ラベル

AI Forecast 8306MUFG (319) Social (79) life (75) Financial (69) IT (62) Proposal (57) idea (51) Fund Management (49) AI (48) Trading (43) economics (43) Covid-19 (40) Hedge Fund (37) Risk (36) Culture (31) BOJ (29) Science (29) hobby (27) accounting (17) Apps (16) Travel (14) career (11) job (7) Hiking (5) emotion (5) music (3) Statarb (2) piano (2)

2025年2月23日日曜日

LLMの成功理由 Reasons for the Success of LLMs

 LLMの成功理由

Reasons for the Success of LLMs

(English Follows)


人工知能を作ろうとして、多くの言語アプローチなど様々な論理的な研究がなされてきたが、LLMに勝る結果は得られなかったのは事実。それをLLMが次の言葉を推測するのみで実現してしまった根本原因は、論理的に解明されて論文にされているか?
ChatGPT:

LLM(大規模言語モデル)が、単に次の単語を予測するタスクによって高度な「知性」のような振る舞いを実現していることは、現在も多くの研究者によって分析・議論されています。しかし、その「根本原因」が完全に論理的に解明されたとまでは言えません。いくつかの主要な見解や論文は存在しますが、LLMの「知性」や「理解」の本質を巡る議論は、依然として進行中です。

なぜLLMが成功したのか?

  1. スケーリングの法則:

    • OpenAIのScaling Laws for Neural Language Models (2020) などの論文では、モデルサイズ、データ量、計算資源を増やすことで、予測性能がほぼ一貫して向上することが示されています。
    • この「スケーリング」が、特別なアルゴリズム変更なしに、言語理解や推論能力の向上につながったことが一因です。
  2. 自己回帰モデルの力:

    • LLMはトランスフォーマー(Transformer)アーキテクチャを使用し、自己注意(Self-Attention)メカニズムによって文脈を捉えることができます。
    • これにより、単なる次の単語予測以上に、文脈や意味、さらには因果関係のようなものまで「推論」しているように見えることがあります。
  3. 言語の統計的性質の活用:

    • 言語には強い統計的なパターンが存在します。LLMは膨大なデータからこのパターンを学習し、人間が手作業で設計するよりも優れた「ヒューリスティック」(heuristic)を見つけています。
    • これにより、知識ベースに基づく手法や論理ベースのAIでは捉えきれなかった曖昧さや多義性にも対応できています。

しかし、根本的な「理解」はあるのか?

  • 「シミュレーション仮説」:

    • LLMはあくまで「次の単語を予測する」ことを学んでいるだけであり、実際に「理解」しているわけではないという見解があります(例: Emily Benderらの「Stochastic Parrots」論文, 2021)。
    • この見解では、LLMは「シミュレーション装置」に過ぎず、人間のような意味や意図を持っているわけではないと主張されています。
  • 「出現的能力 (Emergent Abilities)」:

    • 近年、特にGPT-4やClaude、PaLM 2などのLLMで、学習タスクに設定されていない推論や計画、さらには創造的なアウトプットが観測されています。
    • Wei et al. (2022) の「Emergent Abilities of Large Language Models」では、モデルサイズが閾値を超えると、事前に期待していなかった能力が「出現」することが報告されています。
    • しかし、この「出現」のメカニズムは、理論的に完全には解明されていません。

論理的な解明は進んでいるか?

  • 暗黙的な知識 (Implicit Knowledge):

    • LLMが言語データを通じて「常識」や「専門知識」を獲得しているのは、主に分布的意味論(Distributional Semantics)に基づいています。
    • 例えば、Petroni et al. (2019) の「Language Models as Knowledge Bases?」では、LLMが特定の質問に対して、まるで知識ベースのように正確な回答を返すことが示されています。
    • ただし、これが「知識の理解」なのか、単なるパターンマッチングなのかという点は曖昧です。
  • 計算神経科学・認知科学の視点:

    • 一部の研究者は、LLMの挙動を人間の脳内の予測符号化(Predictive Coding)やベイズ的推論(Bayesian Inference)と関連づけようとしています。
    • 例: Karl Fristonの予測処理モデルや、Chris Summerfieldの「自然言語処理における予測処理」など。
    • これらは、LLMが単に「次の単語を予測する」だけでなく、ある種の内部モデルを構築している可能性を示唆していますが、決定的な理論的証明には至っていません。

結論

現時点では、LLMが「知性」や「理解」のような高度な能力を、なぜ予測タスクのみで示せるのかについて、完全に論理的に解明されたとは言えません。しかし、多くの理論的アプローチや仮説が提案されており、今後の研究でさらなる解明が期待されています。



Many logical approaches, including various language-based methods, have been explored in the quest to create artificial intelligence, but none have achieved better results than LLMs. The fact that LLMs accomplished this merely by predicting the next word is undeniable. Has the fundamental reason behind this been logically explained and documented in academic papers?

ChatGPT:

The success of large language models (LLMs) in achieving what appears to be "intelligence" by merely predicting the next word is a topic of ongoing research and debate. While several papers and theories attempt to explain this phenomenon, a complete and universally accepted logical explanation has not yet been established. However, there are several key insights and hypotheses that shed light on why LLMs have outperformed many traditional approaches to artificial intelligence.

Why Did LLMs Succeed?

  1. Scaling Laws

    • Research such as Scaling Laws for Neural Language Models (OpenAI, 2020) demonstrated that increasing the size of models, datasets, and computational resources leads to consistent improvements in performance.
    • This "scaling" effect allowed LLMs to achieve complex language understanding and reasoning capabilities without fundamental changes to the model architecture.
  2. Power of Autoregressive Models

    • LLMs, particularly those based on the Transformer architecture, leverage the self-attention mechanism to understand context and relationships between words.
    • While the task is technically just "next-word prediction," the model learns to capture patterns, semantics, and even some logical structures within language data.
  3. Leveraging Statistical Properties of Language

    • Natural language has strong statistical regularities, and LLMs excel at identifying and exploiting these patterns.
    • Unlike rule-based or symbolic AI approaches, LLMs do not rely on handcrafted logic but instead develop heuristics from vast amounts of text data, often outperforming more "logical" methods in real-world applications.

Do LLMs Truly "Understand"?

  • "Simulation Hypothesis"

    • Critics argue that LLMs do not actually "understand" language but rather simulate understanding through sophisticated pattern matching.
    • Emily Bender et al.'s Stochastic Parrots paper (2021) suggests that LLMs generate plausible outputs without genuine comprehension, as they lack grounding in the real world or a model of meaning.
  • "Emergent Abilities"

    • Recent studies, such as Emergent Abilities of Large Language Models (Wei et al., 2022), show that as model size increases, LLMs exhibit unexpected skills, including reasoning, planning, and creative output.
    • These "emergent abilities" suggest that LLMs may be leveraging complex internal representations, though the exact mechanisms remain unclear.

Is There a Logical Explanation?

  • Implicit Knowledge Hypothesis

    • Research like Language Models as Knowledge Bases? (Petroni et al., 2019) demonstrates that LLMs can answer factual questions as if they possess stored knowledge.
    • While this suggests some level of "understanding," it might still be the result of high-dimensional pattern recognition rather than logical reasoning.
  • Cognitive Science and Predictive Coding

    • Some researchers draw parallels between LLMs and human cognition, particularly the predictive processing model of the brain (e.g., Karl Friston's predictive coding theory).
    • There are attempts to frame LLM behavior as a form of Bayesian inference or predictive modeling, indicating that the models might be constructing internal representations of the world.
    • However, these theories are still in their early stages and do not yet constitute a full logical explanation.

Conclusion

While significant progress has been made in understanding why LLMs are so effective, there is not yet a complete and universally accepted logical theory that fully explains their success. Current research spans scaling laws, emergent behavior, and cognitive science analogies, each offering valuable insights but not a definitive answer. Further studies, both empirical and theoretical, are needed to unravel the deeper principles behind LLMs' surprising capabilities.

2025年2月21日金曜日

AI 株価予測 20250221の状況 8306MUFG Stock Price AI Forecast -20250221

AI 株価予測 20250221の状況
8306MUFG Stock Price AI Forecast -20250221


週初の仕掛け買いで2000円台に突入。今後は、2000円台でのレンジに入ると考え、MUFGのみ買い増し。ところが、米株がトランプ関税のためか下落を開始し、日本株も下落した。金曜日の米株及びMUFGのADRは大きく下落。Entered the 2,000-yen range with buying momentum at the start of the week. I expected the stock to trade within the 2,000-yen range going forward, so I only increased my position in MUFG. However, U.S. stocks began to decline, possibly due to Trump’s tariffs, leading to a drop in Japanese stocks as well. On Friday, both U.S. stocks and MUFG’s ADR fell sharply.

現状、1/4程度の配当重視(配当3.1%平均)のポジション損益は、若干マイナス。MUFG(3.21%)、長谷工(4.31%)、三菱商事(4.07%)、村田製作所(2.29%)、トヨタ(3.23%)、金ETF。Currently, a dividend-focused position (with an average dividend yield of approximately 3.1%) is at a break-even level. Holdings include MUFG (3.21%), Haseko (4.31%), Mitsubishi Corporation (4.07%), Murata Manufacturing (2.29%), Toyota (3.23%), and a gold ETF.

来週は月曜日が休日のため、米株の動向と日経先物を見てからの判断となる。週末の米株の下落が今年来るであろう大暴落の開始なのか、それとも、一時的な下落でしばらくするとまたぞろ上昇を再開するのかはわからない。現状経済指標は若干の不安要素を持つものの、いまだ景気の持続を示唆しているため、本格的な経済指標の悪化と景気後退、暴落の開始はもう少し先であると考えられる。Since Monday is a holiday next week, the plan is to monitor the movement of U.S. stocks and Nikkei futures before making any decisions. It is unclear whether the drop in U.S. stocks at the end of the week marks the start of a major crash expected this year, or if it is merely a temporary decline before the market resumes its upward trend. While current economic indicators show some uncertainties, they still suggest continued economic strength. 
Therefore, it seems likely that a genuine deterioration in economic indicators, an economic downturn, and a market crash are still a bit further down the road.


今回下落ポイントからのエントリーではないが、どこから入ったとしても最初の一手はリスクを伴う。このため、最初の一手はそのまま保持。下落した場合は、1000円に向かって段階的に新値ごとに買い増していく方針。Although this entry wasn’t made from a low point in the market, any initial move carries risk regardless of the entry point. For this reason, I will maintain the initial position as is. If the market declines, the plan is to gradually increase holdings at each new price level, aiming toward the 1,000-yen range.






































来週は13日水曜日のインフレ率が焦点。



IMMは円買いに転換。IMM switched back to JPY buy.




海外勢は売りが継続、個人が買う悪い展開に。Foreigh investors are selling while Japanese investors are buying. This is not a good pattern.










新値日足は下落。AI予測は日足が下落。週足が上昇。月足は下落。The new daily candlestick is down. AI predictions indicate the daily candlestick will decline, the weekly candlestick will rise, and the monthly candlestick will decline.


メガバンクの上昇は利上げの効果というより、バリュー株の底上げという性格が強い。メガバンクはことごとく一律に上昇してきた。爆上げしたのはメガバンクだけでなく、建設、重工業など、割安に放置されてきたセクターだ。ただ、このまま上がり続けるには無理があり、どこかでメガバンクのように一旦頭を打つだろう。The rise in mega-banks is characterized more by the uplift of value stocks than the effect of interest rate hikes. Mega-banks have uniformly risen, as have sectors such as construction and heavy industries, which have been undervalued. However, it's unlikely that they will continue to rise indefinitely, and at some point, they are likely to experience a downturn, similar to mega-banks.


米株は、リモートワーク定着による商業用不動産の暴落が、いまだ株式市場に織り込まれていないのが最大の懸念材料。The crash of US commercial real estate due to the establishment of remote work has not yet been factored into the stock market.

根底には中国の不動産崩壊もある。中国マネーの縮小が、世界の商業用不動産の下落に拍車をかける。The underlying factor is also the collapse of Chinese real estate. The reduction of Chinese money will further accelerate the decline of commercial real estate worldwide.

米貸し出し延滞率もコロナ支援金の枯渇に伴い急速に上昇しつつある。The delinquency rate for US lending is also rapidly increasing due to the depletion of COVID relief funds.

米株の急落とともに世界株式が暴落するのも、ここ数カ月以内に起こるだろう。The crash of US stocks will likely lead to a global stock market crash within the next few months.



現状日銀は700兆円近くの国債を買い入れており、これは国債発行残高の6割近く。このうち6兆円が毎月満期になるため、月額6兆円を買い入れている。このため、月額買入を3兆円に減額すると1年間に36兆円のQTとなる。全保有額を放出するには20年近くかかるため、それまでは廃止したと言いながら継続しているYCCが続くことになる。これが国債が売られた背景。 Currently, the Bank of Japan holds nearly 700 trillion yen in government bonds, which is close to 60% of the total government bond issuance. Of this amount, 6 trillion yen matures every month, leading the BOJ to purchase 6 trillion yen monthly. Therefore, reducing the monthly purchases to 3 trillion yen results in a QT of 36 trillion yen annually. It would take nearly 20 years to completely unwind their holdings, so despite announcing the end of yield curve control (YCC), it effectively continues. This is why government bonds are being sold off. 


PSAVEは若干持ち直し基調。ドル円は介入開始か。PSAVE shows a slight recovery trend. Could intervention begin with the USD/JPY pair?






QTは一旦停止のようだが、マネーストックは下落基調。QT seems to be on hold, but money stock is declining.


ベースマネーの増大は、地銀に対する政府の無制限支援によるもの。The increase in base money is due to unlimited government support for local banks.







長谷工は下落トレンドから復帰。Haseko showed exit from down-trend.





長谷工の決算はインフレによるコスト高を反映した減益となった。Haseko's financial results showed a decrease in profit, reflecting higher costs due to inflation.



















トランプはウクライナ問題の早期解決をプーチンと話し合う予定。ヨーロッパとウクライナは蚊帳の外。最悪の展開が待っている模様。Trump is going to talk Putin on the early resolution of Ukraine problem. Neither Europe nor Ukraine were invited. Worst case scenario is now expected. 

米銀株、USTU10とSPX, NKY  US Banks, UST10&SPX, NKY

米国株はサームルール発動により景気後退入りが確実となった。

今後のシナリオとしては、最初の利下げまでは株価は上昇、最初の利下げを境に下落開始となるシナリオ。The upcoming scenario is that stock prices will continue to rise until the first rate cut, after which they will start to decline.



中国株は反発するも未だ復活していない。不動産不況から始まるバブル崩壊はまだ始まったばかりかもしれない。Chinese stocks have rebounded but have yet to recover. The bursting of the bubble, starting with the real estate recession, might have only just begun.



NY株は景気後退のシグナルが其処ここに出ているのを無視して、パウエルの利下げとM7のみの好況を理由に上昇しているが、これがいつまでも続くわけはない。Despite various signs of an economic downturn, NY stocks are rising, driven by Powell's interest rate cuts and the strong performance of the M7 companies. However, this cannot continue indefinitely.

VIXは上昇。Vix has gone up.


FEAR&GREEDはへ強欲へ FEAR&GREED dived into greed.


米金利の今後 US FF rate forecast

11月の利下げは0.25%となりそうだ。


前回Previous



最新Latest





米インフレの鎮静化は一時停止状態で、失業率は低下し、景気はまだ強く、インフレ圧力の復活の可能性も秘めているように見える。これがパウエルが利下げを早期に行えない原因。The pacification of US inflation remains suspended, and with a decrease in the unemployment rate and continued strong economy, there seems to be a possibility of a resurgence of inflation pressure. This is the reason why Powell cannot cut interest rates early.


[今予測20250221]Forecast at this time





[前回予測20250214]Previous Forecast 





Midterm Range

長期的には1年後をめどに暴落開始だが、それまでは上昇を続けるだろう。Long term forecast is negative expecting market crash after 1 year however, market may still going up until then.



MUFGの長期チャートを見ると、まだ伸びそうではある。By looking at the long term chart of MUFG, it looks MUFG still has room to grow.