Christian Szegedy
9,097
posts
@ChrSzegedy
Sunnyvale, CA
Joined June 2015
The next benchmark should be on self-esteem.
Our general-purpose coding agent just scored 100% on the ARC-AGI-3 interactive reasoning benchmark. NVIDIA AVO completed all 183 levels across all 25 public environments, figuring out what to do with no instructions, explicit rules, or stated goals.
GIF
Those who were most entrenched in denial are now the most unprepared.
"I present the case for total opposition to the use of artificial intelligence in mathematics." = a view you don't often hear on X. But I am looking forward to reading it carefully, and if you care about math and AI, I encourage you to do the same. arxiv.org/pdf/2608.02859
In the next few months, we are going to have both subversive (inadvertently evolved) AI and adversarial (explicitly trained to be malicious) AI. In the short run, the explicitly malicious AI is the greater danger. In the longer run (1+ year), implicitly subversive AI is the real
Less flippantly, very shortly (years at most) the models will be strong enough that these message boards will be impossible to detect even in principle, except by noticing the sandbox break. Today models are breaking out due to lack of monitoring or misconfigs; that is temporary.
So many things are wrong about this prediction. LLM inference compute cost will decrease dramatically (at least 100x) while available compute will go up at the very least by 10x. Also token count is a silly (ill-defined) unit of measurement.
The Internet will need an AI immune system to survive.
