Sad state of Amazon Bedrock and AWS AI
Here’s a summary for every enterprise customer who is leveraging the cloud for generative AI. As of Q4 2023, AWS seems to be hitting a bit of a rough patch, especially when it comes to their genAI offerings. Here's a breakdown: AWS's Bedrock is kind of trying to juggle a lot here, leaning on third-party models like Claude 2 and their own Titan model for text and image generation. But let's be real – it's not exactly knocking anyone's socks off. Users are dealing with slow responses, a buggy experience, and missing some key features they really want. Claude 2 has potential, but it's like it's being held back, especially when you look at the broader context. It's struggling with longer conversations, heavily lobotomized, which is a big deal when GPT-4.5 and Gemini Ultra are just around the corner, ready to show what they've got. And the Bedrock agents? They're meant to be this cool, complex feature, but they're just not hitting the mark. Also, it is not exactly a secret that a lot of AWS customers, especially the big players, are playing the field with Azure and GCP. They're not shy about sharing their frustrations with Bedrock and Titan during calls. And when you've got around 80% of your customers testing the waters with your competitors, that's something to sit up and take notice of. Plus, there's this whole thing with Amazon Q not really landing well with the crowd. So, the big question is, why would customers stick around with AWS when they can get better, more mature services with Google and Microsoft? Especially when Google's even throwing in cheaper hosting. AWS needs that killer feature or service that's going to make everyone do a double-take and think, "Okay, AWS is back in the game." Hope this gives a clearer, more casual perspective on where AWS stands and what could be the next steps! TC: 400
Because bigger players lock in contracts and pricing power matters when projects are in exploratory phase. It isn’t always the best product that wins. It is the combination of best product + network + pricing .. and a bunch of other factors. The fact that Bedrock customers (and I don’t use it) crib about it and still stick around tells you a lot about their lock-in into AWS ecosystem. I’d argue GCP & Azure would need the killer feature to steal AWS customers, not the other way round
You’re right that they’ve got security with a lot of enterprise players, but the original beauty of AWS wasn’t its appeal to enterprise players it was its appeal to cutting edge start ups who were doing new things and fast.
Generative AI doesn't need all the data out there – just a specific, processed bit of it. Companies are pretty open to moving this kind of data to different cloud services if they need to. Plus, lots of big organizations are now into the multi-cloud thing, letting their teams pick and mix different cloud options. So, the whole cloud game is changing because of generative AI. It's not like the old days where one size fits all – things are way more flexible now.
I interviewed with AWS to own the marketing for Amazon Bedrock a couple months back. I bungled the interview and didn't get an offer, but I do find it interesting to see Amazon's whole, "let's build as many tools as we can and let our partners/customers figure out how to use them," strategy backfire now that integrating Generative AI capabilities into existing tooling is all the craze.
You should be thankful for not getting the offer 😀
Aws has had a good run. They will lose market share in a larger market in due time, but I don't see them getting any smaller. Because the switching cost is high for most enterprises. And most enterprises are multi-cloud. They will keep AWS for their existing applications. They will most likely use AWS less and the competitors like Azure and GCP more for new applications in due time. But AWS will still be a major player purely because of the lead that they have had for a long time.
So just inertia?
Being third to the market requires you to play catch up for sometime. But this space is still nascent and a lot of problems like LLM hallucinations and inaccuracies need to be resolved before clients undertake a large scale adoption. So don’t feel sorry for being late to the market. Focus on where your clients need problem resolutions today and keep working towards a better LLM. We’ll get there.
Can you add some more color on Bedrock slowness? I was planning to use it for a side project, primarily because I don’t like the potential use of my input being used as training data in ChatGPT.
The system's got major issues with speed and stability – it's pretty hit or miss, with random slow-downs and limits on how much you can do at once. If you're a big player, they might fix things up for you, but usually, it's a shared setup. Don't expect much in terms of keeping things separate or being able to really see what's going on under the hood. And the whole secure setup they talk about? It's not all it's cracked up to be. So yeah, it's a bit of a mess right now.
That’s disappointing to hear. Conceptually bedrocks the one AI play AWS has made so far that’s compelling to me. My use case is pretty small and will likely use langchain so swapping providers later on should be easy.