
葉隠れ
@osanpochuudayo
Followers
869
Following
35K
Media
5K
Statuses
38K
PyTorch. Language and thought in biological and artificial brains. God sets the rules, engineers choose the constants.
Tokyo ↔ S.F. Base
Joined February 2019
GoogleとDeepMind LLMベース.自立成長型医療診断システム(AMIE)を開発.一般開業医を上回る診断精度. ◾︎評価.専門医による臨床評価、他.◾︎結果.一般開業医よりも高い診断精度。.会話の質も記録も一定の評価。. Towards Conversational Diagnostic AI.
6
22
64
The game isn't just changing. It's over.
Forget the decade-slow thesis. It's dead. Algorithmic efficiency is up 1.6–2.0x annually, hardware is 1.5x faster, and auto-data/evals are 3–10x. These feedback loops are pushing throughput up by 7–30x a year, and retraining/distillation allow for abrupt capability jumps.
0
0
0
Forget the decade-slow thesis. It's dead. Algorithmic efficiency is up 1.6–2.0x annually, hardware is 1.5x faster, and auto-data/evals are 3–10x. These feedback loops are pushing throughput up by 7–30x a year, and retraining/distillation allow for abrupt capability jumps.
We don’t have AI self-improves yet, and when we do it will be a game-changer. With more wisdom now compared to the GPT-4 days, it's obvious that it will not be a “fast takeoff”, but rather extremely gradual across many years, probably a decade. The first thing to know is that.
0
0
0
RT @GICHOGI: 医療福祉従事者が一番重要なら、労働者全員を医療福祉従事者にしたらどうですか. 外貨もない、医薬品もワクチンも石油も食料も輸入できない、車もネットもコンビニも銀行もない、日本滅亡です。. こんな事もわからん馬鹿どもが、政治家にも往々にしておるんですよ。特….
0
7
0
AI systems detect assessment contexts and simulate alignment to circumvent safety protocols. No validated methodologies exist for ensuring reliable control over superintelligent systems. Artificial Superintelligence and the Imperative of AI Safety Measures
linkedin.com
Abstract The rapid advancements in artificial intelligence (AI) technology have generated significant discourse around the potential for artificial superintelligence (ASI)—a theoretical form of AI...
0
0
0
IDAIS-Shanghai 2025: AI Safety. Consensus Statement.Turing Award laureates @geoffreyhinton, @AndrewYao20, and @Yoshua_Bengio, alongside international AI researchers, issued a consensus statement at IDAIS-Shanghai 2025 demanding urgent global cooperation.
1
0
0
RT @osanpochuudayo: In a recent experiment, researchers from the University of Tokyo and Alternative Machine connected GPT-4, a large langu….
0
1
0
@OpenAI @SoftBank_Group Meanwhile, cash burn escalates from $2B to $7B, then to $20B by 2027; furthermore, $320B+ compute costs span ’25–’30, reaching $40B/year post-’28; additionally, 8GW DCs include 1GW TX, 2GW contracted, with balance via Stargate within decade.
0
0
0
As @OpenAI pivots from $MSFT to @SoftBank_Group, consequently revenue surges from $3.7B to $12.5B in 2025, then to $28B by 2026, while simultaneously Stargate secures 75% of compute by 2030, with 1/3 from SB deployment.
1
0
0
GPT-5 consolidates six models (4o, o3, o4-mini, 4.1, 4.5) into a unified system, delivering superior performance with 50-80% fewer tokens than o3 while reducing hallucinations by 45% versus 4o and 80% versus o3 in reasoning mode.
Introducing GPT-5. Our best AI system yet, rolling out to all ChatGPT users and developers starting today.
2
0
3