Karthikeyan Saravanan

Building @KuralitHQ - making every app conversational. 5 yrs production AI. Prev: built @Kuralynx. Shipping in public now.

[Karthikeyan Saravanan @itskeyans]

@sugarjammi Orca is what I settled with

[Naval @naval]

When you look back, you’ll wish you’d just done the long-term version of everything.

[Karthikeyan Saravanan @itskeyans]

Much expected this, waiting for Fable beating Opus soon!

[Claude @claudeai]

Sonnet 5.5 improves on Sonnet 5 across benchmarks, in some cases dramatically.

It’s a faster, lower-cost complement to Claude Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.

[Karthikeyan Saravanan @itskeyans]

@claudeai @ChatGPT ps: here comes the @ChatGPT

[Karthikeyan Saravanan @itskeyans]

@claudeai has improved a ton, getting back after few days. I asked questions to answer, it divided, trying to actually get with varieties of concepts for 3 minutes and trying to come up visually.

On the other hand, I tried same with @ChatGPT, it answered plainly in 10seconds.

btw, I really like LLMs takes his own time to come back with answers after actually knowing answer, it's like someone taking effort to learn something and come back with real answers.

[Karthikeyan Saravanan @itskeyans]

Can't agree more!! OMP is being my daily driver these days, from my view, multiple ways AI slop get created, one is through lack of human review. OMP is trying to reduce the second part with its great pack of fat harness & multi agents.

[Ahmad @TheAhmadOsman]

My 3 favorite harnesses now are DeepSeek Harness, OMP, and ODS Harness (in Beta)

We’ve come a long way from the slop that is Claude Code thank God

[Karthikeyan Saravanan @itskeyans]

Relative preference is not always absolute acceptability. And this changes how we interept Jev's zero hallucination claim.

0% hallucination number is not empirical, it comes from the fact that schema matching is guaranteed. So Choices and State we provide to the JEV should be in align always to expect absolute acceptability.

[Kumar @barveen_kumar_]

I keep thinking about this with Jev and i'm no expert or tech guy.

What if I give it 5 options and all 5 are not relevant?

Will it pick one with 60% confidence and technically never “hallucinate”?

But it's still wrong lol.

For agents and UI decisions, I think “none of these are good enough” is actually a pretty important answer.

Am I right here @itskeyans?

[Karthikeyan Saravanan @itskeyans]

@barveen_kumar_ JEV only works based on given options, this is where it fails for UI decisions if options are worse. here's a quick exp 🤔, jev designed to chose the least bad answer, not a good answer.

Technically speaking, this comes directly from the relative choice architecture in RLCD, utilities become meaningful through comparison within a candidate set.

[Karthikeyan Saravanan @itskeyans]

btw if anyone wonders we've paper already released 2 years back on RLCD

You are not missing a secret Jev paper. You are looking at two different methods that happen to share the letters RLCD.

@typesafeai JEV is Reinforcement Learning for Calibrated Decisions (*RLCD). This @arxiv paper is Reinforcement Learning from Contrastive Distillation (*it's LLM 🙂).

[Karthikeyan Saravanan @itskeyans]

My whole feed is full of JEV today, really @typesafeai has not exposed any information on what is RLCD, but we got lot of open contributions until JEV opens up about the architecture.

If anyone wants to know reality of building the JEV lookalike models, Sharing the resources I found it useful here:
https://t.co/rDRCvJGL9F
https://t.co/kCBTnCpaJj

[Karthikeyan Saravanan @itskeyans]

Admiring @JeffBezos on his writing so cleanly

four element to high standards: they are teachable, they are domain specific, you must recognize them, and you must explicitly coach realistic scope.

[Hemant Mohapatra @MohapatraHemant]

In 2 pages @JeffBezos teaches you more about high standards than you'll learn from reading 2 books. The art of summarizing that which resists summary is critical to achieving greatness. The world will forget you, except what you've written down, then what you've written down well.

[Karthikeyan Saravanan @itskeyans]

@thsottiaux 5 hour limit 😅

[Karthikeyan Saravanan @itskeyans]

If you think Terra medium is the best tradeoff between intelligence and cost. You're wrong!

Luna max is best in intelligence scores and cost. Impressive! catching up this lately

[Karthikeyan Saravanan @itskeyans]

@HarveenChadha herdr is good

[Karthikeyan Saravanan @itskeyans]

@claudeai is so silent recent days, but they come up with just counter update model, sadly no pro plan preview release 😟

[Claude @claudeai]

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.

They're the world’s most advanced models for coding and knowledge work.

[Philip Kiely @philipkiely]

Teach your agents inference.

The full text of Inference Engineering is now available to LLMs everywhere. No more parsing PDFs.

3,000 tokens | links to each chapter: https://t.co/ggaII5NwTr

70,000 tokens | full text: https://t.co/1wa2w1cCnf

[Karthikeyan Saravanan @itskeyans]

if anyone wants to build presentations with codex/claude code, here's the gem i found: https://t.co/jRis17k5cp

@zarazhangrui good curation!!

[Karthikeyan Saravanan @itskeyans]

3 papers we’re discussing at Distill Zero this Saturday in Chennai:

• Self-Harness - agents improving their own harness
• Refuses the Shape, Serves the Substance - why many AI failures aren’t attacks
• Ask, Don’t Judge - a different way to evaluate LLM outputs

If you’re into AI/ML research and like going beyond the abstract, come join the room.

https://t.co/xBkGFNQWQD

[Karthikeyan Saravanan @itskeyans]

Chennai AI/ML folks - this one’s for you 👀

Distill Zero #02 is happening this Saturday.

A small room of researchers, engineers, builders & students getting together to actually read and discuss AI/ML papers - not just skim the abstract and move on.

We’ve locked the presenters for this edition.

If you’re into:
• LLMs / NLP
• Agents
• ML research
• understanding papers beyond the hype come join us.

📍 Chennai
📅 Aug 29

https://t.co/xBkGFNQWQD

#Chennai #MachineLearning #AIResearch