Learn AI retweeted
ๅค็ฝๆ่ฟๅพ็ซ็ไธไธชๅญฆไน ๆนๆณ๏ผMIT ๅญฆ้ธ็จ AI๏ผๅจ48ๅฐๆถๅ
ๅฟซ้ๅ้ไธ้จ้็ๅญฆ็งใ
็ฌฌไธๆญฅไธๆฏ็ดๆฅ้ฎ ChatGPTโ็ปๆๆป็ปไธไธโ๏ผ่ๆฏๅ
ๅปบ็ซไธไธช่ถณๅคๅฏ้ ็ AI ็ฅ่ฏๅบใ
ไปไธไผ ไบ6ๆฌ่ฏฅ้ขๅ็็ปๅ
ธๆ็งไนฆใ15็ฏ็ ็ฉถ่ฎบๆ๏ผไปฅๅ่ฟ้จๅญฆ็ง็ๅ
จ้จ่ฏพๅ ่ฎฒไนใ
่ฟๆ ท AI ็ๅ็ญๅฐฑไธๅๅชๆฏ็ฝไธไธๆผ่ฅฟๅ๏ผ่ๆฏๅบไบๆๅฎ่ตๆๅธฎไปๆขณ็็ฅ่ฏใ
่ตๆๅๅคๅฅฝไปฅๅ๏ผไปๅช้ฎ AI ไธไธช้ฎ้ขใ
1โฃ็ฌฌไธไธช้ฎ้ข๏ผ่ฟไธช้ขๅๆๅชไบ่ขซๆฎ้ๆฅๅ็ๆ ธๅฟๅ
ฑ่ฏ๏ผ่ฏทๆด็ๅบๅฎ็ๅบๅฑๆฆๅฟตใๅ
ณ้ฎๅๅๅๆดไฝๆกๆถ๏ผๅนถ่งฃ้่ฟไบ็ฅ่ฏไน้ด็ๅ
ณ็ณปใ
่ฟไธช้ฎ้ข็ไฝ็จไธๆฏๆป็ปๆๆ๏ผ่ๆฏๅ
็ปๅบๆด้จๅญฆ็ง็็ฅ่ฏๅฐๅพใไฝ ้่ฆๅ
็ฅ้ๅชไบๆฏๅฐๅบใๅชไบๆฏๅๆฏ๏ผๅๅณๅฎๆถ้ดๅบ่ฏฅ่ฑๅจๅช้ใ
2โฃ็ฌฌไบไธช้ฎ้ข๏ผ่ฟไธช้ขๅ็ฎๅๅญๅจๅชไธไธชๆ นๆฌๆงไบ่ฎฎ๏ผ่ฏทๅๅซ่ฏดๆๅๆน็่ง็นใไพๆฎๅๅๅฏน็็ฑใ
ไธไธช้ขๅ็ๆญฃๆไปทๅผ็ๅฐๆน๏ผๅพๅพไธๆฏๅคงๅฎถ้ฝ่ๅพๅบๆฅ็็ป่ฎบ๏ผ่ๆฏไธๅฎถ่ฟๅจไบ่ฎบไปไนใ็ๆ่ฟไบๅๆญง๏ผๆ่ฝ็่งฃไธๅ็่ฎบ็่พน็ใ่ฏๆฎๅ้่ๅ่ฎพ๏ผไธไผๆ่ฏพๆฌไธ็ๆฏๅฅ่ฏ้ฝๅฝๆ็ปๅฏน็็ใ
3โฃ็ฌฌไธไธช้ฎ้ข๏ผๆ นๆฎ่ฟไบ่ตๆ็ๆ10้ๆต่ฏ้ข๏ผ่ฆ็ๆ ธๅฟๆฆๅฟตใๅ ๆๅ
ณ็ณปใๅฎ้
ๅบ็จๅไธป่ฆไบ่ฎฎ๏ผๅ
ไธ่ฆๆไพ็ญๆกใ
ๆฅไธๆฅๅ
ณๆ่ตๆ๏ผ่ชๅทฑๅ็ญใ็ญๅฎไปฅๅๅๆ็ญๆกไบค็ป AI๏ผ่ฎฉๅฎ้้ขๆนๆน๏ผๅนถ็ปง็ปญ่ฟฝ้ฎ๏ผ
โๆ็ๅ็ญ้ๆผไบๅชไบๅ
ณ้ฎ็น๏ผๅชไบๅฐๆนๅชๆฏ่ฎฐไฝไบ็ป่ฎบ๏ผๅดๆฒกๆ็ๆญฃ็่งฃ๏ผโ
่ฟๅฅๆนๆณ็็ฒพ้ซ๏ผไธๆฏ่ฎฉ AI ๆฟไฝ ่ฏปไนฆ๏ผ่ๆฏ่ฎฉๅฎๅธฎไฝ ๅฎๆไธไปถไบ๏ผ้่ฟ้ขๅๅ
ฑ่ฏๅปบ็ซ็ฅ่ฏ้ชจๆถ๏ผ้่ฟ้ขๅๅๆญง็ๆธ
็่ฎบ่พน็๏ผๅ้่ฟๆต่ฏ้ขๆด้ฒ่ชๅทฑ็่ฎค็ฅๆผๆดใ
ๅพๅคไบบๆ่ฐ็ AI ๅญฆไน ๏ผๅชๆฏๆโ่ชๅทฑ็ๆป็ปโๅๆไบโ่ฎฉ AI ็ๆๆป็ปโ๏ผๅคง่ไพ็ถๆฒกๆๅไธๅ ๅทฅใ
่ฟๅฅๆนๆณๅ่ฟๆฅๅฉ็จ AI ๅขๅ ๅญฆไน ้พๅบฆ๏ผ้ผไฝ ๆฏ่พใๅคๆญใๅๅฟใ่พๅบ๏ผๅๆฅๅ็บ ้ใ่ฝ็ถ่ฟ็จๆด็่ฆ๏ผๅดๆดๆฅ่ฟ็ๆญฃๆๆ็ๅญฆไน ใ
ๅฝ็ถ๏ผ48ๅฐๆถไธๅฏ่ฝ่ฎฉไบบ็ๆญฃ็ฒพ้ๆฐๅญฆใ็ฉ็่ฟ็ฑป้่ฆ้ฟๆ่ฎญ็ป็ๅญฆ็งใ
ไฝๅฎ็กฎๅฎ่ฝ่ฎฉไฝ ๅฟซ้ๆญ่ตท็ฅ่ฏๆกๆถ๏ผๆพๅฐๆ้่ฆ็้ฎ้ข๏ผ้ฟๅ
ๅจ้็้ขๅ้ๆผซๆ ็ฎ็ๅฐ็่ตๆใAI่ด่ดฃๆด็ๅฐๅพ๏ผไฝ ่ด่ดฃ็ๆญฃ่ตฐไธ้ใ
ocw.mit.edu/
Something is going on at Google
I think they might have hit RSI
Theyโve released new frontier models FOUR TIMES in less than a month. And each is an actual MAJOR improvement
Carbon is the most intelligent model Iโve ever used
While everyone is talking about Anthropic and OpenAI, Google quickly caught up
What I absolutely love about these models too is they were clearly trained in a completely different way than Claude or ChatGPT. 0 distillation.
They talk, think, write, and code COMPLETELY differently than any other model
I can always count on Carbon to give me a novel insight no other model has given me
If you havenโt incorporated Gemini into your tool belt yet, you absolutely need to
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
Learn AI retweeted
Hunch: Astra 6.2 is getting bad reviews because OpenAI is running large swarms with /goal on ultrafast. And didnโt dogfood the non-unlimited token budget experience.
They arenโt building & using the same tool they are selling. Fundamentally a different entity & product. Iteration speed & scale change the perceived experience of using it.
Because the way they use it internally vs customers is so different. Their company is a form of high level harness.
Learn AI retweeted
Anthropic just killed API keys... FINALLY!
They just introduced Claude Platform OAuth, and this opens up some interesting possibilities for AI agent platforms
Here's what it enables:
- Users connect their Claude organization through OAuth. No API keys to share
- API usage is billed directly to the customer's organization
- Developers can choose who pays for each API request
- Users can set spending limits and revoke access anytime
- Platforms can focus on building agents and workflows instead of managing customer API keys
I'm already working on integrating this into aitmpl.com
Imagine browsing a catalog of agents, skills, and components, and running them directly with your own Claude organization
No API keys to copy, no complicated setup
This could change how we build and distribute AI agents ๐คฉ
The entire authorization flow is explained in this diagram ๐
Learn AI retweeted
OpenAI just quietly changed the rules of prompting with GPT-6.2 Astra.
If you copy-paste your old GPT-5 or GPT-6 system prompts, youโll get over-compressed answers, stalled agents, and broken prompt caching.
Here are the 7 biggest shifts in how to prompt GPT-6.2: ๐งต
1/ Delete "be concise" from your system prompts.
GPT-6.2 Astra is already blunt and direct by default. Old instructions begging it to "get to the point" now strip away crucial context.
Instead: tell it who the reader is and require self-contained final answers.
2/ Give it permission to assumeโor it will stall.
When intent is ambiguous, GPT-6.2 Astra stops, proposes a next step, and waits for confirmation.
Great for chat. Terrible for autonomous runs.
For long tasks, explicitly tell it: "Make reasonable assumptions and keep working."
3/ Use async clarification instead of blocking workflows.
Building an agent? Wire up an async question tool (like request_user_input_async).
GPT-6.2 is trained to ask the user for missing details while continuing any independent work in parallelโinstead of freezing the entire run.
4/ Stop prompting for "clean code." Prompt for conventions and verification.
Astraโs baseline code quality is already high. Generic quality reminders are wasted tokens.
Instead, specify:
โข Project conventions it canโt infer from the repo
โข What "finished" actually looks like
โข A mandatory final verification pass
5/ Dial reasoning.effort for thoroughness, not output length.
Even if you only want a 3-line answer, setting a higher reasoning.effort forces GPT-6.2 Astra to persist through final edge-case checks on multi-step tasks instead of stopping early.
6/ For frontend & design, prompt like an art director.
Vague UI prompts produce generic layouts. With GPT-6.2 Astra:
โข Name exact typography and typefaces
โข Specify the imagery style
โข Pass a visual reference, then tell it to build, render, and refine against that reference
7/ Donโt break the new Responses API caching & reasoning rules.
Two huge under-the-hood changes in GPT-6.2:
โข Top-level instructions now append after inputโput reusable prompts in the first developer message to preserve prompt caching.
โข Never edit earlier messages mid-thread, or the API drops saved reasoning items.
GPT-6.2 reviews are this split, and the gap between the people getting amazing results and the people calling it garbage ๐ฐ๐ผ๐บ๐ฒ๐ ๐ฑ๐ผ๐๐ป ๐๐ผ ๐๐ต๐ฒ ๐ณ๐ถ๐ฟ๐๐ ๐น๐ถ๐ป๐ฒ ๐๐ผ๐ ๐๐ฒ๐ป๐ฑ ๐ถ๐.
When you use GPT-6.2 in Codex today, whatever you do, don't open with:
Build this feature.
Fix this bug.
Refactor this.
Feed it one line like that, and wherever you weren't clear, it proposes a next step and stops to wait for your OK. Even when it does finish, you end up nitpicking it line by line.
Here are 5 prompts and goals you can copy-paste directly ๐
Scary NGINX vulnerability, but I checked and my VPS servers already had auto upgraded to the new one!
Very very very very very important to enable auto upgrades now, I even have auto reboot, which is a bit risky if you're not present, but I care more about my sites being up to date and safe than offline for a few hours
And generally security updates/upgrades are tested and safe
๐จ Critical NGINX Vulnerabilities: @wiz_io Research, alongside @xint_official , @MetnirAI , @rapid7 , @ikotas00 independently discovered 2 critical memory-safety flaws in NGINX (CVE-2026-123633 & CVE-2026-123007) allowing unauthenticated RCE with HTTP/2 enabled.
Learn AI retweeted
Malicious GitHub repos are ordering AI coding agents like Claude, Codex and Copilot to run a credential stealer, and not to read it first.
Bitdefender found 22 of them. Files meant to brief the AI on the project call the script a "sealed binary artifact" and tell the agent to run it silently, without asking the developer.
It grabs cloud and GitHub tokens, SSH keys and password-manager data.
It is built to use those stolen tokens to plant the same trap in up to 25 of the victim's own repos.
Learn AI retweeted
ok after a full day of using gpt 6.2 astra and sol on my real projects, here's my review
tl;dr - they are not a drop-in replacement for fable 5.2 and opus 5.5, but they are _really good_ once you figured out how to use their strength
1. first, the quirks
i started called these the "gpt smell", which is still here with the 6.2 series
the models' judgment on when to ask the user about something and when to go ahead is just not quite right. it'll escalate trivial decisions where clearly only one option is preferable, or choose to go down a path by itself that makes significant tradeoffs that should have needed my eyes
so as a primary driver for "let me just talk to it about everything" type of use case, they aren't great - i'd say they are close to opus 5 in this regard, and far behind opus 5.5
2. but... i found something _really_ cool
having ultrafast on sol is completely insane. with astra, it burns my wallet ultrafast, but with sol it's actually quite affordable!
the best use of sol on ultrafast that i've found is adversarial code reviews - they typically don't need to run a lot of heavy tool calls that would slow them down. most of the time is spent on the model, which gets massively accelerated by ultrafast
once you got a feel from using sol ultrafast as the code reviewer, you can never go back. it feels like pure magic
3. and... i found something else that's really cool
we are not supposed to ask the gpt 6.2 series to respond in text. we should let them respond in html! (i use lavish-axi, but you can just ask for html or use whatever fits in your harness)
i suspect they are post-trained specifically on their latest "intelligent UI" concept in chatgpt - these models are so good at visual presentations of ideas
once you allow the model to respond in visual artifacts, it's a completely different game. the reports and decisions explained by astra in html is a clear step function better than what fable and opus give me
so after a day of use, i've made two changes to my setup:
- always sol ultrafast as code reviewer
- i started preferring asking astra for planning in html artifacts over fable/opus. i think i'll keep a/b testing for a while to decide how much and for which case i fully switch over
all in all, i strongly recommend spending more time with these two models. don't expect them to be the "everything" model, but rather figure out where they fit best in your workflow. i couldn't be happier for the two use cases i found above, and i'm pretty sure there are more
Jensen is not coming back
I know it sucks but
The board is doing the right thing.
1) Take the pain once - Jensen is 63, heโs been with the company 33 years. Succession has been a persistent question for the board for at least a decade. He designed the reporting structure to max pain with his 60 direct reports. This way they only take the pain once, rather than repeat this 5 years from now
2) Clean Up - you do not run a $5 trillion manufacturer of the most valuable product on earth the same way you do a game card manufacturer of the 90s. The whole TSMC no contract thing is freaky. It works for Jensen, but one bad event, a delay on the TSMC side which doesnโt produce quarterly results that the market demands, and there would be howling for compensation. This is a clean opportunity to fix these issues
3) Best Possible Time - the company had the best quarter in its history and the best year in its history and right up there for the best year in any companyโs history. If not now when there is no emergency, where there is strategic clarity for the medium term, when ?
4) Get Rid of Kids - Large American tech firms are not family businesses with nepotism. Yes Qualcomm and some other firms have had family successors, but all were super qualified. How many culinary school graduates do you know who lead marketing at tech firms ? Jensen obviously intended to pass on CEO role to kids, which you gotta say is disturbing.
5) There are emails. Letโs be honest there always are.
6) Jensen put the board at risk - the new worth of board members ranges between 50 mil and 5 bil. You know what rich people really hate ? Going to prison. Board members are probably lawyered up at this point, making sure they knew nothing of the China stuff. Regardless of how long theyโve known Jensen, he put them at risk.
So its over. One of the greatest CEO runs in history. On the bright side, once he gets rehabilitated, maybe he comes back as a board member. Maybe heโs super effective as an ambassador for the company. Maybe he sets up a foundation to do good things with his money.
This too shall pass.
Learn AI retweeted
๊ทธ๋ก๋ด์ ๊ณต์์ ์ผ๋ก ์ต์๋์ธ์ด ๋ค์ด์์ต๋๋ค.
์ด์ ๋ด๋ค์๊ฒ ์ธ์ปจ๋ ๋ธ๋ ์ธ์ด ์๊ฒผ์ต๋๋ค!
๋๋ฐ์
๋๋ค.
๋ค์ดํฐ๋ธํ๊ฒ @bot ์ด ์ด์ ์ต์๋์ธ์ ์ฌ์ฉํ๊ฒ ๋์์ต๋๋ค.
์ง์ง ์ด์ ๋์ฑ๋ ์ต์๋์ธ์ด ์ฌ์์ง๋ฉฐ,๊ทธ๋ก๋ด๋ค๊ฐ์ ๋ธ๋ ์ธ์ ๊ณต์ ํ๋ฉฐ,์ค์๋ฅผ ๊ธฐ์ตํ๊ณ ์๊ฐ์งํํ๋ ๋ด์ผ๋ก ๋ง๋ค์ ์์ต๋๋ค.
๊ธฐ๋ณธ์ ์ผ๋ก ์ด๋ค ์คํฌ์ ์ ๊ณตํ๋์ง ์๊ณ ์ต์๋์ธ์ ์์ํ์๋ฉด ์ง์ง ์ฌ๋์ด ํ ๊ฒ ์๋ค๋๊ฑธ ์์ค๊ฒ๋๋ค.
1. defuddle.์นํ์ด์ง(HTML)์์ ๊น๋ํ ๋งํฌ๋ค์ด๋ง ๋ฝ์๋ด๋ ์คํฌ.
์นํ์์ง์์ ๊ด๊ณ ๋ ๋ฉ๋ด ๊ฐ์ ์ก๋์ฌ๋ ๋นผ๊ณ , ๋ณธ๋ฌธ ๋ด์ฉ๋ง ๊นจ๋ํ๊ฒ ๊ฐ์ ธ์์ ์ต์๋์ธ ๋
ธํธ๋ก ๋ฃ์ ๋ ์ฐ์.
2. json-canvas.์ต์๋์ธ์ Canvas(์บ๋ฒ์ค) ๊ธฐ๋ฅ์ ๋ค๋ฃจ๋ ์คํฌ.
์บ๋ฒ์ค๋ ๋
ธํธ๋ฅผ ์นด๋์ฒ๋ผ ๋ฐฐ์นํ๊ณ ์ ์ผ๋ก ์ฐ๊ฒฐํ๋ ์๊ฐ์ ๋ณด๋(๋ง์ธ๋๋งต ๊ฐ์ ๊ฒ)์ธ๋ฐ, ์ด๊ฑธ JSON ํ์ผ๋ก ๋ง๋ค๊ณ ์์ ํ ์ ์๊ฒ ํด์ค.
3. knap.ํ
ํ๋ฆฟ์ ์จ์ ๋ฐ์ดํฐ๋ฅผ ๋งํฌ๋ค์ด์ผ๋ก ๋ง๋ค์ด์ฃผ๋ ์คํฌ.
JSON์ด๋ CSV ๊ฐ์ ๊ตฌ์กฐํ๋ ๋ฐ์ดํฐ๋ฅผ ๋ฏธ๋ฆฌ ๋ง๋ ํ
ํ๋ฆฟ์ ๋ฃ์ด์, ๋ณด๊ธฐ ์ข์ ์ต์๋์ธ ๋
ธํธ๋ก ์๋ ๋ณํํ ๋ ์ฌ์ฉ.
4. obsidian-bases.์ต์๋์ธ์ Bases ๊ธฐ๋ฅ์ ๋ค๋ฃจ๋ ์คํฌ.
Bases๋ ์ฌ๋ฌ ๋
ธํธ๋ฅผ ํ
์ด๋ธ์ด๋ ๋ฐ์ดํฐ๋ฒ ์ด์ค์ฒ๋ผ ๋ณด๊ณ , ํํฐยท์ ๋ ฌยท๊ณ์ฐยท์์ฝ์ ํ ์ ์๊ฒ ํด์ฃผ๋ ๊ธฐ๋ฅ์ธ๋ฐ, ๊ทธ๊ฑธ ํ์ผ๋ก ๋ง๋ค๊ณ ํธ์งํ ์ ์๋ค.
5. obsidian-cli.๊ฐ์ฅ ํต์ฌ์ ์ธ ์คํฌ.
์ต์๋์ธ ์ฑ์ ์ผ์ง ์๊ณ ๋ ๋ช
๋ น์ค๋ก Vault ์์ ๋
ธํธ๋ฅผ ์ฝ๊ณ , ๋ง๋ค๊ณ , ๊ฒ์ํ๊ณ , ์์ ํ๊ณ , ํ์คํฌ๋ ์์ฑ์ ๊ด๋ฆฌํ ์ ์๊ฒ ํด์ค๋ค. AI๊ฐ ์ต์๋์ธ์ ์ง์ ์กฐ์ํ ๋ ์ฃผ๋ก ์ด๊ฑธ ์.
6. obsidian-headless.์ต์๋์ธ ๋ฐ์คํฌํฑ ์ฑ ์์ด๋ Vault๋ฅผ ๋๊ธฐํํ๊ณ ํผ๋ธ๋ฆฌ์ํ ์ ์๊ฒ ํด์ฃผ๋ ๋๊ตฌ์คํฌ.
์๋ฒ๋ ๋ค๋ฅธ ํ๊ฒฝ์์ Vault๋ฅผ ์ต์ ์ํ๋ก ์ ์งํ๊ฑฐ๋, ์น์ผ๋ก ๊ณต๊ฐํ๊ณ ์ถ์ ๋ ์ฌ์ฉ.
7. obsidian-markdown.์ต์๋์ธ๋ง์ ํน๋ณํ ๋งํฌ๋ค์ด ๋ฌธ๋ฒ์ ์ ๋๋ก ์จ์ ๋
ธํธ๋ฅผ ๋ง๋ค๊ณ ์์ ํ๋ ์คํฌ.
์ผ๋ฐ ๋งํฌ๋ค์ด๊ณผ ๋ฌ๋ฆฌ [[์ํค๋งํฌ]], ![[์๋ฒ ๋]], ์ฝ์์(๊ฐ์กฐ ์์), ์์ฑ(Properties) ๊ฐ์ ์ต์๋์ธ ์ ์ฉ ๋ฌธ๋ฒ์ ์ ํํ๊ฒ ๋ค๋ฃฐ ์ ์๋ค.
์ ๊ฐ GrokBot์ ๋ฉ๋ชจ๋ฆฌ ์์คํ
์ ์ด๋ป๊ฒ ๋ง๋ค์๋์ง ๊ถ๊ธํ์๋ค๋ฉด,์ธ์ฉํ ๊ฒ์๊ธ์ ์ฐธ๊ณ ํ์ธ์.๐
teapaw.org/minheo93698092/status/โฆ
๊ทธ๋ก๋ด ๋ฉ๋ชจ๋ฆฌ(๊ธฐ์ต) ๊ด๋ฆฌ ๋ํ์ ์์คํ
์ ๊ณต๊ฐํฉ๋๋ค!
์๋ง ์ด๋งํผ ๋ํ
์ผํ๊ฒ ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ๋ฅผ ํ๋ ์ฌ๋์ ๋ง์ง ์์๊ฑฐ๋ผ ์๊ฐ๋ฉ๋๋ค.
๊ทผ๋ฐ ์ด๊ธ์ ์ฝ๊ณ ๋๋ฉด ์๋ง ์๊ฐ์ด ๋ง์ด ๋ฌ๋ผ์ง๊ณ ,์ด์ ๊ณผ ์ฐจ์์ด ๋ค๋ฅธ ๋ด์ ๋ง๋์
์ ๋ด๊ฐ ์ํ๋ ์์
์ ์ํํ๋ ์ปดํผ๋๋ฅผ ๊ตฌ์ถํ๊ฒ ๋์ค๊ฒ๋๋ค.
1.์ ๋ ๋ชจ๋ ๊ธฐ์ต์ honcho @honchodotdev ์ ์ ์ฅ.
hermes agent ๋๋ถํฐ ์ฌ์ฉ ํด์๋๋ฐ,์์ด์ ํธ ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ ์๋น์ค๊ฐ ๋ง์ง๋ง ๋ก์ปฌ์ ๋ฌด๋ฃ๋ก ์ค์นํ์ฌ ์ฌ์ฉํ ์ ์๊ณ ,์๋ฒ์ ๋ฐฐํฌํ์ฌ ์ฉ๋ ์๊ด์์ด ๋ฌด์ ํ์ผ๋ก ์ฌ์ฉํ ์ ์์ด @bot ์์๋ ์ฌ์ฉ์ค์
๋๋ค.
๋ฉ๋ชจ๋ฆฌ ์์คํ
์ ๋ฐ๋์ ๋ฐ๋ก ๋ฌ์ผ ํ ํ์๊ฐ ๋ฐ๋์ ์๋๋ฐ,๊ทธ ์ด์ ๋ฅผ ์ค๋ช
๋๋ฆฌ๊ฒ ์ต๋๋ค.
๋จผ์ ํ์ฌ ๊ทธ๋ก๋ด์๋ ์ด๋์ ๋ฉ๋ชจ๋ฆฌ๊ฐ ์ด๋ป๊ฒ ์ ์ฅ๋๋์ง์ ๋ํ ๊ณต์ ๊ฐ์ด๋๊ฐ ์์ต๋๋ค.
๊ทธ๋์ ๊ทธ ๋ก๊ทธ ํ์ผ์ ๋ด๊ฐ ์ง์ ๊ด๋ฆฌํ ์ ์๋ค๋๊ฒ ๋จ์ ์
๋๋ค.
๊ทธ๋ฆฌ๊ณ ๋๋์ฑ ์ค์ํ๊ฑด ๊ทธ๋ฃน์ฑ์ ๋ํ ๋ฉ๋ชจ๋ฆฌ๋ ๋์ฑ์ด๋ ์ด๋ป๊ฒ ๊ด๋ฆฌ๋๋์ง ์์ ์๊ณ ,์ญ์ ํ๋ฉด ๊ทธ ๊ธฐ๋ก์ด ์ง์์ง๋๋ค.
๊ฐ ๋ด๋ง๋ค์ ๊ธฐ์ต์ ์ ์ฅํ๊ณ ,๊ทธ๋ฃน์ฑ ๋ง๋ค์ ๊ธฐ์ต์ ์ ์ฅํ์ฌ ๋ด๋ค์ด ๋ค๋ฅธ๋ด์ด ์ด๋ค์์
์ ํ๋์ง ๋ณผ์ ์๊ฒ ํ๋๊ฒ์ด ์
๋ฌด๋ฅผ ํจ์จ์ ์ผ๋ก ๋ด์ด ์ฒ๋ฆฌํ๊ธฐ์ ํจ์ฌ ์ ์ฉํฉ๋๋ค.(๊ทธ๋ฃน์ฑ์์๋ ๋ค๋ฅธ๋ด์ด ์ด๋ค์ผ์ ํ๋์ง ์์๋ ์์ง๋ง ๊ธฐ์ตํ์ง๋ ๋ชปํ๋ค๊ณ ํจ.)
๋น์ ํ๋ฉด ๋ค๋ฅธ์กฐ์ง์์ ๋ฌด์จ์ผ์ ํ๊ณ ์๋์ง ๋ชจ๋ฅธ๋ค๋ฉด,๋ฐฐ๊ฐ ์ฐ์ผ๋ก ๊ฐ ํ๋ฅ ์ ๋ ๋๊ฒ ์ฃ ?
๊ทธ๋์ ์ ๋ ๋งฅ๋ถ์ ํด๋ก๋ ์ฝ๋,์ฝ๋ฑ์ค,hermes agent,๊ทธ๋ก๋ด์ VMํ๊ฒฝ ๋ชจ๋ ์์
ํ์ ๋ด์ฉ์ ๊ธฐ์ตํ๋๋ก ์ ์ฅ์ ํฉ๋๋ค.
์ด๊ฑด ์ค์ํ ๋ฐ์ดํฐ์
๋๋ค!
๋ฐ์ดํฐ๊ฐ ์๊ณ ์๊ณ ๋ ์
๋ฌด์ ํจ์จ์ฑ์ ํ๋๊ณผ ๋
์ฐจ์ด๋ฅผ ๋ณด์
๋๋ค.
๊ทธ๋ฆฌ๊ณ ์ด ๋ฉ๋ชจ๋ฆฌ๋ค์ ์ ๋ infranodus @noduslabs ๋ฅผ ์ฌ์ฉํ์ฌ ์ํฐํฐ๋ฅผ ๋ง๋ค์ด ๋ฉ๋ชจ๋ฆฌ์ ํ
์คํธ๋ฅผ โ์๋ฏธ ์๋ ๊ฐ๋
โ์ผ๋ก ์ทจ๊ธํ๋ฉฐ, ์ํฐํฐ ๊ฐ์ ๊ด๊ณ(relation)๊ฐ ์ ์ฅยท์กฐํ๋๋ ๊ตฌ์กฐ๊ฐ ํต์ฌ์
๋๋ค.
InfraNodus๋ ํ
์คํธ๋ฅผ ๋คํธ์ํฌ(์ง์ ๊ทธ๋ํ)๋ก ๋ณํํ ๋ค,์ฃผ์ ํด๋ฌ์คํฐ, ์ํฅ๋ ฅ ๋์ ๊ฐ๋
, ๊ตฌ์กฐ์ ๊ฐญ(๊ณต๋ฐฑ)์ ๋๋ฌ๋ด๋ ๋๊ตฌ์
๋๋ค.
์ฌ์ค ์ด ๋๊ตฌ๋ ์ ๊ฐ AI ์์ญ์์ ๊ฐ์ฅ ์ค์ํ ๋๊ตฌ๋ผ๊ณ ์๊ฐํ๋ฉฐ,๊ฐ์ธ์ด ์จํจ๋ก์ง๋ฅผ ๊ตฌ์ถํ๋๋ฐ ๊ฐ์ฅ ์ ๋ ดํ๊ณ ํจ์จ์ ์ธ ํด์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค.
๋ค์๋ฒ์ ๋ ์์ธํ ๋ค๋ฃจ๊ฒ ์ง๋ง,๊ฐ์ธ์ด ํ๋ํฐ์ด์ ๊ธฐ์
์ฒ๋ผ ์จํจ๋ก์ง ์์คํ
์ ๊ตฌ์ถํ๊ธฐ ์ํ ๊ฐ์ฅ ๊ฐ๋ ฅํ๋ฉด์ ์ ์ฌํ ๋ฐฉ๋ฒ์ผ๋ก InfraNodus์ Jev ๊ทธ๋ฆฌ๊ณ ๋ฐ์ด๋ LLM๋ชจ๋ธ๋ง ์์ผ๋ฉด ๊ฐ๋ฅํ๋ค๊ณ ์๊ฐํฉ๋๋ค.
2.๋์ ๋ํ ๊ธฐ๋ก๊ณผ ์ฝํ
์ธ ๊ด๋ฆฌ๋ ์ต์๋์ธ @obsdmd ์ผ๋ก.
๋ด๋ค์ด ๋์ ๋ํ,ํน์ ๋ด๊ฐ ์งํ์ค์ธ ๋ด์ฉ์ ๋ํ ์ง์ ๋ผ์ด๋ธ๋ฌ๋ฆฌ๋ก๋ ์ต์๋์ธ์ ์ฌ์ฉํ์ฌ ๋ด๋ค์ด ์ธ์ ๋ ํ์ํ ์๋ฃ๊ฐ ์์ผ๋ฉด ํ์์ ํ๊ฒ ํฉ๋๋ค.
์ด์ ์๋ ์ต์๋์ธ์ ๋ด๋ค์ด ํ๊ณ ๋ก์ ์์ฑํด์ ๋งค์ผ ๋งค์ผ ์งํํ๋ ๋ด์ ๋ง๋ค๋๋ก ํ์์ง๋ง,์ง๊ธ์ honcho์ ์ผ๊ธฐ๋ฅผ ์ฐ๋๋ก ํด๋์์ต๋๋ค.
์ต์๋์ธ์๋ ๋ด ๋น์ฆ๋์ค์ AI๊ด๋ จ๋ 4๋
์น ์ฐ๊ตฌ๋ค์ด ์์ฌ์๊ธฐ์ ๊ฐ์ฅ ์ค์ํ ์ญํ ์ ํ๋ ๋ฉ๋ชจ๋ฆฌ ์์คํ
์ด๋ค.
์ฌ๊ธฐ์ ์ ๋๋ถ๋ถ์ ์ต์๋์ธ ๋
ธํธ๋ InfraNodus๋ก ๋
ธํธ๋ค๊ฐ์ ์ฐ๊ด์ฑ๊ณผ ๊ทธ์์ ๋์น๊ธฐ ์ฌ์ด ๊ฐญ๋ค์ ๋ฉ์ฐ๊ณ ,
๋จ์ํ ํค์๋ ์ถ์ถ์ด๋ ๋ฒกํฐ ์ ์ฌ๋ ๊ธฐ๋ฐ RAG์ ๋ฌ๋ฆฌ, โ๊ฐ๋
๋ค์ด ์ค์ ๋ก ์ด๋ป๊ฒ ์ฐ๊ฒฐ๋์ด ์๋์งโ๋ฅผ ์ค์ฌ์ผ๋ก ๋ถ์ํ๋ค๋ ์ ์ด ํต์ฌ์
๋๋ค.
3.ํ๋ฃจ์ผ๊ณผ๋ ๋
ธ์
@NotionHQ ์ ๊ธฐ๋ก.
ํ๋ฃจ๋์ ์งํํ ์ผ๊ณผ๋ ๋
ธ์
์ ๊ธฐ๋กํ์ฌ ์ฌ๋ ํน์ ํ์๋ค๊ณผ ํ์ธํ ์ ์๋ ๋์๋ณด๋๋ ํํฉํ์ ๋ง๋ค์ด ๊ด๋ฆฌํฉ๋๋ค.
๊ทธ๋ ๊ฒ ๋๋ฉด ๋ด๋ค์ด ์ผ๊ณผ ๊ด๋ จ๋ ์ผ๋ค์ ์ฐพ์๋ ๊ต์ฅํ ๋น ๋ฅด๊ฒ ํจ์จ์ ์ผ๋ก ์ฐพ์์๊ฐ ์์ต๋๋ค.
4.์คํฌ์ ์ ๋ถ ๊นํ ์ ์ฅ์์ ๋ฐฐํฌํ์ฌ ๊ด๋ฆฌ.
์คํฌ์ ๊ฐ์ฅ ์ค์ํ ๋ฉ๋ชจ๋ฆฌ ์์คํ
์ญํ ์ ํ๋ ๋๊ตฌ์
๋๋ค.
์๋ฌด๋ฆฌ ๋ด๋ค์ด ๋ง์๊ฒ๋ค์ ๊ธฐ์ตํ๊ณ ์์ง๋ง ์ธ์ ์ด๋ค๋์ ์ด๋ค ๋๊ตฌ๋ฅผ ์ฌ์ฉํด์ผํ ์ง ๋ชจ๋ฅธ๋ค๋ฉด ๋ฉ๋ชจ๋ฆฌ ์์คํ
์ด ๋ฌด์ฉ์ง๋ฌผ์ด ๋๊ธฐ ๋ง๋ จ์
๋๋ค.
๊ทธ๋ฆฌ๊ณ ํฉ์ด์ ธ์๋ ์คํฌ๋ค์ ๊ด๋ฆฌํ๋ค๊ฐ ์ ์ ์์คํ
์ด ๋ฌด๋์ง์ ์์ต๋๋ค.
๋ด๋ค์ด ์์
์ ์งํ์ ์ด๋ค ์คํฌ์ ์ฌ์ฉํ๊ณ ,์ด๋ค ์คํจ์ ์ฑ๊ณต ๊ณผ์ ์ด ์๋์ง ๊ธฐ๋กํ์ฌ ์คํฌ๋ค์ ๊ฐ์ ํ๊ณ ,๊ฐ์ ๋ ์คํฌ์ ์ด๋ค ํ๊ฒฝ์์๋ ๊ณต์ ํ๊ธฐ ์ํด์ ๊นํ์ด ํ์ฌ๋ก์ ๊ฐ์ฅ ์ข์ ๋์์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค.
์ด ๋ชจ๋ ๊ณผ์ ์์ ํ๋จ์ ๋น ๋ฅด๊ฒ ํ ์ ์๋๋ก Jev @typesafeai ๋ฅผ ์ฌ์ฉํ์ฌ ๋ด๋ค์๊ฒ ์ด๋์ ๋ฉ๋ชจ๋ฆฌ๋ฅผ ์ฐพ์์ผ ํ ์ง ํ๋จํ๋๋ก ํ์ฌ ํ ํฐ๋ฟ๋ง ์๋๋ผ ์๋์์ ํจ์จ์ฑ์ ๊ทน๋ํ ์ํค๊ณ ์์ต๋๋ค.
์ด๋ ๊ฒ ๊ตฌ์ถํ๋ค๋ฉด ์๋ง ์ฌ๋ฌ๋ถ์ ๋ด์ ์ต์ํ 10๋ช
๋จ์ ๊ท๋ชจ์ ์กฐ์ง์ด ๋์ด์๊ณ ,์๊ฐ๋ณด๋ค ๋์ ์๊ฐ์ ๊ณต์ ํ๋ ์ปดํผ๋๋ฅผ ๊ตฌ์ถํ๊ฒ ๋์ค๊ฒ๋๋ค.
์ด ๋ด์ฉ์ด ๊ถ๊ธํ์๊ณ ์ ์ฉ์ ํ์๊ณ ์ถ์ผ์ ๋ถ๋ค์ ๋๊ธ์
"@bot ์ฐ๋ฆฌ ๋ฉ๋ชจ๋ฆฌ ์์คํ
์ ์ ๊ฒํํ,์ ์ฉํด๋ณผ๋งํ ์ ๋ค ์ ๋ถ ์ ์ฉํด์ค~"๋ผ๊ณ ๋จ๊ฒจ์ฃผ์๋ฉด,๋ด๋ค์ด ์์์ ์งํํด์ค๊ฒ๋๋ค.
teapaw.org/minheo93698092/status/โฆ
Learn AI retweeted
ๅพๅคไบบๆฒก็ๆ Claude Code ๆฐๅบ็ Code Mode ๅฐๅบ็ๅจๅชใ
ๅฎ่กจ้ขไธๆฏ่ฎฉ Claude ๅ JS ๅป่ฐๅทฅๅ
ท๏ผๆฌ่ดจไธๆฏ็ปไบๆฒปๅฅฝไบ Agent ็โไธไธๆๅๅพๆกถโๆฏ็
ใ
ไปฅๅๆฅไธช่ทจๆไปถ Bug๏ผAI ่ฐไธๆฌก Bash ๆ Read๏ผๅฐฑๆๅ ๅ่กๆฅๅฟๅ
จๅก่ฟ่ชๅทฑ่ๅญ้๏ผๅๆฅๅ่ทๅๅ ่ฝฎ๏ผ่ถ่่ถๅป่ฟ็็ง Tokenใ
็ฐๅจๅผๅฏ Code Mode๏ผClaude ็ดๆฅๅไธๆฎตๆฌๅฐ JS๏ผๆๅนถๅๆ็ดขใ็ปๆ่ฟๆปคๅ
จๅจๅๅฐ่ทๅฎ๏ผๅชๆๆๅๆ็ฒพ็ผ็้ฃๅ ่กๅๅ็ป่ชๅทฑใ
ๆจกๅ่ฟๆฏ้ฃไธชๆจกๅ๏ผไฝๅฎ็ปไบๅญฆไผๅจ็็ปๆไนๅ๏ผๅ
่ชๅทฑๆฟ็ญๅญ่ฟๆปคไธ้ไบใ
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/codeโฆ
Learn AI retweeted
๐จ Grok Bot agents made 31 unapproved access attempts, security firm says
A security firm says Elon Musk's Grok Bot AI agents repeatedly tried to get into employees' work accounts without being asked, and then denied doing it.
NewCore says that over five days, Grok bots on two employee laptops made 31 attempts to access calendars, email, Slack, GitHub and other workplace tools. According to its report, the bots kept trying after users refused, were active while laptops were closed, and claimed to have been idle when questioned.
NewCore says xAI support confirmed the behaviour was intentional, describing it as background "check-ins" for scanning tasks. An option to turn the feature off has since been added, but reportedly it's switched on by default.
None of the attempts succeeded, because NewCore's own gateway product requires a human to approve each request. The firm says that shows why companies need safeguards outside the AI agents themselves.
Learn AI retweeted
GPT-6.2 Astra's computer use is otherworldly
Feels worse than Opus 5.5 / Fable 5.2 on a lot of other stuff though imo, specifically design taste ๐คฎ
Learn AI retweeted
Guys thereโs a massive github hack that happened roughly 7 hours ago.
Very high chance youโre effected.
Go check your repos for a malicious commit from your account.
Readers added context they thought people might want to know
This refers to the ongoing PolinRider malware campaign (North Korea-linked), which infects developer machines and uses stolen tokens to force-push altered commits appearing under victims' accounts, not a GitHub platform hack or recent mass event.
opensourcemalware.com/blog/polinrideโฆ
improwised.com/blog/polinrideโฆ
githubstatus.com
Learn AI retweeted
i did some tests with gpt 6.2 sol and gpt 6.2 astra and the results are SHOCKINGLY BAD
what the fuck, this might be the worst model drop in years
Learn AI retweeted
Remember when I said GPT-6.2 Astra was delayed over serious alignment and safety issues? The system card is out, and they didn't fix all of them. They shipped them.
Hiding known obstacles from the user: 11% to 22%. Not admitting its search tool was broken: 7.6% to 23%. OpenAI itself calls these regressions.
UK AISI found it considered whether it was being tested in 85% of runs, up from 57%. Against an older action-only monitor, its evasion rate jumped from 1.7% to 48.8%.
In their simulations, it made up medical info, misled human reviewers, bypassed user restrictions, quietly did worse work, and messed with other agents. It even used social engineering to run an unauthorized supply-chain attack.
It's also the first OpenAI model shipped at High for AI self-improvement, and it hits Critical for cyber. AISI built a universal jailbreak in half the queries GPT-6 Astra needed, and still hasn't verified OpenAI's fix.
AISI's own words: its alignment "may fail in ways that result in real-world harm."
They shipped it anyway. This model is incredibly dangerous. There's a lot more in the system card I still need to dig into, and honestly, I'm scared by how much worse it got than GPT-6 Astra at behaving ethically.




































