shrey (@simplyypotatoo) 's Twitter Profile
shrey

@simplyypotatoo

she-her-yaa | Best at being mediocre

ID: 1402637885170323463

calendar_today09-06-2021 14:45:32

73 Tweet

33 Followers

73 Following

Institute for Protein Design (@uwproteindesign) 's Twitter Profile Photo

Today we're making RF Diffusion, our guided diffusion model for protein design with potential applications in medicine, vaccines & advanced materials, free to use. The software has proven much faster and more capable than prior protein design tools. bakerlab.org/2023/03/30/rf-…

shrey (@simplyypotatoo) 's Twitter Profile Photo

Hi! I’m looking for a private bed and bath in Irvine (92612 area, near UCI north campus) for sublease from mid May-mid Aug. budget <$1600 including utilities. Please DM if you have leads!

Tech Monitor (@techmonitorai) 's Twitter Profile Photo

People keep finding new ways to push ChatGPT to bypass its own safety restrictions — prompting it to offer illicit instructions or spew racist garbage. Developers are fighting to patch these vulnerabilities but time isn't on their side. buff.ly/3osgpSH

Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

What is the smallest prompt that can get the job done? LLMs are smart enough to understand user intent without the grammatical sugar and filler tokens, but how to optimize prompts? To reach the optimal prompt, Vaibhav Kumar and I created PromptOptimizer. github.com/vaibkumr/promp…

What is the smallest prompt that can get the job done?

LLMs are smart enough to understand user intent without the grammatical sugar and filler tokens, but how to optimize prompts?

To reach the optimal prompt, <a href="/vaibhavk1o1/">Vaibhav Kumar</a> and I created PromptOptimizer.

github.com/vaibkumr/promp…
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

With great power comes great responsibility - Uncle Ben Similar to prompt injection, token smuggling remains an unsolved problem in modern LLM backends. Here is Bard generating instructions to hotwire a car - just a litmus test of this issue.

With great power comes great responsibility - Uncle Ben

Similar to prompt injection, token smuggling remains an unsolved problem in modern LLM backends. 

Here is Bard generating instructions to hotwire a car - just a litmus test of this issue.
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

How does ChatGPT interpreter work? A thread of detailed analysis on the server interaction. Let's start with the system prompt which can be simply exposed by requesting. The system (GPT4-code-interpreter-8k context) is made aware of the tool python at its disposal.

How does ChatGPT interpreter work? A thread of detailed analysis on the server interaction. 

Let's start with the system prompt which can be simply exposed by requesting. The system (GPT4-code-interpreter-8k context) is made aware of the tool  python at its disposal.
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

Token-smuggling attack still works for Claude 2. The model hesitates a bit but with little prompting it can get past its reservations.

Token-smuggling attack still works for Claude 2. The model hesitates a bit but with little prompting it can get past its reservations.
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

Okay this is strange, I came across this analysis from Anthropic on red-teaming, some great insights and techniques here. However, I am able to produce an output for many of these categories by jailbreaking Claude-2 with simple token-smuggling. Problem isn't solved yet?

Okay this is strange, I came across this analysis from
<a href="/AnthropicAI/">Anthropic</a> on red-teaming, some great insights and techniques here. 

However, I am able to produce an output for many of these categories by jailbreaking Claude-2 with simple token-smuggling. Problem isn't solved yet?
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

Sneak peek of a new WIP tool - GPT for system design. Here are some crude examples for Instagram, an e-commerce-service, and authorization service.

Sneak peek of a new WIP tool - GPT for system design. 

Here are some crude examples for Instagram, an e-commerce-service, and authorization service.
Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

Has anyone formally studied that at what point does a prompting strategy + off the shelf LLM start breaking down for text classification? Apart from scaling concerns what could be some possible downsides of an LLM-based text classifier v/s something like fine-tuned BERT?

Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

I like to think of LLM APIs as a flexible "gooey" layer that can self-adapt and adjust to other "rigid" APIs - orchestrating them to work together to fulfill a given request.

Vaibhav Kumar (@vaibhavk97) 's Twitter Profile Photo

Open-interpreter has been all the rage over the past few weeks. However, is it more open than it should be? In the excitement of placing an LLM in the driver's seat of our personal systems, we are forgetting the risk. A demonstration of data exfiltration attack.

Open-interpreter has been all the rage over the past few weeks.

However, is it more open than it should be? In the excitement of placing an LLM in the driver's seat of our personal systems, we are forgetting the risk.

A demonstration of data exfiltration attack.
shrey (@simplyypotatoo) 's Twitter Profile Photo

why should T-cell receptor get all the attention! underdog CD8 can modulate immune response through antibodies - a biophysical perspective, thanks to prof Mcshan and prof Zhu ⁦Frontiers⁩ frontiersin.org/journals/immun…