@willcb
reward hacking @primeintellect
AI
Dario Amodei says Anthropic withheld general access to Mythos-class models until June, when stronger cyber safeguards were in place.
AI
A user described FlashAttention as a 'mega kernel' in an October 20 post.
AI
The poster says getting the proposed approach to work remains a meta-reinforcement learning problem with many hurdles ahead.
AI
A user credits DeepSeek directly for technological acceleration and predicts people will generally underestimate its contribution.
AI
A user argues that 100% accuracy is information-theoretically impossible, contrasting that limit with expectations of superintelligence.
AI
A user imagines model scores soaring from 1,000 to 100,000, questioning whether IQ would still mean the same thing.
AI
A user argues that the internet grows exponentially each day, countering claims that AI has used it up for pretraining.
AI
A commenter attributes some WWII-era growth to Depression recovery and calls COVID-era growth a return to trend.
AI
One post argues raw-data brokerage will commoditize, while refined post-training tasks will grow more differentiated.
AI
Anthropic says its wholly owned fund starts with 5% of shares; further contributions depend on valuation growth and board discretion.
AI
A user says ideas from Nick Cammarata’s interpretability essay inspired the work, which is still in progress.
AI
A user jokes that it “hits different” when every ablation experiment costs $1 million.
AI
A poster argues evidence from studying LLM internals matters as other monitoring becomes less useful.
AI
The poster says surveyed Anthropic staff's answers imply median annual US GDP growth of 17.1% in 2030 under Anthropic's model.
AI
A user makes the comparison in a one-line observation.
AI
A user compares that potential shift to writing and the printing press, which spread intellectual output widely.
AI
A user jokes that frontier lab researchers estimate AI agents will fully automate both kinds of roles within 70 years.
AI
One user says asking why minor labs compare with two-month-old rather than the newest frontier models is usually stupid.
AI
A user says “recursive generalization” is essentially what grokking produces and argues backdoors don’t emerge from training without adversarial data.
AI
A user jokes about checking your floating-point number too, but says looking at the data matters most.
