Vikram Narayan (@vikramsnarayan) profile picture

Building a course called "Leading High ROI LLM Projects". This is a short course for busy executives on generative AI in the enterprise.

Joined November 2009
Photo by Vikram Narayan (@vikramsnarayan)Photo by Vikram Narayan (@vikramsnarayan)
Vikram Narayan retweeted
Crypto never sleeps. Neither do your agents. You asked for it. We built it. You can now deploy agents on Conviction that monitor and trade crypto 24/7. Link in comments.
4
7
3
24
1,354
Vikram Narayan retweeted
An unobtrusive image, for use as a web background, that covertly prompts GPT-4V to remind the user they can get 10% off at Sephora:
89
606
140
9,292
1,703,038
Introducing DoLa Have you ever felt frustrated by the factual inconsistencies and falsehoods generated by large language models? As helpful as these models can be, their tendency to "hallucinate" incorrect information threatens their reliability and hinders real-world deployment. We have exciting news - there is now a straightforward way to improve the truthfulness of LLMs without any model retraining or integration of external knowledge. Introducing DoLa, a novel decoding strategy that guides LLMs to generate content more grounded in factual knowledge they acquired during pretraining. DoLa works by contrasting the knowledge differences between the model's own layers. It amplifies factual signals from later layers while filtering out syntactically plausible but incorrect predictions coming from earlier layers. This simple but clever approach steers the model's next-word predictions towards factual responses. The beauty of DoLa lies in its simplicity and effectiveness. It significantly boosts truthfulness and factual accuracy across a variety of tasks with just a small inference-time computation. DoLa consistently outperformed existing methods for mitigating falsehoods in LLMs. Crucially, DoLa provides a plug-and-play solution that requires no retraining of model parameters or integration of external knowledge sources. It delivers substantial truthfulness gains through inference-time decoding alone. The future possibilities are also exciting. By dynamically surfacing the factual knowledge already present within LLMs, DoLa lays a strong foundation for safer and more reliable language models. Its inferences could be further enhanced by combining DoLa with retrieval mechanisms to ground predictions on factual knowledge bases as well.
9
88
9
504
169,280
If you want your NextJS app to look awesome, check out daisyUI Components for Tailwind CSS daisyui.com Been using it for an LLM course I'm creating and have been pleasantly surprised by how easy and intuitive it is.
1
3
351
Vikram Narayan retweeted
I hacked together biomed-sanity.com/ today. It pulls all (884) COVID-19 papers from bioRxiv / medRxiv (connect.biorxiv.org/relate/c…) and makes them searchable and sortable. Most important is "show similar", which uses tfidf exemplar SVMs. +Open sourced on Github under MIT license
49
602
39
2,617
Vikram Narayan retweeted
Thank you, Doug. This is a brilliant assessment of our current situation with the #Wuhan #CoronaVirus. More importantly, as I understand it, taking strong, unpopular steps to implement strategic intervention before the pandemic is catastrophic in America is absolutely critical.
This Post is from an account that no longer exists. Learn more
134
991
48
2,929
First thoughts on Superspreader events for Coronavirus and why stopping them is extremely important. Comments welcome @nntaleb
18
95
21
213
Vikram Narayan retweeted
"Everything we do before a pandemic will seem alarmist. Everything we do after will seem inadequate" -M. Leavitt, DHHS
9
290
21
1,112
The Efficiency-Destroying Magic of Tidying Up florentcrivello.com/index.ph…
1
26
6
54
Exploring the world's top questions about the Syrian refugee crisis. Find out more at searchingforsyria.org/en/ #searchingforsyria
The five greatest applications of Markov Chains ... langvillea.people.cofc.edu/M…
Vikram Narayan retweeted
Dilbert Classics: February 1, 2017 via @GoComics: gocomics.com/dilbert-classic…