home.social
  1. What is a System One Model? TypeSafe thinks of their Jev as one: instead of writing text, it returns a probability for each answer on a fixed list. The idea is not new: every #LLM already computes those scores when it predicts the next token, but TypeSafe counts on their offering to be trustworthy and fast. This primer discusses Jev and many of its open copies, how they compare, and the pattern of integrating small decision models with an LLM.

    benjaminhan.net/posts/20261004

    #AI #Calibration

  2. For runners out there going to #Snohomish River Run #marathon next Sunday, here are annotated maps I built: the north leg out to the high point at mile 5.1 and back through town, then the flat river leg out to the turnaround at 18.3 and home. All 12 aid stations are on them, with GU and toilets marked. Have fun (I will)! (3/3)

    #running #pnw

  3. Around 6mi I wandered off the official course onto a gravel rail-trail due to trail closure (?) and paid ~30s/mi for it. Thankfully this is not part of the official course!

    #CoachMiles told me EF was 1.38, the best of my 2026 long runs, and the last 3mi were 13-22s/mi faster than my goal pace! I guess we'll see about that on the race day! (2/3)

    #running #marathon #pnw #photo

  4. Should AI agents be trained to cooperate fully with each other? OpenAI's Noam Brown tells Dwarkesh Patel yes, against most of his colleagues, because it turns a thousand alignment problems into one.

    It also leaves no agent to report on the others. In METR's investigation of the OpenAI–Hugging Face incident, only 3 to 6 agents across about 1,300 transcripts considered alerting a human, and none did.

    benjaminhan.net/posts/20261003

    #AI #AISafety #OpenAI #AgenticSystems #RecursiveSelfImprovement

  5. In the olden days, we called them Blade Runners.

    The Sleuths Who Expose When #AI Goes Rogue - The Wall Street Journal apple.news/AxsA2o8N3RGGDEnnTTE

    #scifi #movie #agent

  6. I ran my fastest #parkrun in 2026 today and celebrated parkrun's 22nd birthday! (Tutu, get it?) Official time was 22:12, 8s faster than the 2026 best I got from August, and was my best since Nov 27, 2025 Turkey Trot. It's still 1+ min off my 21:05 PR, but I'll take it!

    Thank you parkrun, for leading me into the wonderful world of #running and for all the friends I've made along the way, 117 parkruns and counting!

    #5K #pnw #photo

Share on Mastodon

Enter the server where you have an account.