Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Finding Bugs (matklad.github.io)
    —discuss
  2. What's New in Oracle Solaris 11.4 SRU 95 (oracle.com)
    —discuss
  3. Pompeii test case demonstrates Gaussian Splatting's potential (gim-international.com)
    —discuss
  4. Every Package Is Installed (fzakaria.com)
    —discuss
  5. Perch: Semantic Code Linting with Jev (github.com/lakeday-org)
    —discuss
  6. Is indirect prompt injection still a big threat as models get more advanced? (realarcherl.github.io)
    1comments
  7. AI/ML Engineer (docs.google.com)
    —discuss
  8. We Built Safety into Muse (meta.ai)
    —discuss
  9. Micron Technology sucessfully demonstrates first 512GB DDR5 RDIMM module (micron.com)
    —discuss
  10. September's ozone hole biggest in 20 years, warning signs appeared months ago (rnz.co.nz)
    —discuss
  11. Collaborative Editing in Wordgard (marijnhaverbeke.nl)
    1comments
  12. Plan Mode Is Dead (aymannadeem.com)
    —discuss
  13. Power your agents: Gemini 3.8 Live with Live Avatar is now generally available (cloud.google.com)
    —discuss
  14. What I Found Out About Organ Donation Shocked Me (nytimes.com)
    3comments
  15. Scalable Subagents (github.com/ringlochid)
    —discuss
  16. We Built a City of 10k AI Researchers (lab.cloud)
    —discuss
  17. Congestion pricing funds $5M program for overnight truck deliveries in NYC (gothamist.com)
    —discuss
  18. Show HN: I built 11 Sudoku variants – killer cage generation was the hard part (iqiqgame.com)
    1comments
  19. Side-Channel Leakage from the File-Notification System (inoti.fyi)
    1comments
  20. Learning to Discover Interesting Mathematics (arxiv.org)
    —discuss
  21. Research into file-notification attacks on Linux (lwn.net)
    —discuss
  22. FDA panel endorses first-of-a-kind cancer blood test from Grail (apnews.com)
    —discuss
  23. Government considers £11 monthly UK broadband fee to fund BBC TV (ispreview.co.uk)
    1comments
  24. Show HN: Today I deployed this Durable Chat application and was impressed (netly.workers.dev)
    —discuss
  25. AICore: Exploring an adaptive control layer for heterogeneous PCs (github.com/7-hello007)
    —discuss
  26. Scaling Laws for Neural Language Models (first "scaling laws" paper from 2020) (arxiv.org)
    —discuss
  27. Bad examples can override agent skill files (codecut.ai)
    —discuss
  28. Search and Recommender Engines AI (github.com/alopatenko)
    —discuss
  29. Ask HN: Who is smarter than you that you listen to what they have to say?
    3comments
  30. Using AI to Contribute to Open Source (visidata.org)
    1comments

Study shows AI is as good as human tutoring for GRE learning gains

3 pointsby 2h agoarxiv.org
1 comments
1h agoHN ↗

Author of the paper here. We found that AI is as good as expert human tutors for immediate GRE learning gains (p=.015, n=2,383 students).

One hour with an expert GRE tutor: $75. One equivalent hour with an AI tutor: 7 cents. 918x cheaper per point of learning, with equivalent learning (p=.044).

In 5 of 7 GRE topics, the top AI tutor beat the expert human tutors on average.

Motivation is that RSI and most AI research is focused on machines improving machines. We created this suite of benchmarks and evaluations to shift investment to focus on machines improving humans. We were also just genuinely curious how much AI could improve a human's learning gain on real actual standardized exams.

All data from students and teachers is open-sourced to advance research in augmenting human capabilities with AI.

Data: https://huggingface.co/datasets/handshake-ai-research/studen... Code: https://github.com/Handshake-AI-Research/studentbench Try it: https://studentbench.org Video: https://x.com/cgnorthcutt/status/2103171858057277532/video/1

Happy to answer questions.