Sonnet 3.7 vs GPT 4.5
4.5 is huge but the "big model smell" is elusive. The RLHF holds it back, making it boring and predictable. Sonnet is less creative with words but more steerable. I'd befriend Sonnet and avoid GPT IRL. Unless OAI overhauls their RLHF, even a 100T model wouldn't change this.
4.5 is huge but the "big model smell" is elusive. The RLHF holds it back, making it boring and predictable. Sonnet is less creative with words but more steerable. I'd befriend Sonnet and avoid GPT IRL. Unless OAI overhauls their RLHF, even a 100T model wouldn't change this.
March 3, 2025 at 4:33 PM
Sonnet 3.7 vs GPT 4.5
4.5 is huge but the "big model smell" is elusive. The RLHF holds it back, making it boring and predictable. Sonnet is less creative with words but more steerable. I'd befriend Sonnet and avoid GPT IRL. Unless OAI overhauls their RLHF, even a 100T model wouldn't change this.
4.5 is huge but the "big model smell" is elusive. The RLHF holds it back, making it boring and predictable. Sonnet is less creative with words but more steerable. I'd befriend Sonnet and avoid GPT IRL. Unless OAI overhauls their RLHF, even a 100T model wouldn't change this.
Yoo LLMs are fun again. This moment feels just like the early llama 1 days. Thanks Deepseek 🫶
February 1, 2025 at 8:04 PM
Yoo LLMs are fun again. This moment feels just like the early llama 1 days. Thanks Deepseek 🫶
TIL smolagents is gaslighting models
January 21, 2025 at 4:31 PM
TIL smolagents is gaslighting models
Am I the only one that feels like Claude Sonnet is the only LLM that actually gets what you mean when talking to it? Like all other chatbots feel to me like text2slop machines that are useful to the economy, but usually default to outputting closest matching SEO-optimized idle talk in conversations
January 5, 2025 at 2:43 AM
Am I the only one that feels like Claude Sonnet is the only LLM that actually gets what you mean when talking to it? Like all other chatbots feel to me like text2slop machines that are useful to the economy, but usually default to outputting closest matching SEO-optimized idle talk in conversations
Dear people who write documentation: we do not need the dollar sign. We know it's shell command. Just let us double click and copy/paste the whole line.
December 11, 2024 at 9:28 PM
Dear people who write documentation: we do not need the dollar sign. We know it's shell command. Just let us double click and copy/paste the whole line.
So we have pre-training and post-training. When do we do... *training*?
November 25, 2024 at 9:16 PM
So we have pre-training and post-training. When do we do... *training*?