余ってるAMDグラボがあれば嬉しいのか…？！
なんかまたゲームチェンジャーな技術が出てきてそう。「アダプティブKVストリーミング」だって。要するにKVキャッシュをメインメモリにオフロードできるという。今まではコンテキスト長を大きくするとKVキャッシュがVRAM領域を食いつぶしてしんどかった。RedditでLlama.cppフォークにこの技術を実装した人によれば16GBのVRAMで普通だとQwen3.8-27Bでコンテキスト6.5万までしか使えてなかったのがアダプティブKVストリーミングだと13.1万に増やしても速度低下はちょっとで済んでる。7.4GBのKVキャッシュを2.1GBのVRAMでやりくりする形。26.2万だとさすがに速度1/3くらいになるけどまだ一応動く。しかしこれだと1トークン推論するたびにKVキャッシュ全スワップ発生してPCIe帯域は食われそうだね。Strataではすでに似たようなKVオフロードが実装されてるらしいけどエキスパートのスワップとPCIe帯域を食い合うハメになるでしょう。それにしてもこのスレでも最適化されてるのはROCm。Strataもそうだが最近はAMDグラボの救済されっぷりがすごい
じゅんぴー retweeted
Context about how I got CVE-2026-99999: it wasn't assigned by incrimental luck. There were nearly 2k unassigned CVEs before it. The Linux maintainers felt it deserved the special number because of impact+new attack surface for KVM escapes and offered it. It was a prize! I was just telling someone it feels like graduation from cybersecurity. So yeah it's intentionally assigned, and I'll be eternally grateful for it.
じゅんぴー retweeted
official now: anthropic is preparing the release of their new model(s). Via The Information
Anthropic is reportedly preparing a model that, when asked to fix one bug, can identify the broader class of problems behind it and ask whether it should address those too. That account comes from one anonymous tester; the model family and release date remain unclear.
But looks like it will be a significant jump compared to their released models.
ちなみに、クォータが若干少ないって免責事項に書いてあってくそおもろいんだよね
one.google.com/offer/student…
最終更新日: 2026年8月19日
Google might have pulled a prank on its Pro users.
If you got Google AI Pro through:
- Pixel purchase
- Jio bundle
- Student offer
You will not be able to access Gemini 4 Argon in Antigravity.
And the catch? You apparently can't switch to a regular paid Pro subscription with the same account until your existing offer expires.
2 questions for Google:
1. If Pro is Pro, why are users on these offers excluded from features advertised to paid Pro users?
2. How can existing users upgrade to paid Pro without waiting a year?
This is ridiculous.
ハンドラに不具合が〜みたいなことをOpusが言ってて、これpatch当ててPRを出しましょうみたいに言ってくるんだけど、それって本当に不具合か？って聞いたらそうじゃないって言い出して、あーもうこうやってコードベースは汚れていく
こういうモデルが欲しいです
Asked Opus 5.5, GPT-6.1 Sol, and Muse Wonder Full to fix a bug
Opus found the root cause and filed a pr
GPT-6.1 Sol asked a question, crashed from the async question tool interrupt, found the bug, fixed it, wrote 300 lines of tests, and explained it in impossible to understand terms.
Muse Wonder Full found a small PR someone else had filed and said "you should probably merge this instead"
I really like this model.
