Balaji

Yesterday, your AI wasn’t enlightened enough, so it needed to be centralized.

Today, your AI isn’t safe enough, so it needs to be centralized.

And tomorrow, your AI won’t be patriotic enough, so it will need to be centralized.

Want a version of this where centralized power is the real shoggoth all along. And the masks it wears are:

terrorism
climate
COVID
AI

The upcoming mask is the flag. And the masks being tossed out are older ideologies that also advocated for total centralization, like communism. https://t.co/nxAITErtJs

Immutable money, infinite frontier, eternal life. #Bitcoin

This is the Linux move that’s proven to counterbalance Microsoft. https://t.co/l9rDB3cMF2

Time for a “CopyLeft AI Data License” - if this data is used to train an AI model, then the model must open its source code and its weights.

Naval

Note that I deeply respect what Microsoft and OpenAI have built. They deserve profit and props.

But we also don’t want all of tech forced to use one API again; we saw that movie in the 2000s. Need to decentralize.

So: GPL for data, not just code? Could use BitTorrent to sync…

The fact that LLMs are bootstrapped from human language — and not from, say, sensor data — means they arguably are more human-like than a genuine alien.

It’s trained on Reddit posts, not born on a remote planet.

Many theories about what genuinely alien life looks like.

Perhaps you can get something like the complexity of carbon-based life with silicon. Or you just find “vanilla” carbon-based life in oceans like that of Europa.

[1]: https://t.co/elbzEpPvq5

[2]: https://t.co/6WvGI3R3RK https://t.co/247YAV8diz

Oh, it’s a huge difference.

Human language encodes so many assumptions based on our biology.

A small example might be — why don’t we have many different words for what magnetic fields feel like, but do for what foods taste like?

We have a dedicated sense for taste. https://t.co/VM4Gz5Zs1e

A better San Francisco is possible.

- Start with 1-2 month experiments around the world
- Set up pop up towns to recruit curious people
- Have them contribute so it's break-even
- Choose inexpensive locales to cut costs
- And if it works, we "just" scale it up 5-10X https://t.co/jYqQPCVArP

🤔 Should we make a Longevity Network State?

1️⃣ month to go until the world’s first pop-up mini-city experiment with longevity at heart

📅 Mar 25 - May 25 (2 months)

We are at capacity but we have a few spots left for longevity workshop facilitators
https://t.co/nBkR22CFtN

🧵

VitaDAO is a community owned collective funding early stage longevity research, governed by $VITA-holders 🧬 👉 https://t.co/uq15UrsgHO

VitaDAO 💛

Focus in particular on:

1) Recruiting talent from around the world that's boxed out by US immigration

2) Pulling together an integrated package that hyperdeflates cost-of-living in return for walkable communities and a more standardized lifestyle

On-campus, anywhere?

San Francisco is the worst-run city in the world.
But perhaps the world can give it a run for its money.
https://t.co/Bnnkj2Nzbq

See this excellent post that decompresses a tweet.
https://t.co/6sBzDZnRwf

Will the new Al Qaeda be Artificial Intelligence Qaeda?

It looks the same: AI Qaeda

And also wants to stop progress because it fears a hypothetical all powerful being... https://t.co/v4H5EEGSbY

Viral doomerism is itself a far more plausible threat to human progress than self-improving AI.

e/acc originator || growth lead @ civilization || once banished from the Twitter realms || currently somewhere in intergalactic space

Beff Jezos — e/acc

We don't yet know what AI looks like, but we do know what doomsday cults look like.

That's a known human phenomenon — and a different kind of prior probability.
https://t.co/jkCxkMxIsw https://t.co/LH8wTyitF0

Against Clippy, consider clipping.

Clipping happens when something that theoretically exists on the computer hits a physical world roadblock.

In theory, your model goes exponential! But in practice it can’t deliver a signal over one volt. https://t.co/N1RPqZEpG0 https://t.co/mh2T3WZtrj

Marc Andreessen

@pmarca okay but then i read 

https://t.co/gbjAa23rM6

and

https://t.co/OQfZ3gNUPC

now i can’t figure out how a self improving exponentially growing superpower of pure intelligence, widely available, isn’t the equivalent of handing out nukes to random 

i know how i sound but …

🏗️ Love to build stuff (@runwayco, @sandboxvr, @postmates, @zynga) people love. 💸 Investor @amplitude_hq, @moonpay, @sorarehq, @bankmercury, @italic

Siqi Chen

The digital-to-analog interface is just completely handwaved in most of these AGI doom scenarios.

And yet it’s highly nontrivial, and imposes real bounds on what any program can do.

A duality to play with.

The internet was originally developed to be resistant to nuclear weapons. Digital defense against physical attack.

Assume AI can eventually do anything a human can do digitally. What are potential physical defenses to a digital attack?

The fundamental concept is that the digital/analog interface is highly nontrivial and introduces a lot of friction.

Basically, the physics of the physical world bounds AI. And so does math in the form of cryptography. https://t.co/rWT4gvEL7r

See also @SergeyNazarov’s thread, where he discusses how cryptographic verifiability can put fences around AI.

Even a very smart thing can’t break out of a cryptographic cage. No amount of computation can feasibly solve certain kinds of math problems. https://t.co/UVVDcnfW6N

There’s an obvious solution: you’ll know you’re talking to a human if and only if they have sufficient proof-of-human.

It’s the authenticity industry.

AI makes everything digital easy to fake.
Then cryptography makes it again hard to fake. https://t.co/NqhRq09LKf

Recovering economist turned biotech VC. Lossy compression not recommended. No, we have not picked all the low-hanging fruit yet. My other half: @diviacaroline

William Eden

In the 20th century people relied on centralized institutions — the government, the newspapers — to tell them what was real.

In this century, we’ll need decentralized cryptography to tell us what is real. https://t.co/ESM4rb9xzy

There will be an arms race, but you can use AI to detect uses of AI.

Moreover we already use a combination of cryptographic, statistical, and human signals to verify things online. https://t.co/Tuystk4ABa

AI is like a bug light pulling people back into San Francisco, California.

But what happens when a city and state facing massive deficits decides to predictably turn upon its last high profile industry?

They’ll attack them for automation.
Then, punitive regulation and taxation. https://t.co/UYFkLZ9uR4

Ron Conway made an interesting point recently: AI startups are concentrated in the Bay Area in the classical pattern. Like Internet startups in the 90s. If the San Francisco city government met them halfway, SF could be the AI capital of the world. It's theirs to lose.

Paul Graham

We’re seeing the scorpion and the frog play out in real time.

Of course the state will sting all these AI startups again. That’s its nature. https://t.co/Fp6y0dulvH

California’s population fell by 500k.

And deficits soared to $24B.

There is however one very high profile group of people still in state.

And very publicly receiving millions of dollars.

And automating all the jobs! 

Or so it’ll be portrayed.

Then will come the attacks. https://t.co/pLGuvc72mL

Also, AI directly threatens the Democrat base: college-educated white collar workers.

Automating their jobs while headquartered in San Francisco, California is like preaching atheism in Vatican City.

You may have no need for their services. But they won’t appreciate the logic. https://t.co/8e5jNNRo93

The concept of a low-probability-but-extreme-downside scenario is not some new thing.

Dick Cheney used it to justify invading Iraq. If there was even a 1% chance of nuclear terrorism...well, we need to treat it as a certainty. https://t.co/KSqL32YR0d

Thing is...right after 9/11 that seemed plausible!
And even in hindsight, you wanted *some* reaction.
But you wanted a surgical reaction.
A scalpel, not a sledgehammer.
So too for climate, COVID, and other apocalypses.
And AI risk.
The cure can be worse than the disease.

Dick Cheney was an x-risk pioneer, along with Zhengli Shi of the Wuhan Institute of Virology.

He justified invading Iraq based on the 1% risk of nuclear terrorism. She justified gain-of-function research to prove the risk from deadly coronaviruses was real.

Do we have a third? https://t.co/i7AVgBkkbq

@balajis's Latest X Threads

Read threads by Balaji and get notified about future ones.