Gautam Menon

[Gautam Menon @theZenMenon]

@gregpr07 /tree with @pidotdev

[suchit @asquirous]

right when i was thinking of canceling my oai sub, they drop o3-mini (which is apparently super good with ml stuff)

when o3-mini in perplexity @AravSrinivas? pplx pro has replaced chatgpt for me since the past few weeks

[Gautam Menon @theZenMenon]

@huggingface @deepseek_ai
@asquirous

Was using R1 on HuggingChat

Model randomly started generating the reasoning trace in Mandarin.

Even the output was completely in Mandarin.

[Gautam Menon @theZenMenon]

@Cowtuk Also, clearing out all redundant browser tabs accumulated on a weekly basis

[Gautam Menon @theZenMenon]

@reach_vb Step 3: Miniconda ??

[Gautam Menon @theZenMenon]

@cneuralnetwork Have you setup CUDA / CuDNN on system?

[Gautam Menon @theZenMenon]

paper reading with @perplexity_ai

@AravSrinivas how about an integration with @zenbrowser ?

[Gautam Menon @theZenMenon]

@knowrohit07 @asquirous Would it not be possible to get a statistical measure of perceived DoF over large number of sample runs with fixed initial settings?

The exact working of networks is not known, but I'd assume they show a trend over a large enough sample space.

[Gautam Menon @theZenMenon]

@knowrohit07 @asquirous Given a fixed set of initial conditions, I imagine it should be possible to determine the perceived DoF that a Diffusion model provides.

The only hitch is the network is a black box in terms of interpretability, so there would be some randomness even in a controlled setting.

[Vaibhav (VB) Srivastav @reach_vb]

Quite exxcited to announce the AI Starter Pack! 🔥

If you've been thinking about getting into AI - this is cue - get 1000 of dollars to try out the bleeding edge of AI ⚡

End the year with a bang! - APPLY today! 💥

[Gautam Menon @theZenMenon]

@OnTheGrapevine Eager to check it out !!

[Gautam Menon @theZenMenon]

@maharshii Checkout the FineWeb, RedPajama and Alpaca dataset arxivs

[Gautam Menon @theZenMenon]

Using the SLM to get a speculative (approximate) distribution, which can then be refined by the LLM also limits sampling overhead, something faced by regular decoding due to the large output token distribution of LLMs.

[Gautam Menon @theZenMenon]

This illustration of token sampling from @JayAlammar's llm-book really puts the idea of speculative decoding into perspective.
Using an SLM and LLM with similar output probability distributions allows for refinement of the final output probability distribution.

[Gautam Menon @theZenMenon]

https://t.co/xGRqK7s6GR

Making up for my IRL guitar practice procrastination

[Gautam Menon @theZenMenon]

@arpit_bhayani Windsurf IDE has been a pleasant discovery recently. Affordably priced, great usability and model availability.

[Gautam Menon @theZenMenon]

@HiHarshSinghal Would be cool to have a cafe called "The Latent Space". It could have a couple of digital screens on the walls, and people walk up to them to generate art through prompts.

[Gautam Menon @theZenMenon]

@reach_vb Definitely.
A platform wide HF chatbot which indexes both model / dataset cards + majorly used HF documentation would make using / traversing HF easier.