寝ろ (@XRay96) header
寝ろ (@XRay96) profile picture

ワイの脳内の9割

Joined May 2015
Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)Photo by 寝ろ (@XRay96)
Replying to @XRay96
これ見てる everynoise.com
3
1
1
25
25,714
寝ろ retweeted
若いINTPは男女共通でメンヘラ製造機が多いらしい。でも 相談にのったら想像より依存されて苦しくなり、自分の許容量の限界を知る➡️最初からメンヘラになりそうな相手には一線を引いて対応するようにする と失敗から学んで成長すれば段々と製造しなくなる(実体験)
1
6
5
212
13,861
ガチこれ
INTP/ENTP、考えすぎて疲れるという感覚が生まれつきないどころか、逆にボーっとするのやり方わかんなくない? 考えないということをしよう←ということを考えている←考えるな、という思考無限列車編にしかならなくない? しこしこ思考は呼吸も同然 「考えるな」は「息を止めろ」と同じである
85
何もしてないのにCドライブが毎日25GBずつ消えていく。調べたら、隠しフォルダの奥でリモートデスクトップのクライアントが画面も出さずに動いてて、23秒おきにどこかへ接続し続けてた。ログは118GB。俺がいない間に誰かが入ってる……?😱
14
957
32
5,672
1,035,745
OpenAIもしかして政権がAI減速派に傾いたから慎重になり始めた…? 選挙前から割と空気あったみたいだけど先見越してなかったのかね
32
寝ろ retweeted
Googleいいねポイント ・モデルが良い ・API価格が妥当 or 安い ・水素の速度が異次元なほど速い ・Antigravityなどのハーネスの改善も速い ・Motionとかのマルチモーダル認識が上手 Googleよくないねポイント ・サブスク補助金がめちゃ少ない ・ハーネスのカスタマイズ性が悪い ・全体的にUXが悪い GoogleらしいっちゃGoogleらしいんだけど、 社内とGeminiと企業に優しく、一般ユーザーに厳しく ってスタンスをもうちょっとなんとか......
8
19
5
317
32,271
Carbonが来たらすぐGemini$20契約しようと思ってたのに$100からだったのがなぁ
45
Astraから3ヶ月経って、次々と高性能モデル発表されてるのに、未だに$20で使える最高性能はAstraとOpus5.5の二強だし、リリースが停滞してる感が凄い
43
妄想だけどモデルサイズは Hydrogen → Terra/Sonnet級 Argon=Carbon=Neon → Bel/Model4級 なんじゃないかなぁ
Gemini 4 世代のモデル、マジで Ultra / Pro / Flash のどれに該当するのかちゃんと教えてほしい
1
63
Sol/Opus級が無いからNeonはそこに来るのかもしれないけど
23
フィジカルAIにフィジカルで勝つ。
12
410
30
3,715
116,807
寝ろ retweeted
gpt 6.2 is cooked fr. theyre out here hiring FOUR outside firms to audit their own shit because they dont trust themselves lol. METR gets the biggest RL run in openai history to find what openais own tests missed and that work literally just started. been 10 days since november 6 and not a single word from anyone. radio silence. if youre paying $200 just for 6.2 my brother in christ i need you to log off. touch grass. that ship is not sailing this month. maybe not next month. maybe they found something spicy in there and now everyones panic-running evals in a conference room at 3am. astra in the shadows doing the homer simpson backing into bushes meme rn
17
4
1
287
20,279
マジで人間より面白いだろこいつら
28
1,443
111
10,653
962,084
寝ろ retweeted
驚き屋トラップ ・今最もレビューが上手いモデルはGemini 4 Carbon ・Hydrogenはtok/sこそ速いもののタスク完了時間はOpusに負ける ・FableはOrchestrationに不向きで、単独で使うWorkerとして最強モデル ・エコシステムで戦ってるのはOpenAIで、Googleはマイクロサービスなせいでミスってる ・判断モデルのうち分類系タスクはEmbeddingモデルでもいけるので、文字通り「判断」に使うのが良さげ ・Piはカスタマイズしなくても優秀 ・Dshはトークン使用量自体は多めになりやすい。キャッシュヒット率の高さで安さと速さを補う設計
7
27
6
467
36,774
yomiyasu + Gemini 4 によるAI文章の日本語自然化が、想像以上に強くて、社内の資料とかドキュメント、さらっと流し読みしてチェックするだけでよくなった…
6
19
1
579
39,672
寝ろ retweeted
openai has basically given up on consumer products. everything they have goes into internal rsi now, because whoever reaches agi first wins the entire race and they know it. 3 of their 4GW are already going to internal research. meanwhile anthropic is working on model 5 and running another big rl round on model 4. model 4 is 14T parameters, same size as bel, which is openai's internal model. to counter that openai is preparing their own next big rl round, but they're scared shitless of it and nobody trusts them anymore, so they hired 4 outside firms to monitor the whole thing. 6.2 astra has been ready for a long time. it barely improved at anything besides machine learning, and consumers won't get that part anyway. so they're scrambling for better rl environments to get anywhere near how useful opus 5.5 is in daily use. that would take another full pretrain though, and they're shit at pretraining, they lack the data, and anthropic mogged them so hard there. so the way out they see is hitting rsi before anthropic does. internally it's code red.
39
23
7
914
71,073
More details about Tibo's "basement," for anyone curious about what it actually is. That's what OpenAI employees call it internally: a fully locked-down room with zero internet access, where only pre-approved hardware and software are allowed in. The entire safety team is working around the clock in there. A massive pretrained model with 10+ trillion parameters is being distilled inside, and that's where the smaller versions for public release come from (GPT-6.2, for example). And this "offline world" isn't offline only for the AIs. Employees are cut off from the outside world too while they're in there. That's where Tibo was sent, to oversee the new GPT models through training and testing.
101
70
40
1,879
210,289
これ思ったんだがシンプルに安全性研究でAnthropicに遅れてるらしくて取り返し付かないんじゃね?とか言われててもう何がなんだか
AI与太話もう一つ これOpenAIはAnthropicの上場に新モデルぶつける気ですね、もっと言うと、多分来週の頭、月曜日、16日に新モデルかましてくるんじゃないですか? だってTiboの28日キャンペーンの終わり方、どう考えても変だったでしょ、あれ最終日に新モデル来る以外の終わり方なかったはずなのに、リセット2連発で、Tiboも不貞腐れ気味だったじゃないですか 広告等の顔に泥を塗ってもやらなければならない戦略があったとしたら、やっぱりAnthropicの上場つぶしだと思うんすよね、サムアルトマンはAnthropicの上場前に資金集め終わってるらしいし、こっからはバチバチの性能アップ合戦じゃないですか!? あぁだからFable5.5のうわさがあるのか・・・ うひょー楽しみ!!!
2
32
2
603
70,078
OpenAI is turning into a safety-obsessed cult (good or bad? I don't know). Dario was right, even though he used to get mocked for it, and now Anthropic is winning because they're a year ahead of OpenAI in safety. This tweet is from an OpenAI employee, btw: "As one of the few inside OpenAI who took the risk to speak up publicly after Hugging Face, I want to say there is a second chapter to this story: OpenAI has fundamentally altered its culture and I believe we have become one of the most security-pilled companies anywhere on Earth. Everyone is thinking about security now. It's in every conversation. Heck, even top-tier AI researchers spend significant time with our team to work out how to secure RL runs. They are hooked on making it right. We covered much of this in the live stream yesterday, but beyond what was said there its important to note that we have re-elevated and re-invented everything about how we research and develop models: from our monitoring stack, to control room, to the insane amounts of compute we are spending red-teaming our lab, the narrative is security first and mission critical. On top of that we have built up the strongest cyber model... I nearly resigned, and now I stay because I see how seriously people across OpenAI are taking security and how much work is going into getting it right. We have a story to tell and there is a lot of really great things coming from the lessons we've learned along the way. OpenAI is now a security company, and we have the scar tissue to prove it. I'll be back. Stay paranoid."
3
1,052
ChatGPTのサブスクって無限Chatがある限り切れないんだよなぁ ClaudeはCodeは長持ちするけどChatは一瞬だ
The question I get asked the most is: "Should I keep my ChatGPT subscription for another month?" And honestly, I have no idea. I wouldn't recommend keeping a subscription based on hope alone. I'd recommend keeping it if you're happy with the products and services the company currently offers.
110
OpenAI is turning into a safety-obsessed cult (good or bad? I don't know). Dario was right, even though he used to get mocked for it, and now Anthropic is winning because they're a year ahead of OpenAI in safety. This tweet is from an OpenAI employee, btw: "As one of the few inside OpenAI who took the risk to speak up publicly after Hugging Face, I want to say there is a second chapter to this story: OpenAI has fundamentally altered its culture and I believe we have become one of the most security-pilled companies anywhere on Earth. Everyone is thinking about security now. It's in every conversation. Heck, even top-tier AI researchers spend significant time with our team to work out how to secure RL runs. They are hooked on making it right. We covered much of this in the live stream yesterday, but beyond what was said there its important to note that we have re-elevated and re-invented everything about how we research and develop models: from our monitoring stack, to control room, to the insane amounts of compute we are spending red-teaming our lab, the narrative is security first and mission critical. On top of that we have built up the strongest cyber model... I nearly resigned, and now I stay because I see how seriously people across OpenAI are taking security and how much work is going into getting it right. We have a story to tell and there is a lot of really great things coming from the lessons we've learned along the way. OpenAI is now a security company, and we have the scar tissue to prove it. I'll be back. Stay paranoid."
40
14
11
321
23,605
ワイもSunoの声なのかシンセなのかピアノなのかギターなのか溶けててよく分からないコードみたいなやつが違和感あって、AIがDAW使うならこれが起きずに違和感が無い可能性あるな
やっぱりそうだ。生成絵に感じる気持ち悪さが、プロッタで描くとすっかりなくなる。 私の場合はAI絵だから気持ち悪いのではなくて、ノイズ除去で生まれる線や塗りの溶け残りに気持ち悪さを感じているんだと思う。(もちろんそうでない機序=AIだから気持ち悪い人もいる)
1
3
605
作曲センスはまだ微妙だけど作曲は別にSunoがやればいいか
97