> Cerebras is built for speed in running AI applications, which matches India’s focus on deploying AI across healthcare, agriculture, and public services
Oh no not this again. Using domain specific models for critical things like healthcare is probably the worst thing you can do. There's this stupid notion that you can just lean into AI slightly without committing and you can pay 10% and get 10% of capabilities - like just for healthcare, just for agriculture. That's not how it works.
I hope they don't cheap out and force industries to use some cute domain model on this Cerebras thing - this is the last thing India needs. India should partner with proper AI companies instead of half-assing here. If I could do something about it, I would recommend go all in on proper data centers and encourage hosting companies (that may host open models like Deepseek) as well as OpenAI/Anthropic to get their models here.
I also did rough maths on the throughput these chips support - 64 cerebras chips support around 500 RPS which is pretty low and insignificant IMO.
Cerebras powers the fastest publicly available models, so whether or not the chips are slow, the end result is quite fast. Groq provides the other fastest models.
> There's this stupid notion that you can just lean into AI slightly without committing
I don't see anything in this article that suggests how much they're going to commit to it. Domain specific models work great at specific things, that's the whole point of fine tunes.
In any case this is completely unavoidable. AI is now munitions, in addition to an economic driver/tool. Depending on another country for your defense, economic development, governmental data/compute, etc is a non-starter for any developed nation. They will all eventually have sovereign AI, it's a certainty. It will probably take 20 years for most nations to get there due to hardware shortages but the genie is out of the bottle.
>Using domain specific models for critical things like healthcare is probably the worst thing you can do. There's this stupid notion that you can just lean into AI slightly without committing and you can pay 10% and get 10% of capabilities - like just for healthcare, just for agriculture. That's not how it works.
In short: there’s a reason OpenAI doesn’t have health care model, chemistry model, mathematics model etc. If they want a smaller model they nerf all domains together like in GPT nano. Why? It’s because intelligence compounds and if you take away capability of mathematics from a model, you remove capabilities in all dimensions.
I completely agree. India seems to be talking about building models for specific use cases like language translations and voice translations etc. which I think is not the right direction for LLMs.
That being said India’s built from scratch (not just finetuned) sovereign LLM Sarvam is actually the right direction. I’ve played around with the 26b parameter model and it’s pretty solid.
This is just a bid to secure the infra, what they are going to do with it is anyone's guess.
But I reckon boosting state surveillance would be first on their list of priorities before they spend time on difficult things like healthcare and agriculture.
UAE feels significantly ahead of the curve on AI sovereignty, within the constraints of what a small but wealthy country can do. Between TII (state-owned) and its very credible Falcon models, G42 as a commercial / integrator offering, and a lot of time and energy being poured into the local AI university, they’re clearly taking AI Very Seriously in a way that puts other rich countries to shame.
Every country is slowly discovering that "just use the API" gets awkward when the API becomes infrastructure.
> Cerebras is built for speed in running AI applications, which matches India’s focus on deploying AI across healthcare, agriculture, and public services
Oh no not this again. Using domain specific models for critical things like healthcare is probably the worst thing you can do. There's this stupid notion that you can just lean into AI slightly without committing and you can pay 10% and get 10% of capabilities - like just for healthcare, just for agriculture. That's not how it works.
I hope they don't cheap out and force industries to use some cute domain model on this Cerebras thing - this is the last thing India needs. India should partner with proper AI companies instead of half-assing here. If I could do something about it, I would recommend go all in on proper data centers and encourage hosting companies (that may host open models like Deepseek) as well as OpenAI/Anthropic to get their models here.
I also did rough maths on the throughput these chips support - 64 cerebras chips support around 500 RPS which is pretty low and insignificant IMO.
Cerebras powers the fastest publicly available models, so whether or not the chips are slow, the end result is quite fast. Groq provides the other fastest models.
> There's this stupid notion that you can just lean into AI slightly without committing
I don't see anything in this article that suggests how much they're going to commit to it. Domain specific models work great at specific things, that's the whole point of fine tunes.
In any case this is completely unavoidable. AI is now munitions, in addition to an economic driver/tool. Depending on another country for your defense, economic development, governmental data/compute, etc is a non-starter for any developed nation. They will all eventually have sovereign AI, it's a certainty. It will probably take 20 years for most nations to get there due to hardware shortages but the genie is out of the bottle.
>Using domain specific models for critical things like healthcare is probably the worst thing you can do. There's this stupid notion that you can just lean into AI slightly without committing and you can pay 10% and get 10% of capabilities - like just for healthcare, just for agriculture. That's not how it works.
What's your evidence for this claim?
Wrote about it here https://simianwords.bearblog.dev/why-domain-specific-llms-wo...
In short: there’s a reason OpenAI doesn’t have health care model, chemistry model, mathematics model etc. If they want a smaller model they nerf all domains together like in GPT nano. Why? It’s because intelligence compounds and if you take away capability of mathematics from a model, you remove capabilities in all dimensions.
So basically you r saying diversity matters
Interesting though ..
I completely agree. India seems to be talking about building models for specific use cases like language translations and voice translations etc. which I think is not the right direction for LLMs.
That being said India’s built from scratch (not just finetuned) sovereign LLM Sarvam is actually the right direction. I’ve played around with the 26b parameter model and it’s pretty solid.
This is just a bid to secure the infra, what they are going to do with it is anyone's guess.
But I reckon boosting state surveillance would be first on their list of priorities before they spend time on difficult things like healthcare and agriculture.
[dead]
A full dedicated cluster only for government use?
UAE feels significantly ahead of the curve on AI sovereignty, within the constraints of what a small but wealthy country can do. Between TII (state-owned) and its very credible Falcon models, G42 as a commercial / integrator offering, and a lot of time and energy being poured into the local AI university, they’re clearly taking AI Very Seriously in a way that puts other rich countries to shame.
Which rich countries do you think the UAE is putting to shame?
My native UK, for starters