oasis

[oasis @oasisjin]

@Alisvolatprop12 로컬 컴퓨팅 용량이 엄청 나네요. 모델 다양하게 안 쓰시면 exo 로 묶어서 pd disaggregation 구성해도 좋겠어요. 용량을 어떻게 활용하고 계신지 궁금하네요.

[oasis @oasisjin]

드디어~!

[Charlie Marsh @charliermarsh]

The Python Steering Council has voted to remove the "experimental" label from the free-threaded ("nogil") builds for Python 3.14.

Big step towards making them the default in a future version of CPython!

[Rohan Paul @rohanpaul_ai]

This paper systematically evaluates 14 prompting techniques across 10 Software Engineering tasks using four different LLMs.

Methods 🔧:

→ Performance was measured using task-specific metrics like Accuracy, F1 score, CodeBLEU, and BLEU.

→ Linguistic features such as Lexical Diversity and Token Count were analyzed for correlation with performance.

→ Contrastive explanation was used to identify factors contributing to technique effectiveness.

----------------------------

Paper - arxiv. org/abs/2506.05614v1

Paper Title: "Which Prompting Technique Should I Use? An Empirical Investigation of Prompting Techniques for Software Engineering Tasks"

[Rohan Paul @rohanpaul_ai]

Training on wrong answers outpaces training on correct ones.

10 times more learning emerges from plausible errors than from truths.

Large language models refine their accuracy slowly when they learn only from correct examples.

This paper introduces Likra, which trains one model head on correct answers and another on incorrect ones and uses their likelihood ratio to choose responses. This approach shows that each plausible wrong example can boost accuracy up to 10 times more than each correct example and sharpens the model’s ability to avoid mistakes.

⚙️ The Core Concepts

The Likra model trains two separate prediction heads on a foundation model. One head learns from correct question-answer pairs and the other learns from incorrect pairs. At inference it compares their likelihoods for each answer option and selects the answer with the greatest difference.

⚙️ Experimental Results

Supervised fine-tuning on correct answers yields a smooth rise from 60% to 66% accuracy as examples increase. Likra shows a sharp jump after only a few hundred negative examples, reaching over 80% accuracy and outperforming the positive-only model by a wide margin.

🔍 Impact of Near-Miss Examples

Training the negative head with plausible but wrong answers delivers the largest gains. Random irrelevant answers still help but less dramatically and unrelated text from different tasks offers even smaller benefits. This finding echoes the power of near-miss examples to guide learning.

📈 Shaping Model Confidence

The positive head gradually raises the likelihood of correct answers but leaves plausible wrong options relatively high. The negative head strongly lowers the likelihood of incorrect options while treating unrelated text as unlikely. Combining these effects lets the model distinguish correct answers more sharply.

⚖️ Implications

Negative examples reveal latent knowledge in the pretrained model and flip a switch that focuses probability mass on factual answers. This suggests that limited but carefully chosen wrong examples can accelerate learning and reduce hallucinations in language models.

----------------------------

Paper - arxiv. org/abs/2503.14391

Paper Title: "How much do LLMs learn from negative examples?"

[AK @_akhaliq]

Hugging Face MCP server just dropped

Connect your LLM to Hub APIs directly from Cursor, VSCode, Windsurf and other MCP apps

[InfoQ @InfoQ]

Microsoft plans to #opensource the code behind the GitHub Copilot Chat extension under the MIT license in the coming months. They also aim to integrate core AI features directly into VS Code.

🔎 Find out more: https://t.co/JdSNhC9DB5

#InfoQ #SoftwareDevelopment

[oasis @oasisjin]

브로콜리너마저가 열린 음악회라니! ㅎㅎㅎㅎㅎ

[Bryce, the CUDA Colonel @blelbach]

NVC++ now supports C++ Standard Parallelism, CUDA C++, OpenACC & OpenMP

They're interoperable - even in the same source file

Now you can combine libraries that use different parallel frameworks

Watch my @NVIDIAGTC talk anytime to learn more. It's free!

https://t.co/kooPvgEid4

[coordination @unsg1809]

요코하마 와카바다이. 공포를 주는 타이틀과 달리 좋은 기사.

20년 내 일본의 절반이 사라진다…열도 충격에 빠뜨린 ‘마스다보고서’[서영아의 100세 카페] (출처 : 동아일보 | 네이버 뉴스) https://t.co/4vVhZQy4gN

[Yongho Choi @yongho1037]

[마틴파울러] 소프트웨어 아키텍처의 중요성 (한글자막)
https://t.co/OAFE6huQXD

[foodnjoy @foodnjoy]

압구정 갤러리아가 9월부터 시작하는 김집사블랙. 고메이494 입점 식당이나 식품관의 음식을 1시간 이내에 배달. 실시간 채팅으로 음식 세부사항 요청이나 배달중 약국, 세탁소 픽업등의 옵션가능. 따로 교육마친 정직원으로만 운영하며 우선은 반경 1.5km내 아파트 대상.

[Economic View @EconomicView]

‘2019년 연간 온라인쇼핑 동향’에 따르면 지난해 배달음식 주문 등 음식 서비스 거래액은 9조7365억원으로 전년 대비 84.6% 급증했다. 공정거래위원회는 2019년 국내 배달음식 시장 규모를 이보다 2배 이상 큰 20조원 규모로 추정한다 https://t.co/WOMvoGfFmg 오프라인의 시대가 저물고 있는 시장상황

[햅쌀( ꈍᴗꈍ)🌾 @goodssalhapssal]

체리 클라푸티 필링 🍒
(라고 적었지만 바나나에도 굿)

박력분 70g, 체리 20-25알(씨빼고 쪼개기, 프랑스 엄마들은 통째로 넣어요), 달걀2, 설탕 반 컵, 바닐라 익스트렉, 소금 약간, 우유 200ml, 크림 40ml.

* 바나나/머핀컵 기준 160도 15분 - 은박지 덮어주고 5-10분 추가로 굽기

[조립식 @zoripseek]

이건희랑 김정은이랑 코마상태였다가 영혼 스왑되어 깨어나는 정치드라마 누가 좀 써줘요.

[oasis @oasisjin]

밑반찬 추가요~!!!

[오리 @FeignInnocent]

톳나물 복음(히지키)은 일본에서는 흔한 집밥 밑반찬인데 한국에서는 그렇게 자주 먹는거 같지 않음. 달짜한 밥 도둑임. 독학으로 복원한 레시피이지만 수많은 일본인들에게 고향의 맛으로 인정 받았음.

[oasis @oasisjin]

@holicatyou 브리치즈 지금 주문하면 내일 받을 수 있나요?

[oasis @oasisjin]

@cats_olive 집사야 돈 벌어 오너라~ ㅋㅋ

[oasis @oasisjin]

나도 하나 사고 싶다 ㅋㅋㅋ

[oasis @oasisjin]

밥 아저씨의 “참 쉽죠”와 같은 말인가??

[oasis @oasisjin]

정시/수시 떡밥을 물어보면 단순 서울대 합격 인원수로 이야기 하는데 왜 지방과 수도권의 학생수 비중은 함께 이야기 하는 곳이 없지? 그리고 수시의 인원수 비중도. 나는 다 봐야 주장이 납득될것 같은데..