Vin Bhaskara

With the new Roaring Kitty, it's no GameStop but GameStart. #luna #skye

discipline >>>>>>>> inspiration or motivation

consistency >>>> bursts of productivity

a palet breton a day keeps the therapist away...

An youtube comment reading: "Art is how we decorate space. Music is how we decorate time." with 51 likes

Oh I love this

Distrobox is so much cleaner and lightweight, I love it! Finally time to move away from the Docker memory hog

A place of worship to reflect on what matters. Frederick Banting lives on through his gift of Insulin to the world that he sold to UofT for $1. Some live once, and some others live forever through their actions... #BantingAndBest #ThisIsCANADA #Never51

GitHub - vinbhaskara/ClinTrialsGPT: ClinTrialsGPT: ChatGPT grounded in the latest US Clinical Trials data ClinTrialsGPT: ChatGPT grounded in the latest US Clinical Trials data - vinbhaskara/ClinTrialsGPT

GitHub: github.com/vinbhaskara/...

Disclaimer: This tool is for informational purposes only and does not provide medical advice. Use at your own risk. Always consult a healthcare professional.

No setup needed. Just install Docker and run:

`docker run -p 8501:8501 vinbhaskara/clintrialsgpt:latest`

(🧵 3/3)

Here's a simple app that lets you ask these kinds of questions — and it automatically infers relevant search terms, queries the ClinicalTrials.gov API, and responds with grounded answers and citations.

👇 (🧵 2/3)

ChatGPT grounded in Real-TIme US Clinical Trials Data:

Want to find out which clinical trials are ongoing for a drug you take?

Curious about the sample size, inclusion criteria, or outcomes for a condition affecting someone you know?

👇 (🧵 1/3)

I don't always post on Twitter. But when I do...

my experience with XFCE has been way better than LXQt haha! as simple as choosing an audio output device on LXQt default sound management was painful...

Xubuntu >> Lubuntu factoring in both performance and basic usability on a low-resource machine. My desperate attempt to bring a old Chromebook 4 to life... finally worked!

"if it's not calibrated it's not probability but just a number between 0 and 1"

The modern ‘aha!’ moment for Masked Language Models (MLMs) trained on large-scale data is few-shot learning—the ability to generalize well with very few task-specific labeled examples. This is analogous to how generalist conversational AI became the defining breakthrough for LLMs.

AOC and Bernie in front of the rally crowd in Denver

Zoomed out crowd shot of 34,000 people in downtown Denver with Colorado mountains in the background

Tonight an unbelievable **34,000 people** gathered for our Denver rally to take on billionaires and win our country back.

This was the largest political gathering in Denver since Obama in 2008.

Also bigger than the 2024 DNC.

And the largest ever rally in Bernie’s career (and obviously, mine too).

# python wizardry:
> a = [[]] * 5
> print(a)
[[], [], [], [], []]
> a[0].append(1)
> print(a)
[[1], [1], [1], [1], [1]]

# instead use list comprehension for expected result:
> a = [[] for i in range(5)]
> a[0].append(1)
> print(a)
[[1], [], [], [], []]

tip: how to use heaps with objects in python? pass them as tuples with the first index being the heap value and the second being the object! E.g. min heap usage:

> import heapq
> obj_list = [(3, obj1), (1, obj2), ..., (<val>, <object>)]
> heapq.heapify(obj_list)
> heapq.heappop(obj_list)

(1, obj2)

Not every task is calling an OpenAI API or prompt-engineering the life out of an LLM. Especially when it's not a generative task. Sure, you can do 1+1=2 with a LLM, but also do you not love the Arctic and the polar bears? #WrapperEngineer

Tokenization can be a target for an adversarial attack -
bsky.app/profile/zoes...

Tokenization and Decoding (or, Sampling) are two of the most underrated aspects of an LLM. Follow this live thread as I link relevant works that focus on these as I come across... 🧵 (1/♾️)

@lorne-schell.bsky.social - you might find this interesting! Tokenization from an adversarial perspective

What happens if we tokenize cat as [ca, t] rather than [cat]?

LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.

#AI #LLMs #tokenization #alignment

Noticing more likes on my recent posts. My takeaway is the #computervision community is growing here.

When AI Thinks It Will Lose, It Sometimes Cheats When sensing defeat in a match against a skilled chess bot, advanced models sometimes hack their opponent, a study found.

Early signs of deception, cheating & self-preservation in top-performing models in terms of reasoning are extremely worrisome. We don't know how to guarantee AI won't have undesired behavior to reach goals & this must be addressed before deploying powerful autonomous agents.
time.com/7259395/ai-c...

"Can I have a latte avoine?" "Hey man, ça va?" #Overheard #BonjourHi #J'<3MTL #OnlyInMTL

Bonus: Correlation doesn't mean causation. A specific region with heightened activation on certain stimuli (like faces) doesn't mean that lesioning a part of it will affect its processing. How to establish causation (without risky invasive procedures)? An open question. (11/10)