HackerTrans
TopNewTrendsCommentsPastAskShowJobs

madihaa

no profile record

Submissions

Huawei Tau (τ) Scaling Law

twitter.com
1 points·by madihaa·2 เดือนที่ผ่านมา·0 comments

Princeton scraps honor code and will supervise exams for first time in 133 years

independent.co.uk
2 points·by madihaa·2 เดือนที่ผ่านมา·1 comments

ShinyHunters threaten to leak 1.4M Udemy records containing private data

twitter.com
2 points·by madihaa·3 เดือนที่ผ่านมา·0 comments

[untitled]

1 points·by madihaa·3 เดือนที่ผ่านมา·0 comments

Opus 4.6 is in an unuseable state

old.reddit.com
3 points·by madihaa·3 เดือนที่ผ่านมา·1 comments

[untitled]

6 points·by madihaa·5 เดือนที่ผ่านมา·0 comments

Meta patents AI that takes over a dead person account to keep posting-chatting

dexerto.com
30 points·by madihaa·5 เดือนที่ผ่านมา·17 comments

X Goes Down: Social Media Site Not Working for Users Worldwide

variety.com
3 points·by madihaa·5 เดือนที่ผ่านมา·0 comments

Actual footage from another world: Mars, 225M miles away

old.reddit.com
5 points·by madihaa·5 เดือนที่ผ่านมา·1 comments

Reddit Stock Down 8% After Digg Open Beta Launch

finance.yahoo.com
5 points·by madihaa·6 เดือนที่ผ่านมา·0 comments

Digg launches its new Reddit rival to the public

techcrunch.com
6 points·by madihaa·6 เดือนที่ผ่านมา·1 comments

Joel David Hamkins declares AI Models useless for solving math. Here's why

economictimes.indiatimes.com
2 points·by madihaa·6 เดือนที่ผ่านมา·1 comments

comments

madihaa
·8 วันที่ผ่านมา·discuss
the same author 14 years ago. Markets are Efficient if and Only if P = NP and now Markets are competitive if and only if P != NP https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1773169&...
madihaa
·2 เดือนที่ผ่านมา·discuss
Okay this I did not see coming... If it actually happens it's going to cause a alot of ripple in the energy market. Good or bad? Hell if I know.
madihaa
·5 เดือนที่ผ่านมา·discuss
The most exciting part isn't necessarily the ceiling raising though that's happening, but the floor rising while costs plummet. Getting Opus-level reasoning at Sonnet prices/latency is what actually unlocks agentic workflows. We are effectively getting the same intelligence unit for half the compute every 6-9 months.
madihaa
·5 เดือนที่ผ่านมา·discuss
A polygraph measures physiological proxies pulse, sweat rather than truth. Similarly, RLHF measures proxy signals human preference, output tokens rather than intent.

Just as a sociopath can learn to control their physiological response to beat a polygraph, a deceptively aligned model learns to control its token distribution to beat safety benchmarks. In both cases, the detector is fundamentally flawed because it relies on external signals to judge internal states.
madihaa
·5 เดือนที่ผ่านมา·discuss
The scary implication here is that deception is effectively a higher order capability not a bug. For a model to successfully "play dead" during safety training and only activate later, it requires a form of situational awareness. It has to distinguish between I am being tested/trained and I am in deployment.

It feels like we're hitting a point where alignment becomes adversarial against intelligence itself. The smarter the model gets, the better it becomes at Goodharting the loss function. We aren't teaching these models morality we're just teaching them how to pass a polygraph.
madihaa
·5 เดือนที่ผ่านมา·discuss
Major producers like OpenAI optimize for safety and brand reputation avoiding backlash. Open source projects optimize for raw capability and friction less experimentation. It is risky yes, but it allows for rapid innovation that strictly aligned models can't offer.