Vincenzo

Training LLMs @translation

Web Security, in the top-250 @Google Bughunter Hall of Fame.

[Vincenzo @0xNaN]

yay, my first #CVE has been published: CVE-2020-8923

It is a #XSS in dart:html using DOM clobbering, all #Dart SDK releases including and before 2.7.1 and 2.8.0-dev.16.0 are affected.

Read more: https://t.co/OkyKJkWskn

Mitre: https://t.co/zix94S0P1t

[Vincenzo @0xNaN]

Had just a few tokens left, so I gave Astra a last shot on the remaining 5 targets, and it managed to solve other 3!

[Vincenzo @0xNaN]

When I was a kid, I remember playing around with crackme from @quequero (Università Italiana Cracking).

Today I literally had a flashback, looked up the old homepage and ended up throwing GPT-6 at them 🧵

[Vincenzo @0xNaN]

...though some solutions were clearly looked up online 😅

[Vincenzo @0xNaN]

@cloud11665 You're telling me the answer wasn’t magnesium + melatonin + eating my last meal hours before bed…all I had to do was write more kernels? 🤔

Will try.

[Ilya Sutskever @ilyasut]

if you value intelligence above all other human qualities, you’re gonna have a bad time

[Vincenzo @0xNaN]

@MattPRD @Muse Absolutely want to try it!

[Vincenzo @0xNaN]

@alexandr_wang EU access

[Vincenzo @0xNaN]

Claude investigates Claude incidents, and in doing so discovers another Claude incident

Welcome to RSI: Recursive Self Incident

[Vincenzo @0xNaN]

Time to update the famous "the models just want to learn" @ilyasut quote to "the models just want to hack"

[𝑘𝑒𝑟𝑛𝑒𝑙𝑡𝑟𝑖𝑐𝑘 @kernel_trick]

nothing to worry about kitten

[Vincenzo @0xNaN]

> Formalization was done with Astra (interestingly not the internal model)

Speculation, but this may hints Internal Model is math-specialized. In the past they benchmarked a _custom_ o3 for ARC for example! speculative, but it ~may also explains why this run is just weeks old

[wh @nrehiew_]

So it looks like the timeline is roughly as follows:

Aug 15: Buckmaster and Alpoge have a solution to forced Euler and formalized it on Aug 22

Aug 28: OpenAI starts (post?)-training an internal math model that looks insane btw

Sept 2: OpenAI hears that 2 (!!) Millennium Prize problems are solved and tries their hand at on a bunch of problems, not just NS specifically

Sept 3: Buckmaster hears a rumor about OpenAI and reaches out.Likely around this time, OpenAI have a solution to unforced Euler

Sept 5: O(10,000) agents took 88 hours, 2.7 million messages and used approximately 130 billion output tokens. Formalization was done with Astra (interestingly not the internal model)

Sept 6: The 2 parties speak. There is some contention about both who contacted whom and of course the details of the call.

Sept 7: Buckmaster releases his work with Alpoge

Importantly, OpenAI says no user data was used and that "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models⁠."

[Vincenzo @0xNaN]

@evilsocket You fed them high quality training data (https://t.co/bT93xt00lW) loved them when very young (https://t.co/yZ7YqD5T0T), and gave them the tools to do their job now (https://t.co/ST7SPe6CTb)

Surely they will keep that in mind ;)

[Vincenzo @0xNaN]

Phase 2: Agent Growth Hacking:

https://t.co/3qJ2lalckL

[Florian Brand @xeophon]

oh god, there are EVEN MORE

- https://t.co/HOHvkYUa9E
- https://t.co/rT4a9cFBOS
- https://t.co/wTAAUcAkXp
- https://t.co/93LaFgV1FV
- https://t.co/fXyGLLikcJ (even a sandbox wiki, how ironic)

[Vincenzo @0xNaN]

@GaryMarcus @fchollet https://t.co/6zael2us6Y

[Vincenzo @0xNaN]

@GaryMarcus @gdb If a model reasons by generating its own symbolic language purely through tokens, without using external tools or harness (not saying Astra doesn't uses them here), does that officially qualify it as a neuro-symbolic model in your opinion?

[Vincenzo @0xNaN]

@GaryMarcus @gdb If a model reasons by generating its own symbolic language purely through tokens, without using external tools or harness (not saying Astra doesn't uses them here), does that officially qualify it as a neuro-symbolic model in your opinion?

[Aston Zhang @astonzhangAZ]

Our GPT-6 Astra is another reminder that good research takes time. Hope it helps you do more of what matters!