Control your models, control your destiny. I think there are two winners in the OpenAI drama: Microsoft and competitive models as well as open-source AI models. For builders, it shows the importance of choosing your API provider carefully.
Microsoft got an incredible team. Competitors get a moment of opportunity to gain traction while the market leader is in crisis.
Which of the two take this advantage all the way is anyone's guess. But OpenAI has ceded their lead over to these players. And builders have lost their default API option and endpoint.
🐟🐟🐟
The lesson? De-risk your suppliers.
There's a real risk that all of openAI engineers walk over to Microsoft to join Sam Altman and Greg Brockman. There's 700/770 signatures already suggesting they will resign. It's highly commendable the engineers at OpenAI are committed to keeping their APIs running. But in case they all walk and OpenAI is left with just the board, what alternatives do AI developers have?
👉 First, many companies probably have moved their model serving directly to Microsoft’s Azure AI APIs. We did the same thing.
👉 Second, many developers who are spooked about centralizing dependency on just one player, are now likely thinking of orchestrating across multiple models – like Google’s PaLM, Anthropic’s Claude2 or the open source model Llama.
👉 Third, if you can stomach the cost of training your own models, then the open source algorithms like Llama are your best bet.
🐟🐟🐟
Open source isn’t the easy option
One major limitation of working with open source models that doesn't get talked about often, is serving them in real time.
Don’t underestimate the super power of OpenAI or Azure or Google Cloud – they have world class serving infra that can host these large language models that require huge amounts of RAM or custom GPU chips like Nvidia A100s or H100s, which are in acute shortage.
🐟🐟🐟
But open source de-risks builders and gives much more optionality to builders
Luckily there are some great efforts out there to scale-up the serving of open source models.
For example, instead of hosting the entire model across hundreds of gigabytes of RAM, techniques like LoRA (Low Rank Adaptation) allow you to vector quantize your models ie reduce their size at the cost of lowering accuracy.
Further approaches like Fal.ai allow one to serve these models at scale across multiple GPUs without actually having to purchase these GPUs outright.
It all boils down to:
Control your models, control your destiny
- Control your models, control your destiny. I think there are two winners in the OpenAI drama: Microsoft and competitive models as well as open-source AI models. For builders, it shows the importance of choosing your API provider carefully. Microsoft got an incredible team. Competitors get a moment of opportunity to gain traction while the market leader is in crisis. Which of the two take this advantage all the way is anyone's guess. But OpenAI has ceded their lead over to these players. And builders have lost their default API option and endpoint. 🐟🐟🐟 The lesson? De-risk your suppliers. There's a real risk that all of openAI engineers walk over to Microsoft to join Sam Altman and Greg Brockman. There's 700/770 signatures already suggesting they will resign. It's highly commendable the engineers at OpenAI are committed to keeping their APIs running. But in case they all walk and OpenAI is left with just the board, what alternatives do AI developers have? 👉 First, many companies probably have moved their model serving directly to Microsoft’s Azure AI APIs. We did the same thing. 👉 Second, many developers who are spooked about centralizing dependency on just one player, are now likely thinking of orchestrating across multiple models – like Google’s PaLM, Anthropic’s Claude2 or the open source model Llama. 👉 Third, if you can stomach the cost of training your own models, then the open source algorithms like Llama are your best bet. 🐟🐟🐟 Open source isn’t the easy option One major limitation of working with open source models that doesn't get talked about often, is serving them in real time. Don’t underestimate the super power of OpenAI or Azure or Google Cloud – they have world class serving infra that can host these large language models that require huge amounts of RAM or custom GPU chips like Nvidia A100s or H100s, which are in acute shortage. 🐟🐟🐟 But open source de-risks builders and gives much more optionality to builders Luckily there are some great efforts out there to scale-up the serving of open source models. For example, instead of hosting the entire model across hundreds of gigabytes of RAM, techniques like LoRA (Low Rank Adaptation) allow you to vector quantize your models ie reduce their size at the cost of lowering accuracy. Further approaches like Fal.ai allow one to serve these models at scale across multiple GPUs without actually having to purchase these GPUs outright. It all boils down to: Control your models, control your destiny
