Soumyaranjan Sahu

Researching on How to build an Intelligent Machine
Creator of @Terminal44_ Experimenting with AI and weather

Building @Yapdotmarket
Ex- WEB3 SDE at @True_ZK

[Soumyaranjan Sahu @Samrn888]

I am currently working on 2 different projects which have literally 2 different concepts but the collision is so exciting I want to document it.

[Soumyaranjan Sahu @Samrn888]

@Tim_Hua_ Not everything is AI

You should checkout
https://t.co/jUoLZa7hUA

https://t.co/Lh7CtM1F95

Here is a sandbox to play around too
https://t.co/lg7XGKK8D0

[Fei-Fei Li @drfeifei]

https://t.co/Kt50ttQRMJ

[Soumyaranjan Sahu @Samrn888]

https://t.co/BE35qAt0AP

[Soumyaranjan Sahu @Samrn888]

@hamptonism Its just getting cheaper
Buy moar

[tetsuo @tetsuoai]

I can’t sleep at night because my mind races with all the cool shit I could be building. AI has turned my workdays into 24 hour grind sessions. I code until I literally collapse from exhaustion 7 days a week.

[Soumyaranjan Sahu @Samrn888]

I still believe LLMs and agents are just impressive mimics of biological intelligence

LLMs need massive datasets and still hallucinate and struggle to generalize. A brain learns, predicts, and minimizes surprise from just a handful of examples.
The learning and predicting never stops.

The brain treats all sensory input as fundamentally the same sparse, noisy, fault tolerant signals. Swap an ear for an eye and the cortex would adapt, because it doesn’t process vision vs sound as different data types.

It efficiently runs your entire biology (breathing, digestion, balance, speech, movement in 6 degrees of freedom) while multitasking, all without you even aware of.

It operates in a true closed loop input, update beliefs / learn / generate action, take action on environment and the get feedback through input.

[Chen-Hsuan Lin @chenhsuanlin]

Very excited to introduce Cosmos 3: Omnimodal World Models for Physical AI! 🚀

https://t.co/xwrFwHsXY6

Cosmos 3 works across language, image, video, audio, and action . It brings together capabilities that often live in separate systems: multimodal reasoning, image/video/audio generation, action modeling, world simulation, and robot policy learning, all within a single unified omnimodal world model 🌎. This creates a more direct path from perception to simulation to control 🤖.

Cosmos 3 also ranks #1 among open models on multiple reasoning and generation benchmarks and leaderboards. This was a huge one-team collaboration across NVIDIA. It brought together research, engineering, data, simulation, infrastructure, deployment, and many other efforts. I am deeply grateful to everyone who contributed and proud to have been part of this journey.

Technical report: https://t.co/5qBo1Js5yi
Models: https://t.co/qBi2j9gvkH
Code: https://t.co/ZlIeDxW3lx
Website: https://t.co/xwrFwHsXY6

#nvidia #cosmos #physicalai #worldmodels #robotics

[Soumyaranjan Sahu @Samrn888]

So everybody starts their prompts with "Ok now"

[Nous Research @NousResearch]

We have been working closely with @nvidia to ensure Hermes Agent works smoothly on their new @NVIDIARTXSpark superchip and integrates with the new OpenShell runtime, which connects Hermes to @Microsoft's security primitives.

Watch our feature in the big announcement at Computex:

[Soumyaranjan Sahu @Samrn888]

@nikitabier @levelsio that looks like a suitcase you would smuggle top secret nuclear codes

[Soumyaranjan Sahu @Samrn888]

@WorldAndScience You mean like this
https://t.co/N8jqdZsoXT

[Soumyaranjan Sahu @Samrn888]

Do you remember couple of months ago an Australian research lab @CorticalLabs grew 200,000 brain cells in petri dish and trained it to play Doom ?

Before that in 2022, the released a paper How they use In vitro neurons to play the arcade game Pong
https://t.co/AIcq20BMwB

I will be posting more breakdown on this very soon

[Soumyaranjan Sahu @Samrn888]

@SakanaAILabs This paper can unlock new potential of efficiency in model training

This way small teams and researchers can now fine-tune or train much larger models without massive GPU clusters

I have given performance breakdown on this thread

https://t.co/n860y0aLB3

[Soumyaranjan Sahu @Samrn888]

DiffusionBlocks is one of the best paper from @SakanaAILabs that can help small teams and startups get access to foundational models without spending huge costs on GPUs and compute.

Curious to know how DiffusionBlocks works I ran the provided code (https://t.co/kzP7jcw9Zm) on VRAM constrained old laptop GPU.

Paper https://t.co/6OW5VMD1gt
Blog https://t.co/9KOo6GX4Um

I trained 3 different configuration
- Baseline ViT (https://t.co/oHCREf1dEM)
- DiffusionBlocks (default 3 blocks)
- DiffusionBlocks (default 6 blocks)

Since DiffusionBlocks trains each block independently, the total epochs were auto adjusted so that each block goes through the same amount of training.

Meaning if num_blocks = 3 and num_epochs = 8, total_epochs = num_blocks * num_epochs = 24

Each run uses the same training dataset as provided in the original repo CIFAR-100 dataset (https://t.co/id1AbwnRRz)

Results are below 👇

[Soumyaranjan Sahu @Samrn888]

@heynavtoor I have explained the science behind it in this thread

https://t.co/N8jqdZsoXT

[Soumyaranjan Sahu @Samrn888]

Do you remember couple of months ago an Australian research lab @CorticalLabs grew 200,000 brain cells in petri dish and trained it to play Doom ?

Before that in 2022, the released a paper How they use In vitro neurons to play the arcade game Pong
https://t.co/AIcq20BMwB

I will be posting more breakdown on this very soon

[Soumyaranjan Sahu @Samrn888]

@olsenbdnr 😂

[Soumyaranjan Sahu @Samrn888]

The general population is divided in 2 ways.

People who wants to develop intelligence (Providers)

People who wants to use intelligence (Consumers)

[Soumyaranjan Sahu @Samrn888]

Repetition strenghtens synaptic connections over time (Hebbian Learning) through Long Term Potentiation.

But it doesnt guarantee a stable long-term memory.

Memory consolidation heavily relies on the hippocampus. During sleep it generates sharp-wave ripples, periodic oscillation waves that replay the days experiences. These replays broadcast the new information to the neocortex helping it distribute and stabilize its across neural networks and long term storage

Without sleep and these offline reactivation processes even repeated information get fades away.

The paper highlights how repetition boosts cortical reactivation but it still depends on this hippocampal-cortical exchange during rest or sleeping period.

Simple thought experiment to prove my point will be

You can remember a car accident for the rest of your life when and where it happened, and about the car as well even though the event only happened once and maybe unexpectedly

But even if u repeat a random 10 or 12 digit number every day 100 times, still it would be very hard for you to correctly say it if someone asks it when you are not ready.

Memory formation depends on lot of factor such as attention, meaning, emotions, active recalls etc...

[Soumyaranjan Sahu @Samrn888]

Summary
Entropy 12.03 (+1.37) ~96% of theoretical maximum
DeadColumnRatio 7% (down from 51%)
MaxColumnUsage 718 (down from 1239) lower is better
TemporalOverlap 61% (down from 65%) still in ideal range
OutputSparsity 2%

PCA has much denser manifold now
t-SNE has many organized regions
UMAP has many compact clustered

Massive improvement in column participation as almost entire SP is participating in learning

[Soumyaranjan Sahu @Samrn888]

Huge improvement on Spatial Pooler algorithm

I will make a post about the updates soon.

[Soumyaranjan Sahu @Samrn888]

For intelligence to emerge from a system

First the system has to be a closed loop feedback system.

Second there are 2 interrelated processes required for sentient behavior

- System must learn how external state influence internal state (self-belief)
- How system's action effects external state and system must infer from its sensory state when it should adopt a particular activity and how its action influence environment.

[Soumyaranjan Sahu @Samrn888]

@hardmaru this is where they figure out "bigger is better"

Happy birthday to GPT3

[Soumyaranjan Sahu @Samrn888]

Huge improvement on Spatial Pooler algorithm

I will make a post about the updates soon.

[Soumyaranjan Sahu @Samrn888]

entropy : 10.002141
dead_column_ratio : 0.656494
mean_temporal_overlap : 0.744768
mean_active_columns : 128.000000
max_column_usage : 1091.000000

Interpretations

Entropy is ~10 indicated SP learned diversed patterns, but not evenly enough

65% columns almost never participated which is bad

Temporal overlap is in ideal range for METAR data of 10 parameters

Active columns number is healthy and SP achieved a sparsity of ~3.1 % (128 * 100 / 4096)

Max column usage of 1091 indicated a subset of columns learned quickly and dominated the space which is not good either