Daksh Varshneya

Struggling to turn LLMs into practical, reliable #AI solutions? Our experts @alanmnichol, @hugobowne, & @dakshv22 will show you how to overcome challenges like hallucinations, high costs, & latency for cost-effective & fast AI assistants. 👉 Register now: hubs.ly/Q02NJh2f0

llama 3 70b beamed to my phone from my M1 Max ~7.6 tok/s with mlx. your own little gpt-4 at home

Practical lessons from building an enterprise LLM-based assistant: 1/ There's a common misconception that you can just finetune an LLM on your company's data. However: - Finetuning is better suited to teaching specialized tasks, than to injecting new knowledge - Finetuning is slow and expensive so finetuned models won't have the freshest information - Finetuning doesn't respect data permissions. Not everyone has access to the same document, but you also can't finetune one model for each user

Overheard: “People who know nothing about machine learning are now paradoxically advantaged in LLMs because they don’t immediately reach for overly sophisticated ideas and spend a lot more time hacking prompts” When hacking prompts feels below your dignity but it works :’|

@peakbengaluru spot the single room

@fishnets88 @Aflah02101 @mervenoyann Absolutely! Check out the DIET implementation 😉

@Aflah02101 @mervenoyann Interesting! Can you also feed these tensors as inputs for Keras layers?

interesting benchmarks from sklearn NN vs Keras: convergence is faster in Keras inference time and consumed memory is way lower in sklearn article by @fishnets88 calmcode.io/til/scikit-neu…

@tylerjdunn 🤦‍♂️😂

what if that satoshi whitepaper was written by a generative model and propagated by the recommender systems

#musingsinbenares now available on #universalbooksellers lucknow. The ghats of varanasi....and much more

Bring in your questions and we'll make an attempt at answering them together! 🤝💪

You know the craze was real when an Irish bartender in Netherlands consoles you for yesterday's loss! #INDvPAK

@tylerjdunn Why do I think this is GPT-3 talking?

it’s hard to exist, let alone function, if you don’t do at least that

my opinion about pretty much everything: it’s hard for me to know because of my simple brain

Young folks in India are happy working from their native home & saving cost of rent and getting home food. This is a fantastic short term benefit but the cost of this is very high. Learning being around smarter folks & absorption by osmosis and overhearing has reduced to zero.

Labeling new data can be a challenge when there is *lots* of data to review. In this blogpost, @fishnets88 explains how you can improve your labeling strategy by using our new UnexpecTEDIntentPolicy which was released in Rasa Open Source 2.8. hubs.ly/H0Tt1pL0

@vibhavagarwal5 Congratulations! Interesting topic, curious to read more about it 😀

6 months of grind is finally over now. I've successfully defended my Master's Thesis on Code-Mixed Dialog Generation y'all. 🎉 Inch away from graduation.🤞🏻

What if the last 10 years was a turbulent bull market all along the way and now's when the bubble actually bursts. 💩 Jk #HODL !

Pls help contribute here: app.cred.club/k63y/d68f93f4

@paper_nl i would like to read a recent paper on evaluation of sentence embeddings.

I don't think this data paints the complete picture. In a market where supply is limited, absolute number of vaccinations also need to be taken into account. Some approx numbers - Israel - 6 mil UK - 24 mil Germany - 7 mil India - 26 mil