Sabitlenmiş Tweet
Tom Holz
206 posts

Tom Holz
@tomholz
G4 challenger @tomusually - personal shenanigans https://t.co/zImytGLDeW - apps https://t.co/r8ugK3HUal - the boring stuff
San Diego, CA Katılım Ocak 2019
164 Takip Edilen83 Takipçiler

@bushidocodes It once took us 4 months to buy a network switch.
But six weeks for something like that feels reasonable.
English

@tomholz 😅 Is this fast? I feel like I’m on the verge of a breakdown from the waiting. Maybe my colleagues sense this?
English

Before @gauntletai I was saying I was a 3/10 on AI.
People thought I was underselling myself, because I tend to undersell, but in hindsight I was maybe 1.5/10.
Some people in our cohort were up there, but everyone is getting better. The weekly progress is wild.
Austen Allred@Austen
By the way, one of the most common things we hear coming into Gauntlet AI now is, “Oh I’m already a pro at using AI.” Every single time two weeks in everyone is saying, “Turns out I was barely scratching the surface.”
English

@bushidocodes In that context, security's job is to keep things secure, full stop.
It's a hard paradigm shift from the private sector where security is almost always secondary to other business objectives.
English

Personal Log: @gauntletai Week 5, Friday Night
It's 10:24 PM in Austin and I started a new partner project an hour ago.
I thought it would be boring and impossible, but now I'm hooked and I don't want to leave the office.
Currently Playing: Pentatonix
youtube.com/watch?v=q7ZlLN…

YouTube
English

Personal Log: @gauntletai Week 3, Friday Night
It's 10pm and downtown Austin is buzzing with SXSW.
Aside from the low thump of distant music, nobody in the gauntlet workspace seems to have noticed.
People are heads-down coding or talking agents and models and demos.
English

Personal Log: @gauntletai Week 4, Hello Austin
The energy here is exactly as advertised.
Overloaded and loving it.
Currently Building: a progressive web app with a low-latency computer vision pipeline.
Now Playing: Mongolian throat-singing folk metal:
youtube.com/watch?v=QkWX1q…

YouTube
English

@nakoerschner There's more coming and the impacts will be worse as agents are given more control over the real world.
English

So everyone has been clear about prompt injection in theory. I think this is the first high profile case where I’ve seen it in action.
Sash Zats@zats
> The attacker got the npm token by injecting a prompt into a GitHub issue title, which an AI triage bot read, interpreted as an instruction, and executed.
English

AI is in its “let’s put windows 95 online” era.
People are so busy chasing functionally that very few are thinking about security and liability.
I predict a several large, public incidents followed by tighter governance.
Sash Zats@zats
> The attacker got the npm token by injecting a prompt into a GitHub issue title, which an AI triage bot read, interpreted as an instruction, and executed.
English

Personal Log: @gauntletai Week 3, Final Submission
-- The Shape of Code --
On the left is LAPACK, a Fortran linear algebra package.
On the right is gnucobol, a... Gnu COBOL compiler.
Running live here:
gitnexus.smallcatlabs.com
You're looking at a graph representation of each code base rendered on GitNexus, where agents can access this graph as through it were a RAG, but one based on symbolic chunks instead of semantic ones:
github.com/abhigyanpatwar…
However, GitNexus supports neither Fortran nor COBOL.
For my week 3 @gauntletai project, I challenged myself to vibe-hacked both languages into my own fork of GitNexus:
github.com/gerwaric/GitNe…
That's what's running on Small Cat Labs right now.
And aside from all the package dependency and deployment nightmares, this was terrible fun.
And made in 3 days...


English

@bushidocodes If it's the verbosity of the keyword, would it help to replace them (and maybe variables and symbols) with non-ascii unicode characters, so the agents can speak to each other with a dense private language?
English

@bushidocodes That's a tough choice, but congrats on making the call when it was needed.
English


@ChShersh @0xleegenz C++ is currently in the middle of saving my career
English





