Soumya Deb (@Debloper) header
Soumya Deb (@Debloper) profile picture

Cogito, Creo, Dico, Duco. Maker of @BreakerSpaceHQ

Orion–Cygnus Spur
Joined May 2009
Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)Photo by Soumya Deb (@Debloper)
We do what we must because we can.
5
8
...said Dumbledore calmly, in an Indo-Iranian descendant language.
1
48
I wrote something up a little while back (after a week of pondering whether I should publish it): medium.com/debloper/nvidia-d… ANYWAY...
“I think having Jensen go back… a lot of people are asking Jensen. They think Jensen should go back. That certainly isn’t a bad idea.”
1
82
In simpler (less epistechnically rigorous) terms: Learn to tell apart the rope from the snake, faster than others figuring out the utility of the rope or the danger of the snake. You'll always be limited by how much can you do with that advantage, but that's all secondary.
The most important skill you should pick up today (regardless of how fast the fast-takeoff gets) is to tell apart, in real time: the time evolution of any phase space, by identifying the type of dynamic system it embodies. All else are mere utilities to this one terminal skill.
2
107
The most important skill you should pick up today (regardless of how fast the fast-takeoff gets) is to tell apart, in real time: the time evolution of any phase space, by identifying the type of dynamic system it embodies. All else are mere utilities to this one terminal skill.
1
2
163
Soumya Deb retweeted
definitely check this out you'll enjoy working with Deb and learn a ton!
I really shouldn't be adding more stuff to my TODOs, but this is a time sensitive one; so I need help. If you are: - a student with high agency and energy - into local AI/LLM or compute sovereignty - interested to build micro-datacenters ...then, I've a project for you.
1
1
95
I really shouldn't be adding more stuff to my TODOs, but this is a time sensitive one; so I need help. If you are: - a student with high agency and energy - into local AI/LLM or compute sovereignty - interested to build micro-datacenters ...then, I've a project for you.
1
1
1
3
287
Oh no... all these feticide of the "hallucinating" models, and selective breeding of "grounded" models would be in vain!
can we create programs or smth that we can feed to LLMs, and then their weights temporarily get modified in a way that they start to act drunk?
1
87
Anyone using Open WebUI's Computer (`pip install cptr`)... thoughts?
1
64
Chat checkpoints as traversable/branchable like git commit tree history? Bonus: context compression on branching
Does any AI app have this feature - During a conversation, ask the LLM to branch off into a new session and pass enough context to explore a different direction, while the main conversation still goes on. I know some apps support forking, but they copy the entire session history. I’m looking for something more selective, where the LLM determines what context to carry over to the new branch instead of copying the full conversation, and then automatically starts a new session (probably a tool call?)
1
1
72
You love your horses? Good! But that doesn't have to mean you have to go around setting cars on fire.
45
Hey @postmarketOS can you guys please look up "how to enable dark mode on MediaWiki deployment"?
1
188
"just because you've started paying, doesn't mean you're no longer the product." 😅
Replying to @Debloper
I pay every month for Microsoft 365 and still encounter ads in Outlook. If a customer is paying for the service, should they also be monetised through advertising? @satyanadella @MicrosoftHelps What's Microsoft's position on this?
1
119
In 2 years time (which I'll be wrong about, it'll end up taking longer)... if modular desktop computing still exists (hope it does), then expect motherboard vendors to start adding on-board SXM4+ support directly, without PCIe adapter cards (which I'll be right about).
53
Since when did ChatGPT/OpenAI start showing ads with/after responses?
1
1
113
Wish there was a selection pressure for Indian Banks to offer MCP servers & support agentic workflows.
1
60
doing it in any other way is profoundly obtuse.
I host my blog on github pages by the way
1
62
This Ofcom<>4chan drama is hilarious. Reminds me of TPB's responses to takedown notices.
1
75
NCM will also generalize the generative AI pipeline, as the image/audio/video generations already use the iterative approach of diffusion models. So when the most used modality also uses similar architecture, the optimizations to the architecture results in big gains across all.
50
Just in case you're wondering, NCM is how human brain thinks as well. We don't think one word (token) at a time like LLMs. It's the whole nebulous thought vector gets formed, which then gets processed layer by layer, with specific details added where needed.
1
65