Aarya Banthia

building private, reliable ai for better business decisions. managing partner, karya. decision engineering × local ai. enabling humanity.

[Aarya Banthia @AaryaBanthia]

@ZeroOneCreative @henrijsdan love the visuals and concept and the way you've brought it to life

[Aarya Banthia @AaryaBanthia]

How much data is Instinct harvesting?
Pre-revenue + the valuations = data is all I can infer

we need confidential intelligence :)

[Aarya Banthia @AaryaBanthia]

@TslShahir @googlegemma @GoogleDeepMind this is superb. all these recent examples show the power of finetuning. embedding gemma, fine tuned has showcased so many unique and stellar results. all running locally

[Aarya Banthia @AaryaBanthia]

Forward Deployed Engineers won't exist in a few years.

Context gathering agents embedded within each enterprise will gather and interact on behalf of FDEs and allow engineers to work with multiple enterprises parallelly.

[Aarya Banthia @AaryaBanthia]

@allietheicon We've extensively tested with Jev and this is an interesting use case. allows you to parse through huge documents: cut down context. could help with compression algorithms within harnesses. you're looking at the entire prompt and seeing what is relevant as you're moving forward.

[Aarya Banthia @AaryaBanthia]

@tomgreenwald The future is going to be a lot of confidential intelligence and running models locally. Just as a question when do you think there will come a time when we can run models locally on CPUs or consumer-grade hardware, not necessarily M4 Maxes but what hardware the vast majority use

[Aarya Banthia @AaryaBanthia]

@omarsar0 The future is all about confidential intelligence and this is a great step towards that. Having tool calls and being able to partition and have things running locally versus on ZDR-compliant cloud inference: that is the optimized way

[Aarya Banthia @AaryaBanthia]

Agreed!
Looking over Anthropic's IPO filings, and seeing 85% gross margins. I think the time is near when we don't have 5-hourly/ daily/ weekly subscriptions.
we have monthly quotas (splendid amounts) depending on the subscription.

Anthropic wants to push for usage based pricing. But as consumers, people want subscriptions and fixed costs.

[Aarya Banthia @AaryaBanthia]

@0xfJuan https://t.co/2uOeivaoBs

[Aarya Banthia @AaryaBanthia]

@0xfJuan launch

[Aarya Banthia @AaryaBanthia]

We've moved on from graph and loop harnesses to terrain harnesses.
Much more capability and fluidity while added determinism.
white paper, will release later this week

[Aarya Banthia @AaryaBanthia]

@hazhubble @pally This is absolutely fantastic Haz. This has been a huge pain point for us. And today, warm intros work more than anything else.
Intrigued… will try it out… with your network first :)

[Aarya Banthia @AaryaBanthia]

@liquidai This is an interesting use case. Let’s chat

[Aarya Banthia @AaryaBanthia]

@reallygabriella @goodhartproof Not best day x 365 💀💀

[Aarya Banthia @AaryaBanthia]

We don't need prompts anymore. We need references and detailed explanations and instructions. By the end of 2027, we won't have to hand-hold AI as much. It will have access to every resource and piece of data needed. tool calling abilities (at the harness and reasoning level)

[Aarya Banthia @AaryaBanthia]

Interesting phenomenon (not surprising). The people used to spoon-feeding and hand-holding others are seeing AI treble productivity.
They understand how to breakdown tasks into its fundamental pieces.

[Aarya Banthia @AaryaBanthia]

Most people give vague instructions and abstract conversations, expect the AI to fill in the pieces on its own, and expect frontier responses. Yes, the better a model can reason, allows it to fill in more of these gaps and use tools and resources to better implement the goal.

[Aarya Banthia @AaryaBanthia]

@joanrod_ai @QuiverAI Replying for early access

[Aarya Banthia @AaryaBanthia]

@daymakerguy @MonacoGTM @NowadaysAI @Rippling @salesforce @marketercom @posthog @retellai Aarya Banthia

[Aarya Banthia @AaryaBanthia]

Just got off the call with @subquadratic. We had the longest prompt recorded for SubQ snap.

Interesting times…