Ernie Aase

Humanity has everything it needs, right now, to achieve a large part of the value proposition for AGI in science.

Does this help me build a better solution? Does this help me communicate more honestly? If not, it might be noise worth filtering out.

AGI capability, once possible, inevitably proliferates As capability rises, the entry barrier lowers More actors with capability leads to statistical certainty of misalignment Therefore, defensive systems become necessary, not just alignment.

Designing robust interaction protocols between models means AGI quality output before AGI tier threat.

The ability to create AGI systems proliferates over time The barrier to entry continually lowers Perfect coordination or control is impossible without totalitarianism Therefore the emergence of misaligned or "bad" AGI approaches inevitability over a long enough horizon.

We're likely entering an era of permanent coexistence with both aligned and misaligned systems.

After we build "safe AGI," the threat never disappears. The capability to build unsafe AGI will always exist from that point forward. This creates an inescapable situation: Either we establish perfect control (essentially a totalitarian state, likely run by the "good AGI") Or bad AGI will eventually emerge. Since the totalitarian option is both undesirable and unlikely, the emergence of bad AGI becomes inevitable. The only solution isn't preventing bad AGI (impossible), but creating widespread protective systems - essentially "good AGI bodyguard systems" distributed to defend humans/society against the inevitable bad actors. This isn't about winning once, but establishing an ongoing defensive capability that can constantly detect and counter new threats as they emerge. The race never ends - it just shifts to maintaining a defensive advantage in perpetuity. We're likely entering an era of permanent coexistence with both aligned and misaligned systems.

The combinatorial complexity of language and huge AI context windows means it's mathematically impossible we've found optimal input patterns yet. It's like we've been handed an incredible mathematical instrument, and we're still just pressing buttons to see what happens.

@maxwinga @elonmusk We can produce systems capable of AGI quality output before AGI level risk. These same systems can automate alignment research. Working on it rn.

This is an AI safety researcher from OpenAI, who quit because he's terrified of what they're doing. With no hope for alignment, their race to build superintelligence is a race to kill every man, woman, and child on planet Earth. @elonmusk what are you doing to keep us safe?

@lessin Surprised that youre taking the cost claims at face value. tech in china literally is an arm of the ccp. most prolific totalitarian power in history = sure ill just take their word for it. ?????????????

$6m in training destroys $500b+ in meme driven marketcap in 2 business days. Highest ROI play in history (hope they shorted before release!) ... and this is the problem with belief driven investing... fickle.

@Akshay_VAK Indias biggest issue will be getting their top talent to stay, this is nothing new. Everything else, infrastructure, quality of life (pollution, medical care) all revolve around this. Cant build the genius businesses when all your geniuses opt out of working in the country.

@capitalspidey This is so pathetic. "he doesnt represent America" just say you hate our country, grow some balls.

Americans using this to bash Mackie when hes just saying Captain America should represent what it means to be a good person is really telling.

@emollick The best personality is one that adjusts to enhance whatever youre trying to prompt it into doing. (since it predicts off its own responses in the context window too)

A likely incorrect lesson that the AI labs took from the Sydney/Bing debacle two years ago was that AI assistants should have any personality removed. I suspect personality not only makes AI pleasant to use (eg DeepSeek, Claude) but also makes people understand they are fallible

absolute fucking banger of a pitch deck coming through Were Building The Mega Factory of Desci

@tszzl use the crutches, all of them.

feel this way about text based chains of thought it’s probably a massive kludge that still scales all the way to superintelligence

What people are missing about Deepseek R1 : Even if the output from a system is 5% better, if you cant guarantee privacy, its value drops drastically for many use cases. Theres a reason anthropic advertises "Privacy-first AI that helps you create in confidence" I wouldnt let any data with a modicum of importance touch chinese servers for a nanosecond.

People are simultaneously worried AI will solve every issue in the world, and also fail to solve an issue of nihilism it will create. Are you attributing maximum problem solving abilities to it or are you not? Be consistent. This is unclear thinking. AGI. AI. ASI.

Either AI masters psychology or it doesn't If it doesn't -> No universal solution -> No universal nihilism problem If it does -> Full understanding of meaning-making -> No nihilism problem Your pessimism has no grounds.

@nealkhosla Well said and strongly agree. Taking the numbers at face value is naive. Perhaps theyre not a flexible enough power structure to withstand what AGSI would mean. We usually model ccp as benefiting from ai. Going to have to revisit this premise.

for everyone up in arms: 1. i'm not my father 2. i'm not saying open source AI is bad 3. it's a great model all i'm saying is this is ccp propaganda + strategic undercutting as they do in every high value industry (EVs, drones, solar, etc.) x.com/avichal/status…

deepseek is a ccp state psyop + economic warfare to make american ai unprofitable they are faking the cost was low to justify setting price low and hoping everyone switches to it damage AI competitiveness in the us dont take the bait