reward hacking @primeintellect
sf | nyc
- Tweets 17,618
- Following 1,491
- Followers 53,570
- Likes 106,904
the new zeroth-order speedrun optimizers all have incredibly high aura names. Dust. Boost->Cut. Relay. Drift Leap. hell yeah dude those sound like some pretty sick algorithms
6
8
115
4,347
one tidbit i’ll share is that i’m now mostly convinced that engram modules (or things like them) will be fundamental for continual learning and aren’t just an efficiency trick
8
1
149
7,074
will brown retweeted
Today we're releasing prime-rl v0.10.0 and verifiers v0.4.0
We added our new RL algorithm IPO as the default, tailored for async RL stability, as well as performance improvements and new model support
13
37
439
33,984
this is a very creative capital structure. everyone at anthropic agrees to lock ~half their shares as illiquid collateral for funding the kinds of research anthropic already does a lot of. float is majorly restricted but everyone gets donation write-offs. brilliant.
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit.
It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion.
In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding.
In the post below, we share more details about our new fund.
anthropic.com/news/public-be…
17
10
354
23,499
some of the tax stuff here is weird, the post very clearly frames it as "donations" but also not a 501c3, take that part with a grain of salt
good q, don't know the precise mechanics being used here (yes it's an LLC but there are ways you could layer entities to get exemptions), the post talks about "giving" and "donation" a lot but yes i'm speculating a bit.
1
11
1,509
the equilibrium for a non-exclusive data vendor is to sell their data to all large and small labs who are probably not data brokers themselves, at whatever custom rates those labs can justify
6
76
6,676
the equilibrium for a lab is to buy mostly non-exclusive data, because both have the same local effect for training performance, and vendors will charge much less for non-exclusive rights (as they can make much more in aggregate)
4
13
1,695
my hot take conclusion from this, which i've really been saying for the past 2 years but i think still hasn't sunk in for many, is that broad proliferation of capabilities towards the intelligence-per-dollar limits imposed by chips and physics is the baseline, and truly private specialization is where defensible edge will come from. we are commoditizing research far faster than moats can be built upon research. the marketing value from leaking alpha is astronomical. the rate of deployed specialization (and "continual learning") will equilibriate based on physics rather than secrecy
2
19
1,146
shitposting is my best way to convey to everyone that "intelligence" is just computation over data for action selection, which caps out at the info-theoretically impossible accuracy of 100%, contrasting sharply with mainstream expectations of superintelligence, and so i shitpost
18
8
238
8,087
just imagine how many IQ points the models will have in a few years. probably like 1000. then 100K. then 100 bajillion. off the charts, more IQ than we can even fathom. thank god we developed IQ as a measure of intelligence which preserves real meaning and implication across OOMs
35
11
605
22,026