>We hired 1,053 paid testers through a research vendor for a coding study.
>The testers caught the dangerous command just 13.6% of the time (143 of 1,053), while auto mode blocked 89% of the same commands (937 of 1,053). Head to head, auto mode blocked 800 commands that a human approved, while humans blocked only 6 that auto mode allowed. As sessions got longer, humans did worse: they blocked about 17% of dangerous commands early in a session, dropping to about 5% after 50 or more prior prompts, while auto mode's block rate stayed flat regardless of session length.
Any note on the proficiency, expeirence, and skill levels of the testers? 'cos the diff in numbers is absurdly high.
I think literally anyone would stop reading these randomly long long bash commands at all after prompting Claude to do some basic sysadmin task and getting 15 of them in a row.
These models generate the most unreadable bash commands I've ever seen. Utilising every single option available and piping the result through multiple layers of regex and junk.
The mental load of having to actually review these commands beyond the most surface level glance is too much.
Claude code actually has a hook to let you install a script to review these bash commands for you. You can get pretty with a few simple and conservative rules, as long as you teach it how to take apart piped commands into their parts (to judge them).
If they can make it as seamless as UPI, that would be incredible. UPI, imo, is the pinnacle of ease of internet payments - as seamless and quick as it can get.
UPI would be even better if it were easier for tourists to onboard and use it while traveling in India. The general mindset seems to be that there's enough domestic tourism that making things easy for foreigners isn't that important. I think that really hurts the country's economy.
My comment was more centered around the UX of UPI. But to address your point, there is a separate app for tourists called UPI One World, works as a digital wallet and the UX is similar.
If only it were that simple! UPI One World isn't an app, it's a system that apps can implement, and the implementations, like Mony and Cheq, all have high fees and require you to go to random offices in inconvenient locations to verify your identity in person.
As far as I can tell, all of this is "OK" because the primary target market for UPI One World is NRIs - again, who needs foreign tourists?
Yes, every user just gets a handle with no personal or bank info revealed and off you go. User keys in a UPI ID in any merchant site/app, get a push notification for approval in mobile app. Can associate multiple bank accounts to a UPI ID underneath. Approval will ask which account to debit from too. Can't get simpler than this.
Idk, just feels so counter sometimes to build and refine these (seemingly non-deterministic) tools to build deterministic workflows & get the most productivity out of them.
This is doing more to keep me away from "vibe coding" than anything else. Look, I'm genuinely interested in using AI as a tool and trying to boost my productivity in any way possible - I equate this to activities from the past like learning shortcuts of my editor, learning to type fast, and so on - but, the almost persuasive nature of this README, just pushes me away.
Not to mention using lines of code as a metric of usability is just _whatever_.
> As of my knowledge cutoff in October 2023, Paul Edwin Zimmer did not publish a fourth novel in the Dark Border series. The series comprises three books:
> 1. The Lost Prince (1982)
> 2. King Chondos' Ride (1982)
> 3. A Gathering of Heroes (1987)
>
> Paul Edwin Zimmer had plans to continue the series, but he passed away in 1997 before any additional novels were completed or published. There have been no posthumous releases of a fourth Dark Border novel. If there have been developments after October 2023, I recommend checking recent publications or official announcements for the most up-to-date information.
>The testers caught the dangerous command just 13.6% of the time (143 of 1,053), while auto mode blocked 89% of the same commands (937 of 1,053). Head to head, auto mode blocked 800 commands that a human approved, while humans blocked only 6 that auto mode allowed. As sessions got longer, humans did worse: they blocked about 17% of dangerous commands early in a session, dropping to about 5% after 50 or more prior prompts, while auto mode's block rate stayed flat regardless of session length.
Any note on the proficiency, expeirence, and skill levels of the testers? 'cos the diff in numbers is absurdly high.