@Snowyprints
Joined March 2018
snowyprints retweeted
Something big is brewing with @EclipticaOS & @rei_labs
8
8
28
881
snowyprints retweeted
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
370,547
"If you are an AI power user, you will probably be as excited as I am." base:0x6b2504a03ca4d43d0d73776f6ad46dab2f2a4cfd
Just Plumbing... Already cleared 8-fig trades on 2/3 of these tokens since this post. VVV, I hold my last 10%. That was a monster trade for me because it let me double tap it.....Thanks Dragonfly crew for putting the fear of GOD in people by talking nonsense. NEAR, I had been accumulating since it was below 1. It took its sweet ass time, but as soon as ZEC started to take off, I knew the entire privacy sector was going to reprice, and NEAR was the best bet with the biggest spread against ZEC. Add to that all the new ships and the lack of supply overhang, and I knew it had to be a winner. I still hold 60% of my position and will keep selling every time the chart breaks key resistance. AERO, I still hold 100% of my position. Not a bad trade, but not the best one either. RH motion stole the mojo on this one, I still think it gives me my 2.5x tho.... What trades are next? One that excites me, Variational. I have been farming it non-stop and, as I have been saying, this team is special. It ain't just another perps protocol, they have a very creative and original approach to what they are building, and we already saw a glimpse of it with the swaps market. If you have been following me for a while, you know I put two wallets in the top 30 on HYPE after Jeff rejected my investment check, and I am about to do the same on VAR. VAR can play out two ways.....It can behave similarly to HYPE, where we just go up and up until it reaches the comp valuations everyone is already charting (HYPE, LIT, ASTER), or it can do what I call the LONG U, where everyone sells it into oblivion and, once that process is exhausted, it shoots up (see LIT). If we get the LONG U, you will see me buying a shit ton more on top of my airdrop around the bottom of that U. If we don't, you will see me buy a shit ton on top of my airdrop when they inevitably face their dragon. One that intrigues me, TIBBIR. I was able to accumulate a ton of it around the 110M range. I am intrigued by what Micky Malka will cook with this one.....I respect that dude a ton and like the way he thinks, and that alone was enough for me to size what I thought, and hoped, was the bottom. This is a dangerous trade because expectations among holders are so high that whatever gets revealed might simply fail to meet them. I have been debating cutting a ton here (50% of the position), but I have decided to play ball. If the reveal sucks, I will most likely roundtrip the trade....buuut if it doesn't......I could 5x this with size. One that is nerdy, REI. I have been accumulating a good amount of it below 15M. This one is pure excitement about what this team might achieve by making AI adaptive after deployment instead of largely freezing intelligence at inference. If you are an AI power user, you will probably be as excited as I am. One that is rogue, ASTER. I have been building a size on it....Yep...I just can't fade 280+ days of accumulation and a lovely LONG U chart. As you can see, every trade above has the same underlying structure....Setups where I can reasonably underwrite 20–30% of downside while having a path to 2–3x on the upside. Bear markets make these setups easier to find because prices can compress far more violently than the underlying opportunity, but they exist in every market.....This is my arena. You see.... When I was finally able to play with size, I started thinking differently about how to play the game and eventually found myself applying the logic behind the Kelly Criterion to the way I trade....It's simple, once you have size, maximizing the return on any single trade becomes less important than maximizing the number of times you can bet that capital to favorable asymmetry without meaningfully impairing the base. If you have read all the way to here, stick with me, you will probably get something out of this. It’s weird but many never care to think that wealth does not grow arithmetically, it’s all multiplicatively. If you start with 100, lose 50%, then make 50%, you are not back at 100. You go: 100 → 50 → 75. So even though the average return between those two trades is 0%, your wealth is down 25%. Ouch, right? Geometric compounding problem..... Every return you make is being earned on whatever capital base the previous trade left behind, which means large losses do much more damage than they first appear to....Which matters enormously once you start playing with size because the base itself becomes the most valuable part of your game. When your base is small, you are forced to look for insane multiples because even a great percentage return won't change your life in the ways you dream of....So if you have 20K and want to make 2M, you need a 100x. That necessity naturally pushes you into chasing insane upside....buuuut....most of those trades come with a nasty fkn failure rate attached to them too. When you are finally able to play with size, that necessity is no longer there, and you need to readjust. You need to be a tad more boring because now, with 20M on hand, a 10% return makes you the same 2M, which buys the same things as the 100x 2M. Which changes the hunt completely.....you can stop digging through shit for 1000xs and start hunting fat asymmetry with size. Now, you can spend your time looking for trades where you think you can lose 20–30% if you are wrong and make 2–5x if you are right, and then keep doing that over and over again with meaningful size. When the base gets large enough, every percentage point you earn on it becomes meaningful in absolute terms, and every time you protect it, you preserve the ability to keep compounding from a larger denominator. 10M becomes 20M. Now another 2x is not another 10M, it is 20M. Do it again and it is 40M. Which means you do not need increasingly crazier trades, just favorable setups applied to an increasingly larger base. I’ll tell you this, size does make the game easier, but only if you stop trading like you are still trying to escape the basement.....If you already have size and you are still mentally trapped in the 10,000x cage, eventually you are going to blow yourself tf up. Stop chasing miracles, if you are already trading size there is a very good chance you already had one or two break your way.....take the gift and stop going back to the market begging for another one.
2
12
301
snowyprints retweeted
"bro wat is imd?" An explanation of imd.fun, in (very) simple terms. 🧵
21
35
12
191
29,237
snowyprints retweeted
Adapt-1 Developer Docs v1.0 is live. We’ve rebuilt the docs around getting started with Adapt-1 and understanding what you can build with it. Quickstarts, API reference, mechanisms, traces, integration guidance, and use cases. docs.reilabs.org/docs/neuroa…
12
40
4
126
6,034
snowyprints retweeted
Me watching CT try to figure out what they missed as rei goes to $5 in a straight line next year
It's interesting reading the replies here. "what can I do with this?" "how is it useful?" "it can't generate code. whats this even for?" lol kinda makes you wonder about how conditioned we've become to thinking useful AI has to look like gen ai meanwhile at @rei_labs lol
7
10
2
34
2,470
snowyprints 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.
29
107
12
649
41,442
snowyprints retweeted
Bullmaxxing
10
7
1
104
6,480
snowyprints retweeted
Can the fly turn sensory evidence into a decision? In this short experiment it has to earn bananas and solve the captcha. We follow its retinal-contrast sensory readout as it inspects images.
7
11
2
47
2,058
WE ARE ABOUT TO WALK INTO THE MOST ABUNDANT BALANCED WEALTHY AND SUCCESSFUL PERIOD OF OUR LIFE. WE NATURALLY ATTRACT GOOD FORTUNE, AND WE ARE WEALTHY IN MORE WAYS THAN ONE. WE GIVE OURSELVES PERMISSION TO PROSPER, AND WE HAVE THE POWER TO BUILD THE LIFE THAT WE DESIRE.
WE ARE ABOUT TO WALK INTO THE MOST ABUNDANT BALANCED WEALTHY AND SUCCESSFUL PERIOD OF OUR LIFE. WE NATURALLY ATTRACT GOOD FORTUNE, AND WE ARE WEALTHY IN MORE WAYS THAN ONE. WE GIVE OURSELVES PERMISSION TO PROSPER, AND WE HAVE THE POWER TO BUILD THE LIFE THAT WE DESIRE.
2
128
snowyprints retweeted
Can Adapt‑1 Preview stay in an interactive decision loop while thousands of outcomes keep changing its persistent decision state? We finished migrating to the infrastructure intended for public access. Here’s what latency looks like as retained experience accumulates.
16
42
4
133
10,183
"Hello I got a few questions for the team, is Adapt-1 built around RL? A friend of mine works at a chemistry lab and is thrilled with its learning speed, he gave it reaction conditions and measured yields, it started broad and retained which combinations had been productive. Any plans to highlight and guide scientists on how to use it?"
3
5
10
638
snowyprints retweeted
Launching an Ethereum game soon, Like and retweet if you want early access.
Launching an Ethereum game soon, DM me if you want early access tomorrow.
942
1,270
43
2,047
65,649
snowyprints retweeted
Big implications I think Google DeepMind's Symbolic Alchemy Test: Adapt-1 Preview w/ authored ontology: 251.6 Adapt-1 Preview w/ Discovery Enabled: 239 Discovery retained 95.0% of the authored-ontology score after removing the learner interpretation supplied before use.
Replying to @rei_labs
We had to stress-test this on Google DeepMind’s Symbolic Alchemy. Adapt-1 sticks to the task rules and opaque policy IDs. Potion identities, slot semantics, transition inputs, causal variables, action-state bindings, effect hypotheses and chemistry priors are removed.
1
7
39
1,640
n00b critics tried telling you : "it only learns because a human sets up the task structure for it." REI new release: they removed that structure and let it figure the structure out itself ... and it kept 95% of the performance. [mic drop]
Introducing Emergence. It removes learner ontology from Adapt-1’s online learning setup. Observation paths, temporal context, causal variables and action-state bindings can now be induced online from interaction and committed as revisioned state. Blog : reilabs.org/blog/emergence-t…
3
14
70
2,808
snowyprints retweeted
Introducing Emergence. It removes learner ontology from Adapt-1’s online learning setup. Observation paths, temporal context, causal variables and action-state bindings can now be induced online from interaction and committed as revisioned state. Blog : reilabs.org/blog/emergence-t…
18
71
14
411
47,604
“Sometimes an AI does not need to learn a better answer. It needs to learn a better representation of the problem.” base:0x6b2504a03ca4d43d0d73776f6ad46dab2f2a4cfd
6
25
679
snowyprints retweeted
With the latest feature unlocks of Adapt-1 we wanted to further solidify its structure. Temporal Context Projection (TCP) and Counterfactual Utility Plasticity (CUP) are rules that you can toggle on/off. Temporal Context Projection allows the substrate to learn which parts of past experience are relevant to the present decision. So, instead of treating every observation the same way, delayed evidence now becomes explicit and revisable context. Counterfactual Utility Plasticity instead asks which signals become valuable specifically when combined. Recording predictions before outcomes are known makes possible to evaluate each combination by the predictive utility it contributes beyond its components. That way useful structures are reinforced, harmful ones inhibited, and stale ones allowed to decay under distribution shift. This has nothing to do with episodic memory, instead CUP and TCP change how retained evidence is transformed into decision state and organized into persistent predictive structure. TCP and CUP combined make rules whose temporal reach and internal composition can both reorganize online as consequences arrive. TCP learns when the past matters, CUP learns what becomes meaningful together. Docs : docs.reilabs.org/docs/neuroa… docs.reilabs.org/docs/neuroa…
A reward can tell an RL/online learner that something worked without telling it which combination of internal signals made it work. Today, we’re unlocking two learning rules in Adapt-1 Preview: Counterfactual Utility Plasticity (CUP) and Temporal Context Projection (TCP).
12
36
1
105
6,147
snowyprints retweeted
against all odds and the knowledge larping pundits it'll win, good luck reversing their architecture and its behavior, maybe web2 in like 1 year or more but no one in this industry will be able to in their wildest dreams they need to figure distro and it's total victory
A reward can tell an RL/online learner that something worked without telling it which combination of internal signals made it work. Today, we’re unlocking two learning rules in Adapt-1 Preview: Counterfactual Utility Plasticity (CUP) and Temporal Context Projection (TCP).
4
13
52
2,801
so I think what Conan is pointing to here with base:0x6b2504a03ca4d43d0d73776f6ad46dab2f2a4cfd is how quickly Adapt-1 seems to acquire the relevant temporal structure and reach the task ceiling compared with trained sequence models like Mamba and vTransformer…. not just with fewer samples, but how friggin fast it gets there and how consistently.
Adapt-1 TCP and vTranformer matched the ceiling, Mamba was close. But the sample budget is the story: Adapt-1 TCP: 510 online interactions per run vTransformer: 1,000,000 environment steps Mamba: 1,000,000 environment steps
13
52
2,349