RT @DimitrisPapail: We should be seriously asking, how a 1.5B model that can't answer basic questions can also be that good at competition…

Original thread:

Original thread

@mblondel_ml @andrew_n_carr +1, but if you torture rejection sampling a bit, you set temperature tending to zero and approximate the scaling factor from the proposed samples, it ends up giving best-of-N (RSO paper makes this point)

RT @andre_t_martins: Good to see @EU_Commission promoting OS LLMs in Europe. However (1) "OpenEuroLLM" is appropriating a name (#EuroLLM) w…

RT @LChoshen: Not released yet, but @karpathy leaked our gym like environment plus model competition...

@dearmadisonblue You mean distilling an existing diffusion into visual CoT, or training one from scratch. The latter seems the one that's hard. You will end up with a VAE anyway. There were some models like this such as DRAW.

RT @Yikang_Shen: It's good to see Deepseek v3 draw everyone's attention to reducing the training cost of LLM. Over the last two years, we…

RT @seirasto: 🌟New Benchmark! 🌟 Do you work on RAG? Are you interested in Multi-Turn conversations? Very excited to share the new MTRAG be…

@ch402 ☝️ So TLDR, IMO this all feels organic and tech/market driven, rather than a conscious change of style. These forces may last until AGI or maybe we hit a serious winter and we return to the old ways of more exploration less exploitation.

@dirk_hovy Did not know you were there already! Found you 🙂

RT @ZachWeiner: Bluesky: Mastodon: Better yet, leave all social media and get comics and…

@creatorscan Thanks!

Even onion's owner is in bsky!

@Teknium1 ☝️That does not mean that the model never saw human CoTs. It probably saw a huge amount of high quality ones, because the initial policy must be some GPT. But, ofc, if you do some STaR-like search to that model, you are going to get better ones.