스누피

소소하게 합니다

[Anatoli Kopadze @AnatoliKopadze]

Anthropic engineer:

"You're not supposed to prompt Claude. You're supposed to build a system that prompts itself."

In 45 minutes she shows exactly how Anthropic builds agents that remember, fix their own mistakes and get smarter with every run.

This beats any paid course on agents I've seen.

Watch it, then read the guide on building loops below.

[ミサキ @misakin883]

長押しで翼が現れるAIイラストが大反響！のこれですが
みどりさんアプリ「カクシエMAKER」で簡単にできます🪽

[Claude @claudeai]

The Claude Security plugin for Claude Code is now available in beta.

Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal on the Claude inference you already run.

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Boris Cherny @bcherny]

I talk to engineers at other companies every day and hear the same thing: one person is 10x'ing their output with Claude but the rest of the org hasn't caught up.

Watching teams adopt AI, I keep seeing the same 4 steps.

I mapped them out here: Steps of AI Adoption https://t.co/kQnRAUMKpP

[Muhammad Ayan @socialwithaayan]

anthropic just open sourced the tool they use to read what claude is thinking before it says anything 🤯

it's called the jacobian lens (j-lens). they used it to find something they're calling j-space, a small set of internal directions that hold the exact concepts a model is about to talk about, before it writes a single word.

why it works: most interpretability tools make you choose. sparse autoencoders are accurate but expensive to train. the old logit lens is free but turns to noise in early layers. j-lens is one matrix multiply per layer, cheap, and stays readable everywhere.

what you get for $0:
→ a lens that decodes any layer's activity into ranked vocabulary tokens
→ works on any open weights huggingface decoder, examples use qwen
→ a full walkthrough notebook, load model, load or fit lens, render the layer view
→ a live no-setup demo on neuronpedia

what it replaces:
→ training your own sparse autoencoder for every model you want to study
→ guessing what a model "meant" from output alone

how to set it up (2 minutes):

1. pip install -e .
2. load your model and tokenizer with transformers
3. wrap it with jlens.from_hf(hf, tok)
4. apply a pretrained lens or fit your own on ~100 prompts

the honest part: no claude weights or corpora are bundled, this reads open models on your own gpu, not claude directly.

899 stars, apache-2.0, free and open source.

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[스누피 @snoopy__lee]

@cfyj00 쿠팡 주문 완료했어요 ㅎㅎ 이번 여름은 모기물림 가려움에서 좀 해방될지 기대되네용.

[스누피 @snoopy__lee]

Fable을 3분 남겨놓고 97% 썼군. 뿌듯해.

[구루 @xguru]

와 진짜 대단. 정부에 뭔가 신고하거나 할때 어떤 순서로 이뤄지는지를 일목요연하게 보여줌. 총 100개인데 궁금했던거 몇개만 보면

1. 전세사기 피해지원
2. 예비타당성 조사
3. 정보공개청구
4. 재건축·재개발 정비사업

AI 를 제대로 활용하는 공무원이신듯

[스누피 @snoopy__lee]

gpt-5.6 모델은 어떤 성능을 뽑아줄지!!

[OpenAI @OpenAI]

GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday.

We’re expanding preview access globally now.

[할로묘묘 @Hellomeow0]

와...방금 엄청난거 봄...

어떤 면접 프로그램이었는데
면접관이 탁자 위에 종이컵
1,000개를 쫙 깔아놓고

"자, 여기 1,000잔의 물이 있음.
이 중 딱 한 잔만 설탕물입니다.
가장 빠르고 효율적으로 그 한잔을
찾는 방법은 무엇입니까?"

지원자:
지나가는 사람 1,000명한테 이
물을 한 잔당 1달러씩 받고 팝니다.

그리고 '설탕물 당첨자에게는
상금 500달러를 드립니다!'라고
이벤트를 거는 겁니다.

그럼 설탕물은 손님들이 알아서
빠르시간내에 찾아줄거고
남은 수익금 500달러는 회사
마케팅 비용으로 쓰겠습니다."

결국 그 사람은 그 자리에서
바로 프리패스 합격함.

남의 돈으로 문제 해결 + 리스크
전가 + 회사 수익 창출까지...

요즘 본 브레인중에 제일 신박했음.

[스누피 @snoopy__lee]

으아아아 감사합니다아

[Claude @claudeai]

We're extending access to Claude Fable 5 on all paid plans through July 12.

[구루 @xguru]

https://t.co/bmqwcJXh2C

[스누피 @snoopy__lee]

https://t.co/zPSC9Cm96n

[스누피 @snoopy__lee]

"AI가 일자리를 가져가기 한참 전에 이미 조용히 가져가는 것은 '몰입할 기회'다. 되찾는 법은 아직 모르지만, 손으로 글을 쓰는 지금 오랜만에 조금 몰입하고 있다고 말한다."

[스누피 @snoopy__lee]

@nowlovepan 코엑스라니 !! 회사 코앞이라 일단 신청했습니당 ㅎㅎㅎ

[스누피 @snoopy__lee]

이거 너무 웃김 ㅋㅋㅋㅋㅋㅋㅋㅋㅋㅋ

[유우 @yuu_uu_0]

말을 예쁘게 하는 사람이 좋아 >> 이 말의 뜻은
예쁜 단어를 쓰는 사람이 좋다는 말이 아니라
상대방의 말을 자기 기준으로 해석하지 않는 사람이구나

결국
사람들은 말을 잘하는 사람보다
자기 맘을 정확히 이해해주는 사람에게 호감을 느끼는구나

내가 그동안 어떻게 말했지
앞으로 어떻게 말해야할지 생각해보게 되는 영상