Hsc (@hsrvc_) profile picture
Joined March 2021
Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)Photo by Hsc (@hsrvc_)
Pinned Tweet
Can Adapt-1 by @rei_labs learn to play Super Mario Bros without demonstrations? We tried it on World 1-1, starting untrained. It learned a reactive policy from its own play in about 36 minutes of gameplay, then cleared the level with learning off.
11
29
5
101
371,380
Update on World 2-1. I froze the 1-1 player built on @rei_labs Adapt-1 and added two small learners, the correctors, that can replace one of its moves. It now clears 2-1 and still clears 1-1.
Can Adapt-1 by @rei_labs learn to play Super Mario Bros without demonstrations? We tried it on World 1-1, starting untrained. It learned a reactive policy from its own play in about 36 minutes of gameplay, then cleared the level with learning off.
5
9
1
16
582
Machina, Adapt-1's sequence engine, also clears 2-1 when it starts from its own 1-1 sequence. But that sequence no longer clears 1-1. The correctors keep both levels.
1
2
68
How it works: my harness tries single moves at the spot where Mario dies, and Adapt-1 learns when to use each one. It's one level so far. Logs, video and the caveats are here: github.com/hsrvc/adapt1-mari…
2
64
The source code is on GitHub, now with clips of both ways Adapt-1 played: its decision loop and its Machina control sequence. github.com/hsrvc/adapt1-mari…
Can Adapt-1 by @rei_labs learn to play Super Mario Bros without demonstrations? We tried it on World 1-1, starting untrained. It learned a reactive policy from its own play in about 36 minutes of gameplay, then cleared the level with learning off.
2
11
2
88
200,773
Here's the decision loop of the policy it learned from its own play, starting untrained. The bars are Adapt-1's score for each move, and the highlighted one is its pick. Learning is off.ok
1
7
300
And Machina's control sequence: one fixed plan of 244 commands, played to the flag without looking at the screen. The strip is the whole sequence; the bars show how long each button is held.
8
281
Hsc retweeted
Introducing Adapt-1 Machina: Learning Continuous Control Sequences With Adapt-1 Machina, we explore how continuous sequences of machine controls can be learned through reinforcement learning, starting with coarse outcome feedback and without demonstrations or a critic.
32
109
12
658
44,826
Hsc retweeted
the next deepseek moment for AI will be crypto native
19
4
2
76
11,122
Three Predictions: 1. Some form of AI, probably neurosymbolic in nature, will come that is far more economical and data- and energy-efficient than LLMs, and it will make an absolute fortune. 2. LLMs, on the other hand, will never be all that profitable (aside from the chip companies selling shovels in the gold rush). 3. Today’s gigantic bets are premature, and most won’t pay off.
144
134
37
1,055
90,977
Hyperliquid’s secret to success was as simple as it was genius: Nail the product → attract the liquidity and users → then let builders build on top of that. Product becomes platform. The exact same thing is happening with @Collector_Crypt right now, with projects like @renaissxyz, @slabcash, @ComicBook, @MagicEden others building on top of their infrastructure and inventory - horizontally expanding both their distribution and revenue. What they are cooking in-house on that front will take all of this to a whole new level. If you only knew.
23
19
1
141
21,276