Learning Library

← Back to Library

Beyond Chatbots: Deployable AI Intelligence

Key Points

  • The hype around chat‑based interfaces overstates AI’s true potential; we should view large language models (LLMs) as general intelligence that can be embedded throughout applications, not just as a chat window.
  • LLMs represent “deployable intelligence,” meaning they can be assigned tasks much like a high‑performing employee, with future versions gaining more autonomous, agent‑like abilities.
  • Even as LLMs become more autonomous, ultimate accountability remains with the human delegator, especially because AI still lacks true business judgment, which relies on nuanced, context‑dependent decision‑making.
  • Current LLMs excel at processing explicit information but struggle with the implicit, political, and cultural cues that underpin effective business decisions, limiting their ability to replace human judgment.
  • Designers need to broaden their mindset to integrate LLM capabilities across diverse workflows and interfaces, moving beyond the narrow chatbot paradigm to unlock richer, more functional AI deployments.

Full Transcript

# Beyond Chatbots: Deployable AI Intelligence **Source:** [https://www.youtube.com/watch?v=ODLyBQd4VHU](https://www.youtube.com/watch?v=ODLyBQd4VHU) **Duration:** 00:17:29 ## Summary - The hype around chat‑based interfaces overstates AI’s true potential; we should view large language models (LLMs) as general intelligence that can be embedded throughout applications, not just as a chat window. - LLMs represent “deployable intelligence,” meaning they can be assigned tasks much like a high‑performing employee, with future versions gaining more autonomous, agent‑like abilities. - Even as LLMs become more autonomous, ultimate accountability remains with the human delegator, especially because AI still lacks true business judgment, which relies on nuanced, context‑dependent decision‑making. - Current LLMs excel at processing explicit information but struggle with the implicit, political, and cultural cues that underpin effective business decisions, limiting their ability to replace human judgment. - Designers need to broaden their mindset to integrate LLM capabilities across diverse workflows and interfaces, moving beyond the narrow chatbot paradigm to unlock richer, more functional AI deployments. ## Sections - [00:00:00](https://www.youtube.com/watch?v=ODLyBQd4VHU&t=0s) **Beyond Chatbots: Deployable AI** - The speaker argues that AI should be seen as deployable intelligence rather than merely a chat interface, urging designers to envision LLMs as agents integrated across applications. ## Full Transcript
0:01AI is not a chat box it's just not and I 0:05want to talk about that because I think 0:07that we have overstated the value of the 0:09chatbot because we have seen how 0:12powerful the release of chat GPT 3.5 was 0:16in the world landscape and that specific 0:18release changed our perception of what 0:20generative AI is capable of that's great 0:24but one product release does not make a 0:26user interface and so I actually want to 0:28talk through where I think the 0:30capacities of llms are going and then I 0:34want to revisit and say how does this 0:36not fit a chatbot interface because I 0:39want you to walk away from this with 0:40some more open Horizons around where we 0:44could be designing uh llms into our 0:47applications we need more designers 0:51thinking about how llms could be more 0:55effectively deployed outside that little 0:57chat window that you see in the corner 0:59okay 1:00so the first thing I want to call out 1:02there will be five of these I want to 1:03call out on the capacities of AI by the 1:05way so number one AI should be thought 1:08of as Deployable intelligence in other 1:11words we've never really had 1:14intelligence that we can deploy 1:15management Theory would say that you can 1:17deploy intelligence among your employees 1:19but if you're working with fairly 1:20high-powered knowledge workers they are 1:23deploying their own intelligence 1:24autonomously and they should be that's 1:27great AI is directly Deployable by the 1:30person managing it and that means on the 1:33one hand that you have to put time into 1:35managing it and I think that we're going 1:36to start to see llms develop more 1:38agentic properties so they'll be able to 1:40do things autonomously more over the 1:42next couple of years which means 1:44eventually you're going to get to a 1:45point where you can delegate in a way 1:47that's sort of similar to the way you 1:49would delegate to a high-powered 1:51employee and they will just go off and 1:53and go after a particular task or 1:55particular objective 1:56independently that's not where we are 1:58today but I I think even if we get there 2:01Deployable intelligence is still the 2:03right frame because at the end of the 2:05day they are going to be accountable to 2:08the person who is driving that 2:10delegation in a way that an employee is 2:12not an employee is expected to exercise 2:15business judgment and an 2:17llm may give you business perspective 2:21but I think their ability to exercise 2:23business judgment is going to be one of 2:24the last things that AI is able to 2:27conquer so to speak and the reason for 2:30that is that business judgment is a very 2:31context dependent art and it's very very 2:34difficult to get right and I actually 2:36haven't seen a lot of evidence that llms 2:38are getting better at it despite their 2:39progress in a lot of other areas they 2:41are certainly getting better at talking 2:44business but business judgment is 2:47actually an act of decision against a 2:49particular set of data and I think part 2:51of why llm struggle with that is they 2:55can't really take in a lot of the 2:56implicit data that business judgment is 2:58grounded in they take in explicit data 3:00and so if you're looking at the 3:02complexity of the politics in a 3:04particular organization as part of your 3:05business judgment llms can get that if 3:08you exhaustively describe it to them but 3:10that's still a low Fidelity picture of 3:12what's really going on because they 3:14can't experience it that's just one 3:16example so wrapping that back to the top 3:18fundamentally if we think of AI as 3:20Deployable 3:22intelligence we will have a much more 3:24flexible view of where it can be 3:26deployed it's not just in a chatbot all 3:28right number two AI is fast and 3:31therefore cheap and I think that is one 3:33of the fundamental attributes that we 3:35have forgotten chatbots are designed to 3:37be callon response right like you're 3:39supposed to have an interaction what we 3:41miss there is this idea that llms are 3:45fundamentally much much faster at doing 3:47a bunch of tasks once they're trained 3:50than humans are and so even if they do 3:52them at lower Fidelity if we are 3:54tolerant of a lower Fidelity task done 3:57with additional 3:58guidance we you're going to go way 4:01faster on a whole range of knowledge 4:02working tasks and that's in anecdotally 4:04what I see AI being used for in 4:09organizations people are using it to do 4:11a bunch of tasks that would previously 4:13take a long time and they're just 4:15getting it done 4:16faster all right uh number 4:19three AI is opening up new business 4:22models because of that cheapness and 4:25speed and I want to call that out 4:26separately because we are just at the 4:28beginning of uncovering what that looks 4:29like 4:30I think that what we see with the uh 4:32layoffs at McKenzie is a good example of 4:34this McKenzie basically functioned on 4:36this idea that human intelligence is 4:39something that is hard to replicate 4:41anywhere else and so if you have human 4:43intelligence as a 4:44consultant you sort of have for lack of 4:47a better term a monopoly or a corner on 4:51the ability to provide consultant 4:52services and I think what McKenzie and 4:54others are discovering is that at the 4:56end of the day a lot of people are using 4:58chat GPT as their cheap MBA 5:03consultant and even if it's not quite as 5:06good as McKenzie uh depending on your 5:08opinion of 5:09McKenzie it's still good enough and 5:11that's really the fundamental issue a 5:13substitute doesn't have to be 100% 5:16equivalent it can be good enough and 5:19McKenzie is finding out that a lot of 5:21companies in an era of belt tightening 5:23when the cost of cash is 5:25higher they're going to be looking for 5:27cheaper options and there's this new 5:29technology right at the CEO's fingertips 5:31that basically functions as an MBA on 5:34top 5:36so the reason I'm calling that out is 5:39that we have gotten to a point where 5:41we're starting to see the old business 5:43models disrupted but we have not yet 5:45gotten to a point where we're seeing the 5:47new business models come into play and I 5:49think one of the interesting hints at 5:50new business models is the insistence 5:53companies have on training their own 5:56models they want to see their own models 5:59ins inside the house trained on their 6:01data that are effectively private models 6:04that they can use the way they see fit 6:07and I think that that may be a hint that 6:09we are seeing a move toward for lack of 6:12a better term vertically integrated 6:14intelligence so there's this idea in 6:16business that if you vertically 6:18integrate you gain efficiencies right 6:19you reduce costs up and down the supply 6:22chain what we're seeing here is 6:24potentially that businesses want to 6:25vertically integrate the intelligence 6:27stack and they therefore can do more 6:30with what they have if they are building 6:32inside the house they save costs but 6:35they also save on risk right like they 6:37know what that AI will produce they know 6:39what that model is capable of and I 6:41think the risk piece is bigger 6:43anecdotally than the cost savings piece 6:46right now no one wants the debacle from 6:48Air Canada where the chatbot started to 6:51issue fake policies and people see 6:53vertical integration is a way to handle 6:55that so my guess is the new business 6:59model disruption may actually look a lot 7:01like vertical integration and what's 7:04interesting is that may Empower smaller 7:06firms who are tightly vertically 7:08integrated and who have a compelling 7:10value proposition to compete really 7:11really effectively because they can move 7:13fast as a business and they're moving 7:16fast as a business because like AI is at 7:18the core of everything they do we will 7:20see it's a guess uh but I think 7:23fundamentally we're at a stage where 7:24we're sort of at that crossover point 7:25where we're starting to see the 7:26disruption of the old models like 7:27McKenzie and we're starting to just 7:29begin to see hints of the new business 7:31models okay number four smart deployment 7:36is a function of Habitual access and so 7:38this is skipping over to the UI side of 7:40things if you are building for llms you 7:44need to build assuming that you have to 7:46meet a very low friction bar for any 7:49kind of success and I see a lot of 7:50people assuming that a low friction 7:54bar means what it doesn't for lack of a 7:57better term people people assume a 7:59degree of commitment to using 8:01AI that is 8:04absurd I've seen people say oh they'll 8:07go to the special page on my Wiki and 8:09they'll use my special model they're 8:11just going to use chat GPT because 8:13that's the default in their heads like 8:15that's one of the consequences of having 8:17a highly successful external product is 8:19that people have been wired to use this 8:21UI and even if it's not an ideal UI 8:25they're still using it because that's 8:26the default habit and so if you want to 8:29deploy an llm you have to be thinking as 8:32a product person how do I reduce the 8:34friction in this space so that it is so 8:38easy to be habitual with this where are 8:39people already in my space and where can 8:43I put the llm right there where they're 8:45where they're at so they have the habit 8:47of using it in the spaces they're 8:48already in I think that's really 8:50important to drive effective usage okay 8:53fifth one generative AI means 8:56hallucination it means it because you're 8:59gener generating data and I think that 9:01when people talk about hallucination 9:02they keep talking about it like a defect 9:05and it's not it's built into the system 9:07you're designed to generate data well 9:08it's going to generate 9:10data and generating data means 9:12inherently some of the data is not going 9:14to be factual and so we need to get past 9:18the point where we assume Hallucination 9:20is a defect from a technical point of 9:22view and we look at it as an undesirable 9:24outcome from a business point of view 9:26and there's two comments I have there 9:28number one I think there's a billion 9:30dollar opportunity out there for a third 9:32party who can validate the factual 9:34accuracy of llm outputs maybe that's a 9:37service on tap where you just call them 9:39and they validate it I have no idea I'm 9:41not building it but somebody should 9:44somebody should think about the problem 9:46of factual accuracy as if it's a first 9:48class problem because it's about to be 9:50something that people are going to pay 9:52big bucks to be sure of and that is a 9:54lot of the reason why companies are 9:56putting all this effort into fine tuning 9:58and building those internal models as I 9:59touched on earlier they want to be in 10:01control of the risk imagine a world 10:03using the Air Canada example where the 10:06llm had to call a factchecking service 10:10before it could put that policy language 10:13out there in front of the customer it 10:15would be a different world we wouldn't 10:16have that story in the news because the 10:18llm would have given the correct 10:20response and so businesses May build 10:23this internally using tools like Lang 10:24chain but fundamentally I think that 10:27factual accuracy as a service is going 10:30to be an interesting opportunity and now 10:32I I grant you learning the facts inside 10:35the business is something that a third 10:37party service May struggle with so 10:38there's still going to be that internal 10:41element but we have a lot of widely 10:43accepted facts around the world that are 10:46publicly available that are in 10:47newspapers that are in scientific 10:49Publications those are all things that a 10:52third party service could act as a Data 10:54Bank for and fact check against Food For 10:57Thought I will also say the second Point 11:00here on generative Ai and hallucination 11:02I have seen this personally get better 11:04and I would bet that even though we may 11:06need a thirdparty industry or service of 11:09some sort to to validate the outputs of 11:11generative AI we are still going to be 11:15smart to bet on progress from the man 11:18model manufacturers here open Ai and 11:20others are going to get better at 11:22producing more factual AI outputs and 11:24they're not going to talk about how they 11:26did it but they're going to get better 11:27and I've seen that since starting to use 11:29chat GPT 3.5 is that it start it's 11:32getting better every generation at 11:34producing outputs that are factual and 11:36that is reducing my risk radar which is 11:40dangerous in one sense because I 11:41probably should be more thoughtful about 11:44those edge cases that still emerge than 11:46I am but as a human being I have to make 11:48risk assessments all day and I am 11:50starting to find that it is Dependable 11:52enough that I am relaxing and trusting 11:55it more and that's a very intuitive 11:56thing and maybe that's just me but my 11:58sense as those models are getting better 12:00over time as open Ai and others who are 12:03building those models start to get more 12:05and more emphatic about driving clear 12:09factual output so that they don't get in 12:12trouble anyway let's recap five things 12:16that militate against this idea of a 12:18chatbot and I'm going to kind of get 12:19into how chat 12:20Bots are ineffective in the next section 12:23but this is the first section so the 12:25five things where AI is Deployable 12:26intelligence which means it's much more 12:28widely deployed able than a chatbot AI 12:31is fast and therefore cheap which means 12:33that it doesn't have to be limited to a 12:35chatbot right AI opens up new business 12:38models that's more of the business model 12:39side of things AI is all about smart 12:44deployment as a habitual access point 12:47and so that's a UI thing where where I 12:49was talking about the fact that you have 12:50to put the llm where people are and not 12:52just assume people will go to a special 12:54page or whatever to use it and finally 12:57generative AI means hallucination and we 12:59should stop pretending it doesn't and 13:00that also has chatbot implications 13:03because chatbots and we're just getting 13:05into this second section here right the 13:07this second section of the video is all 13:09about the three flaws that I think are 13:12really killing the chatbot as a UI for 13:18llms so the first one that I want to 13:20call out uh and I just jumped into it is 13:23that chatbots simulate a human 13:26conversation and that means that they 13:28cause hum humans to overstate the 13:30veracity or the accuracy of llm outputs 13:33the problem is that humans believe 13:36chatbot outputs more because it looks 13:38like the kind of conversation we're used 13:40to having in text messages with other 13:42humans all day 13:43long and that's a big problem and I 13:45think if we had a different UI we might 13:48be tuned to believe chatbot outputs less 13:52compellingly right like we would be less 13:54likely to believe them second reason or 13:58second issue with llm uh in 14:00chatbots llm capacity continues to 14:04evolve but the chatbot interface largely 14:06is not I think anthropic has done a 14:09little bit of a push on that by S of 14:10starting to develop a canvas where the 14:12chatbot can paint something on the side 14:14like a chart and that's a great step but 14:17in 14:17general the chatbot interface is staying 14:20static as models evolve and so people 14:22are actually not really keeping track of 14:25how models are evolving because there's 14:26no indicator to think to say in the 14:28interace face that chat chat GPT 40 is 14:31smarter than chat GPT 3.5 if you are 14:34like me and you study this all day then 14:36you know that they are but if you're 14:38just a casual user using the chatbot it 14:40looks exactly the same and apple figured 14:44out that you have to make the iPhone 14:45look different car manufacturers figured 14:47out you have to make the car look a 14:48little bit different in order to sell 14:50the new car and I think there's 14:52something some insight there around 14:54human psychology in sort of how we 14:56handle interfaces with evolving models 14:58and I kind of want to put a pin in that 15:00for a future 15:01conversation okay the third reason that 15:03chat Bots are an issue chat Bots require 15:06Advanced llm knowledge because it's just 15:08a text input box it's up to you to know 15:12enough about the llm to use it well and 15:13I have literally seen sidebyside 15:15examples where I've been prompting next 15:16to someone else at the same time and I'm 15:18getting different results and it's not 15:20just that the generative AI is producing 15:22hallucinations it is that the generative 15:24AI is something I am more able to prompt 15:27because I know more about it 15:29and I know more about how it works than 15:32the person sitting next to me and that 15:34is fundamentally inequitable and I mean 15:37that in the sense that if the Deployable 15:39intelligence is there we should be 15:41developing user interface principles 15:45that allow anyone to access that 15:47Deployable intelligence it should not be 15:50dependent on the ability to read an 15:52exhaustive number of Twitter posts and 15:54read a bunch of articles about how llm 15:56works and prompting in order to drop an 15:59effective conversation with a chatbot 16:01you should be able to Max the capacity 16:03without that esoteric or hidden 16:06knowledge you just should okay so where 16:09did we end up with all this I wanted to 16:10call out sort of that those five themes 16:12that I did in the first half of the 16:13video around how AI is changing in ways 16:19that make it bigger than a chatbot and 16:21then I wanted to spend the second half 16:23talking about the three issues I see 16:25specifically with chat Bots because I 16:26think we don't talk about them enough so 16:28so I talked about how chat Bots require 16:31Advanced llm knowledge I talked about 16:33how llm capacity is evolving in ways the 16:36in for interface is not and I also 16:38talked about the fact that chatbots 16:40simulate human conversation in a way 16:42that probably causes us to overstate 16:44their intelligence and the veracity of 16:45facts that they give us I think those 16:47are all concerning so where I net out on 16:49this is I think we're in an inflection 16:52point where we need both new business 16:53models and also new UI models for large 16:56language models like I I just think we 16:58do and I know overused the word models 17:00here but it's an AI talk so what do you 17:01want um so that's where I'm netting out 17:04and I would be interested to hear 17:06examples in the comments of where you 17:10see Mo business models or user interface 17:13models that you think are a direction 17:16that the future could build on as far as 17:18how we deploy large language models in 17:20Tech because I'm always looking for new 17:23examples all right this has gone on long 17:25enough cheers