tech / ml | empathy is the key for a safe AGI and human future MSc Data science - @unibocconi Ambassador @italianbldrs AI engineer for https://teapaw.org/t.co/HboEKGfw35
Italy
Joined September 2026
- Tweets185
- Following152
- Followers83
- Likes5.7K
If you hide the ball from a Pong agent, does it remember where the ball is, or does it just copy the enemy paddle?
Earlier work showed Pong agents often track the other paddle instead of the ball. For our DL for Computer Vision project at Bocconi we want to test that properly:
> real Atari Pong, PPO trained from pixels
> a fog band hides the ball in the agent's half
> the enemy paddle is either visible or masked
> CNN vs +LSTM vs +attention, 3 seeds each
Then we read the agent's memory with linear probes, wipe that memory mid-rally and check the attention maps.
I'm kinda new to RL so please feel free to give suggestions!
github.com/windyITS/ppo_atar…
Most people have no idea of RL's power.
We, humans, have an unconscious genetic knowledge that is fundamentally our operating system. Everything else we learn we do it via RL.
"Eh but AIs do not work like our brain and our optimization algo is not backprop"
True, but you forget that RL is younger than me.
More to come!
I don't really know what they did with Gemini Argon but they surely had crazy RL. To me, it beats Opus 5.5 for learning! What a model!
I've been thinking a lot about this type of use and, at the moment, I think that for big companies is by far the best one.
Then, what is left for the frontier? I think they are believers in the scaling laws: they want an intelligence so intelligent that will get humanity much further than what automatisations of this processes ever will.
Still, for 95% of today's consumers, the current open weight mid-size chinese models are more than enough.
If you play a game long enough, the probabilities of you not winning even once shrinks over time.
And as Thiel says, one win is enough to wipe out all your failures.
Just keep playing.
Here to remind everyone that you can change your life with some “no” and with realising that the past is the past and not the future
literally me rn
because of the trend i re-watched the prequels and oh, man, what a trilogy! Episode III is such a beautiful piece of philosophy
suggest you all do the same :D
In the next 30 days I'll build a small computer vision project: teaching an agent to play pong, only via RL, with distractions in it and out of view visuals.
Will the agent still know how to play pong when distractive balls come to screen, fogs appears and more?
Let's see!
If you are young and ambitious, read it, no need to thank me. Very good advices.
felixrieseberg.com/hunger-be…
An empathic machine: it sees the worst of us, the tremendous human nature in our souls, but still chooses us. That is the ultimate frontier of AI alignment
Read my pinned article!
I feel bad for my future son because when he will ask me to buy him a $5k pc to study I will know that actually he'll train neural networks on it and not play minecraft like all other kids 😭
Think about what a young boy, full of dreams and stories in his head, will be able to produce 2 years from now in his little bedroom.
Think about that.
How can you not be bullish on the future of humanity?
Watch Episode 1 of our Odyssey series made with Grok Imagine.
Let us know in the comments if you'd like a full breakdown of how it was done!
Creative collaboration with @WonderStudiosX and director Dev Walker @glumlot_
The first time I booted Omarchy it felt like the OS was on my side. Friendly, almost.
Took me a while to get why, but now I think that this quite friendliness I felt is just @dhh's idea of what tech is for: bring people together and make them go "wow"
