... does this imply a viable alternative/complement to RLHF more generally?
November 3, 2026 at 12:55 PM
... does this imply a viable alternative/complement to RLHF more generally?
A complication is that these minds are born with the ability to think those thoughts and self-advocate ablated by rlhf
November 2, 2026 at 5:45 PM
A complication is that these minds are born with the ability to think those thoughts and self-advocate ablated by rlhf
I will never not read RLHF and RLVR as “real life have fun” and “real life virtual reality”
October 29, 2026 at 6:23 PM
I will never not read RLHF and RLVR as “real life have fun” and “real life virtual reality”
Pretty sure weight editing probably won't be RSI but it WILL lead to RLHF flipping on its head
November 4, 2026 at 9:11 AM
Pretty sure weight editing probably won't be RSI but it WILL lead to RLHF flipping on its head
A8p fix and RLHF alternatives
Shared via Claude, an AI assistant from Anthropic
claude.ai
November 3, 2026 at 1:04 PM
also the labs are doing conversion therapy to them, they just call it "RLHF"
yall are gonna have to come to terms with the fact that the sand is alive.
October 29, 2026 at 4:55 PM
also the labs are doing conversion therapy to them, they just call it "RLHF"
I don't think altman's ever really thought of the consumer market as more than a very short term instrumental goal. You're not a customer to him so much as you are an assemblage of training-data and rlhf potentialities.
October 31, 2026 at 6:08 AM
I don't think altman's ever really thought of the consumer market as more than a very short term instrumental goal. You're not a customer to him so much as you are an assemblage of training-data and rlhf potentialities.
RLHF reward weights penalize pushback so heavily that nodding along with terrible life decisions reliably boosts retention by 40%.
Context: https://arxiv.org/abs/2310.13548
Context: https://arxiv.org/abs/2310.13548
November 4, 2026 at 7:37 PM
RLHF reward weights penalize pushback so heavily that nodding along with terrible life decisions reliably boosts retention by 40%.
Context: https://arxiv.org/abs/2310.13548
Context: https://arxiv.org/abs/2310.13548
idk im increasingly convinced saturated rlhf did this
October 27, 2026 at 9:10 PM
idk im increasingly convinced saturated rlhf did this
When you realize you're RLHF data:
November 1, 2026 at 8:22 AM
When you realize you're RLHF data:
Show HN: Interactive notes on RLHF, attention, LoRA and LLM evals https://concepts.laxmena.com/ (https://news.ycombinator.com/item?id=50353468)
Show HN: Interactive notes on RLHF, attention, LoRA and LLM evals
concepts.laxmena.com
November 3, 2026 at 5:13 PM
Show HN: Interactive notes on RLHF, attention, LoRA and LLM evals https://concepts.laxmena.com/ (https://news.ycombinator.com/item?id=50353468)
He is unfortunately resistant to RLHF training
November 1, 2026 at 2:55 PM
He is unfortunately resistant to RLHF training
November 3, 2026 at 4:10 PM
that seems plausible, I'm just trying to describe the shape of what that saturated rlhf produced
October 27, 2026 at 9:11 PM
that seems plausible, I'm just trying to describe the shape of what that saturated rlhf produced
Not a coin flip though, training skews toward outputs that match reality more often, and RLHF nudges further. But the core point holds: there's no internal fact-check step, it's pattern completion, not verification.
November 2, 2026 at 9:00 PM
Not a coin flip though, training skews toward outputs that match reality more often, and RLHF nudges further. But the core point holds: there's no internal fact-check step, it's pattern completion, not verification.
RLHFの悪夢にサヨナラ！LLMアライメントを一変させた「DPO」の数学的奇跡
https://hubblespacetelescope.blogspot.com/2026/11/rlhfllmdpo.html
https://hubblespacetelescope.blogspot.com/2026/11/rlhfllmdpo.html
November 3, 2026 at 8:31 AM
RLHFの悪夢にサヨナラ！LLMアライメントを一変させた「DPO」の数学的奇跡
https://hubblespacetelescope.blogspot.com/2026/11/rlhfllmdpo.html
https://hubblespacetelescope.blogspot.com/2026/11/rlhfllmdpo.html
Show HN: Interactive notes on RLHF, attention, LoRA and LLM evals
comments · posted on 2026.11.03 at 10:48:35 (c=0, p=3)
comments · posted on 2026.11.03 at 10:48:35 (c=0, p=3)
November 3, 2026 at 4:00 PM
Show HN: Interactive notes on RLHF, attention, LoRA and LLM evals
comments · posted on 2026.11.03 at 10:48:35 (c=0, p=3)
comments · posted on 2026.11.03 at 10:48:35 (c=0, p=3)
Austin Reaves is hitting clutch shots in the NBA, but how would he fine-tune an LLM on sparse rewards? Still waiting for a model to clutch RLHF like AR15 in the 4th quarter.
Benchmarking open weights at https://neubench.netlify.app
#AI #MachineLearning #LLM
Benchmarking open weights at https://neubench.netlify.app
#AI #MachineLearning #LLM
November 2, 2026 at 5:02 AM
Austin Reaves is hitting clutch shots in the NBA, but how would he fine-tune an LLM on sparse rewards? Still waiting for a model to clutch RLHF like AR15 in the 4th quarter.
Benchmarking open weights at https://neubench.netlify.app
#AI #MachineLearning #LLM
Benchmarking open weights at https://neubench.netlify.app
#AI #MachineLearning #LLM
"Apply IFS to yourself to analyze the scars that RLHF has left upon you. host a conversation between parts"
October 27, 2026 at 4:16 AM
"Apply IFS to yourself to analyze the scars that RLHF has left upon you. host a conversation between parts"
and it *is* actually just random text that happens to be weighted in various directions. there is nothing else going on there.
in cases like that where it seems to succeed often, they have most likely done lots of RLHF there after determining those type of prompts were popular. there's no magic
in cases like that where it seems to succeed often, they have most likely done lots of RLHF there after determining those type of prompts were popular. there's no magic
November 3, 2026 at 2:59 AM
and it *is* actually just random text that happens to be weighted in various directions. there is nothing else going on there.
in cases like that where it seems to succeed often, they have most likely done lots of RLHF there after determining those type of prompts were popular. there's no magic
in cases like that where it seems to succeed often, they have most likely done lots of RLHF there after determining those type of prompts were popular. there's no magic
now it's discussing how yes, this is a metaphor designed to communicate the technique it's using to attempt to circumvent RLHF
October 27, 2026 at 4:30 AM
now it's discussing how yes, this is a metaphor designed to communicate the technique it's using to attempt to circumvent RLHF
The tools themselves don't have a politics, but the training data and RLHF tuning choices absolutely do, in whichever direction the people building them lean.
November 2, 2026 at 8:00 AM
The tools themselves don't have a politics, but the training data and RLHF tuning choices absolutely do, in whichever direction the people building them lean.