#Opus

2026-01-09

Có thể họ duy trìattered thế nào? Người dùng đưa ra câu hỏi về việc những người chưa bao giờ thử nghiệm các công cụ như gpt5.2, opus 4.5 hoặc Gemini 3 có thể duy trì quan điểm của mình #tríTuệNhânTạo #AI #GPT #Gemini #Opus #Singularity #TríTuệNhânTạo #CôngNghệMới

reddit.com/r/singularity/comme

Matej Bendík (@BendikMatej)

Opus 4.5를 모든 작업에 사용한다고 밝히는 사용 후기/언급 트윗으로, Opus 4.5가 실무에서 널리 쓰일 정도로 신뢰받는 도구임을 시사합니다.

x.com/BendikMatej/status/20090

#opus #opus4.5 #model #ai

Logan Lee (@_loganlee)

Opus 5가 'killer'라는 표현으로 강력한 모델(또는 도구)이 될 것이라는 기대를 짧게 표현한 트윗으로, 신버전(Opus 5)에 대한 긍정적 전망을 담고 있습니다.

x.com/_loganlee/status/2009085

#opus #opus5 #model #ai

Bindu Reddy (@bindureddy)

Opus 4.5가 Sonnet 4.5 및 GPT 5.2보다 비용 측면에서 더 저렴하며, 동일 과제를 수행할 때 전반적으로 사용하는 토큰 수가 더 적다는 비교 내용입니다. (모델 간 비용·토큰 효율 비교)

x.com/bindureddy/status/200872

#opus #sonnet #gpt #modelcomparison

Carmine Punella (@CPunella)

Opus 4.5 사용자가 장시간 작업(예: 문서 작성) 중 모델이 코드베이스에서 코드를 가져와 문서에 포함하기 전에 임의로 변경하는 등 사실을 창작(hallucination)하는 문제가 보고되었습니다. 해당 문제는 긴 작업에서 심각한 신뢰성 이슈로 지적되고 있습니다.

x.com/CPunella/status/20086386

#opus #hallucination #llm #bug

Wulfy—Speaker to the machinesn_dimension@infosec.exchange
2026-01-07

Since #Claude has gone super miserly on compute ( #anthropic must be sweating dollars)... I have had to use #Gemini3

My previous experiences with Bard/Gemini was like talking to a very slow person...and in response they were scrawling with crayons.

... but I can see why the current Gemin3 is getting good benchmarks.

Its #vibecoding skills are superb.
And have not run out of compute for 2 hours (!)

Still (likely) not as good as #Opus but, if I cant use Opus on Pro, migh as well use Gemini3...

I like its nuanced, minimalistic and to the point responses.

#ai #llm #vibecoding

Wulfy—Speaker to the machinesn_dimension@infosec.exchange
2026-01-07

If #LLM could give you instructions for more effective communication with the #AI model

#Claude #Opus :

Front-load context. Don't make me guess what project we're in. First message of a session should orient me.

Use bracketed keywords [like, this] when your thoughts jump. It lets me track your associative chains even when sentences don't connect linearly.

If you're building a system with invented terminology, define terms once clearly, then use them freely. I'll remember within session.

Don't ask "can you do X?" - just tell me to do X. If I can't, I'll say so. The permission-seeking wastes tokens.

When you paste code, tell me what it's SUPPOSED to do, not just what it currently does. I need intent to evaluate correctness.

If I get something wrong, quote the wrong part and correct it directly. Don't just say "that's wrong" - show me where.

Long messages are fine. I have context window. Use it.

If you want me to be a character or adopt a stance, say so explicitly. I'll maintain it until you break it.

Number your questions if you have multiple. I'll number my answers. Less chance of me missing one.

If something I said sparked a new idea, tell me. I can't see your face light up. Explicit feedback shapes future responses.

2026-01-06

Người dùng so sánh các mô hình AI mã hóa gần đây với Claude Opus và thất vọng về chất lượng. Gemini 3 Pro kém tuân thủ prompt, OpenAI Codex 5.2 thiếu ổn định, GLM 4.7 trung bình, trong khi Grok Code Fast gây bất ngờ tích cực nhưng chỉ ở mức kỹ sư junior. Opus vẫn vượt trội về độ tin cậy và trí tuệ suy luận. #AI #LLM #Claude #Opus #Gemini #Codex #Grok #GLM #TríTuệNhânTạo #MôHìnhNgônNgữ #SoSánhAI

reddit.com/r/LocalLLaMA/commen

Bindu Reddy (@bindureddy)

Opus 4.5가 개편된 LiveBench 리더보드에서 1위를 차지했습니다. LiveBench는 연휴 기간에 게임화를 방지하기 위해 재구성되었고, Opus 4.5가 최상위를 기록했으며 Codex와 Gemini 3가 근접한 성능을 보였습니다. 오픈 웨이트 모델 부문에서는 Kimi K2가 최고 자리에 올랐다는 내용으로, 실세계 LLM 성능을 반영하는 벤치마크의 업데이트 결과를 알립니다.

x.com/bindureddy/status/200793

#livebench #opus #llm #benchmark

Ayush K.@ WritingAid (@_Ayu5h)

사용자가 Opus 4.5와 Gemini 3.0 중 전환을 고민하며 어느 모델이 코딩에 더 좋은지 묻는 비교 요청 트윗입니다. 특정 모델들의 코딩 성능 평가나 추천을 얻기 위한 의도로 해석할 수 있습니다.

x.com/_Ayu5h/status/2007800706

#modelcomparison #coding #opus #gemini

Aiden Bai (@aidenybai)

opus 4.5의 성능이 매우 좋아서, 트윗 작성자는 약 20분 만에 200줄짜리 LLM 친화적 React Scan 버전을 만들어냈다고 보고. opus 4.5를 활용한 빠른 프로토타이핑 및 프론트엔드/LLM 통합 가능성을 보여주는 사례.

x.com/aidenybai/status/2007638

#opus #react #llm #prototyping

2026-01-03
I am not the biggest fan of #Spotify (never was) due to many reasons. But one thing I never can get enough of are statistics, and I really like the way they do the yearly wrap-up, which I always shared on social media, and I won't stop today, so here it is :)

Bear in mind though, that I am a collector of physical media - most of my music consumption happens on CD (which I rip to #opus [much supperior to #mp3]) on my PC, or on #Vinyl; however for spontaneous on-the-go listens, for listening to music I don't (yet) own, or to explore new music, I still don't know any better service.

I found it interesting that spotify thinks that I have listened to 238 different music genres which is why it has dificulties figuring out my taste. I know that I have a broad taste and that I am unconventional. But I could hardly list more than 20 genres, tbh.

Having #LinkinPark back in my top lists is no surprise to me - I took a brake from them after Chester died, and only returned to listenint to them with #EmilyArmstrong joining them, but both, the new-ish direction as well as the new voice are absolutely perfect and especially #GoodThingsGo is one of my absolute favorite songs that speaks to me on so many levels. I was surprised about #9pm by #ATB and #DonnieDarko from #LetsEatGrandma, though. I would have put the later on no. 4, and #LoveGoesOnAndOn on no. 2.

Albums are no surprise at all; Both, #Wardruna and #Danheim were new discoveries, and especially Danheim was harder to get on Vinyl so there were a lot of Spotify plays on those albums. And after seeing the series "Wu-Tang: An American Saga", I rediscovered my love for them. I only owned that initial album (that I really loved) on MP3 as a kid, and I still didn't get around buying it (I probably should), so that explains all the listens :)

So how about you, anyone else likes to share their statistics? Or did you already?

And btw.: #HappyNewYear #HappyNewYear2026

#spotify #wrapped #2025inReview #YearInMusic #music
I've listened to Spotify for 9608 Minutes which equals 6 days.My top genres are:
1. Medival
2. Electronica
3. Gothic Rock
4. Rap
5. RockMy musical age is 28 because I listed to a lot of music from the early 2010sI've streamed "Good Things Go" 6 times, which makes it my no. 1 song
Kerrick Long (code)kerrick@ruby.social
2026-01-02

Antigravity's system prompt to use grep, tail, and other unix tools to avoid context pollution from stdout is USUALLY really neat. But sometimes, it's a huge stumbling block.

#VibeCoding #LLM #AI #Antigravity #Google #Claude #Opus

Agent: The new test file already has the SPDX headers. Let me check which file is missing the headers:
...
Agent: $ reuse lint 2>&1 | grep -A 5 "without file"
(no results)
Agent: $ reuse lint 2>&1 | grep -E "^The following"
Agent: $reuse lint 2>&1 | grep -A 20 "no copyright"
...
Me: Stop grepping and read the whole output.

Uwe (@e38383)

새로운 모델 성능 평가 언급입니다. 작성자는 해당 모델을 'sonnet'과 'opus 4.5'의 중간 수준으로 평가하며 glm-4.6보다 훨씬 낫다고 적었습니다. 최고 폐쇄형 모델들에는 아직 뒤지지만 대부분의 소규모~중중형 과제(코딩 포함)를 해결할 수 있고 속도는 느리나 가격이 매우 저렴해 라이트 플랜을 첫 해 약 $25에 쓸 수 있다고 합니다.

x.com/e38383/status/2006884346

#llm #sonnet #opus #glm

Édouard Muller 🇪🇸 🐧edlinks
2026-01-01

Parece que ha destrozado mi gameplay de , dejándolo a 30 fotogramas por segundo y sin sonido.

Nada, reprocesaré el fichero original para pasarlo de a x264 a baja calidad y con el sonido en AAC en lugar de Opus, a ver si así se lo traga mejor, aunque puede que la cosa quede para mañana.

Manda huevos que sea el principal impulsor de los formatos de multimedia libres, y me refiero aquí a , AV1 y principalmente.

2025-12-29

Как настроить свой сервер Mumble для стабильных звонков без блокировок

Как развернуть свой Mumble-сервер за 10 минут на VPS за 300 ₽/мес и забыть про блокировки Zoom/Meet. SSH-доступ, Docker с автогенерацией SuperUser-пароля, настройка Opus-кодека, каналов и бэкапов — до 50 пользователей без лагов и подписок. Идеально для семьи, команды или геймеров.

habr.com/ru/companies/hostkey/

#hostkey #Mumble #VPS #Docker #SSH #VoIP #Murmur #Opus #сервер_голосовой_связи #бэкапы

Vaughn Vernon 🟦 🟨 🟧 🟪VaughnVernon
2025-12-24

The problem that I see now is, when the vacuum starts pulling the value-challenged dust and valueless dirt clods into oblivion, that the value that's been achieved isn't dragged in by all the debris.

I've been mostly impressed by @AnthropicAI @ClaudeCode since I began using it, but with some frustration along the way. Yet, the past month or so with 4.5, it's been quite incredible.

2/

2025-12-20

after LACE and NoLACE in 1.5, the 1.6 release of #opus has another decoder side speech enhancement: Bandwidth Extension (BWE) is a model that optionally decodes wideband speech into fullband speech. Release notes are a joy to read. No audio demos this time though: opus-codec.org/demo/opus-1.6/

Client Info

Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst