, it's an incredibly important problem.
00:35:09.840 -->
00:35:14.400Uh, and I I think there are are so there are lots of sectors that have this problem.
00:35:14.480 -->
00:35:16.320Like you said, finance is very similar, right?
00:35:16.559 -->
00:35:29.760It's it's in the end, actually, when you when you think about what the future looks like, what makes a company unique or what makes a company a company is its data and its expertise and its intellectual property, right?
00:35:29.840 -->
00:35:38.400So beyond like in the future, maybe there there might be like three people working at at JP Morgan, and the rest is all AI, right?
00:35:38.480 -->
00:35:43.280Uh like I don't know if that's a real uh good example, but uh you get my point, right?
00:35:43.360 -->
00:35:50.320It's like people uh make a company right now, but in the future, if if you extrapolate, data makes the company.
00:35:50.559 -->
00:35:55.280Um, and um uh you have to be extremely careful with that data.
00:35:55.360 -->
00:36:05.840And one of the worst things you can probably do right now is give all that data to OpenAI or Anthropic so that they can train on your data and then uh solve all these problems in the future, right?
00:36:06.000 -->
00:36:09.840So that's one of the things that we we focused on when we started the company.
00:36:09.920 -->
00:36:17.039We knew that RAG is really a way where you can give language models this information without the language models training on it.
00:36:17.280 -->
00:36:25.920Um so our entire system is deployable um inside the customer's VPC, which means that no data ever leaves your security perimeter.
00:36:26.079 -->
00:36:34.000Everything stays inside, it doesn't get leaked so that your competitor also benefits from um from the the work that you put in.
00:36:34.079 -->
00:36:39.920Um so we think that's very important and one of the the things we do differently from a lot of our competitors.
00:36:40.159 -->
00:36:40.719Yeah, yeah.
00:36:40.800 -->
00:36:52.000Like you mentioned, I totally agree that you know we don't want to share our design data with open AI or other large language models so that they can build a better tester than we do.
00:36:52.239 -->
00:36:53.840So it's same, yeah.
00:36:53.920 -->
00:36:54.800So so same thing.
00:36:54.880 -->
00:36:58.079Um or that your competitor can use open AI to build a data.
00:36:58.400 -->
00:37:00.719Yeah, and then build a better machine that's better than Avan Test.
00:37:00.800 -->
00:37:01.840We don't want that to happen.
00:37:02.079 -->
00:37:16.159Um, so as you talk about RAG contextual AI, you know, being able to extract the data and keeping it keeping all our sensitive data within our secure uh parameters, how do we know RAG like contextual AI is not learning from us as well?
00:37:16.480 -->
00:37:20.400Yeah, so so the system gets deployed in your data center.
00:37:20.559 -->
00:37:22.719So we don't even see the data.
00:37:23.280 -->
00:37:23.599Right?
00:37:23.679 -->
00:37:34.960So so it's really your deployment there, and and that's really uh I think important is that you control what data goes into VAF and you control kind of the artifacts that are created as a part of that system.
00:37:35.280 -->
00:37:43.280So, what about going across companies, going across uh different companies right now along the semiconductor value chain?
00:37:43.360 -->
00:37:47.760So, because we start to see that there is a need to share information.
00:37:48.079 -->
00:37:54.719So the question becomes how can I share information with my customer, with my vendor, but not losing my IP?
00:37:55.199 -->
00:38:00.320So, and and I suspect RAG and contextual AI could be a way of doing that.
00:38:00.480 -->
00:38:02.159So, can you talk a little bit about that?
00:38:02.559 -->
00:38:05.199Yeah, so I mean there are different ways to do that, right?
00:38:05.280 -->
00:38:13.679I I think honestly a lot of us more broadly are trying to figure out kind of like where where do these data handoffs happen, right?
00:38:13.760 -->
00:38:28.719If you if you think about like all the different like file storage providers from your like Dropbox to like your SharePoint or something, it's like where like who owns the data and who owns the context that comes out of that data.
00:38:28.800 -->
00:38:31.679I I think that's still a lot of people are figuring that out, right?
00:38:31.760 -->
00:38:39.199Like Slack and Salesforce, they've actually tried to lock it down so that it's not possible to do your own search on it.
00:38:39.280 -->
00:38:40.800You have to use their search.
00:38:41.039 -->
00:38:52.800Um, so there are there are good protocols like MCP and A2A uh that are getting developed now in AI when it comes to making different data accessible to different uh parties, right?
00:38:52.960 -->
00:39:17.119So maybe in the future for for some of your like internal manuals at at FNTS, you might have an MCP server that is powered by contextual AI that has all of your manuals so that if one of your customers wants to ask a question uh on a manual inside their AI system, then they can just very quickly call uh the MCP server and get the right answer with the right attributions.
00:39:17.679 -->
00:39:28.159Um so yeah, we're we're still figuring that out, but I think that's one of the key benefits of having context as kind of the connective tissue between intelligence and data, right?
00:39:28.239 -->
00:39:35.199If you wanted to take all the data and just train the intelligence on it so that it knew everything, then you could never really solve that problem.
00:39:35.360 -->
00:39:39.440So uh the only way to really do that is with RAG and context engineering.
00:39:39.760 -->
00:39:43.199We've heard a lot from Dow and thank you very much for joining us.
00:39:43.440 -->
00:39:52.559So before we end this, can I just ask you one last insight that you have for our audience, for our guests, for in this in the semiconductor?
00:39:52.639 -->
00:40:00.079How can contextual AI, how do you envision contextual AI in the next five years for the world of semiconductor?
00:40:00.320 -->
00:40:06.800Yeah, I think we we are extremely excited about semiconductor as a sector and the potential that it has.
00:40:06.960 -->
00:40:10.320Uh, it's extremely contextual as a domain.
00:40:10.480 -->
00:40:13.760Um, so you need contextual AI to make it work.
00:40:13.920 -->
00:40:23.280And um, yeah, I I think the the real interesting problem with AI is about delivering real-world value.
00:40:23.519 -->
00:40:26.320And that means not solving the easy problems, right?
00:40:26.400 -->
00:40:31.920Like we used to have AI demos, and then you could ask it, like, how many vacation days do I get?
00:40:32.159 -->
00:40:35.679It's like, okay, sure, like AI can answer that, but that's not valuable, right?
00:40:35.760 -->
00:40:39.679Like you could do that before already, and it's it doesn't really add value.
00:40:39.840 -->
00:40:49.519But if you can actually do some of the things we talked about in this podcast, if you can really do that with AI, that is so much value that we're adding to society.
00:40:49.840 -->
00:40:53.840Um uh, you know, the so we're we we live in very exciting times.
00:40:54.079 -->
00:41:14.400So, sort of my my my parting thought, I guess, for for folks listening in is really try to be ambitious and and think about what you are going to be able to do in the next couple of years, and and don't settle for the boring, easy use cases and like you know, basic question answering on some documents or vacation days or things like that.
00:41:14.480 -->
00:41:25.840Like try to solve really, really hard, valuable problems where you have high enough accuracy that you can actually put this in front of your own customers and and uh and rely on AI uh doing a good job.
00:41:26.079 -->
00:41:27.599So we're we're heading there.
00:41:27.679 -->
00:41:29.440Uh it's gonna happen in the next few years.
00:41:29.599 -->
00:41:30.880So it's exciting times.
00:41:31.199 -->
00:41:32.239Exciting, exciting.
00:41:32.320 -->
00:41:32.960I love that.
00:41:33.119 -->
00:41:40.800And and it's extremely important that uh when we do AI, we think about actual and real ROI, and that's why it matters.
00:41:41.039 -->
00:41:48.960Well, today we have seen how the semiconductor industries provide one of the toughest, most fascinating task beds for AI.
00:41:49.119 -->
00:42:04.800From teaching machines the language of semiconductors to bridging structure and yes logs with unstructured design nodes to avoiding costly error snowballing, the idea of a unified context layer shows us a path forward.
00:42:05.119 -->
00:42:19.679Uh, we've also explored how context-aware AI could serve as the brain of future lights out factories and how all this innovation must be built on the foundation of trust, security, and data ownership.
00:42:20.079 -->
00:42:27.519It's clear that semiconductors don't just need more data or faster model, they need contextual intelligence.
00:42:27.679 -->
00:42:31.840Uh, that is exactly what contextual AI is bringing to the table.
00:42:32.159 -->
00:42:37.679So a big thank you to Dow for from Contextual AI for joining us and sharing all these insights.
00:42:37.760 -->
00:42:41.760And a big thank you for our listeners for tuning in to Advanced Talk Semi.
00:42:42.480 -->
00:42:45.199And I look forward to seeing you in the next episode.
00:42:45.519 -->
00:42:46.400Thank you for having me.
00:42:46.960 -->
00:42:47.440Thank you.
