Suresh Kumar R

It’s astonishing to know that some of us have not even been born If they could do it back, then we should be able to do even more Of course for startups, hard work is unavailable as Narayana Murthy says regardless of how much people hate him for saying that

Neuromorphic computing with biological neurones on chips is the biggest concern I have in the direction of AI becoming conscious. If we use biological neurons on chips, they are bound to become conscious as a unwanted side effect even if it’s being done for efficiency and better results Personally it’s disgusting to see these chips have been developed and they demonstrated successful game control by those chips as it’s more likely to lead to conscious AI sooner than without using biological neurons

Absolutely No one would believe in you if you don’t believe in yourself

I think you’re missing the point of the original thread having two parts in it of two games being improved by editing model weight which has nothing to do with the model card mentioned improvement you are talking about **No. The game examples are not based on a model card.** In the original thread by Rikiya Takehi, the two game-related cases are: - **Car policy fix** (post 3/6): A small PPO-trained policy for a ramp task that kept getting stuck. Astra was given the weights, a scene description, and one failed trajectory, then produced a one-shot numerical edit to ~3.5% of the weights. This made the car succeed 100% of the time (by effectively adding a reverse-then-accelerate behavior). No model card or documented prior issue was involved. - **Writing Breakout (and other MinAtar games) from scratch** (post 4/6 and related blog experiments): Astra generated the full weight tensors for small networks based on the game rules and source code description, in a single reply with zero training or environment interaction. These hand-written policies scored dramatically higher than PPO networks trained for hundreds of millions of steps (e.g., Breakout ~20× higher than 500M-step PPO). Again, this is pure reasoning over the weight space from first principles, not patching a known documented flaw. Only the **Qwen3.5-0.8B math experiment** (post 5/6) explicitly leverages the model card. The researchers note that Qwen’s own documentation already warned about reasoning loops, and Astra located and subtracted the corresponding direction in two MLP layers (editing 0.016% of weights). The game cases use different setups (tiny RL policies without published model cards) and rely on direct inspection of weights, trajectories, or game mechanics rather than any pre-existing official warning.

It seems you are missing the point of “weight space programming” Here is Grok’s answer on game improvement part of thread when asked if it’s actual weight change It is actual model weight editing — direct numerical changes to the existing weight tensors (the parameters), not layer removal or high-level node/operation graph edits.2 For the Qwen3.5-0.8B case: •Astra identified a “looping direction” in the MLP layers 15 and 19. •It subtracted that direction from 64 output columns in each of those two layers. •This changed exactly 131,072 individual weight values (≈0.016% of the model’s ~800M parameters). •The edit is a precise numerical operation on the weight matrices (conceptually similar to (W_{\text{new}} = W - \text{strength} \times) outer product of a direction vector with selected columns). No layers were removed and the architecture stayed identical.5 For the car policy example in the post you asked about: •It zeroed specific weight ranges, adjusted biases, and set a small number of new connections (e.g., specific entries in actor layers). •This modified ~6,073 weights out of ~171k (3.5%). Again, pure numerical edits to the existing weight tensors.2 In short: the model architecture and computational graph remain the same; only selected numerical weight values are rewritten based on the LLM’s reasoning about what those weights compute. This is the core of what they call “Weight-Space Programming.”4

These emerging capabilities to edit model weights seems to be similar to reverse engineer binary code of apps @grok what is the nuance here not being said? Here is another who tried same approach teapaw.org/alexkrentsel/status/21…

Bhagavad Gita 2.47 Sanskrit: कर्मण्येवाधिकारस्ते मा फलेषु कदाचन । मा कर्मफलहेतुर्भूर्मा ते सङ्गोऽस्त्वकर्मणि ॥ English: You have a right to perform your prescribed duties, but you are not entitled to the fruits of your actions. Never consider yourself to be the cause of the results of your activities, nor be attached to inaction. Tamil: செயலில் மட்டுமே உனக்கு உரிமை உண்டு, பலன்களில் எப்போதும் இல்லை. செயலின் பலனை நோக்காதே, செயலற்று இருப்பதிலும் பற்று வேண்டாம்.

These days some engineers become middle managers removing CEO or tech lead from coding agents so it’s important that all the logs of conversations with coding agents are recorded, analysed and engineers and their work reviewed to ensure that they are not being sloppy and that they are using coding agents intelligently. This will help train them.

@grok is it real as claimed?

The posts like this are danger to Society, you are twisting Stop Russia phobia

Hope molecules Yeah, it goes well with the study that said when we work out, there are hormones secreted that are collectively called hope molecule @grok is this that study? What are those harmonies?

Ha ha. I hope it fixes all Siri and other dictations tools across the board including AI chat apps misrecognising words which starts shouting match we have. Thanks god I live in a village so no one thinks I am weird seeing me shouting while walking 😭😂

Congratulations team and coach and founders and organisers Well done

I am talking about using embedding to get all spec rules and coding files instead of grep for search and edit