Недавно делал быстрый прототип асинхронного приложения в котором требовалось вызывать много синхронного кода. Да, я знаю, что это не лучший дизайн, но нужно было быстрое решение на один процесс и без очередей. Поэтому я выполнял код в потоках.
Выглядело это примерно так:
from fastapi.concurrency import run_in_threadpool
async def execute(data: DataRequest) -> DataResponse:
try:
result = await run_in_threadpool(sync_function, data)
return DataResponse(data=result)
except Exception as e:
return DataResponse(
error=str(e),
success=False,
)
В общем работает нормально. Для всех вызовов под капотом используется общий тредпул, всё работает предсказуемо.
Но потребовалось изменить количество запускаемых в пуле потоков (по умолчанию создается 40 воркеров).
Так как дело происходит с FastAPI, делается это через lifespan используя настройки anyio:
import anyio
@asynccontextmanager
async def lifespan(app: FastAPI):
limiter = anyio.to_thread.current_default_thread_limiter()
limiter.total_tokens = 100
yield
# если вдруг нужно вернуть обратно
limiter.total_tokens = 40
Зачем менять количество воркеров?
- уменьшить, если оперативки мало (один тред занимает ~8мб)
- увеличить чтобы выдержать нагрузку
Если есть предложения получше при тех же вводных - предлагайте😉
#async
🇺🇸Anthropic Reached Agreement with Music Publishers on AI and Copyright
Anthropic has reached a settlement with major music publishers, including Universal Music Group and ABKCO, over allegations of copyright infringement by its Claude AI model. At the heart of the dispute were claims that Claude had been trained on lyrics from over 500 protected songs and reproduced substantial portions of these works without authorization. The publishers argued that Anthropic bypassed copyright protections by using data scraped from licensed lyric platforms.
As part of the agreement, Anthropic committed to maintaining and expanding its AI guardrails to prevent copyright violations in future models. The company also established a process for publishers to address concerns directly. This settlement underscores a broader tension between generative AI and copyright law, as Anthropic maintains its practices fall under fair use. Meanwhile, publishers have requested a preliminary injunction to prevent further training on copyrighted lyrics, with a ruling pending.
#AIandCopyright#GenerativeAI
🇪🇺European Parliament Pushes New Copyright Safeguards for AI
The European Parliament’s Committee on Legal Affairs has approved an own-initiative report proposing new measures to strengthen copyright protection in the context of generative AI. The proposals include mandatory transparency obligations on training data practices for all generative AI systems placed on the EU market.
MEPs also call for remuneration mechanisms for rightsholders and the development of voluntary, sector-specific collective licensing agreements. The report is not legally binding but signals political direction and will be submitted for a vote by the full Parliament during the March plenary session.
#AIandCopyright#EUAI#GenerativeAI#CopyrightLaw#AIRegulation
SFWA Advocates for Author Rights Amidst Copyright Challenges in AI Era
The Science Fiction Writers Association (SFWA) is taking a stand for authors' rights in the digital age. In a recent submission to the US Copyright Office on December 7, the SFWA emphasized its role in representing authors who have generously shared their work on the open internet. This move comes as technology, particularly AI, exploits content in ways that challenge the rights and compensation of creators.
Over the past two decades, many science fiction and fantasy authors have embraced the open internet to share their work freely, believing in the societal and cultural benefits of accessible art. However, the SFWA argues that making a work freely available doesn't imply abandoning authors' moral and legal rights. It stresses the importance of authors' rights, including the obligation to enter into legal contracts that ensure fair compensation and define how their work should be used.
The SFWA's submission coincides with the US Copyright Office's inquiry into copyright and artificial intelligence launched in August 2023. This inquiry, designed to gather public comments, has seen a significant response, with over 10,000 comments received.
#SFWA#AIChallenges#DigitalCulture#IntellectualProperty#AIandCopyright
🌟 AI Sunday Wonders: New Tool Empowers Artists in the Fight Against Unauthorized AI Training
Hello everyone! This Sunday Here's the latest from the world of AI and art.
Meet "Nightshade," a groundbreaking tool enabling artists to safeguard their work from being used in AI model training without their consent. By subtly altering the pixels in their artwork before posting it online, artists can introduce invisible changes that, when incorporated into AI training data, can cause unpredictable distortions in the AI-generated output.
This tool is designed to challenge AI companies that exploit artists' creations without authorization to train their models. The result? Future iterations of image-generating AI models may produce bizarre results: dogs might become cats, cars could transform into cows, and so on.
Nightshade aims to shift the balance of power back to artists. It serves as a potent deterrent against AI companies that disregard artists' copyrights and intellectual property.
#AIandArt#ArtProtection#AIInnovation#NightshadeTool#AIandCopyright