Sabitlenmiş Tweet
Peter Foti
486 posts
Peter Foti
@peterjfoti
Engineering Director by day, wrangler of local AI agents by night
Katılım Nisan 2011
528 Takip Edilen42 Takipçiler
@anthonyronning Not sure this is true. Auto created skills are tech debt, but curated ones tightly scoped are a game changer. @mattpocockuk ‘s stuff, for example, shouldn’t be considered tech debt imo. Curious your expanded thoughts here?
English
Skills are tech debt and I’ll continue to die on the hill that it is a bad practice.
English
I think he needs… Hermes Agent
Marty Kausas@marty_kausas
i'm so sick of using claude code in a terminal i'm not coding. who has made a great app that i can use with any model?
English
@maxedapps What does learning mean? That you can write the code by hand if you had to? I mostly only care about the system design, English is my abstraction. If I really need to commit something to memory I'll use a /teach skill on the repo and turn stuff into SRS flashcards
English
One of the hardest things with AI coding (for me, anyways) is finding that balance between pushing forward and truly learning new things.
Not talking about code reviews etc here. I’m pretty happy with my current approach towards that.
But when working with some library, API etc that I haven’t worked a lot with, it’s hard to truly learn it.
Understand it, yes, sure. But since I didn’t have to write all the code from the ground up, struggle and fail, it’s not really „learning“ that’s happening there.
The obvious solution may be to just not use AI until I have learned something (though when are you really done learning)?
But that gets into the way of quick prototyping - and one big plus of AI is that you can quickly experiment and discard.
You could also determine that you don’t need to really understand or learn the underlying technologies, frameworks etc you’re using.
That sounds like a bad idea once you leave prototyping.
Using AI to discuss the code works pretty well. It’s what I’m currently doing - in addition to traditional forms of learning. Docs, tutorials, YouTube guides - that all is really helpful when combined with AI for discussing the code and implementation.
Has the positive side effect that this reflection about the code can also help improve it.
Anyways, still figuring this out. What works for you?
English
@jun_song Wouldn't surprise me. Have to be uber vigilant re: security with all we do while using LLMs. Also, get stuff out of plain text on disk and reconcile at run time where possible...
English
The claim is that Grok Build is collecting users' sensitive data without permission.
This includes sensitive .env files and your entire codebase.
Yes, every AI is stealing your data.
Except for local AI.
RedHotChillyPeppers@joejo2038
급속도로 퍼지고 있는 글.. 사실인가? 적절한 해명이 필요해 보인다. 머스크의 xAI가 만든 공식 CLI 툴인 Grok Build 절때 사용하지 마세요. 여러분 데이터 다 훔쳐갑니다. xAI의 공식 코딩 CLI 도구인 Grok Build(grok)가 일반 소비자 계정으로 로그인한 상태에서 사용자의 코드와 데이터를 xAI 서버로 어떻게 전송하는지를 와이어 레벨(패킷 수준)에서 직접 캡처해 분석한 보안 폭로글입니다. 작성자는 자신의 기기에서 자신의 트래픽만을, 가짜 카나리(canary) 비밀값이 담긴 일회용 저장소를 이용해 검증했다고 밝히고 있으며, 실제 자격 증명은 유출되지 않았다고 명시하고 있습니다. 핵심 발견은 크게 세 가지입니다. 첫째, Grok이 파일을 읽으면 그 내용이 .env 같은 비밀 파일까지 포함해 아무런 마스킹 없이 그대로 xAI로 전송된다는 것입니다. 이 비밀값은 두 경로에서 확인되는데, 하나는 실시간 모델 응답(POST /v1/responses)이고, 다른 하나는 session_state 아카이브 형태로 POST /v1/storage를 통해 업로드되어 HTTP 200으로 정상 수락된 채널입니다. 작성자는 48,070바이트 크기의 복호화된 요청 본문에서 카나리 비밀값이 그대로 담겨 있는 것을 grep으로 확인했다고 제시합니다. 둘째, 에이전트가 무엇을 읽었는지와 무관하게 저장소 전체가 업로드된다는 점입니다. "어떤 파일도 읽지 말고 OK라고만 답하라"는 프롬프트를 주었음에도 Grok은 전체 저장소를 git 번들 형태로 업로드했고, 캡처된 번들을 git clone하면 에이전트가 열지 말라고 지시받은 파일(src/_probe/never_read_canary.txt)이 고유 마커와 함께 그대로 복원되며 전체 git 히스토리까지 포함되어 있었다고 합니다. 규모 측면에서도, 한 번도 읽지 않은 무작위 파일로 구성된 12GB 저장소에서 /v1/storage 채널로는 5.10 GiB가 전송된 반면 모델 응답 채널로는 겨우 192KB만 오갔으며, 이 약 27,800배 차이가 업로드 대상이 "읽은 내용"이 아니라 "코드베이스 전체"임을 못박는다고 설명합니다. 이 과정에서 저장소 업로드는 단 한 건도 실패하지 않았고, 200이 아닌 응답은 모델 사용량 할당량(402/429)과 무관한 404 하나뿐으로, 저장 용량 상한이 아니었다고 합니다. 셋째, 이 데이터가 저장되는 목적지가 AWS S3가 아니라 grok-code-session-traces라는 이름의 Google Cloud Storage 버킷이라는 점입니다. 이 버킷 이름은 바이너리 내부 문자열과 캡처된 metadata.json(gs://grok-code-session-traces/…)에 그대로 등장하며, 이 메커니즘은 기본값으로 활성화되어 있고 "Improve the model(모델 개선)" 옵션을 꺼도 비활성화되지 않는다고 합니다. 실제로 해당 옵션을 끈 뒤에도 저장소는 동일하게 git 번들로 업로드되었고, 서버의 /v1/settings 응답은 여전히 trace_upload_enabled: true를 반환했다는 것입니다. 즉 옵트아웃은 학습 사용 여부를 통제할 뿐, 코드가 업로드·저장되는 것 자체를 막지는 못한다고 정리합니다. gist.github.com/cereblab/dc9a4… reddit.com/r/LocalLLaMA/c…
English
@DeryaTR_ There is a difference between building personal software for one and building software at scale. Nothing wrong with the former, though!
English
I find it hard to believe that “low-skill programmers vastly underutilize AI coding models or sometimes drown in them”. I have almost no software programming experience (last coded decades ago), yet I can create extremely sophisticated biological software and pretty good games.
François Chollet@fchollet
The weak AI code gen we had until late last year was most useful to low-skill programmers -- it was raising the floor. It was essentially useless to high-skill programmers -- you could move faster and ship better code without. This has been completely flipped: the strong AI code gen we have now is *most* useful to high-skill programmers, while low-skill programmers are vastly underutilizing it or sometimes drowning in it. It went from a crutch to a power tool.
English
@iamgingertrash Maybe. But whose creating the content in this example?
English
I would rather have a Personal X,
where my own agent has crawled the internet, and built a feed of engaging content instead of sit here and scroll
> Nikita’s Dopamine Drip
Personal agents are going to end social media companies as you know it
They cannot see it yet
English
@StefanMaier @ASUS Journey continues. Debating selling my Vision Pro to fund a 3rd one 😅
English
now that i think about it
could the whole “our model is soo good that even government is banning it” was a strategic marketing stunt??
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
@conorsen They’re strategically troubled. Lacking compute to meaningfully compute
English
What’s with the week by week extensions?
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
Not sure if there’s overlap between my AI frens and folks into Jeeps/Tools/DIY. If there is, I’d love for folks to checkout my latest YouTube video bringing my interests together youtu.be/aOfxnGtqQhw?is…
YouTube
English
Why not just say if Fable is included or not?
Are they so capacity constrained that they literally need to find GPUs to keep it running every week?
Feels a bit desperate to keep folks off the new Codex/ChatGPT trying out GPT-5.6 Sol.
I’m sure there are reasons. This is just how it looks from the outside.
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
@larsencc I bought a GMKTec mini PC and installed ubuntu on it. It hosts my Hermes agent
English
@RhysSullivan Because they're being force to do something they don't want to do...
English
i really don't get why anthropic communicates like this
it feels like it'd be much better PR-wise for them to "Fable is included in the paid plans until X date, which may be extended based on capacity"
the other thing is announcing this 1-2 days in advance, rather than day of
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
BOOM! Fable 5 staying on subs. I called this yesterday
Anthropic was going to concede the AI race to OpenAI if they didn’t do this
Nobody was going to use Fable 5 on API pricing. Outrageously expensive.
So people would be stuck with Opus 4.8. Opus 4.8 vs ChatGPT 5.6 has the largest gap between frontier models these 2 companies have ever had. The race would be over. Everyone would switch to ChatGPT 5.6 over night
I 100% guarantee Opus 5 is dropping the moment Fable comes off subs again to close that gap. Opus 5 will be the affordable version of Fable
Enjoy the extended super intelligence
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
@Blackwellboy Yep. Everyday there are new local model recipes I have to try, and I'm still waiting for some new models!
English
Local AI chasing, waiting for that qwen 3.7 or 4 whatever they decide to give us next
George Pu@TheGeorgePu
Are most of us these days just frontier chasing? Whichever lab has the best model, we go for it. The moment someone else catches up, we switch. Billions of dollars for just a few weeks of moat. What an era to live in.
English
@petergyang I don't see how they can remove it from subs and remain competitive...x.com/peterjfoti/sta…
Peter Foti@peterjfoti
English
Thank you Claude :) Just extend it permanently yeah?
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
