Learning Library

← Back to Library

AI Foundations for Non‑Tech Professionals

Key Points

  • The talk is aimed at non‑technical professionals who work with AI daily (e.g., marketing, sales, product, leadership) and will cover the basics of how AI works and its broader implications.
  • Core technical foundations are explained in plain language, focusing on neural networks (pattern‑recognizing artificial neurons, back‑propagation) and tokenization (breaking text into manageable “building‑block” units).
  • These fundamentals are then linked to how machines learn—covering concepts like fine‑tuning, context windows, and the emergence of behaviors such as hallucinations when AI operates in real‑world settings.
  • The presentation concludes with a look at AI’s social impact in 2024, exploring how large language models are reshaping careers, companies, and society at large.

Full Transcript

# AI Foundations for Non‑Tech Professionals **Source:** [https://www.youtube.com/watch?v=u7GayC4XTJ0](https://www.youtube.com/watch?v=u7GayC4XTJ0) **Duration:** 00:32:04 ## Summary - The talk is aimed at non‑technical professionals who work with AI daily (e.g., marketing, sales, product, leadership) and will cover the basics of how AI works and its broader implications. - Core technical foundations are explained in plain language, focusing on neural networks (pattern‑recognizing artificial neurons, back‑propagation) and tokenization (breaking text into manageable “building‑block” units). - These fundamentals are then linked to how machines learn—covering concepts like fine‑tuning, context windows, and the emergence of behaviors such as hallucinations when AI operates in real‑world settings. - The presentation concludes with a look at AI’s social impact in 2024, exploring how large language models are reshaping careers, companies, and society at large. ## Sections - [00:00:00](https://www.youtube.com/watch?v=u7GayC4XTJ0&t=0s) **Foundational AI Concepts for Everyone** - The presenter offers a non‑technical overview of core AI technologies, machine‑learning mechanisms, emergent phenomena such as hallucinations, and the 2024 social impact of large language models across roles like tech, marketing, sales, product, and leadership. ## Full Transcript
0:01hello today we're going to talk about 0:03foundational Concepts in artificial 0:05intelligence and I'm aiming this 0:07presentation at people who perhaps 0:09haven't had technical Concepts in 0:11artificial intelligence explain to them 0:14but they work with AI every day so if 0:16that's you if you work in Tech if you're 0:19in marketing if you're in customer 0:21success sales product management this 0:24presentation is for you even if you're 0:26in leadership and you want to understand 0:27what is AI so let's dive right into it 0:32here's what we're going to cover today 0:34number one we're going to talk about the 0:35foundational Tech that enables AI so 0:38things like neural Nets Transformers 0:40attention ETC number two we're going to 0:43use that knowledge of foundational Tech 0:45to understand how machines learn so 0:48fine-tuning context Windows all of that 0:51number three we're going to talk about 0:53the emergent behaviors that happen once 0:55these machines start to actually operate 0:58in the real world these are things that 1:01get us excited and or frightened so 1:03Hallucination is one example that we'll 1:06cover and finally we'll talk about the 1:08social impact of AI what can we see here 1:10in 2024 that shapes the way we 1:13understand how llms could influence our 1:15careers how they're influencing our 1:17companies how they're influencing The 1:18World At Large all right let's get into 1:20it foundational Tech what is inside that 1:24magic box and no it's not a ghost it 1:26just sometimes feels like 1:28it all right we're going to we're going 1:30to start with neural Nets neural Nets 1:32are fundamentally machines that we've 1:34built that develop pattern recognition 1:37by mimicking neurons they're artificial 1:39neurons that learn from 1:41experience so like neurons they use 1:43networks to spot patterns learning 1:45happens when they tweak connections 1:47which is what happens in our own brains 1:49and that simple mathematics that tweaks 1:52connections actually yields emergent 1:55understanding if you want to get 1:56technical about it multi-layered 1:59networks of artificial neurons process 2:01input data and they adjust connection 2:04strengths through something called back 2:06propagation to minimize error now back 2:09propagation isn't is that simple 2:11mathematics we talked about is where you 2:13you just have very simple adjustments to 2:15the connection strength and that 2:16minimizes error and it improves pattern 2:19recognition over time as neuron Nets are 2:22trained on large amounts of data now 2:24this foundational concept underlies 2:26everything else that we're going to talk 2:27about especially trans for our 2:33architecture another key concept I want 2:36to talk about is 2:38tokenization tokenization is this idea 2:41of a vocabulary of small units that 2:45helps the machine read relationships at 2:47a larger range of scales so tokenization 2:51actually chops text into little legol 2:53likee building blocks it helps machines 2:56recognize patterns and Tents and other 2:58parts of the word suffixes prefixes and 3:01it balances in between letters and whole 3:04words if you want the technical 3:06explanation algorithms segment text into 3:10discrete units or tokens like words or 3:13subwords or even characters which are 3:15then converted into numerical 3:18representations so that the model can 3:20process them tokenization is a 3:23prerequisite for everything in text 3:25based Ai and it directly impacts things 3:28like context window size which we'll 3:30talk about later in this 3:33presentation so Transformers we talked 3:36about this earlier when we mentioned 3:37neural Nets fundamentally Transformers 3:40enable machines to consider an entire 3:43block of text at once that means the 3:45machine reads the entire sentence it 3:47doesn't just discover the text word by 3:49word that helps it figure out 3:51relationships between words and juggle 3:54those relationships as far as I know jet 3:56GP jet GPD is not autonomously Contender 4:00although I hear there's a lot of AI 4:01enabled dating right now the technical 4:04description here is that Transformers 4:08utilize self attention mechanisms to 4:11weigh the importance of different parts 4:13of input simultaneously as they pull the 4:15whole text in and that enables them to 4:18process in parallel across that 4:20relationship set and to capture very 4:22long range dependencies and so part of 4:25why this text processing is important is 4:27if you have a good siiz context window 4:29which will get to you can then 4:31understand how parts of the text that 4:33are somewhat separated actually relate 4:35together humans can read a text and say 4:37oh well this is how the introduction led 4:39to the conclusion that kind of thing is 4:41something that Transformer architecture 4:43enables machines to do as well this is 4:47key to the performance that modern 4:48language models have been able to unlock 4:51and relies heavily on this idea of self 4:55attention which is what we want to get 4:57to next here an llm can 5:00dynamically focus on different parts of 5:03the conversation independently that 5:05means it can listen for connections it 5:07remembers context versus current words 5:10and it can shift Focus as your 5:12conversation with the llm 5:15evolves so technically speaking that 5:17means it's Computing attention scores 5:20between all pairs of input elements 5:24creating a weighted sum that emphasizes 5:27the relevant connections and the textual 5:30information now part of how it does that 5:33on the Fly is based on previous training 5:36where it's done this it's looked at test 5:38texts before it's looked at a lot of 5:41training Data before and it's computed a 5:44lot of these scores in the past so that 5:46it understands this is how this kind of 5:49text typically works these are the 5:51relationships that typically apply and 5:53that allows it to you know behave as if 5:56it's seen resumés before when you input 5:58your resume 6:01all right this mechanism is the crucial 6:05underlying factor for Transformer 6:08architecture in fact there's a very 6:10famous paper in AI called attention is 6:13everything and it's all about why this 6:15unlocks so many capabilities in 6:18llms it contributes to the model's 6:20ability to understand and the entire 6:23piece of context that the chat is giving 6:26it we'll get into chats and context 6:28windows in a little bit all 6:31right so those are some foundational 6:34Concepts when we understand those how do 6:37those help us understand how models 6:38learn and how they adapt for different 6:42applications general knowledge is too 6:44General most of the time is one of the 6:46things I want you to take away here 6:48because we have these huge models that 6:49are very general but a lot of the work 6:53being done in AI is basically taking 6:55those foundational models that are 6:56trained on so much data most of what 6:58Humanity has created created today and 7:01and focus on taking them to a particular 7:03application and tuning them we'll 7:05describe how that works so the first 7:08concept to understand when you're 7:10talking about this is transfer learning 7:13so if you want to talk about how 7:14machines learning fund how machines 7:17learn fundamentally pre-training on lots 7:20of General data enables rapid learning 7:23of specialized 7:26subjects so what that means is that AI 7:30is using Knowledge from one task for 7:32other tasks because it has that 7:34foundational knowledge 7:36set it initially needs a massive amount 7:39of data and when I say everything that 7:41Humanity has written to date I really do 7:44mean that it's everything they can find 7:45on the internet it's digitization of 7:47every book they can get a hold of it's a 7:49massive amount of data this enables it 7:53to understand fundamentally how human 7:56language 7:57works there is a lot of suspicion that 8:02llms have figured out intuitively a 8:06grammatical structure that linguists 8:08can't yet formally describe which is a 8:11nice little tidbit for a cocktail party 8:13sometime uh when it's learning it 8:16preserves learned patterns as it looks 8:19at new data and those patterns are 8:21driving really quick learning when it 8:23grasps new subjects so it looks at 8:25things in the light of what it's learned 8:27before if you want to get technical 8:30about it pre-trained models on large 8:33data sets are fine-tuned on smaller and 8:37task specific data 8:39sets and that allows the transfer of 8:42learned features and knowledge across 8:44those domains 8:46efficiently this is closely related to 8:49fine-tuning which we're going to get 8:52into all right 8:54fine-tuning is when you are adjusting 8:57parameters or weights from your general 9:00learning data set to more effectively 9:03fit a task specific data set so you 9:07start with base parameters from general 9:09knowledge you look at examples of a 9:11specific task and you optimize those 9:13weights I think of this as tweaking a 9:17recipe so technically speaking a model 9:20adjusts pre- trained model parameters 9:23using task specific data often with a 9:28lower learning r rate so it's not 9:30perhaps quite as efficient because it's 9:32a smaller amount of data but it adapts 9:35the model's knowledge to a particular 9:37domain so you get that efficiency back 9:39because there's all of this background 9:41knowledge which is what we talked about 9:43earlier fine-tuning is one practical 9:46application of this idea of transfer 9:49learning and can help mitigate 9:50hallucination and we'll get into talking 9:52about 9:53hallucination so again from from a super 9:56non-technical perspective this is taking 9:58your French omelet recipe and it's 10:01adjusting it just a little bit if you 10:05want to bring in the summer herbs from 10:07the 10:08garden that's the idea of fine tuning 10:10you're fine tuning especially if you're 10:12fine tuning for an audience like you're 10:13bringing someone over for brunch well 10:15you want to optimize your 10:17omlets okay next concept is context 10:22windows so fundamentally llms can only 10:26ingest a certain number of tokens per 10:29turn there is a hard limit and by per 10:31turn I mean per conversational utterance 10:33that we give them and this is the only 10:36text they directly reference every other 10:39piece of text they've read is encoded as 10:41numbers somewhere uh it's not directly 10:44readable as text current limits are 10:47roughly the size of a big novel 10:50technically speaking it's about a 10:52128,000 tokens you might call it um a 10:56little bit over a 100,000 words 11:01300 Page book something like that 11:04there's no direct text reference outside 11:06this window and older inputs in the text 11:09window Decay as new inputs 11:13arrive now there is a hack for this and 11:15I note that llms tend to hack this by 11:18reinges chats and so part of why Claude 11:21will tell you that you can reach a limit 11:24in a chat with Claude is because 11:26fundamentally on the back end Claude is 11:29is taking the entire chat that you have 11:33had about that subject every time you 11:35talk and reinus it as a single piece of 11:38text and so it's sort of like making the 11:41llm reread the entire conversation plus 11:45your new chat every time you 11:48talk and that helps llms to simulate the 11:52kind of 11:54responsiveness that and attention that 11:57humans show without the ability 12:00to have to remember every word of the 12:02conversation first and so when humans 12:05are having a conversation we don't 12:06really think about this I don't exactly 12:08know what our brains are 12:10doing but when we talk all we do is 12:13listen to what's being said and we come 12:16up with a response and we come up with a 12:18response that's contextual right that 12:20matters to the 12:22conversation when llms are responding to 12:25get to that kind of human fluency they 12:27have to reread the ire chat in the 12:30background and if you didn't know that 12:31that's okay you never see that happen 12:33that's something that's submitted on the 12:34back 12:35end so technically speaking a context 12:38window defines the maximum number of 12:40tokens usually in the thousands right 12:43there's there's really none of them now 12:44that are not in the thousands they can 12:46be processed together serving as the 12:49model's working memory for generating 12:51contextually relevant 12:54outputs so context window size is 12:57affected by tokenization 12:59and as I've discussed it impacts the 13:01model's ability to maintain coherence in 13:04longer responses and that also matters 13:06because if you're trying to get it to 13:08sort of do these lengthy responses if 13:10you're trying to get it to write your 13:11document for you it's reinus and 13:14rereading the document and that's why 13:15you sometimes get this idea of like 13:17alphabet soup when you get longer 13:19documents out of AI and that's why I 13:23tend to use AI more often for shorter 13:25tasks for bullets for things that I can 13:27expand into later because it doesn't do 13:30as well with those longer pieces of 13:32text all right let's get into prompt 13:36engineering prompt engineering is all 13:38about guiding an llm toward a desired 13:40response by carefully forming the input 13:44text you're framing the utterance just 13:46to your pattern recall for the llm 13:48you're providing examples to trigger 13:50specifics and by the way those can be 13:51positive or negative both are helpful 13:54and you're laying out very clear output 13:56expectations 13:59so technically speaking prompt 14:01engineering is designing your inputs 14:04with very specific structures or 14:05examples or instructions that guide the 14:08lm's generation process for new text the 14:12next token prediction that the llm 14:14relies on and improves response 14:17relevance and accuracy fundamentally 14:20you're saying I can prompt you with a 14:23set of tokens that are carefully 14:26structured such that you are triggered 14:29into a pattern recognition with 14:31particular relationships you've seen 14:33before between tokens in your general 14:36knowledge and then you're able to come 14:38back and recognize that pattern and 14:40predict a set of tokens which is exactly 14:42what they do when they write out that 14:43text on the 14:45screen that will be more relevant right 14:47that will be more useful to me and so 14:50it's all about being able to be inside 14:53the llm system and part of good prompt 14:56engineering is having a good mental 14:58working model of how llms work and 15:00that's why I'm doing this lesson because 15:02I think sometimes we think of them as 15:03black boxes and that doesn't lead to 15:06helpful 15:08prompts Okay so we've talked about 15:11foundational patterns we've talked about 15:12learning let's talk about emergent 15:15behaviors what are the unplanned 15:17behavioral patterns that are 15:18characterizing llms today in 2024 number 15:21one is emergent problem solving models 15:25trained on huge amounts of text are 15:27showing surprising capabilities like 15:30mathematics 15:31understanding now there are limits to 15:34things like mathematics understanding in 15:36a pure 15:37llm 15:39architecture but there are not really 15:42limits that we found if we adjust an llm 15:45and put it into a larger tool chain 15:48approach which I'll get to and sort of 15:49describe and I want to call that out 15:51because sometimes we see emergent 15:53problem solving that is useful for 15:55everyday tasks like I can now use a 15:58large language model in 2024 and if I 16:01say figure out the tip across six people 16:03it generally does it 16:05fine but if I'm solving International 16:07math Olympiad problems it will not do it 16:10fine because the llm is just not able as 16:13a text model to do that right now that 16:15being said the example I gave is still 16:19relevant because Google just won the 16:23silver medal in the international math 16:25Olympiad by using an llm and a tool 16:28chain approach approach with some other 16:29tools on the back end to solve very very 16:32difficult mathematics problems so this 16:34kind of llm emergent behavior is built 16:37on by humans so that we get more and 16:41more of what we're looking for whether 16:42it's within an llm natively or a larger 16:46architecture so how do llms do these 16:48emergent behaviors like we don't 16:50necessarily design them to be good at 16:52math when we teach them everything we've 16:54written down with human language that 16:55just happened well fundamentally absorb 16:58patterns that llms are good at 17:01understanding have wider applications 17:03we've built this massive pattern 17:05recognition machine and we've given it a 17:08huge amount of data to look 17:10at now when you give it prompts on top 17:13of that data set it can drive a novel 17:15recombination of patterns and so we are 17:18a part of this emergent problem solving 17:20experience too and we may not recognize 17:22that but our unique prompts are 17:25helpful and llms are inherent cently 17:29because of everything we've discussed 17:30previously very skilled at applying the 17:32patterns they know to new 17:35data so fundamentally what's happening 17:40is that llms when they show behaviors 17:42like mathematical reasoning are not 17:45actually doing mathematical reasoning 17:47they're still doing next token 17:49prediction so they still have the same 17:51complex interactions that drive uh 17:54relationships between models and 17:56extensive training data generating 17:59weights that they can go measure and use 18:01all of that foundational stuff we just 18:03talked about for text is exactly what 18:05they're using when you give them the 18:07what tip should I calculate problem all 18:09that's happening is that they weren't 18:11explicitly programmed to do any kind of 18:14math at all but they figured out a 18:16pattern in their data set that they 18:18could 18:19use and that was so successful it 18:22surpassed the initial expectations of 18:24designers designers didn't really design 18:25llms for math they just found out that 18:28they could do it 18:29when they actually built the system 18:30because there's a pattern in mathematics 18:32that is somewhere in the data set that 18:34is good enough that the llm can use next 18:36token prediction to do 18:38math somewhat reliably right it's again 18:42a pure llm not doing the international 18:44math Olympiad but yeah calculating the 18:46tip for dinner they can do that turns 18:49out maybe Humanities have written about 18:51that uh in fact this was reminding me of 18:53the Douglas Adams joke about the 18:55complexity of calculating tips at a tiny 18:59beastro and among six friends nice 19:03little aside there read Douglas Adams 19:04he's 19:06great so what other emergent 19:09capabilities do we have the other one I 19:10want to talk about is sort of on the 19:11other side we might consider emergent 19:13problem solving as good we might 19:16consider hallucinations as bad I use the 19:18same image for both to remind us that 19:20these are Flips of the coin we are 19:22assigning value here these are just 19:24inherent capabilities of the model and 19:26we need to understand how to use them I 19:28think it's honestly more accurate to say 19:31that llms are very very good at 19:33plausible sounding data because we built 19:35them to be and sometimes we interpret 19:38that as 19:39hallucinations but really all they're 19:41doing is using their next token 19:43architecture they don't think in facts 19:46unless we put a tool chain in place 19:47behind them to help them so pure llms 19:50are only predicting the next 19:53word they optimize for language flow 19:55which is why they're such good 19:57conversationalists which is why we been 19:59so uh surprised to see that they've been 20:02used as AI companions first and not in a 20:05lot of other ways that we would expect I 20:06think looking back maybe we shouldn't 20:08have been so surprised about that 20:10because language flow makes them sound 20:12human and again as I was saying 20:15hallucinations can be mitigated using a 20:17toolchain approach that gets into the 20:19fact checking that matters for that 20:20particular 20:22task now we are the 20:26ones that determine what is a 20:28hallucination so the famous example uh 20:32from I think it was Air Canada where the 20:35bot 20:36hallucinated a policy that wasn't 20:40there we assigned the value of 20:42hallucination to the bot's utterance 20:44right the bot just generated a response 20:47that it thought was plausible which is 20:49which it had done every other time and 20:50had a chat with a human in that system 20:53but in that particular case the 20:55plausible sounding response was a policy 20:57that had financial implications for the 20:59airline and we assigned that the value 21:02of hallucination and the airline had 21:06massive news headlines around the world 21:08and had to go and fix it and everything 21:10else the point is we humans need to 21:14think about where we apply the models 21:16that we've designed for next token 21:18prediction because part of this is just 21:20using them the way they should be used 21:21and part of it is building tool chains 21:23that help them to check facts what we 21:25should not be doing is expecting l s to 21:29magically know facts the way we do 21:31because as much as we talk about neural 21:33Nets they're not actually the same as 21:36our brains our brains do have symbolic 21:37logic and reasoning they do understand 21:40what facts are and llms just weren't 21:43built for that and so we shouldn't 21:44expect them 21:46to okay I want to close by reflecting on 21:49how llms are shaping 21:52us this is the first generally 21:54applicable technology in decades so what 21:57are some reliable indicators of how llms 21:59are starting to shape 22:01Society first massive personal 22:04productivity enhancement llms today are 22:08extraordinarily effective accelerating 22:11knowledge workflows I'm going to take a 22:12drink of water 22:15here extraordinarily 22:18effective there was a Danish survey that 22:21came out just a few weeks ago that 22:22talked about something like a 37% 22:25reported Improvement in personal 22:26productivity for knowledge workers in 22:28Denmark based on 22:31llms and it turns out that that's 22:33because Tech worker task knowledge 22:35worker tasks are pattern-driven llms can 22:38tackle many many of these tasks even 22:40without tuning because they're in the 22:42original data set which is why Tech 22:44workers all over the world have a chat 22:46GPT window open or another language 22:48model of their 22:50choice and I would say that that 22:52individual productivity is leading to 22:55based on my anecdotal kind of read of 22:57the situation 22:58team and or productivity one of the 23:01examples here in Tech right now is chat 23:03PRD which started out as an application 23:05for product managers and individual 23:07productivity has recently added on team 23:12plans team uh capabilities and that is 23:15just the tip of the iceberg like what 23:17we're seeing is that llms are following 23:19the path of least resistance where we 23:21start with individual productivity 23:22because it's designed as a call and 23:24response it's designed as an utterance 23:26and a responsive next token prediction 23:28and you have to think about larger 23:30architectures of software to build 23:31effective team productivity off of that 23:34basic flow but it's absolutely 23:37possible uh so watch that space right 23:39you should expect AUM more team in org 23:41productivity enhancements in 23:452025 this is also changing our skill set 23:48patterns so Tech worker skill sets 23:50knowledge worker skill sets have to 23:51shift and we need to think about 23:54allocating intelligence versus just 23:56applying our own intelligence what is 23:58appropriate for an llm as an 24:00intelligence versus our brain as an 24:02intelligence when do we use which is 24:04becoming a hot skill in Tech markets 24:07today so you have to know when to use an 24:10llm and I would argue you also have to 24:13understand that your value is a human 24:15brain has shifted from starting with 24:18drafting in a good draft to editing and 24:21polishing because llms are so good at 24:24drafting you no longer have to face a 24:27blank page problem and that is a 10,000 24:31year old at least problem that is gone 24:35we just don't have it anymore and I 24:37think that's really cool and 24:40fundamentally a lot of the sort of 24:41rubric for allocating intelligence is 24:43having an internal quality dial and 24:45thinking about what are llms good at 24:47what are brains good at where is the 24:49quality meter for this task and what do 24:51I need to get to that quality meter most 24:55efficiently so I would expect that AI 24:59interaction AI oversight AI creative 25:01problem solving are things that we will 25:06continue to need to get better at I 25:09think building and sort of getting into 25:10code is something that we are going to 25:12do more and more because llms have been 25:14the greatest code unlock that I think 25:16I've ever seen they've simplified 25:18people's access to coding as a language 25:21because they look at coding as next 25:22token prediction too that's another 25:24emerging capability that they weren't 25:26necessarily designed for 25:28always check your code coming out of an 25:30llm by the way not always bug free just 25:34like a 25:35human all right another Factor here is 25:38pressure to 25:40monetize and that little guy there he's 25:42back he's thinking about the pressure to 25:44generate return on investment for Wall 25:47Street corporations are on track to 25:49spend roughly a trillion dollars on llms 25:52all in over the next couple of years and 25:53I'm factoring in both the costs that you 25:55have at the end of the stack where 25:57corporations are applying Ai and the 25:59estimate also includes the huge 26:01foundational model Investments and 26:03training Investments for building models 26:05that are are coming up if you're 26:08spending that much money there's 26:09pressure to generate a return on that 26:12investment we have not seen anything 26:15like that return there was a paper that 26:16came out by Sequoia just a month or so 26:19ago in June of 2024 that talks about 26:22this $600 billion shortfall where we 26:24have a gap between what we've spent on 26:28LMS and what we would expect to get as a 26:30return and it's not been 26:33closed and where we have applied llms 26:36our Focus has been on efficiency not 26:38growth and I think that is an issue that 26:41is associated with where we were at in 26:43the economic cycle when llms first broke 26:46we were not at a high point in the cycle 26:49and leaders were looking for things that 26:52helped them drive cost efficiencies and 26:54llms are something that you can get cost 26:57efficiency out of if if you put them in 26:58the right place in the org potentially 27:01and so that's where leaders of pigeon 27:02hold them but Business Leaders need to 27:04understand that creative applications 27:06and growth drivers are probably the 27:08bigger long-term value window here for 27:11llms and we need to think more broadly 27:13about what llms can do in order to 27:15realize a massive return on investment 27:18that would justify the cost of creating 27:20them in the first place right now most 27:23monetization is proxied via the cloud so 27:25if you're using cloud or compute to 27:28drive llm fine-tuning if you're using it 27:31to drive direct usage of large models 27:34like open AI that 27:37charge fundamentally that's turned into 27:40like Cloud Revenue at the Enterprise 27:43level so Enterprises will get a 27:45fine-tuned model that derives from chat 27:48GPT or that derives from some other 27:51model they'll put it on a cloud 27:52somewhere maybe it's on Azure maybe it's 27:54on Google Cloud maybe it's on 27:56AWS and then then the cash register 28:00Rings the way it usually does for cloud 28:02spend every time they use it right and 28:03that's how some of that monetization 28:07happens so I want you to just keep that 28:11one in mind because that one is going to 28:13continue to be a factor in the 28:15background of executive decisions around 28:18AI until we start to see substantial 28:21money back that for example is why Andy 28:24jossi decided to start charging for 28:27Alexa that's something that's apparently 28:29coming very soon and it's because he can 28:32no longer tolerate the cost on the 28:34balance sheet without direct return on 28:37the AI investment for 28:40Alexa all right new security risks LMS 28:44are breaking our mental models of 28:45deterministic security So Bad actors may 28:50never have been as simple as this idea 28:53of someone in a gray hat sitting at a 28:55computer far away typing in and sort of 28:58sending green lines of code charging 29:00across the screen the way we had in the 29:02movies but security risks were 29:05deterministic they were driven by code 29:07and the intent of a particular 29:10hacker these days they're more emergent 29:13we can have things like hallucinations 29:15as drivers of liability we talked about 29:17this with the Air Canada 29:18example we can have voice mimicry that's 29:21autonomous is a new spear fishing Vector 29:24people are using this in the wild 29:26already and it's concerning the ability 29:29to mimic a voice becomes the ability to 29:31mimic identity socially if you apply it 29:33in the right context and that generates 29:35new attack vectors and new 29:37vulnerabilities for all of us finally 29:41the scale at which you can create 29:43information and lm's propensity to focus 29:45on the flow of conversation rather than 29:47the facts means that it's never been 29:49easier to create disinformation and that 29:52risk which has already been probably 29:55100x just from llm text outputs is 100x 30:00again when you throw in the potential to 30:02generate video and images because we are 30:05moving to a much more visual Society Tik 30:08Tok is exploding Instagram exploded back 30:11in the 2010s and people are consuming 30:14stuff via video you're watching this via 30:16video I promise I'm not an AI 30:19Avatar when you look at that General 30:22pattern it's easy to get afraid wow we 30:24have new vulnerabilities we have new 30:27attack vectors there are new security 30:29risks that's all 30:31true 30:33but you have to look at a new technology 30:36in terms of the overall impact to 30:37society and one of the things that I 30:39think is really compelling is that llms 30:41like most general purpose Technologies 30:43are being used both for incredibly 30:46positive applications they're being used 30:47in drug development and also for 30:51applications that are negative for 30:53society like we're discussing here on 30:55this 30:56slide and that's what happens when a new 30:58general purpose technology comes out and 31:00it's up to us as individuals as 31:02communities as companies to make use of 31:05llms in ways that benefit us as a whole 31:08and 31:09that is a nice turning point we are 31:13likely living at this moment in history 31:16that we will look back on as a huge 31:18inflection point for our species and I'm 31:20not talking about that with the 31:21assumption that we're going to get 31:23artificial general intelligence or we're 31:24all going to be like under skynet's rule 31:27shortly I'm actually saying just what we 31:30have today is showing us that large 31:32language model models May well go down 31:35as Humanity's greatest invention when 31:36all is said and done people may look 31:38back on this time and say wow I wished I 31:40lived during that time when everything 31:42was new and everything was on the 31:44horizon everything was something that we 31:46could shape that is the opportunity that 31:49is the challenge of large language 31:52models here in 2024 so there you go I 31:55hope you understand large language 31:56models a little bit better I hope you 31:58have a sense of how you can use them and 32:00uh I'll see you in the next lesson