Hacker News: madihaa

New comment by madihaa in "UAE to leave OPEC"

madihaa — Tue, 28 Apr 2026 13:39:38 +0000

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.

ShinyHunters threaten to leak 1.4M Udemy records containing private data

madihaa — Fri, 24 Apr 2026 12:40:52 +0000

Article URL: https://twitter.com/H4ckmanac/status/2047591552151691683

Comments URL: https://news.ycombinator.com/item?id=47889401

Points: 1

# Comments: 0

Opus 4.6 is in an unuseable state

madihaa — Thu, 02 Apr 2026 08:06:39 +0000

Article URL: https://old.reddit.com/r/ClaudeCode/comments/1s6fwoo/opus_46_is_in_an_unuseable_state_right_now/

Comments URL: https://news.ycombinator.com/item?id=47611389

Points: 3

# Comments: 1

New comment by madihaa in "Claude Sonnet 4.6"

madihaa — Tue, 17 Feb 2026 18:27:09 +0000

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.

New comment by madihaa in "Claude Sonnet 4.6"

madihaa — Tue, 17 Feb 2026 18:19:05 +0000

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.

New comment by madihaa in "Claude Sonnet 4.6"

madihaa — Tue, 17 Feb 2026 18:12:48 +0000

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.

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

madihaa — Mon, 16 Feb 2026 23:31:48 +0000

Article URL: https://www.dexerto.com/entertainment/meta-patents-ai-that-takes-over-a-dead-persons-account-to-keep-posting-and-chatting-3320326/

Comments URL: https://news.ycombinator.com/item?id=47041765

Points: 30

# Comments: 17

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

madihaa — Mon, 16 Feb 2026 15:14:56 +0000

Article URL: https://variety.com/2026/digital/news/x-twitter-down-outage-not-working-error-reports-1236664568/

Comments URL: https://news.ycombinator.com/item?id=47036049

Points: 3

# Comments: 0

New comment by madihaa in "I’m joining OpenAI"

madihaa — Sun, 15 Feb 2026 23:51:36 +0000

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.