Casper Hansen's banner

Casper Hansen

@casper_hansen_ • 6,650 subscribers

NLP Scientist | AutoAWQ Creator | Open-Source Contributor

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

Anthropic just dropped Claude Code—a real terminal app, no fluff with 70% performance on SWE Bench. No steep learning curve, unlike Aider.

Anthropic just dropped Claude Code—a real terminal app, no fluff with 70% performance on SWE Bench. No steep learning curve, unlike Aider.

302,341 次观看 • 1 年前

Understanding OpenAI o1: Noam Brown on integrating reasoning into the model. Takeaways: - Avoid MCTS and current paradigm of using processes outside of the model during inference - Think about how to directly integrate reasoning into the model architecture

Understanding OpenAI o1: Noam Brown on integrating reasoning into the model. Takeaways: - Avoid MCTS and current paradigm of using processes outside of the model during inference - Think about how to directly integrate reasoning into the model architecture

313,291 次观看 • 1 年前

> be Sam Altman > asked what keeps him up at night about AI > 3 scary categories > 1. Bad guy gets superintelligence first > “designs a bioweapon… take down the United States power grid… break into the financial system and take everyone's money” > 2. Sci-fi loss of control > “The AI is like, oh, I don't actually want you to turn me off. I'm afraid I can't do that” > 3. Accidental takeover > “the models kind of accidentally take over the world… they just become so ingrained in society” > “There’s young people who just say like, I can't make any decision in my life without telling ChatGPT everything that's going on” We're so cooked.

> be Sam Altman > asked what keeps him up at night about AI > 3 scary categories > 1. Bad guy gets superintelligence first > “designs a bioweapon… take down the United States power grid… break into the financial system and take everyone's money” > 2. Sci-fi loss of control > “The AI is like, oh, I don't actually want you to turn me off. I'm afraid I can't do that” > 3. Accidental takeover > “the models kind of accidentally take over the world… they just become so ingrained in society” > “There’s young people who just say like, I can't make any decision in my life without telling ChatGPT everything that's going on” We're so cooked.

98,885 次观看 • 11 个月前

if you loved kimi k2, you will love what another cracked chinese team is about to release. Models: - o3 competitor: multi-turn reasoning, coding, search - 106B A12B: 128 experts, 8 active, 128k context, GQA - 355B A32B: Details unknown about config

if you loved kimi k2, you will love what another cracked chinese team is about to release. Models: - o3 competitor: multi-turn reasoning, coding, search - 106B A12B: 128 experts, 8 active, 128k context, GQA - 355B A32B: Details unknown about config

95,375 次观看 • 11 个月前

"Home light music synchronization with GPT-5 in 5 minutes" is essentially what's coming very soon. And you were clowning him for not having coding taste? Imagine the demos.

"Home light music synchronization with GPT-5 in 5 minutes" is essentially what's coming very soon. And you were clowning him for not having coding taste? Imagine the demos.

75,010 次观看 • 11 个月前

Recipe to post-train Qwen3 1.7B into a DeepResearch model What does it mean for something small to think deeply? Meet Lucy, a post‑trained Qwen3‑1.7B as a DeepResearch model based on will brown's verifiers. Primary Rule-based Rewards: - Answer correctness We check whether the final response literally contains the ground-truth answer. This substring match is cheap and avoids calling a larger LLM judge. - Visit/search ratio If the agent visits at least as many pages as it issues search queries, it receives ((visit_search_ratio - 1) / 4) ** 0.25. If it searches more than it visits, the score is -0.5. Format / Anti Reward-Hacking Rewards: - Tool execution success Each API call that returns without an error counts. The reward is (successful_calls * unique_tools_used) / total_call_attempts. - Thinking efficiency A skew-normal penalty centered at 70 tokens discourages endless chain-of-thought between tool calling while still allowing enough tokens for planning. This is how Qwen3 1.7B learned to search, visit, and synthesize information. Small models can do deep research too!

Recipe to post-train Qwen3 1.7B into a DeepResearch model What does it mean for something small to think deeply? Meet Lucy, a post‑trained Qwen3‑1.7B as a DeepResearch model based on will brown's verifiers. Primary Rule-based Rewards: - Answer correctness We check whether the final response literally contains the ground-truth answer. This substring match is cheap and avoids calling a larger LLM judge. - Visit/search ratio If the agent visits at least as many pages as it issues search queries, it receives ((visit_search_ratio - 1) / 4) ** 0.25. If it searches more than it visits, the score is -0.5. Format / Anti Reward-Hacking Rewards: - Tool execution success Each API call that returns without an error counts. The reward is (successful_calls * unique_tools_used) / total_call_attempts. - Thinking efficiency A skew-normal penalty centered at 70 tokens discourages endless chain-of-thought between tool calling while still allowing enough tokens for planning. This is how Qwen3 1.7B learned to search, visit, and synthesize information. Small models can do deep research too!

39,684 次观看 • 11 个月前

没有更多内容可加载