Neural channel | Deep Learning | Datascience | AI Memes – Telegram
Neural channel | Deep Learning | Datascience | AI Memes
1.87K subscribers
4.41K photos
472 videos
18 files
1.54K links
Cutting-edge datascience and shiet
Download Telegram
Most of Perplexity’s GPU expenses are because of free users that they classify as “R&D costs” to improve profit margin metrics 😁😁😁
😁7🍌7🤯2
Claude has the same issues as gpt4o but instead of glazing it tries to snitch at you to authorities

Claude dev posted this then deleted:
🤬10🍌3🤮1
I’m building this unironically
💯174🍌2
AI Wars chronicles: Anthropic cancelled Claude access to Windsurf

This is a significant pivot: Anthropic started as an AI API SaaS provider, but, as everyone else, they’re investing more into making AI apps.

After Claude Code is out, Windsurf turned from a client to a competitor

This is a lesson for everyone using big tech AI: at some point they can decide they want to take over your market sector, and you won’t be able to do anything with that

(Of course, the fact that OpenAI acquired Windsurf also plays a role, but Windsurf itself didn’t have the plans to stop using Claude models)


https://windsurf.com/blog/anthropic-models
🤩5🍌2👍1🤔1
Gonna post a banger soon, stay tuned!
👀16🍌2🍾2
ChatGPT agent reasoning about clicking “I’m not a robot”
😁11🫡2🥰1🍌1
Neural channel | Deep Learning | Datascience | AI Memes
Gonna post a banger soon, stay tuned!
I decided to post this shortly after GPT-5 comes out (which comes out in a few hours/days) to take it into consider when writing. Stay tuned
The scariest part about the dead internet is that it’s more alive than some of the humans
💯15
Keep thinking about this
🍌11
GPT5 stream

They proudly show how ChatGPT:
1) Doesn't answer fireworks-related question
2) Decides how to do a cancer therapy

AI will decide for you how to live but fireworks is too much chud
😁17🍌3
A random thought: now when more and more tools are AI-powered, security by obscurity becomes [more] important

It’s harder for AI to hack you if its pretraining data had nothing similar to your system

Do you agree?
👍12🍌1
Thinking of starting a list of (mostly esoteric) public goods for AI dev/research that I personally would like to work on if/when I have more time:

- a toolkit to benchmark models of different quantizations: usually authors only check for perplexity which IMO is not enough
- open-domain datasets of various extremely niche knowledge, ideally requiring an internal world model, for example: Stalker games knowledge; lifehacks and howtos of my local neighborhood; etc. I would've also kept the validation part private just like the ARC dataset
- "questions requiring reasoning about multiple needles in multiple haystacks" datasets - this one is obvious. Would be interesting to see how all those sparse attention approaches handle them hehe

why do I think those are important?

1) I'm simply curious tbh
2) to turn the Goodhart's law into a tool: if everyone is benchmaxxing, the best way to make sure new models are aligned with your needs is to make benchmarks that measure things you want/need, especially if we're entering the era of "universal verifiers"
🔥10