Re:Invent thoughts

[tippie36]

No surprise that AWS’s focus this year was mainly on Generative AI.

Having worked there I would say that was not the focus at the beginning of the year, and although everything that they presented brings a lot of value and is really cool, some stuff seemed rushed. Even the org that I worked in, completely scrambled to present something that had Generative AI in it and shifted their focused completely from what we were building 12 months ago (FOMO when looking at OpenAI and their recent success).

Anyways, out of the things that did not seem rushed, I noticed Code Whisperer’s new capabilities specifically ones revolving around processing code bases and answering questions about them.

Y Combinator funded a company that literally does the same thing and although I think that’s a great idea, I just don’t see how startups like this one can survive after AWS releases the same thing.

What I’m trying to get at is that it seems to me like bigger enterprises have a very very big advantage since the name of the game for generative AI is data, and lots of it, optimized over time.

Not saying that startups shouldn’t do Generative AI, or that they won’t be successful in doing so, but I’m a bit skeptical when I see all these bets being placed.

Just food for thought, obviously these fellas have a lot more experience than I do in investments.

Also everyone is talking about Open AI because they have in house LLMs developed and trained over time. No one seems to be talking about AWS, Google Cloud & Co - Amazon is literally collaborating with a handful of companies and we all now have access to like 5-6 different LLMs through Bedrock which I assume will be an approach that other companies will take as well.

So another thing to consider for startup founders in my opinion is - why don’t we build another AWS Lambda today? The answer is pretty clear, there are some products out there that are similar from some startups but none of them are billion dollar companies. I think that the same thing will happen shortly with all of this hype, where it will be extremely easy for an enterprise to release these products way faster than a startup would given the amounts of available data.

Speed was one of our main advantages remember? Well not so much in this case anymore.

Use Generative AI to solve real problems but keep in mind that if you just base your entire vision and business on that, you might fail fast since you’re never going to have the same amount and quality of data available to you to customize and improve an existent model (I’m assuming not all of us are PHDs that would be able to create a revolutionary model from scratch).

A lot of these Gen AI companies look like features that AWS can slap out of their existing services in weeks (and have done just that), because AI has become some much more accessible given the right amount of data. They actually probably obliterated 3-6 startups that got funded from the past 1-2 batches just at this reinvent alone…

P.S - you’re gonna tell me that they invest in founders not ideas. That’s cool, I hope those founders who were more successful than I was at getting into YC will consider these points while pivoting to new ideas…

 

[hadler]

@tippie36 All the GenAI stuff AWS announced is a demo ware and preview at best. It’s quite underwhelming if you use it. That’s not to say that it won’t improve over time, it certainly will. They are playing super hard to catch up. And they are pretty good at it. They did it in the containers space (came in after google but managed to get on par). GenAI is different though, it’s more about research and science which AWS is not good at. It’s good at taking shit from open source and product-ionizing it. Bedrock is primarily a distribution channel for AI companies (who may not have the brand recognition they need at the outset) to acquire customers. For customer it would be standardized way to consume, fine-tune and steer models while staying private.

 

[tippie36]

@hadler Yep, all fair points and nothing to object on here. My opinion and worries revolve around the speed given data availability at which these cloud providers for example can move at now. I think the game has changed a lot and the gap to create these tools in a fast and be scrappy way, is increasingly smaller day by day between enterprise and startup.

I think what you just described is what makes Bedrock a very very clever way for Amazon to enter the market. In time they will have more models themselves, I think they just have a text generation model now on there, but I foresee it become a marketplace of models and a playground just for what you described above.

Even if they don’t have in house stuff, their offerings will make it even cheaper and more accessible for any company to fine tune, train and customize existing models.

I think that AWS or any cloud providers don’t even need to get good at their own models in house, because there’s absolutely many many companies who love the Bedrock model - and these companies might not be as good as OpenAI, but with the exposure that AWS is providing them now they will be able to move very very fast in enhancing their model offerings.

Overall, I think I maintain my opinion about ease of access, accessibility and cost, all at which companies like AWS excel. If they want to invest in research and science and develop their own models, they can, but honestly I don’t see that big a sense of urgency for them to do that - these models are mere tools in achieving a bigger purpose.

You’re probably not gonna like the following comparison but I see them like SQL for example, there’s a specific use case or use cases for it, and AWS like other cloud providers have created cloud products around it. They have optimized those products without the need for them to reinvent SQL in house and research that.

Now, why aren’t many startups trying to compete with the cloud storage products built around SQL today? I think it would be highly unlikely that any of those potential competitors would end up succeeding.

I know, not a very good anecdote between the two but I hope it proves my point a little.

 

[outlawstate]

@tippie36 1)

Amazon has traditionally been not great with AI-based services that aren't heavily (at a foundational level) commoditized.

If it turns out that all of these services are 99% LLM and 1% final polish, Amazon (or whichever of the big guys has the best LLM) will win.

If it turns out it is, instead e.g., 60% "base" LLM and then 40% additional voodoo (UX, RAG configuration, multi-LLM stacking, custom data training*, etc.), then there is much more likely to be room for startups to grow up.

(*=Yes, Amazon via distribution could potentially have a volume advantage; but, no, they historically have not been great about exploiting this for cutting-edge products.)

And it simply isn't clear yet which of these two universes we live in (and the answer will presumably be different by vertical/industry/product).

2)

The other piece that is still unclear is how different a "great" versus "good enough" product is going to be (which, again, will differ by vertical). Amazon is, historically, pretty solid at delivering a "good enough" product; if the product gap between "great" and "good enough" is small enough, then Amazon's distribution and size will win.

But Amazon, historically, is much less likely to deliver a "great" product.

Domains where "great" matters will be ones where startups are more likely to thrive.

3)

Lastly...

What things are going to even look like in ~12-24 months is totally a moving target. There is a host is very plausible worlds where something like current gen "Code Whisperer" is barely relevant (you don't need to be a singularity enthusiast to believe that this part of the space could dramatically leapfrog in the near-term). With the underlying tech and capabilities such a moving target, it is even harder to predict where (1) and (2) (from above) land.

 

[tippie36]

@outlawstate Can’t disagree with anything that you’re saying above, although most of the market share is looking for commodity and ease of use - and at that, AWS seems to excel.

Now, working at AWS has really put things into perspective in terms of what I agreed with, disagreed with, liked and disliked about the company. But I’m not sure that we can say with conviction that Amazon is not great, especially AWS.

If they are not great then why are they number 1 in the cloud space? Is it because of commodity and ease of use - oh wait, I think we already talked about that, in which case users will still prefer using that over something else.

Lastly, I started with Amazon and talked about ReInvent but the whole opinion that I have revolved around the fact that I don’t think there’s a lot of room for competition in this space if you’re just focused on dev tools. Gen AI combined with a good vision, non dev tool product? Sure that’s probably gonna have a higher chance of success.

P.S I think AWS is aware of RAG and stacking models as well as the additional voodoo. They talked about that during the conference, and they had a presentation I think which was solely focused on stacking models - again, they’ll make it much easier and convenient to stack them and optimize them (as an example).

 

[outlawstate]

@tippie36

I think AWS is aware of RAG and stacking models

Obviously. Implementing them correctly/maximally is quite hard, however, as I discussed in my post. Historically, AWS has not been great at pushing SOTA on AI. They have basically zero successes that they can point to here.

Alexa alone is an example of the massive amount of money they can pour into something and get highly unimpressive results.

although most of the market share is looking for commodity

The market is looking for quality. No one looks for commodity. They only look for quality and (if available) low price.

Is it because of commodity and ease of use

Again, per my extensive post, it isn't clear yet if the useful genai apps will end up commoditized.

And "ease of use" with AWS is historically...not great. Better than Azure, but that's not saying much. They generally aren't the UX people turn to voluntarily.

I don’t think there’s a lot of room for competition in this space if you’re just focused on dev tools

Again, AWS has historically had very low success in providing ML/AI products which are at bleeding edge/SOTA. (And their history with devtools isn't much more inspiring, either.) It is still TBD whether bleeding edge will be necessary to derive value, but if it is, the historically based bet would be to bet against AWS.

 

[erizza]

@tippie36 Obviously big tech companies have the data, however I think (especially with AGI around the corner), a lot of the focus in future startups is data retrieval.

Startups are lucky in that Big Tech companies are locked into the data decisions they made 20 years, startups can build from-the-ground-up services that perfectly tailor data and data retrieval for AI models, and legacy companies will always be behind in that respect.

 

[tippie36]

@erizza Yeah I see your point and agree with the data collection portion, however there are specific niches where big enterprise is gonna be hard to catch up in terms of data retrieval and collection. For dev tools for example, who is better in terms of data than all these cloud providers?

They literally have a huge advantage in that space - and that’s mainly where I don’t think startups will be able to compete anymore, at least not as much as before.

And my overall though it that with Gen AI you can move way faster than before if you have the right data, and they probably have the best data collection and retrieval mechanisms for dev tools - that’s the backbone of their entire business.

 

[emma82]

@tippie36 Good read. We need a social network focused on privacy. The capabilities of AI are scary and can possibly fingerprint us even on an anonymous app like Reddit. I worked at flexport (YC’a golden boy) and nope startups can’t match the speed AWS is moving lol.

 

[tippie36]

@emma82 Yeah privacy is a very important point. I know we can make the argument on speed on anything that’s happened before, but this is not like everything that’s happened before, because this is dependent more on data than anything else.

While before you could probably compete in the dev tools space and get bought, now the chances are way slimmer if your focus is on generative AI yet we see all these bets placed on feature-like companies.

It’s almost as if two years ago YC started investing in a new DynamoDB or a new S3… I honestly just don’t get it and I’m super confused on how long term thinking is not being considered outside of maybe the founder profiles.

Again, to your point privacy and security are very important aspects, and a good example of that is how many freaking time they’ve used these two words at ReInvent. Everyone knows it, these companies don’t want to be targeted as being irresponsible, and start a piss match around that like OpenAI has.

 

[emma82]

@tippie36 I think YC is placing bets on founders more than the company itself these days. I mean I do get it, they literally funded GPT4 wrappers last year but we can also argue about not reinventing the wheel. It’s like you mentioned S3 or DynamoDB, startups historically used AWS infra to build their companies (they still do), just like how they’re using OpenAI’s infra to base their companies on. The AI space is huge and there’s room for more than one competitor but being honest I don’t see AWS or Google succeeding just yet! OpenAI is just too far ahead (I think)

 

[tippie36]

@emma82 Yeah completely agree on that, OpenAI does seem ahead, and you are also right about the history and how much room there is.

My concern is literally the feature-like companies. Most of them are features of other cloud services already existent, and while this might have been worth a go 2 years ago, if you do it now with Gen AI at its forefront you’ll be eaten alive very very soon. It’s not even worth it for them to buy you out, they have [insert a number here] times the amount and quality of data available that is already relevant for what you scrambled to do.

Not saying it’s impossible I’m saying it is highly unlikely to turn something like that in a company, let alone a unicorn.

 

[emma82]

@tippie36 You’re completely right, look up Mintlify and how they got into YC. They were students at Cornell when they created a web app that uses OpenAI’s API to explain any code input to their text box. Literally that’s how they got into YC. I was super surprised but they did pivot into providing documentation to other companies. They knew they’re not gonna survive lmao but hey at least I think YC’s bets on the founders are paying off? Idk

But I’m glad we had this conversation

 

[tippie36]

@emma82 Likewise, likewise!

 

[lauriesinglemom]

@tippie36 Very good post! Lots to consider