Twitxy Saved

Vinay kumar

Verified account This account is verified. Learn more

@SVinayKumar12

Founder @arya.ai; @lexsilabs (lexsi.ai). Working on the alignment & safety research to build Safe Superintelligence.
London lexsi.ai Joined April 2010
308 Following 450 Followers 277 Posts

Latest posts 5

Vinay kumar Verified account @SVinayKumar12
Right now, there is more value in AI getting cheaper than in AI getting smarter.

Model gains are down to single-digit increments. Agents keep getting extremely capable. What is not moving fast enough is the serving cost. It is falling, but nowhere near the rate needed to make most of this viable.

Capability is not what blocks enterprise ROI today. Cost per run is.

Crash token economics to 1/10000th and you don't get today's workloads cheaper. You get use cases nobody has imagined yet.

Sovereign setups make sense, but frontier-grade model science is not democratised yet and compute/energy availability is the hardest fix in the short run.
Vinay kumar Verified account @SVinayKumar12
Good. Welcome to the beginners journey into MI. But there is lot more to climb on.

There a lot more cheaper and more efficient methods to do the same job.

Agentic exploration makes sense in long term though. The inflection is - cost to serve LLMs should crash way more.
Rikiya Takehi @rikiyatakehi Nov 2
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
View quoted post on X
Vinay kumar Verified account @SVinayKumar12
We would love to pay our annual HubSpot subscription if there is a Vibe coding tool that can deliver the custom CRM!

We are seeing the vibe coding platform becoming "vertically focused". Hopefully, it'll replace the SORs soon.

There is so much more value to be derived from the vibe coding platforms to become vertically strong-skilled rather than being horizontally thin-skilled.

It should happen this year. :)
Vinay kumar Verified account @SVinayKumar12
We are expanding our team in Bangalore. If this aligns with any of you, let's chat:

Product Marketing Manager (Kickass PMM who can be a product advocate and create product materials to help our GTM and DevRels) - app.screenloop.com/careers/aryaxa…

DevRel (Going long on OSS- help us build Dev community): app.screenloop.com/careers/aryaxa…

AI Research Scientist (Planning to go big bang on frontier R&D? Check out our research - RL, Interpretability, MI, AI Safety, and Alignment) - app.screenloop.com/careers/aryaxa…

AI Research Scientist (TFMs) (Do you want to contribute to one of the leading tabular foundational models globally? We are changing data science as we know it.) - app.screenloop.com/careers/aryaxa…

And one last role, but the most important:
AI Research Interns: We continue to hire interns across areas.
app.screenloop.com/careers/aryaxa…

#LexsiLabsBangalore #LexsiLabsResearch #RL #FrontierAILab #FoundationalAIResearch #MechInterp #AISafety #ReinforcementLearning #AgentRL