[HN Gopher] You can't build a moat with AI
       ___________________________________________________________________
        
       You can't build a moat with AI
        
       Author : vsreekanti
       Score  : 28 points
       Date   : 2024-04-11 19:19 UTC (3 hours ago)
        
 (HTM) web link (generatingconversation.substack.com)
 (TXT) w3m dump (generatingconversation.substack.com)
        
       | VeejayRampay wrote:
       | the best model is not GPT-4 it's April of 2024
        
         | cpursley wrote:
         | What's the best model now?
        
           | hypoxia87 wrote:
           | Claude 3 Opus
        
             | minimaxir wrote:
             | Opus is an order of magnitude more expensive and slower
             | than GPT-4.
        
       | mattlondon wrote:
       | This is probably why Google or Meta will "win" over OpenAI in the
       | end.
       | 
       | They have _all_ the data, and no one else even comes close.
        
         | altdataseller wrote:
         | Microsoft has plenty of data too. In Microsoft Teams, LinkedIn
         | posts and messages, and Outlook emails.
        
           | dumbo-octopus wrote:
           | Microsoft 365 (nee Office 365) as well. And Dynamics 365. And
           | GitHub. And OneDrive. And SharePoint. And Power Platform.
           | 
           | Honestly I think they might have more useful data than
           | Google, given Bing knows more or less that same as GoogleBot.
           | Meta doesn't come close, unless you want your LLM to be
           | purely conversational.
        
           | hypoxia87 wrote:
           | Plus every company's files in OneDrive and SharePoint.
        
           | ilovetux wrote:
           | Don't forget they have github as well.
        
           | endofreach wrote:
           | I wouldn't worry about microsoft delivering quality in
           | anyway.
        
           | AnthonyMouse wrote:
           | Nobody has explained how they could _use_ that data without
           | producing a model that would emit private information.
        
             | abrichr wrote:
             | Perhaps de-identification before training could be helpful
             | here.
             | 
             | Microsoft does seem active in this, e.g.
             | https://microsoft.github.io/presidio/
        
         | greenavocado wrote:
         | Large companies like Google will fail at making the most
         | successful LLMs because of internal cultural problems.
        
           | altdataseller wrote:
           | Go away ChatGPT
        
             | sumeruchat wrote:
             | So annoying. I feel sorry for people who actually read that
             | paragraph.
        
       | htrp wrote:
       | > We firmly believe the moat for AI application is in the data
       | and the data engineering today. At some point, the process of
       | building custom LLMs might get so fast and easy that we'll all
       | return to building our own models. That simply isn't the case
       | today.
       | 
       | Customized small models will outperform larger general models for
       | your specific use case.
        
         | sumeruchat wrote:
         | No they wont. Any model you train now will be beat by GPT5
         | easily
        
           | crooked-v wrote:
           | I think the real power of customized small models will be
           | running things on local hardware, except that we're in an
           | awkward phase where the local hardware isn't quite beefy
           | enough to run anything really useful yet. Maybe Apple will do
           | something interesting in that space at WWDC.
        
             | sumeruchat wrote:
             | Also not feasible. A network request to groq type machines
             | will outperform your local hardware by such a huge amount
             | that it wont make sense other than some very niche tasks
        
               | transitorykris wrote:
               | Network availability, latency, privacy, etc. many
               | qualities to consider beyond model size and performance
               | for applications.
        
               | happypumpkin wrote:
               | And cost-efficiency, if I'm using an LLM as an Siri-like
               | assistant on my phone, most of the tasks I'll want it to
               | do won't be that complicated and it would be a waste to
               | send them to some SOTA LLM in the cloud, which I'll have
               | to pay for by a monthly subscription or on a per-token
               | basis.
        
               | littlestymaar wrote:
               | Except nobody but groq has such type of machines, and the
               | economics of cloud AI is very hard to make it works in
               | practice. Offloading the _capital_ cost (which is the
               | hardest kind of cost to swallow for a company) to
               | customers is very compelling business-wise.
        
           | littlestymaar wrote:
           | Not for most human language, or anything that requires
           | business-specific context where what's publicly available
           | lags behind the state of the art your business cares about.
           | 
           | And of course, not if you care about token throughput more
           | than fancy abilities. Or price for that matter.
           | 
           | So for many if not most businesses needs GPT-4 isn't the best
           | tool out there, and GPT-5 is the canonical example of a
           | vaporware right now.
        
           | VHRanger wrote:
           | Past performance is not a predictor of future performance.
           | 
           | While it's possible the gap between GPT5 and 4 is as big as
           | between 4 and 3, it's unlikely. The gap between 2 and 3 was
           | much larger as the one between 3 and 4 (and similarly between
           | 1 and 2).
           | 
           | Also, it's not clear that GPT5 will do this in an *economical
           | way* once the spigot of investor money stops.
        
         | jsemrau wrote:
         | Tell more more about the data moat.
        
         | marcosdumay wrote:
         | I did expect that too, but it isn't happening reliably.
         | 
         | But small customized models seem to perform close to as well as
         | large general ones.
        
       | altdataseller wrote:
       | "What that also means is that you don't need to be an AI genius
       | to succeed in building applications. With thoughtful software
       | engineering and a focus on customer data, you'll build a moat
       | over time."
       | 
       | This sounds encouraging at first glance, but even more
       | demoralizing when you think about it. It doesn't matter how
       | clever or smart you are over your competitors. If you don't have
       | the data, you don't stand a chance. And of course the incumbents
       | have the data not you, the entrepreneur
        
         | killthebuddha wrote:
         | I don't see it that way. If you build a genuinely novel
         | application then the critical data doesn't exist yet (IMO
         | almost by definition). Sometimes it's easier for a startup to
         | do this rather than for an incumbent who's trying to shoehorn
         | the application into a preexisting framework (technical,
         | operational, whatever).
        
           | marcosdumay wrote:
           | If the data doesn't exist, you can't train an (neural
           | network) AI as they exist today.
           | 
           | Whatever thing people build, will be based on data that
           | exists, not data that will be created afterwards. And guess
           | what, _you_ don 't have that data.
        
             | killthebuddha wrote:
             | You could bootstrap a niche model using what's available
             | and go from there. I don't think it's all that different
             | from any other kind of bootstrapping startups tend to
             | require.
        
       | gravitronic wrote:
       | Before I clicked I expected the article to be about LLM's
       | failures in structural engineering applications
        
       | BadHumans wrote:
       | If you can't build a moat with technology you build it
       | politically which Altman is already trying to do.
        
       | minimaxir wrote:
       | ...which is why OpenAI is focusing more on enterprise sales and
       | unique value-propositions such as the GPT Store which can't
       | trivially be imitated by competitors.
       | 
       | This post seems to misunderstand what a "moat" is in a business
       | sense (and unfortunately a lot of AI hypesters on social media do
       | as well). The fact that LLMs are becoming a commodity was the
       | point of the original "OpenAI has no moat" memo by Google, which
       | has proven to be accurate.
        
       | ianbicking wrote:
       | "It might feel like your applications' prompts or prompt
       | templates are a good form of differentiation. After all, your
       | top-notch engineering team has invested days into tuning them to
       | have the right response characteristics, tone, and output style.
       | Of course, giving your competitors your prompts would probably
       | accelerate their progress, but any good engineering team will
       | figure out the right changes quickly. The main reason is that the
       | experimentation (with the right evaluation data!) is quick and
       | easy -- trying a new prompt template isn't much harder than
       | writing it out. All it really takes is a little bit of patience,
       | some creativity, and extra Azure OpenAI credits."
       | 
       | And yet, over and over, I see products with output that clearly
       | comes from simple and frankly lazy prompting. You can do a lot
       | with prompting, but engineerings are not putting in the work! (If
       | an engineer is even the right person... probably not, given any
       | specific application of an LLM.)
       | 
       | Prompting also isn't so reductive that you just write an
       | evaluation and then iterate on the prompt until you satisfy the
       | evaluation. Prompting is a co-creative exercise between the LLM,
       | the domain expert, the product, and the user. And sure "data"
       | fits in there, as well as relationships, comprehensibility,
       | workflows, etc etc... the AI component is just a small piece of
       | any full application.
        
       | hintymad wrote:
       | > The real differentiation lies in your data you feed into your
       | models.
       | 
       | It's more than data. Steven Jobs used to tell the founders of the
       | Segway that their secret technologies would leak sooner or later,
       | even if they set up their factories in the middle of a desert in
       | Nevada. His advice to the founders was that they needed to build
       | a product that users couldn't get away from even if all the
       | competitors had the same technology (or so I remember). That
       | could be a process, an ecosystem, market penetration with an
       | amazing supply chain (think about the largest seller of straws in
       | the world. The unit price of a single straw is close to 0, which
       | is really hard to achieve), and etc.
       | 
       | To me, data will just be one key component of the moat but not
       | all. The moat of AI is the same ol' entire ecosystem: an
       | infrastructure that is so efficient that the company can keep
       | driving down the unit price of hosting the AI models; a fabulous
       | culture to enable the company to keep churning out improvements;
       | an extensive data platform and the associated process and sources
       | that keeps provisioning quality data, a number of killer
       | applications...
        
       ___________________________________________________________________
       (page generated 2024-04-11 23:01 UTC)