테일러 스위프트 공연이 지방 호텔 영업이익과 직결

https://blog.naver.com/geniusjo1984/223140084405

Film and TV Profits Have Collapsed Over the Last Decade 3: Net income at the largest U.S. entertainment companies has dropped by more than 60% since 2013, indicating a significant decline in profitability in the film and television industry.

The film and TV industry has experienced a decline in profitability in recent years due to several factors:

1. Rise of streaming services: Platforms like Netflix, Amazon Prime, and Hulu have revolutionized the way people consume media, leading to a decrease in traditional revenue streams such as ticket sales and DVD purchases**[1](https://substreammagazine.com/2023/04/how-streaming-services-affect-the-film-industry-and-the-popularity-of-casinos-in-movies/)[4](https://everymoviehasalesson.com/blog/2023/1/the-impact-of-streaming-services-on-the-film-industry)[6](https://www.cnbc.com/2019/11/08/the-death-of-the-dvd-why-sales-dropped-more-than-86percent-in-13-years.html)**. As more people choose to watch content through streaming services, the demand for traditional film and TV distribution channels has decreased.
2. Increased competition: The rise of streaming services has also led to an increase in the number of original content being produced, resulting in greater competition for viewers' attention**[2](https://www.filmstro.com/blog/the-biggest-filmmaking-challenges-in-2023)[5](https://kievkelvin.com/blog/challenges-facing-the-film-industry/)**. This has made it more difficult for individual films and TV shows to stand out and generate significant profits.
3. High production costs: The cost of producing high-quality content has increased, making it more challenging for filmmakers to recoup their investments and turn a profit**[5](https://kievkelvin.com/blog/challenges-facing-the-film-industry/)**. Additionally, as streaming services prioritize producing their own original content, they may be less willing to invest in licensing content from other producers, further reducing potential revenue streams for filmmakers.
4. Changing audience consumption habits: The convenience and flexibility of streaming services have altered the way audiences consume content, with many people now preferring to watch films and TV shows at home rather than in theaters**[7](https://blog.filmtrack.com/industry-insights/in-the-press/streaming-platforms-and-their-impact-on-the-film-industry)**. This shift in consumption habits has contributed to the decline in profitability for the film and TV industry.

https://youtu.be/i0Dds8TU_-k

admire his insatiable curiosity.

탁월함은 습관이지만, 실패도 마찬가지

매일 조금씩 개선하면 위대한 결과로 이끌 수 있다는 얘기를 많이 들음
이 유명한 조언은 우리 스스로에게 계속 전진하도록 푸시함
하지만, 이 이야기의 다른 면은 잘 얘기하지 않음
→ 방치/무시와 작은 실수가 계속 쌓이면 부정적인 결과로 이어질 수 있다는 것
내 삶을 되돌아 보면, 실제로 궤도를 벗어난 일들은 점진적인 무시와 수많은 작은 잘못된 선택의 결과였음
나는 약물에 하룻밤 사이에 중독된 것이 아님. 내가 건강과 안전보다 순간적인 즐거움을 우선시 했던 수백번의 순간을 통해서 였음
나는 하룻밤 사이에 과체중이 된 것이 아님. 장기적인 건강보다 즉각적인 만족을 선택한 수백번의 순간을 통해서였음
나는 하룻밤 사이에 관계를 망치지 않았음. 어려운 대화에 맞서고, 내 실수를 인정하거나, 누군가가 나보다 무언가를 더 잘하는 것을 인정하는 것보다 "편안함"을 선택한 수백번의 순간을 통해서 였음
이런 경험을 통해, 나쁜 습관을 피하는 것은 좋은 습관을 기르는 것만큼 중요하다는 것을 깨달았음
방치의 패턴을 인식하고 조기에 대처함으로써 더 큰 문제를 예방할 수 있음

Excellence is a habit, but so is failure
We often hear that making small incremental improvements every day can lead to great things. This popular piece of advice rings true, and it’s a powerful reminder to keep pushing ourselves forward.

But there’s another side to this story that we don’t discuss as often: how incremental neglect and small missteps can accumulate and lead to negative outcomes. Recognizing and addressing these patterns of neglect early can make a significant difference in preventing larger problems down the road.

Reflecting on my own life, I’ve noticed that most of the things that really went off-track were indeed consequences of incremental neglect and numerous small yet poor choices:

I didn’t become addicted to drugs overnight. It happened over hundreds of moments where I prioritized momentary pleasure over health and safety.
I didn’t become overweight overnight. It happened over hundreds of moments where I opted for immediate gratification over long-term health.
I didn’t ruin relationships overnight. It happened over hundreds of moments where I chose comfort over confronting difficult conversations, admitting my mistakes, or even just acknowledging that someone was better than me at something.
From these experiences, I’ve realized that avoiding bad habits is just as important as cultivating good habits.

To address these kinds of issues, we must become aware of our patterns of incremental neglect and then take deliberate steps to counteract them and foster healthier habits.

Written on April 5, 2023

https://awesomekling.github.io/Excellence-is-a-habit-but-so-is-failure/

폴 그레이엄이 환상적인 글을 남겼다. 이 글은 평생에 걸쳐서 읽어야겠다.

"How to Do Great Work"


<번역본>
https://frontierbydoyeob.substack.com/p/frontier-13-how-to-do-great-work?utm_source=post-email-noscript&publication_id=944480&post_id=132707382&isFreemail=true&utm_medium=email

<원문>
http://www.paulgraham.com/greatwork.html?utm_source=substack&utm_medium=email

https://youtu.be/GQNYD9yzIjU

wow…

LIMA 모델을 만든 기법을 실증한 사례들이 생겨나고 있는것 같습니다.

WizardLM 모델 파인튜닝에 사용된, 원본 Evol Instruct 데이터셋은 V1이 70K, V2가 190K 여건으로 구성되어 있습니다.

이번에 공개된 WizardLM V1.1은 1K의 데이터셋 만으로 파인튜닝 되었다고 합니다. 원본 Evol Instruct에서 큐레이션된 데이터셋일지, 새롭게 구성된 프롬프트로 얻은 데이터셋일지는 아직 공개된 바가 없어서 확실치 않습니다 (조만간 데이터셋과 그 방법을 공개한다고 하네요).

하지만 모델(13B)의 가중치는 이미 공개되어 있고, 벤치마크 성능은 WizardLM 30B V1.0과 유사한 수준으로 보입니다. 사실 LIMA 논문이 나왔을 때는 약간 의구심이 들기도 했는데요, 실제로 가능함을 보여주는 실제 사례들이 등장하면서 가능성이 있겠다는 생각이 드는군요.

예전부터 생각해왔듯이 "매우 고 품질"의 데이터가 필요하기는 하며, 사실 이 정도는 기업 차원에서는 쉽게 구축할 수 있을듯 합니다. 이미 LLaMA와 동일한 구조를 띈 "상업적 이용이 가능한" OpenLLaMA 및 XGen과 같은 기반 모델들이 나와있습니다. 1K 정도라면, 인력으로도 충분히 만들어낼 수 있는 수준이기 때문에, 조만간 GPT4의 개입 없이 실제 상업적 활용이 가능한 파인튜닝된 모델들이 속출할 것으로 예상됩니다.

당장 커뮤니티 차원에서도 데이터 만들기 품앗이를 해볼 수 있겠는데요. 일단은 GPT4로 데이터를 다량으로 만들기는 할 것인데, 이것을 사람이 큐레이션 하는 작업이 필요합니다. 먼저 "고 품질"로 1K 여건의 데이터를 추리는 작업이 필요하고, 그 다음 기계 번역합니다. 여기 까지하면 GPT4가 생성한 결과 그대로이기 때문에 라이선스가 애매하죠. 따라서 번역된 1K 여건을 사람이 일일이 검수합니다. 당연히 구조를 뒤틀고, 용어도 좀 더 자연스럽게 교정하는 등 매우 꼼꼼히 1K 데이터셋을 검수해보는 방향입니다. 그러면 사실상 GPT4가 만든 원본이라고해도, 결과물은 원본과는 완전히 다르기 때문에, 활용 가능할 것이라고 생각합니다.
: 혹시 관심있으신 분 계신가요?

https://twitter.com/WizardLM_AI/status/1677282955490918401

https://huggingface.co/WizardLM

Oh… 산전수전수중전까지 겪은 아저씨가 Bob아저씨가 그림 그려놓고 쉽죠 하는 것처럼… Great work란 말이지 하면서 이야기해주는 것 같네 ㅎㅎ

추가로 Patrcik(CEO of Stripe)도 블로그에서 인류가 만들어낸 위대한 일에 대한 기록들을 공유한 적이 있는데 같이 읽어볼만한 것 같다.

Hardy's *A Mathematician's Apology*

Some materials about successful industrial/applied research labs. I recommend all of them. Further recommendations very [welcome](mailto:patrick@collison.ie). Also, does it just *seem* that their heyday is past, or has something structurally changed?

- [Dealers of Lightning](https://www.amazon.com/Dealers-Lightning-Xerox-PARC-Computer/dp/0887309895). The definitive book about PARC.
- [Inside PARC: the 'information' architects](https://spectrum.ieee.org/ns/pdfs/inside-the-parc.pdf) (IEEE Spectrum, Oct 1985). Good article about PARC.
- [Interview with Bob Taylor](http://archive.computerhistory.org/resources/text/Oral_History/Taylor_Robert/102702015.05.01.acc.pdf) (and [another](https://patrickcollison.com/static/files/labs/taylor-markoff-interview.pdf)), who ran the PARC CS Lab.
- [The Idea Factory](https://www.amazon.com/Idea-Factory-Great-American-Innovation/dp/0143122797). The definitive book about Bell Labs. (There should be more...)
- [The Art of Doing Science and Engineering](https://www.amazon.com/Art-Doing-Science-Engineering-Learning/dp/9056995006). Only indirectly about Bell Labs but so good that you should read it anyway.
- [Tuxedo Park](https://www.amazon.com/Tuxedo-Park-Street-Science-Changed/dp/0684872889). Book about the MIT Rad Lab, among other things. (Also worth reading [Endless Frontier](https://www.amazon.com/Endless-Frontier-Vannevar-Engineer-American/dp/0262740222). Broader influence of [NDRC](https://en.wikipedia.org/wiki/National_Defense_Research_Committee) is underestimated, as far as I can tell.)
- [MIT's Building 20: "The Magical Incubator"](https://infinitehistory.mit.edu/video/mits-building-20-magical-incubator). Trannoscript of a talk about [Building 20](https://en.wikipedia.org/wiki/Building_20).
- [Funding Breakthrough Research: Promises and Challenges of the “ARPA Model”](http://mitsloan.mit.edu/shared/ods/documents/?DocumentID=4615). An analysis of what the ARPA model is and why it might work.
- [The Dream Machine](https://www.amazon.com/Dream-Machine-Licklider-Revolution-Computing/dp/014200135X). Book about ARPA, Licklider, and the creation of the internet.
- [The Power of the Context](https://patrickcollison.com/static/files/labs/context.pdf). Alan Kay's reflections on ARPA and PARC.
- [The Making of the Atomic Bomb](https://www.amazon.com/Making-Atomic-Bomb-25th-Anniversary/dp/1451677618/). Book about the Manhattan Project.
- [Skunk Works](https://www.amazon.com/Skunk-Works-Personal-Memoir-Lockheed-ebook/dp/B00A2DIW3C). The Lockheed Martin [facility](https://en.wikipedia.org/wiki/Skunk_Works) behind the U-2, SR-71, etc. (See also: [Kelly Johnson's 14 Rules](https://www.lockheedmartin.com/en-us/who-we-are/business-areas/aeronautics/skunkworks/kelly-14-rules.html), [Kelly Johnson's own memoir](https://www.amazon.com/Kelly-More-Than-Share-All/dp/0874745640/), [Augustine's Laws](https://www.amazon.com/Augustines-Chairman-Lockheed-Corporation-Augustine/dp/1563472406), [Boyd](https://www.amazon.com/Boyd-Fighter-Pilot-Who-Changed/dp/0316796883), and [National Defense](https://www.amazon.com/National-Defense-James-Fallows/dp/0394518241).)
- [Organizing Genius](https://www.amazon.com/Organizing-Genius-Secrets-Creative-Collaboration/dp/0201339897): an exploration of commonalities across the Manhattan Project, Black Mountain College, Skunk Works, etc. [Demis](https://en.wikipedia.org/wiki/Demis_Hassabis) from DeepMind commented that it accords with how he manages the company.
- [Sidewinder](https://www.amazon.com/Sidewinder-Creative-Missile-Development-China/dp/1591149819). A history of the development of the Sidewinder missile and of the [China Lake](https://en.wikipedia.org/wiki/Naval_Air_Weapons_Station_China_Lake) Navy research lab.
- [Scene of Change](https://www.amazon.com/Scene-change-lifetime-American-science/dp/B00005WR83). Personal account from Rockefeller Foundation's Warren Weaver. (Worked with Bush at NDRC during WWII; helped fund Green Revolution; funded most of the Nobel-winning molecular biologists.) Worth a quick skim—some good passages.

- [Alvarez: Adventures Of A Physicist](https://www.amazon.com/Alvarez-Adventures-Physicist-Alfred-Foundation/dp/0465001165). Luis Alvarez's first-hand account of participating in the development of [GCA](https://en.wikipedia.org/wiki/Ground-controlled_approach), radar, and the Manhattan Project.
- [Doing the Impossible](https://www.amazon.com/Doing-Impossible-Management-Spaceflight-Exploration/dp/1461437008). How George Mueller managed the Apollo Program.

Data Freshness in Machine Learning Systems.
When it comes to Machine Learning Systems, we usually plug in on top of Data Engineering Systems where data is already collected, transformed and curated for efficient usage in downstream systems - ML System is just one of them. This does not mean however that no additional data transformations need to happen after data is handed over. We refer to Data Freshness in Machine Learning Systems as Feature Freshness.
When thinking about composition of how data is served to the end user in ML Systems there are two mostly independent pieces, hence also two perspectives on

Feature Freshness:
Feature Freshness at Model Training time: how much time does it take for a generated data point to be included when training a Machine Learning Model which is then deployed to serve the end user. Remember that Machine Learning models are nothing more than Statistical models trained to predict certain outcomes on a given feature distribution. We can’t avoid ML Models becoming stale if not retrained. This phenomenon of ML models becoming stale is called Feature and Concept Drift (you can read more about them here).
Feature Freshness at inference time: how much time does it take for a generated data point to be available when performing Inference with the previously trained and deployed model. Features used for inference are usually decoupled in terms of freshness from the ones that are used while training the model and are less stale.

🥇Top ML Papers of the Week

How Language Models Use Long Contexts

- finds that LM performance is often highest when relevant information occurs at the beginning or end of the input context; performance degrades when relevant information is provided in the middle of a long context. ([paper](https://substack.com/redirect/4e6b797d-9aed-4940-88c7-3af5b63e4f20?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/3a9b6a9f-fc9e-40a0-b172-f779c899bacf?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))

LLMs as Effective Text Rankers

- proposes a prompting technique that enables open-source LLMs to perform state-of-the-art text ranking on standard benchmarks. ([paper](https://substack.com/redirect/7782bdfe-6f9c-4c37-87da-f353da8a7a7f?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/dfbde4a7-b3ae-4df2-ac6d-8283361c3ad3?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))

Multimodal Generation with Frozen LLMs

- introduces an approach that effectively maps images to the token space of LLMs; enables models like PaLM and GPT-4 to tackle visual tasks without parameter updates; enables multimodal tasks and uses in-context learning to tackle various visual tasks. ([paper](https://substack.com/redirect/8377a115-b5c0-4a05-80c3-821099b7ccbf?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/56122f84-fd95-4fb3-bc68-2239ef4ba411?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))

Elastic Decision Transformer

- introduces an advancement over Decision Transformers and variants by facilitating trajectory stitching during action inference at test time, achieved by adjusting to shorter history that allows transitions to diverse and better future states. ([paper](https://substack.com/redirect/2b1b7dcb-4143-465f-9cd4-aba578c73279?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/adc07e2e-3c86-423c-aa6b-6e4ab3ed2a0a?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))

Physics-based Motion Retargeting in Real-Time

- proposes a method that uses reinforcement learning to train a policy to control characters in a physics simulator; it retargets motions in real-time from sparse human sensor data to characters of various morphologies. ([paper](https://substack.com/redirect/d7cf6278-7ebf-42f2-9d21-6f598e29cd1e?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/88b736f9-b052-4aa4-be5a-697533fa2d94?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))

InterCode

- introduces a framework of interactive coding as a reinforcement learning environment; this is different from the typical coding benchmarks that consider a static sequence-to-sequence process. ([paper](https://substack.com/redirect/48889a92-d287-4fd2-87e0-3b72f395c3ed?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI)|[tweet](https://substack.com/redirect/e4f3aeb7-b2d3-49a9-9b32-9d0a842dd7f4?j=eyJ1IjoiMWRheDAifQ.YVDAzsk3G87vydkjsTF3WruaemlL7xgZ83byJs8O8dI))