Vitalik Buterin on Warpcast

Vitalik Buterin pfp

Vitalik Buterin

Some predictions on 2030 AI capabilities. But I think it's too pessimistic in its implications: if AI bug-finding is easy, then *the devs themselves* could use it to strip out bugs first. Average code has 15-50 bugs per 1000 lines; if consumer bug-finders could catch 99%, then quite a few apps could become bug-free.

61 replies

316 recasts

1222 reactions

Vitalik Buterin pfp

Vitalik Buterin

1k lines is a lot: four apps that I wrote for personal use and regularly use (findable on https://github.com/vbuterin ) are a total of 919 lines of code. I feel like people forget too much that the "endgame" of cybersecurity (what happens if defense *and* offense become ultrasophisticated) is quite defense-favoring.

3 replies

3 recasts

23 reactions

aferg pfp

@aaronrferguson

15-50/1000? Where are you getting that statistic from

1 reply

0 recast

0 reaction

Emad pfp

Code models will soon be able to ingest a million tokens of code zero shot without any issue I don’t think it will take to 2030 tbh

0 reply

0 recast

1 reaction

Vladimir Sobolev pfp

Vladimir Sobolev

WDYT about this Sir? https://officercia.mirror.xyz/hsKGwmqe2-B5xR1V-qydYwgLwGLuMliUsudxQ6yXiiE

1 reply

0 recast

0 reaction

miro𓆣 pfp

Interesting! Unfortunately so far I'm also seeing a lot of LLMs' approximate retrievals misleading devs and making their code worse actually! Executable code =/= robust and reliable code published a little paper on that on arxiv if you're interested :)

0 reply

0 recast

0 reaction

Ox Bid pfp

Yes, exactly. I want to think more optimistically, and I can already see significant progress in human abilities through the use of AI. The only thing that concerns me is cybersecurity and personal information. ChatGPT is becoming a vast repository of all info.

0 reply

0 recast

0 reaction

CHAINVIRUS 🏴 pfp

CHAINVIRUS 🏴

0 reply

0 recast

0 reaction

Ignas | DeFi pfp

Hope happens soon. Because DeFi smart contracts could benefit from it more with so many hacks.

0 reply

1 recast

4 reactions

David Ryan (drcoder.eth) pfp

David Ryan (drcoder.eth)

Will AIs ever be able to debug themselves?

0 reply

1 recast

2 reactions

jantans pfp

0 reply

0 recast

0 reaction

americanxradass pfp

americanxradass

@americanxradass

What if the code is correct, but the hackers find a loophole and strategically take advantage of it

0 reply

0 recast

0 reaction

Daniil 🎩 🍖 pfp

Daniil 🎩 🍖

Виталя привет!

0 reply

0 recast

0 reaction

coconut 🎩 pfp

im buying $TAO rn 👀

0 reply

0 recast

0 reaction

indiorobot🎩 pfp

are we doomed? 10000 $DEGEN

0 reply

0 recast

0 reaction

Moughite pfp

Finding vulnerabilities statistically and dynamically will be definitely much more improved than current fuzzers and scanners but one thing I am concerned about is the future models ability to find complex attack paths and vectors that are usually chained with multiple exploits

1 reply

0 recast

0 reaction

kromo.eth pfp

I don’t think most people realize the ripple effects these things will cause. Computation is what drives the world forward and AI is such a general catalyst that it can be used in almost every field. With great power comes great responsibility but I’m bullish on humanity so I can’t wait.

0 reply

0 recast

0 reaction

foxx pfp

I can say that using AI to code can initially create an illusion that it can, 🧠 but eventually it gets stuck in a cycle of self-correction that becomes increasingly erroneous. 🔄🚫

0 reply

0 recast

0 reaction

Matthew pfp

Vitalik u are goaaat

0 reply

0 recast

0 reaction

xivanc pfp

1/3 In the very interesting Into the Bytecode podcast "Making sense in a changing world", you mention your reasoning about building AGI as a result of stirring a 'computational soup' long enough, which (similar to human evolution) would essentially lead to general intelligence.

2 replies

0 recast

0 reaction