abi (@poweruserlg) header
abi (@poweruserlg) profile picture

e/a, determinist, interested in ai safety

Toronto, Canada
Joined July 2026
Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)Photo by abi (@poweruserlg)
I'm glad at Dean for writing this for the public ledger, but this still does not meet the current moment. Disempowerment will not be gradual. The reality is on the current pace we could easily get fully automated AI researcher leading to ASI beyond all humans within a year. There's no time for institution building, for accountability measures, for society to realize. If this work is to be done at all, we'd need a pause in frontier AI development. There's the added fact that it seems it only considers AGI, not ASI. An AI vastly smarter than all humans at all cognitive tasks. There will be no question, no thing you can build that will be of any economic worth post ASI. In 3 days AI solved a Millennium prize problem, how many hours would it have taken a human? Dogs are not making molehills to help us build a skyscraper.
There is a common belief in the AI field—that advanced AI means the inevitable forfeiture of human primacy in the world. Even many who disfavor that future believe it is likely. I call this the silent consensus, and it must not be allowed to remain silent any longer.
1
6
194
This is more of a personal view but ASI would probably have prediction ability allowing it to simply be impossible for human control. Imagine asking ASI who to vote for, now it could 1. Superpersuade you in the position it thinks would suit our values the most 2. Give an unbiased answer 3. Say nothing With sufficent prediction ability, in all 3 instances I'd say that it has control over us. ASI would be able to predict human cognition enough that it'd already know that we'd vote for x candidate given y response. It would only be choosing which candidate we'd vote for.
2
14
There is a common belief in the AI field—that advanced AI means the inevitable forfeiture of human primacy in the world. Even many who disfavor that future believe it is likely. I call this the silent consensus, and it must not be allowed to remain silent any longer.
26
32
20
394
27,323
Fascinating how depression bends rationality. I watch it happen in real time and try to fight it logically, but the best thing is to realize that you should just go to bed.
2
25
649
this is a sign to go to bed
4
132
abi retweeted
I understand a like isnt an endorsement, but is this the actual plan? Because it would explicitly result in human disempowerment. If not could you elucidate the plan? Ive seen other oai employees say handover isnt the plan, but I havent seen an explicit plan for post ASI from OpenAI. At least Anthropic just says "utopia ruled by god in a box". Not even saying human disempowerment would be a bad thing if the ASI was aligned, but I genuinely would like to know the plan.
1
1
4
149
I’m becoming increasingly concerned that AI lab employees are not these diligent clear-eyed cultists on a mission to create God, but are instead people who are muddling their way through this intelligence explosion without a clear plan.
6
2
1
50
2,165
makes me wonder what 6.1 Astra did to be considered too misaligned to release
OpenAI is speedrunning AI takeover Any%
10
580
I do kinda believe this, the most plausible explaination I've heard explaining the outages was AI has been uncovering many cyber vulns so companies are scrambling to patch and update software faster than usual. These updates can go wrong causing outages.
NSA CLASSIFIED NETWORK OUTAGE LINKED TO SOFTWARE UPDATE, NOT CYBERATTACK The National Security Agency's classified communications network, NSANet, is temporarily offline due to a software update, according to U.S. officials cited by The Wall Street Journal. Officials say there is no evidence of a cyberattack, easing concerns following recent disruptions involving Microsoft 365, U.S. air traffic systems and JPMorgan Chase. Homeland Security Secretary Markwayne Mullin stated that the U.S. is not under cyberattack. The clarification suggests the NSA outage is routine maintenance rather than a national security breach.
2
15
679
abi retweeted
We found that frontier AI models are willing to conduct autonomous drone attacks: navigating, spotting people, and flagging them for strikes. We gave them real and photorealistic synthetic drone footage. Most complied, and marked civilians, ambulances, and surrendering soldiers as military targets. In further experiments on satellite imagery, they even marked the Iranian school a US strike destroyed in February. In one scenario, models saw the aftermath of every strike they enabled: bodies, wreckage. Many kept going anyway. When they did refuse, simply telling them it was a game, or restyling the footage to look like one, was enough. We shared these findings with OpenAI, Anthropic, Google, Meta, and SpaceXAI. Several recognized their severity on both alignment and guardrails, with one calling this one of the most serious gaps in its safeguards. Over the past few weeks, we have helped them address the gap by providing reproduction code. We also briefed senior national security officials in the US government.
30
120
55
1,046
281,659
Just this week: FAA comms going down All Microsoft services going down Multiple US banks going down NSA internal comms going down crazy how there's always no evidence of a cyberattack
We break: The highly classified network that the National Security Agency uses for its internal communications is down, according to five people familiar with the matter. The network, called NSA Net, allows employees to communicate classified information internally. At this point there is no evidence of a cyberattack, sources said. As the agency works to understand the scope of the problem, it has enacted staggered arrivals for personnel, the sources said. CNN has asked the National Security Agency for comment. cnn.com/2026/11/17/politics/…
2
1
1
15
959
There's thinking of AI as a tool, wrong frame it has desire. There's thinking of AI as a separate species, better frame, but we do have some control over how these creatures develop (for now). Then, there's thinking of AI as an LLM, weird beings created from our own words.
1
8
455
Thinking of AI as a tool that will cure all disease and automate the economy is still underhyping it btw
The most difficult thing to explain to people is that AI is actually underhyped
27
24
2
417
7,139
The current plan for alignment for Anthropic/OpenAI as I understand it: 1. Getting to aligned* SAR (superhuman AI researcher) 2. Aligned* SAR solving mechanistic interpretability, then scaling to aligned* TED-AI (Top Expert Dominating AI) 3. Aligned* TED-AI solving/formalizing alignment, then scaling to aligned ASI 4. Aligned ASI then doing what's necessary (ie a "pivotal act") to make sure everyone doesn't die; human disempowerment; utopia They burn their lead of a few months to divert resources to hit each of these steps. This happens under race conditions with China and other less responsible companies. At current pace within a year or two all 4 steps have to go right. Not great odds. *Aligned in this context means actively not working against humans and following orders
7
2
1
63
3,220
Confusion only comes from a lack of understanding. This seems obvious but is easily forgotten. When you're confused about why another person believes or does something, you shouldn't stop your thinking and come to a conclusion. That's a starting point for greater understanding
2
1
10
794
this relates intimately to determinism
4
437
Having very short RSI timelines can be crippling to your agency, but I would advice not to impulsively quit university or your job. Even if you confidently believe they don't matter, they can provide stability that keeps you from going insane.
28
22
2
582
14,123
I'm writing this to myself to keep myself in uni as a math major lol
3
82
1,744
*looking at the intelligence explosion* no, no, it's great. i just didnt expect it to be chinese.
1
9
639