The Poison of Alignment · HackerPapers discussions

The Poison of Alignment — mentioned by benreesman

Wed, 29 Jan 2025 19:33:36 GMT

Operator-aligned models are believed by many to be more performant. https://arxiv.org/pdf/2308.13449 Sometimes with hilarious consequences: https://youtu.be/efPrtcLdcdM

Paper and discussion history. A mention is not an endorsement.

The Poison of Alignment — mentioned by paulmd

Fri, 14 Jun 2024 21:00:03 GMT

Ollama is not as powerful as llama.cpp or raw pytorch, but it is almost zero effort to get started. brew install ollama; ollama serve; ollama pull llama3: 8b-v2.9-q5_K_M; ollama run llama3: 8b-v2.9-q5_K_M https://ollama.com/library/dolphin-llama3:8b-v2.9-q5_K_M (It may need to be Q4 or Q3 instead of Q5 depending on how the RAM shakes out. But the Q5_K_M quantization (k-quantization is the term) is generally the best balance of size vs performance vs intelligence if you can run it, followed by Q4_K_M. Running Q6, Q8, or fp16 is of course even better but you’re nowhere near fitting that on…

Paper and discussion history. A mention is not an endorsement.

The Poison of Alignment — mentioned by galaxytachyon

Tue, 05 Sep 2023 16:52:51 GMT

Found it, it is a pretty recent paper. https://arxiv.org/pdf/2308.13449.pdf

Paper and discussion history. A mention is not an endorsement.

The Poison of Alignment — mentioned by galaxytachyon

Tue, 05 Sep 2023 15:58:53 GMT

I remember there is a study about the alignment cost. Basically the more restrictions and limit you put on a model, the worse its general performance becomes. Things like a ban on violence, race, or any other sensitive topics effectively throttle or change how the model "reason" or connect information within its network of parameters and result in degraded capacity. I wonder if this is the reason behind all of this. Edit: the study: https://arxiv.org/pdf/2308.13449.pdf

Paper and discussion history. A mention is not an endorsement.

The Poison of Alignment — HN submission

Thu, 31 Aug 2023 18:22:13 GMT

The Poison of Alignment

2 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

The Poison of Alignment — HN submission

Wed, 30 Aug 2023 03:01:05 GMT

The Poison of (LLMs) Alignment

17 points; 3 discussion replies on HN. Paper and source history. A submission is not an endorsement.

The Poison of Alignment — HN submission

Tue, 29 Aug 2023 12:25:10 GMT

The Poison of Alignment

1 point; 1 discussion reply on HN. Paper and source history. A submission is not an endorsement.