Mithilesh

Research Assistant (#PhD, #InterpretableAI #Fairness) @stonybrooku || #entrepreneur by design, prev. worked @sap, Co-Founded #RhodiumAI

[Mithilesh @MitsVision]

A place where you get to work on cool stuffs and get chances to follow your passion.
Thank you @saplabsindia @HarishB07 @Maheshnayakh for the wonderful #innovation initiatives.

[harish-b @HarishB07]

‘There is an #innovator in all of us, but it takes #perseverance and #passion to create an impact in the real world.’ Congratulations @savanmorya on winning the Quarterly Innovation Award 2019 with @Maheshnayakh @saplabsindia @hoysal_v

[Mithilesh @MitsVision]

@sama Autonomy, maybe? First use after the upgrade genuinely shocked me a bit. Felt like one of those GPT-2-ish “something just crossed a line” moments.

[Anthropic @AnthropicAI]

Our interpretability team recently released research that traced the thoughts of a large language model.

Now we’re open-sourcing the method. Researchers can generate “attribution graphs” like those in our study, and explore them interactively.

[Anthropic @AnthropicAI]

New Anthropic research: Tracing the thoughts of a large language model.

We built a "microscope" to inspect what happens inside AI models and use it to understand Claude’s (often complex and surprising) internal mechanisms.

[Mithilesh @MitsVision]

@_akhaliq It’s not about laws of physics.

[Mithilesh @MitsVision]

@sama Deepseek isn’t about competition, it’s a leap forward, reminding us of the power of constraints to drive evolution.

[Mithilesh @MitsVision]

@alokgoyal1971 @SAP @WalkMeInc @whatfix @KhadimBatti @varakumar_n @tejeshwi_sharma Similar story, we pitched a perfect fit solution with extensive use cases working across SAP umbrellas back in 2019/20. Again, they thought they will do this by themselves.

[Ruben Hassid @rubenhassid]

I wanted to know if gpt-4o can create a graph with the same insights:

And apparently not.

It gave me false data & makes up words.

It might only work for simple tasks.

test #4: write an article

[Zain Kahn @heykahn]

Wild. GPT-4o on ChatGPT can generate full blown charts and statistical analysis from spreadsheets with a single prompt in less than 30 seconds.

This stuff use to take ages in Excel.

[Adam Jermyn @AdamSJermyn]

Some small updates from the Anthropic Interpretability team:

https://t.co/VyZiY6D9nQ

[Mithilesh @MitsVision]

This is infinite and of extreme importance.

By the greats of the field.

Deep Learning for AI https://t.co/f94llkweKi

#interpretable #AI

[Mithilesh @MitsVision]

What happens if premium LLM provides like #chatGPT #Claude #Mistral can’t support your growth? How does this equation works? @perplexity_ai @AravSrinivas

[Mithilesh @MitsVision]

Wow, this is insane!!

First: reverting to pre #RLHF behavior #ChatGPT4 #ClaudeSonnet
Link: https://t.co/awjvhgvpKm

Second: Extracting the training data from models like #ChatGPT #LLaMA
Link: https://t.co/VXDt5CSlkP

[Mithilesh @MitsVision]

@paulg Inner workings of AI models >> interpretable ai.

[Mithilesh @MitsVision]

@lexfridman @sama What does he think about the paper “Are Emergent Abilities of Large Language Models a Mirage”?

[Mithilesh @MitsVision]

@soumithchintala Best heading of the day, probably for few weeks. @AnthropicAI ✊

[Mithilesh @MitsVision]

@AdamSJermyn Most of the times I get more direction from the transfer circuits than any other places. Great work ✊

[Mithilesh @MitsVision]

Exceptional and responsible ✊

[Anthropic @AnthropicAI]

Today, we're announcing Claude 3, our next generation of AI models.

The three state-of-the-art models—Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku—set new industry benchmarks across reasoning, math, coding, multilingual understanding, and vision.

[Mithilesh @MitsVision]

@aryaman2020 @jurafsky @ChrisGPotts Fascinating research! I've often noticed causality emerging seemingly out of nowhere in neuron interactions. Actively exploring methods to minimize these instances in my work. Excited to delve deeper into the insights from your paper!

[Mithilesh @MitsVision]

@randall_balestr Very interesting work. It’s definitely gonna help in trying to understand the role of a single neuron or to avoid superposition of a neuron.

[Mithilesh @MitsVision]

Imagine your smartwatch running out of juice, but fear not! A brilliant idea: a fallback mode that transforms it into a classic timepiece, powered by a regular battery. Smartwatch downtime? Not anymore! ⌚️🔄 #SmartwatchFallback #Innovation @Apple @Google