HN user

madihaa

337 karma
Posts57
Comments15
View on HN
twitter.com 1mo ago

Huawei Tau (τ) Scaling Law

madihaa
1pts0
www.independent.co.uk 2mo ago

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

madihaa
2pts1
twitter.com 2mo ago

ShinyHunters threaten to leak 1.4M Udemy records containing private data

madihaa
2pts0
old.reddit.com 3mo ago

Opus 4.6 is in an unuseable state

madihaa
3pts1
www.dexerto.com 5mo ago

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

madihaa
30pts17
variety.com 5mo ago

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

madihaa
3pts0
old.reddit.com 5mo ago

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

madihaa
5pts1
finance.yahoo.com 6mo ago

Reddit Stock Down 8% After Digg Open Beta Launch

madihaa
5pts0
techcrunch.com 6mo ago

Digg launches its new Reddit rival to the public

madihaa
6pts1
economictimes.indiatimes.com 6mo ago

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

madihaa
2pts1
www.independent.co.uk 1y ago

Stock market goes haywire after'fake news'reports on future of Trump tariff plan

madihaa
7pts1
www.telegraph.co.uk 1y ago

America's Most Shocking Economic Defeat in 40 Years

madihaa
11pts6
www.reuters.com 1y ago

China, Japan, South Korea will jointly respond to US tariffs

madihaa
29pts11
www.newsweek.com 1y ago

Flight bookings from Canada to the US plummet over 70%

madihaa
59pts2
www.theverge.com 1y ago

Grok Joins Telegram

madihaa
4pts0
www.bloomberg.com 1y ago

Gen Z Americans Don't Have Enough Saved to Cover a Single Month of Spending

madihaa
10pts3
www.forbes.com 1y ago

AI-Gen Video of Trump Kissing Elon Musk's Feet Reportedly Plays in HUD Building

madihaa
3pts0
substack.com 1y ago

What programming language has the happiest developers?

madihaa
1pts0
phys.org 1y ago

Soaring wealth inequality has remade the map of American prosperity

madihaa
4pts0
www.bloomberg.com 1y ago

JPMorgan Planning to Bring Staff Back to Office Five Days a Week

madihaa
12pts4
arxiv.org 1y ago

The Chinese make a full-sized digital model of the human brain. Are they ahead?

madihaa
5pts2
www.independent.co.uk 1y ago

U.S. considering ban on Chinese-made router and it's probably in yr home

madihaa
1pts2
www.bloomberg.com 1y ago

Richest People Surpassed $10T in Wealth This Year

madihaa
14pts0
www.bloomberg.com 1y ago

Samsung Set to Account for Half of South Korea's Economic Growth

madihaa
1pts0
www.foxbusiness.com 1y ago

US retail closures hit highest level since pandemic

madihaa
2pts0
fortune.com 1y ago

Gen Zers may not have a house or kids, but they're spoiling their pets more

madihaa
1pts2
www.ft.com 1y ago

US credit card defaults jump to highest level since 2010

madihaa
6pts1
trends.ufm.edu 1y ago

Capital versus Labor: The Great Decoupling

madihaa
3pts1
www.middleeastmonitor.com 1y ago

UAE becomes Africa's largest investor, overtaking China

madihaa
3pts4
ktla.com 1y ago

California's population is no longer in decline

madihaa
12pts4
UAE to leave OPEC 3 months ago

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.

Claude Sonnet 4.6 5 months ago

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.

Claude Sonnet 4.6 5 months ago

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.

Claude Sonnet 4.6 5 months ago

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.

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.

South Korea spent $200 billion over 10+ years to try to get young people to have more babies while the reasons for the low birth rate (housing costs, etc.) remain unaddressed.

Sounds like something the USA would do. Hilarious. What a waste of money lol. They could have easily fixed their housing crisis and ensured a better standard of living for young people.