@atomiqlabs Love Atomiq, use it frequently! Missed out on your raise because ironically, I didn't want to invest in "another bridge". Only afterward did I find out how it actually works. Oof.
Some under-appreciated aspects of the distillation debate happening right now and a few conclusions:
1) The Chinese have essentially been laundering American IP back to Americans. This has been steadily improving competitiveness in our whole economy though it's not fun for the absurdly CapEx-intensive incumbents.
2) If this could have been stopped early on, it would have been impactful. The reality is that the models are too good now. Even if the spigot is cut off, Chinese labs can bootstrap on their models' synthetic data to achieve more improvements. Independent American labs can do the same. There is no force in heaven that can halt this now; the paradigm is not going to change.
2a) In a free country, retaining control of all IP produced by models is not a battle you can win. You can't even rate-limit production without handicapping the usefulness of your tools. Clever schemes for locking down weights don't matter when models can easily be distilled just from sampling of output text.
2b) Relatedly, even if you believed the bipolar world slop put out by Leopold Aschenbrenner, you shouldn't adopt his prescriptions for total AI IP lockdown because they make no sense in a situation where IP naturally diffuses constantly (see: status quo of ML for the past decade) and the American AI race advantage has shrunk to about two weeks.
3) The commentary about data being an exclusively American resource is very silly. Do you really think, if pressed, that the CCP would not ensure that Chinese labs got all WeChat's data, for instance? They are probably retaining all Internet / IoT data anyway. China is a bigger country in population terms and generates more data than the US because of its pervasive surveillance (we have our share of that too, sadly).
5) The Chinese are currently winning major soft power by distributing models for free. American labs are losing goodwill in proportion to their stinginess.
So what do I think?
a) American labs should acknowledge their debt to academia and the public commons and share their own model weights. Then they should decrease their CapEx spend, which is going to bankrupt everybody at this rate. It's better for them because it will increase runway and decrease the massive backlash they are about to face due to their unique combination of bad messaging and miserliness. They have major advantages in capitalization, infrastructure and productization; they don't actually need total weight hegemony to make obscene profits.
b) We should pass laws that require model weights to be shared if data from the commons is used for training models in the future. As pointed out by many others, it is beyond hypocritical for labs which copied the entire Internet (including mostly unpaid contributions to every platform like YouTube) and each other, to suddenly cry foul when they themselves get copied. We should also require that carefully anonymized data from platforms be made available to any lab that asks for it.
c) We should pass laws that codify the good parts of lab stances - like that AI should definitely not be allowed to be used for indiscriminate mass surveillance, and that facilitate the flourishing of open source.
Myriad Markets partners with @TrustWallet to become the first ever native Prediction Market product inside a wallet.
Trusted by over 200 million people, Trust Wallet will make Myriad Markets' on-chain predictions accessible in a way the industry has never seen before.
@DefxOfficial@PanteraCapital@CMT_Digital@GumiCryptos@BaboonVC@CoinShares@robotventures Before depositing any money, go to their Discord and search for "bridge". Not a single message. They have blocked the word, because that's how they're taking people's money. You can't bridge back from DefX to Solana. I was muted for a week for opening a ticket containing "bridge"
@hemi_xyz@0xNomis Seems to be missing a lot of data though! LP (Sushi) and staking (Hemi) data not showing at all. Score is really low because most funds are in use but Nomis seems to only be looking at funds currently in wallet.
99% of teams have zero clue how to execute an Airdrop & how to properly reward free QA and stress testing of their infra / apps.
The baiting, farming, yapping, and eventual Binance allocation - while being coy and in character - all so fucking tiresome
Airdrop responsibly
🫡