Author | Lightnews

Suraj Karakulath @surajkarak.bsky.social · 3d

I know it’s not wise to use medical/human biology terms in relation to #AI but I still find it fascinating how they somewhat mirror the effect of junk information on human minds.

Suraj Karakulath @surajkarak.bsky.social · 3d

It’s another instance of the “garbage in, garbage out” idea but also something more. They found that that even after fine-tuning, the brain rot persisted, showing lingering effects of the junk data used in pre-training. A cognitive decline of sorts.

1

Suraj Karakulath @surajkarak.bsky.social · 3d

The choice of junk data was from Twitter - posts with "high engagement" and short, "shallow" (M1) and low semantic quality, i.e. "posts full of clickbait language" (M2). Yeah that checks out... 😆

1

Suraj Karakulath @surajkarak.bsky.social · 3d

Interesting paper this. Nice clickbaity title too. The study tested the effect of “junk data“ in the pre-training of an #LLM and found a "brain rot" effect in its capabilities.
llm-brain-rot.github.io

1

Suraj Karakulath @surajkarak.bsky.social · 4d

Strange. I could have sworn Bluesky was BlueSky (CamelCase - like HubSpot, YouTube, WhatsApp). I have always used it that way until I realised it was wrong.

Suraj Karakulath @surajkarak.bsky.social · 5d

I like Rick Beato's channel a lot. He has done some amazing videos analysing what makes certain classic rock songs great. This video, despite having a very clickbaity title, needs to be seen, for those interested in #creativity and the use of #AI in #art.

I Fried ChatGPT With ONE Simple Question

In this episode I test the limitations of AI to see how much "expertise" it really has. Vsauce's Video: https://www.youtube.com/watch?v=ObiqJzfyACM&t=956s The Scale Matrix (All 25+ Scales) →…

www.youtube.com

2

Suraj Karakulath @surajkarak.bsky.social · 16d

This is adorable! And what a nice handwriting. You can tell a lot about a person by the way they take notes :)

1

Suraj Karakulath @surajkarak.bsky.social · 18d

1 2

Suraj Karakulath @surajkarak.bsky.social · 18d

Normal human beings when joining a work video call: "Sorry just a sec, my bluetooth earphone is not working"

Me (in completely unnecessary panic): "Sorry just a sec, my eartooth is not working".

Sigh...🤦‍♂️

Suraj Karakulath @surajkarak.bsky.social · 20d

Several things to agree in this article, especially on the degradation of #writing skills and our ability to tough it out in cracking problems. I wonder though if #chatgpt has resulted in #reading and comprehension improving (I mean responses are much longer compared to tweets, generally speaking)

“You have 18 months”

The real deadline isn’t when AI outsmarts us — it’s when we stop using our own minds.

www.theargumentmag.com

2

Suraj Karakulath @surajkarak.bsky.social · 21d

Is it agreeing too quickly to my idea? How much of this response could be influenced by previous questions I may have asked? Does anything seem too perfect? If it’s too good to be true, it should ring alarm bells. 🤨 🤨 🤨

Suraj Karakulath @surajkarak.bsky.social · 21d

E.g. - I want A “but” the with the constraint B or I want X “while also” satisfying conditions Y and Z etc. And then I of course have to carefully review the response with a critical eye:
Does it make sense, given what I know already? Is it confirming my biases in anyway?

1

Suraj Karakulath @surajkarak.bsky.social · 21d

Especially when working through complex problems, being as clear as possible in the prompt makes a huge difference. It doesn’t have to be grammatically perfect, although that helps in other ways. But stating clear goals, providing context, constraints to follow, and caveats go a long way.

1

Suraj Karakulath @surajkarak.bsky.social · 21d

I have noticed that coders tend to have strong feelings about using #GenerativeAI in #coding. For me personally, it's not about the actual code but the habit of writing #prompts. It's been a good way to structure my own thinking - like when I'm rubber ducking a problem.

1

Suraj Karakulath @surajkarak.bsky.social · 22d

"It’s still hard to explain what can be so fascinating about beating your head against the wall for three days, not knowing how to solve something the better way, the beautiful way. But once you find that way, it’s the greatest feeling in the world". #programming

The Beauty of Programming

By Linus Torvalds (creator of Linux)

www.brynmawr.edu

Suraj Karakulath @surajkarak.bsky.social · 23d

This is something I have been trying more often lately. Leaving things incomplete was always hard for me but doing a quick recap + plan for tomorrow has helped me completely switch off and forget everything before sleep.

The Simple Habit That Saves My Evenings | alikhil | software engineering, kubernetes & self-hosting

As a software engineer, I often work on big tasks that require hours of continuous and focused work. However, we have plenty of meetings, colleagues asking us something in Slack, and lunch breaks.…

alikhil.dev

Suraj Karakulath @surajkarak.bsky.social · 24d

I don't think I have any public photos after 2015. I'd like to believe I stopped aging after that 🤣.

(Btw - your hair is very striking, it catches the eye!)

Suraj Karakulath @surajkarak.bsky.social · 25d

Exactly! I had a similar experience. I had spent some time on some personal interests and when I tried getting suggestions/ideas for work-related task it kept bringing up some of those interests.

1

Suraj Karakulath @surajkarak.bsky.social · 25d

There is an option to disable this “Memory” in ChatGPT’s Settings but one has to actively do it. I doubt the average user would be aware of this or remember to do it. And disabling it might mean losing the benefits of personalisation when it is actually useful.

Suraj Karakulath @surajkarak.bsky.social · 25d

This is not always desirable. Sometimes you just want a clean start. I am increasingly having to begin prompts with something like “Forget all other conversations we have had before so as to avoid influencing your answer to this question…” more often, when I need an unbiased response.

1

Suraj Karakulath @surajkarak.bsky.social · 25d

Let’s say you want to brainstorm new ideas for fun sideprojects, but you have previously discussed your interest in travel, or music at length. You may see the suggestions for the sideproject ideas hinting at some mention of travel or music.

1

Suraj Karakulath @surajkarak.bsky.social · 25d

This goes beyond a general “LLM echo chamber” problem and is something that could actually affect day-to-day usage. I have found responses to new questions being subtly influenced by things I had asked it before.

1

Suraj Karakulath @surajkarak.bsky.social · 25d

Having the option to remember previous chats always enabled means that new responses may not be ideal when you want answers that are completely independent, and unbiased by #context from past conversations.

1

Suraj Karakulath @surajkarak.bsky.social · 25d

Does anyone feel like the level of #personalisation in #ChatGPT and other #LLM tools is sometimes too much? These tools can remember previous conversations. The idea is good - to keep responses relevant for you. But this may not alway be desirable.

2 1

Suraj Karakulath @surajkarak.bsky.social · 25d

This article hits so many of the points I have been thinking about. We are seeing a collective offloading of cognitive effort (#creativity, analytical thinking) to #GenAI tools. When everyone chooses the path of least resistance to outdo each other, things are not going to end well.

Friction is necessary for Growth - Jameel Ur Rahman

Why I think comfort leads to stagnation and overcoming friction leads to growth.

jameelur.com