Thread

1. firasd · 1785954252 · focus · HN ↗

   Who knew the real revenue unlock wouldn’t be based on how much paranoid red-teaming the model underwent to resist users jailbreaking its ‘alignment’, and instead more on whether the model is post-trained to use ‘sed’ and ‘git’? Poor Gemini

   Scientist-heavy orgs that want to solve everything in token space may overtook tool use; meanwhile Anthropic has been super focused on MCP, Claude Code etc for over a year

   1. thorncorona · 1785954296 · focus · HN ↗

      Gemini isn't as heavily aligned as OAI / Ant models ..?

      1. firasd · 1785954490 · focus · HN ↗

         If you just talk to it over API (no web search) the Gemini models are extremely resistant to thinking the user may be living in a universe outside their training data. Try to discuss any news etc and they assume it’s fake or fiction
      2. croemer · 1785962098 · focus · HN ↗

         Yes, Gemini let's me do SARS-CoV-2 evolution research (perfectly safe, should never be blocked but is impossible with OAI/Ant)
   2. raincole · 1785954511 · focus · HN ↗

      Out of the three main US AI companies' models, Gemini is obviously the less aligned (read: censored). So I really don't know what you're talking about.

      1. Der_Einzige · 1786004526 · focus · HN ↗

         Shhhhhhhhhh!!!!!

         The more people find out that gemini 3.1 pro on the google AI playground or via API is de-facto uncensored, the more likely google will put up actually working guardrails and end the fun for everyone!

   3. biophysboy · 1785958493 · focus · HN ↗

      What do you mean by scientist-heavy orgs solving everything in token space? I feel like they use them to make or run tools almost exclusively.

      1. firasd · 1785964616 · focus · HN ↗

         I'm saying AI researchers have a bias towards thinking what needs to happen is prompt -> [crunching tokens] -> response rather than prompt -> [orchestrates 5 tools] -> response

         In other words 'just add a calculator tool' is not as sexy research-wise as making the model accurately eyeball arithmetic in its chain of thought. Maybe I'm wrong but that seems to be the case

Unofficial Hacker News client; not affiliated with Y Combinator.