On 30 Apr 2026 03:29, Peter Taraba wrote:
imagine you are president of USA and you have to make decision. knowing answer is 50.000001% (decision is yes) is better than 50% (no clue if decision is yes or no).
50.000001% seems pretty much the same as 50% to me, no matter how important the decision is. What I'd be interested in is a deeper analysis, and a more capable (and not necessarily more precise) model could help with that. Although, as a president, I probably wouldn't rely on AI too much.
*Sent:* Wednesday, April 29, 2026 at 5:26 PM *From:* "Andrey Semashev via Boost" <boost@lists.boost.org> *To:* "Boost developers' mailing list" <boost@lists.boost.org> *Cc:* "Andrey Semashev" <andrey.semashev@gmail.com> *Subject:* [boost] Re: Seeking Endorsement: Boost.int128 On 30 Apr 2026 02:11, Peter Taraba via Boost wrote:
Use? Imagine you have really hard decision and your models tell you 50% result... in that case you need higher precision so you know if you should do it or not. Some decisions are really hard...
Would it make a difference if a more precise model gave a 50.1% vs. 49.9% outcome? Not saying higher precision is not useful, just that the application in AI models seems questionable. My understanding is that having more parameters is generally more preferable to precision/quantization in terms of the model quality.