Laya and Jev

Recently Typesafe AI made claim of a new kind of LLM and released it as Jev - which is termed as "Non Autoregressive" System and as being first to do so. I still do not understand the full math behind that - but the claim, but an Indian from a small town in Kerala - Nandakishor M (https://x.com/Nandakishorm1) made a counter claim that this is not a breakthough at all with references to his paper published with open weight models for an LLM called laya in 2025 (c.f. https://laya.convaiinnovations.com/).

I haven't gone though the claims or the paper myself, but it is interesting to know at least that there are different LLM architectures being evaluated that could potentially result in less compute and memory usage. 

Programming the LLM

I have been dabbling over past 6 months with LLM for different applications. All of them in some way relate to unstructured documents - largely text, and something fundamental seems to come to me: LLM is like an operating system, and we are essentially programming it by writing prompts - with one big cavate - while the traditional programming is all about precision, correctness and predictability - the LLM is kind of fuzzy in all of these characterises - and in that way it is more like how humans actually think.

Next level of automation

Couple of years ago, I had plenty of ideas which I wanted to execute. But the shear amount of work required to write code made it near to impossible to test any of these ideas. FastForward 2025, and you can literally develop an app by just writing a proper spec in proper English and a bit of targeted prompting. This blog is all about my experiments with generative "AI" tech and tools.