HenryNdubuaku

5 ms·

1. 61.▲
   by HenryNdubuaku 2mo ago

   Technically, you could do that, but we trained this one from the ground up!
2. 62.▲
   by HenryNdubuaku 2mo ago

   It does better with clearer tool description, but we are taking note of these complaints for future improvements.
3. 63.▲
   by HenryNdubuaku 2mo ago

   Try the following tool description: "Calculate the sum of two numbers. Use for any arithmetic or math question." instead of "Add two numbers." Let me know how it goes, thanks!
4. 64.▲
   by HenryNdubuaku 2mo ago

   around +60% confidence threshold is cool from experiments, the problem is that you gotta test on your own workload, no existing benchmark could honestly paint the full picture, so we exposed the confidence threshold for everyone.
5. 65.▲
   by HenryNdubuaku 2mo ago

   Thanks, noted!
6. 66.▲
   by HenryNdubuaku 2mo ago

   So, its not a general language model, focused on tool call strictly for tiny edge-devices. There are solutions everywhere for high-capacity devices, Needle is for sub-$200 devices.
7. 67.▲
   by HenryNdubuaku 2mo ago

   thanks!
8. 68.▲
   by HenryNdubuaku 2mo ago

   Thanks, how can we help?
9. 69.▲
   by HenryNdubuaku 2mo ago

   This is exactly why the confidence feature was introduced, the model knows when its wrong, we could hide that part and return a placeholder "sorry I only do function calls", would that be better or you prefer to see everything?
10. 70.▲
    by HenryNdubuaku 2mo ago

    Yes, we ablated Engrams rigorously and found that it returned world knowledge like FFN without without compute expenditure.
11. 71.▲
    by HenryNdubuaku 2mo ago

    Random sentence is not a function call.
12. 72.▲
    by HenryNdubuaku 2mo ago

    Users often stack a transcription model on top to get the voice prompt, then decode to actions. Think of Alexa and Siri.
13. 73.▲
    by HenryNdubuaku 2mo ago

    Thanks, give it a splin!
14. 74.▲
    Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots(cactuscompute.com)

    537 points by HenryNdubuaku 2mo ago|185 comments
15. 75.▲
    by HenryNdubuaku 3mo ago

    This is cool!
16. 76.▲
    by HenryNdubuaku 3mo ago

    haha, interesting analogy
17. 77.▲
    by HenryNdubuaku 3mo ago

    1. Thanks for this! 2. We combine both on Cactus; rolling entropy method helps with per-token signal for early exit, while the probe runs on complete output. 3. Users hate this in production, drains battery on edge devices, but ultimately p
18. 78.▲
    by HenryNdubuaku 3mo ago

    Smaller models in production can produce quite interesting outputs.
19. 79.▲
    by HenryNdubuaku 3mo ago

    YOu are thinking along the right direction, we are going deeper into the signals.
20. 80.▲
    by HenryNdubuaku 3mo ago

    no because we freeze the main weights and only train our added weights.
21. 81.▲
    by HenryNdubuaku 3mo ago

    We are currently working on larger models atm! your suggestion is cool btw
22. 82.▲
    by HenryNdubuaku 3mo ago

    Thats an interesting outlook, loosely similar.
23. 83.▲
    by HenryNdubuaku 3mo ago

    anytime!
24. 84.▲
    by HenryNdubuaku 3mo ago

    Correct, the study is verbose, we will compile into a neat shareable report and publish once we solve the pending caveats. Interesting username btw haha.
25. 85.▲
    Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong(github.com)

    191 points by HenryNdubuaku 3mo ago|44 comments
26. 86.▲
    by HenryNdubuaku 5mo ago

    Interesting, did you add a commit and push tool?
27. 87.▲
    by HenryNdubuaku 5mo ago

    Thanks, try needle and let us know!
28. 88.▲
    by HenryNdubuaku 5mo ago

    We have enterprise users in production with similar workflows actually.
29. 89.▲
    by HenryNdubuaku 5mo ago

    You can actually quickly test with the playground
30. 90.▲
    by HenryNdubuaku 5mo ago

    Haha, great find, this is still an experimental checkpoint, much work is going into this now. These feedback help us improve, thanks!