MetaのオープンソースAIモデル「LLaMA 3」が登場し、AI業界に衝撃が走っています。わずか700億パラメータのLLaMA 3が、1.7兆パラメータのGPT-4に匹敵する性能を示したのです。さらに、LLaMA 3の4000億パラメータ版の開発も進行中であり、その性能がGPT-5を上回るとも予想されています。LLaMA 3のような高性能モデルを低コストで運用できるようになれば、個人でもAIエージェントを開発・運用できる時代が到来するでしょう。さらに、新興チップメーカーのGroqが開発したLPU(Language Processing Unit)は、NVIDIAのGPUよりも高速にLLaMA 3を処理できると注目を集めています。GroqのCEOは、2024年までに世界の推論用AIチップの50%をLPUが占めると予想しています。一方で、AIモデルの性能向上に伴い、政府の規制も強化される見込みです。米国では1x10^26フロップス、EUでは1x10^25フロップスがAIモデルの規制基準とされており、LLaMA 3の性能はこの基準に迫っています。AI業界のリーダーたちは、GPT-5のようなさらなる性能向上を見据えつつ、規制との兼ね合いを模索しています。OpenAIのサム・アルトマンCEOは、AIモデルの性能向上を前提にしたスタートアップ戦略の重要性を訴えています。
公開日:2024年4月20日
※動画を再生してから読むのがオススメです。
リーダーボードのトップまで登り詰めたLLaMA 3です。
GPT-4は上にあります。
Claude 3は、大きなモデルであるOpusを含めて、後方視鏡に映っています。
OpenAIの大きな優位性はなくなりました。
私たちは今、トップのオープンソースの競合相手を持っています。
ここには本当に驚くべきものがたくさんあります。どう呼んでもいいですが。
1つ目は、MetaのモデルであるLLaMAが、今やGPT-4とほぼ同じレベルにあることです。
Facebook/Metaが最も近い競合相手であることです。
2つ目は、それが700億のモデルであり、GPT-4の1.7兆またはそのような数字と比較されることです。
最後に、それはオープンソースですか?
これはどうして起こっているのですか?
もしそこで止まっていたら、それだけでも十分に奇妙でした。
しかし、待ってください、もっとあります。なぜなら、別のモデルが進行中だからです。
LLaMA 3は400億のパラメータを持っています。
しかし、まだトレーニングが完了していないので、結果は改善されます。
NVIDIAのジム・ファン博士は、LLaMA 3、400億モデルがリリースされる前にGPT-5が発表されると予測しています。
この時点で、ほとんどの人々が気づいたかもしれませんが、OpenAIは他の発表に自身の発表で応じる傾向があり、常にAI競争の最前線にいようとしています。
これは、以前はMetaのAIであったHello Robotのクリス・パクストンです。
彼は元Meta AI、元NVIDIA AIです。
700億モデルが1800兆モデルと同じくらい優れていると信じるのは本当に難しいです。
訓練データは本当にすべてだと思います。それは昨日のインタビューでマーク・ザッカーバーグが話したことです。
彼らは非常に厳選された、非常に高品質な訓練データを持っています。
GPT-4と比較して非常に小さいモデルが、今やGPT-4と非常に似たレベルで機能しています。
そして、それが重要な理由です。
なぜそれが非常に重要なのか。
なぜなら、今や自宅にGPT-4があるか、少なくとも非常に近いからです。
基本的には、比較的安価な機械学習リグを作成できる段階にあります。これらのリグは、自宅から実行できる非常に洗練されたAIシステムを駆動し、ビジネス、仕事、個人用途のために自分自身のエージェントを作成できます。
DevOnのようなものは、おそらくGPT-4上で実行されています。
このようなDevOnのようなものを実行できると想像してみてください。コードを書いたり、修正したりできるもの、あるいはウェブサイトを構築し、eコマースストアを運営できるような、このような機械で実行できるもの。これを構築するのに3,000ドルかかるかもしれません、それ以上かもしれません、それ以下かもしれません。どのようなチップ、どのようなGPUを選択するかによって、最新のものか、やや古いものかによって異なります。
録音している最中に、人々が自分のベンチマークや、LLaMA 3をローカルでOLLaMA経由で実行したスクリーンショットを投稿し始めました。
約14トークン/秒でM.2 Ultra 76 GPU上で実行しており、価格は約7,000ドルから8,000ドルで見つけることができます。
GPT-4やChatGPTから期待されるよりも少し遅いです。Groqは300トークン/秒、GPT-4はおそらく20トークン/秒、GPT-3.5 Turboはおそらく100トークン/秒です。
ざっくりとした見積もりですが、ここではLLaMA 3を3ドル,200で実行しています。しかし、秒間約8トークンしか取得できないので、非常に遅いです。ただし、これはラップトップで実行しているので、驚くべきことです。
ここでは、スピードを比較していますが、大きくて重たい8000ドルのコンピューターは約45%速く動作しますが、ご覧の通り、3000ドルから8000ドル、9000ドルのハードウェアで動作しています。使えるけど、少し遅いです。
もちろん、8億モデル、つまり小さいモデルを使用している場合は、はるかに速くなります。
HyperWrite AIのマット・シューマーは、Groqコンピューターチップ上でほぼ300トークン/秒で提供されているCloud 3 Opusを上回るオープンソースモデルを持っていると述べています。
この技術を活用したアプリケーションは、革命的なものになるでしょう。
ここでテキストがどれだけ速く生成されるかをご覧ください。
Groqがリアルタイムの会話エージェントで使用されている例を見てきました。たとえば、AIが電話であなたに電話し、リアルタイムの会話を行うことができるような例です。
例えば、営業電話、予約の取り付け、顧客サービスなど。
私たちは、GPT-4レベルのモデルが基本的にコールセンターを運営し、リアルタイムで顧客と会話し、24時間リアルタイムでアポイントメントを予約することができる領域に入ったばかりです。その上、わずかな費用で。
もちろん、Groqという新しいチップがかなり話題を呼んでいます。
それは異なるのです。なぜなら、それはGPUではなく、グラフィック処理ユニットではないからです。
それはLPU、言語処理ユニットです。
この人、Shamoff Palapatiyaがそれに投資したらしい。
彼はどうして何にでも首を突っ込むんだろう?
まあ、私は嫉妬していないわ。
みんながGroqについて話しているね。あなたはとても幸せそうだ。
私は以前、AI Jasonチャンネルからこれをプレイしたことがあります。
彼は、アポイントメントを設定するためにあなたに電話をかけるテレマーケティングエージェントを作りました。
情報を提出すると、それがあなたに電話をかけて売り込みをします。
そして、あなたのアポイントメントをスケジュールします。
聞いてください。
こんにちは、ジェイソン、私はクレイジーフィットネスから、当ジムの会員に興味を持っているということに関連して連絡を取っています。
必要な情報をすべて提供し、お持ちの質問にお答えするためにここにいます。
それがAIです。
お電話ありがとうございます。
Groqで実行中です。
どのような具体的なフィットネス目標を達成したいですか?
私はただ体重を少し減らしたいだけです。
COVID以来、過去3か月または6か月間、全く運動をしていません。
それはよくわかります。
私たちのパーソナライズされたトレーニングと栄養プログラムは、あなたが体重減量の目標を達成するのを助けることができます。
あなたに最適な会員オプションについて話し合いましょう。
いいですね。
可能であれば、今週の土曜日にメンバーシップを開始することにはかなり満足していると思います。
それを聞いて嬉しいです。
プロセスを完了するために、会員登録を確保するためにクレジットカード番号をお願いします。
おそらくそれは編集されていると思っているかもしれません。
おそらく、彼が話すときとAI電話エージェントが答えるときには長い間隔があるでしょう。
右側の男性は、Groqの創設者で元Googleであり、彼こそ本物です。
ちょっと待って、彼について話しましょう。
ちょっとGroqに尋ねてみてください。
わかりました、それをしましょう。
それをオンにします。
やってみましょう。
私はこれについて答えを持っていませんが、Groq、あなたを特別な存在にしているのは何ですか?
まずは... 別の...
ねえ、私はGroqです。
私に話すときは、私が人間のように扱ってもらえます。
わかりました、そうします。
私が過去に使用したかもしれない大規模言語モデルとの違いは何ですか?
確かに、私は以前に使用した大規模言語モデルとは少し異なります。
私はより自然なコミュニケーションを持つように設計された言語ユーザーインターフェース、またはLUIです。
了解しました!
あなたは人間の脳にどれくらい似ていますか?
私は人間の脳ではありませんが、同様のものを理解し、応答するように設計されています。
私は人間の言語を自然な方法で処理し生成することができます。
了解しました!
ほとんどの人が知らないことを教えてください。
面白いことを教えますね。
タコには3つの心臓があることを知っていましたか?
そして、人々はGroqに気づいています。
再度、これはGroqのCEO、ジョナサン・ロスです。
通常、開発者の指標のようなものをツイートします。
お目にかかったのは今朝ですが、なぜ開発者がとても重要なのですか?
私たちは7万5千人の開発者がいます。
それは、開発コンソールの立ち上げからわずか30日以上です。
私たちの比較では、Vidyaは1,000人の開発者に到達するのに7年かかりました。
そして、私たちは30日ほどで7万5千人に達しています。
気づいていない方のために、これはShamov Palapitiyaです。
彼のことを言及すると、彼が好きでないと感じる人がいつもいます。
おそらく、彼はこれらのSPAC(特定目的会社)を作成し、さまざまなメディアプラットフォームでそれらを大々的に宣伝したようです。
人々はその過程でお金を失いました。
私はそれをすべて追っていないので、ここに注釈として残しておきます。
お金持ちで生まれたわけではなく、スリランカで生まれ、Facebookの初期の幹部となりました。
そして、自分自身の投資会社を設立しました。
そして、その過程でどこかで、彼は数十億ドルの資産家になりました。
彼はゼロから始めて、今ではGroqのCEOと話しています。
しかし、私がここにたどり着くまでにその長く曲がりくねった道を取った理由は、シャモフがAIやテクノロジーに関する予測において非常に正確である傾向があるからです。
彼はしばしば、その時点では明らかでないことを言います。
しかし、彼の意見を振り返ると、驚くほど正確であることがよくあります。
それが常に起こるわけではないと確信しています。
彼が間違っているときもたくさんあると確信しています。
彼がテクノロジーの予測をするとき、私は耳を傾ける傾向があります。
もし革新的なものを作り、それを立ち上げたら、誰もそれをコピーすることさえ4年かかるだろうし、先を行くことなんてないだろう。
それはただ、はるかに良いアプローチに感じられたし、それは原子だ。
それをより簡単に収益化できる。だからその頃...
ここでは、彼はソフトウェアではなくハードウェアを構築し、ビットではなく原子を収益化することについて話しています。
TPUの論文が出た当時、私の名前が載っていて、人々がそれについて尋ね始めました。
TPUはGoogleの独自のチップであり、Tensor Processing Unitです。
NVIDIAはGPU、GoogleはTPU、そしてGroqはLPUs、言語処理ユニットです。
それは違うやり方をするだろう。
その当時、私は公開市場にも投資していました。公開市場でのちょっとした浮気です。
そして、サンダルはプレスリリースでTPUについて話し始めました。
私はとても驚いたんです。
Googleが独自のハードウェアを開発すべき世界が考えられないと思いました。
彼らは私たち他の人たちが知らない何かを知っているはずです。
私たちはそれを知る必要があります。そうすれば、それを世界中で商品化できるようになります。
おそらくその数週間後にあなたに会ったと思いますが、それはおそらく私が今までで最も早く投資をしたものでした。
鍵となる瞬間は、あなたが会社を持っていなかったことです。
私たちは、チェックが書かれた後に会社を設立しなければなりませんでした。それは完全な愚かさの兆候か、15年か20年後には天才のように見えるかのどちらかです。
しかし、後者の可能性は非常に低いです。
そして、彼は自分に非常に高い評価を持っています。
当時のGroqで行っていたデザイン上の決定について教えてください。その時点で知っていたことを考えると。
その時は今とは非常に違っていたからです。
さて、再び、私たちが資金調達を始めたとき、実際にはハードウェアで何かをするかどうかさえわからなかったんです。
でも、シャマスが尋ねたことがあったんですが、それは何を違うようにするかということでした。
私の答えはソフトウェアです。
なぜなら、私たちが抱えていた大きな問題は、Googleでこれらのチップを作ることはできても、それをプログラムすることができなかったからです。
Googleのすべてのチームには、モデルを手作業で最適化している専任の人がいました。
これはまったく狂っています。
ちょうどその頃、NVIDIAからいく人を採用し始めていたのですが、彼らは「いや、いや、あなたは理解していない」と言っていました。
これがただの働き方なんです。
これは私たちも同じようにやっています。
私たちには、カーネルと呼ばれるものがあり、CUDAカーネルを手動で最適化しています。
私たちはただ、そうしていないように見せているだけです。
皆さんはアルゴリズムとビッグOの複雑さを理解しています。
それは線形の複雑さです。
すべてのアプリケーションにはエンジニアが必要です。
NVIDIAは現在、エコシステムに5万人の人々を抱えています。
これらは本当に低レベルのカーネルライティング、アセンブリライティングのハッカーで、GPUやMLなどを理解しています。
スケールしないでしょう。
大規模言語モデルたちはなぜGroqを好むのでしょうか?
何か大規模言語モデルの設計上の決定事項や何かが起こったのですか?
それの一部は明らかにスキルですが、一部は少しの運もありました。
NVIDIAよりもはるかに速い理由は何で、なぜこれほどの進展があるのか、具体的に何が起こったのですか?
それの要点は何ですか?
言語になるとは知りませんでしたが、インスピレーションは、DeepMindのAlphaGoソフトウェアをTPUで動作させることでした。
それを見ていると、推論がスケーリングされた問題になることは非常に明確でした。
他の誰もが推論を、1つのチップでモデルを実行し、それが何かを実行すると見ていました。
しかし、AlphaGoではソフトウェアを移植しました。
そして、170のGPUに対して48のTPUがあったにもかかわらず、48のTPUがまったく同じソフトウェアで100試合中99試合勝利しました。
それが意味するのは、計算がより良いパフォーマンスにつながるということでした。
洞察は、スケールされた推論を構築しましょう。
私たちはインターコネクトを組み込みました、スケールに合わせて構築しました。
そして、これが今私たちがこれらのモデルの1つを実行しているときに行うことです。私たちはAlphaGoと同様に、数百または数千のチップが貢献していますが、これはそれをつなぎ合わせるために構築されました。
しかし、ここで全てをLLaMA 3にリンクさせるのです。
このAIチャットボット、通称大規模言語モデルのAI軍拡がどこに向かっているかについて、彼らがどう考えているかを聞いてみてください。
そして覚えておいてください、これはおそらくLLaMA 3がリリースされる数日前に撮影されたものです、もしかしたらもう少し前かもしれません。
聞いてみてください。
もし今日会社を立ち上げる、ものを作る、ものを構築するのであれば、LLaMAからMistral、Anthropicにできるだけ頻繁に切り替えられる能力を持ちたいと思うでしょう。
最新のものを選ぶ。
これらのモデルが実行されるのを見ている立場として、これらのモデルの品質や、いくつかの企業がどこに向かっているか、または他の企業と比べて何がうまくいっていると思われるかについてコメントはありますか?
彼らはお互いに追いつこうとしており、いくつかの飛び越えを見ています。
GPT-4が先行して始まり、他のすべての人より約1年のリードを持っていました。
もちろん、Anthropicも追いついています。
私たちはMistralから素晴らしいものを見ています。
しかし、全体的に見て、彼らはすべて品質面でまとまり始めています。
特に興味深いのは、Mistralがより小さく、より安価なモデルで品質に近づけることができたことで、これは彼らに大きな利点をもたらしていると思います。
Coheerは、ラグ最適化モデルに興味深いアプローチを取っていると思いますので、人々はニッチを見つけています。
そして、最高水準で全体的に最も優れているいくつかのモデルが出てくるでしょう。
しかし、私たちが見ているのは、これらのモデルを実行するコストに関する多くの苦情です。
アプリケーションをユーザー向けにスケールアップすることはできません。それらは単なる天文学的な数字であり、
OpenAIは公開したり、Metaも同様に、Teslaや他のいくつかの企業も、彼らが購入しているGPU容量の合計を公開しています。
あなた方がスケールを拡大するにつれて、それが本当に彼らに支えられているだけなので、推論市場の規模をどの程度になるかを逆算することができます。
人々が何に対して戦っているのか、その規模を感じてもらえますか?
年末までに、Facebookは65万台のH100に相当するものを持つ予定です。
年末までに、Groqは10万台のLPUsを展開する予定であり、スループットとレイテンシの面でH100を上回ります。
おそらく、我々自身もMetaと同等のものにかなり近づくでしょう。
来年の終わりまでに、我々は150万台のLPUsを展開する予定です。
比較のために、昨年、NVIDIAは50万台のH100を展開しました。
150万台という数字は、Groqがおそらくすべてのハイパースケーラーやクラウドサービスプロバイダーの組み合わせよりも、推論、生成AIの能力が多いことを意味します。
おそらく、世界中の推論計算の約50%です。
それは素晴らしいですね。
彼は、次のカードでクルーラーになれるかどうかを計算しようとしています。それは、すべてのこれらの大手テック企業のAIに焦点を当てた研究所が手に入れようとしている、非常に求められているNVIDIAのGPUです。
おそらく、この美しい写真を見たことがあるでしょう。
その中の1つがこちらです。
これは、顧客別のNVIDIA H100 GPUの出荷数です。
これは2023年の見積もりでした。
ご覧の通り、MetaとMicrosoftが15万台のH100でレースをリードしています。
興味深いことに、マーク・ザッカーバーグは実際には、すべてのそれらのH100を持つことができたという幸運に恵まれていました。
そして、それは大規模言語モデルをトレーニングする準備をしていたからではありませんでした。
おそらく、それはまだ見通しにはなかったようです。
いつでも教えてください。
私たちは、株価が大幅に下落しているような感じでした。
人々は、「このキャペックスに何が起こっているのか」と言っています。
人々はメタバースを買っていません。
そしておそらく、そのキャペックスを使ってこれらのH100を手に入れているのでしょう。
当時、どうやってH100を手に入れるべきかを知っていたのですか?
どうやって私たちがGPUが必要になることを知ったのですか?
私たちはリールに取り組んでいたからだと思います。
私たちは常に、まだ見えていない何かを構築するための十分な容量を持っていたいという状況に陥りました。
そして、リールの状況になり、トレーニングにより多くのGPUが必要でした。
それは主にコンテンツのプレイでした。
彼らはより良いコンテンツを推薦できるようになりたかったのですが、それは大きなビジネスではありませんでした。
彼らはより良いコンテンツを求めていましたが、それは大規模な言語モデルのプレイではありませんでした。
ChatGPT、GPT-4などはまだ登場していません。
ハイプサイクルはまだ始まっていません。
そして、元OpenAI、元Teslaのアンドレイ・カルパシー氏は、今はもっと教えていると思います。
彼はYouTubeや他のさまざまなプラットフォームを通じて、モデルのアーキテクチャについて少し話しています。
しかし、私には気になる2つの点があります。
1つは、具体的に彼らがトレーニングデータについて話していることです。
彼は、このモデルをトレーニングした15兆のデータが、8つのパラメーターしかないモデルをトレーニングするには非常に大きなデータセットであると述べています。
これは通常行われることではありませんが、新しく非常に歓迎されています。
これは他のモデルよりも75倍多くトレーニングしています。
彼は、「私たちは非常に使いやすく、非常に小さな非常に優れたモデルを得るので、非常に歓迎されると思います」と述べています。
そして、Metaは、この時点でもモデルが収束していないように見えると述べています。
つまり、私たちが常に使用している大規模言語モデルは、おそらく100から1000倍以上の要因で十分にトレーニングされていないため、より長期間トレーニングされた小さなモデルをリリースすることができるということです。
私たちはここで壁にぶつかっていません。
これからもっと進む余地があります。
しかし、私が実際に考えていなかったものの1つが目立ったのはこれです。
メタは、FLOPSを見てモデルの強さを決定することについて話しています。
そして、このFLOPSの数が重要な理由は、モデルの強さを1つの数字に減らさなければならない場合、これがその方法だからです。
なぜこれが重要なのかというと、これはモデルのサイズとトレーニングの長さを組み合わせて、それにどれだけの合計FLOPSがかかったかの強さを1つにまとめるからです。
なぜこれが重要なのか?
実際に、FLOPSの数には限界があり、その数がどれくらいになるかのしきい値があります。
たとえば、LLaMA 370億は、その数が9×10の24乗として表されます。次に、400億モデルが出てくる可能性があるので、それが次の大きなものになります。そして、それがオープンソースになることを願っています。
それがオープンソースにしても安全であれば、それは10の25乗の4倍になります。
彼は、これが本当に大きくなり始めると言っています。
バイデン大統領令では、報告要件が10の26乗の1倍に設定されています。
これはその2倍下回ることになります。
ポイントは、新しい、大きなLLaMA 3モデルがリリースされたら、これらのモデルがどれだけ強力になるかの限界に近づいている、または限界に達することになると思われる前に、報告義務が設定される可能性があるということです。
もし私が正しく読んでいるなら、EUはec.europa.euだと思います。
それはシステムリスクの閾値として10の25乗と書いてあります。
EUは10の25乗、LLaMA 370億で制限していますので、現在出ているものは10の24乗です。
米国の報告要件は10の26乗に設定されています。
そして次の大きなLLaMA 3モデルは10の25乗になります。
これらのモデルがその領域に少しずつ侵入し始めると、各国政府がどのような行動を取るか、もしあれば、興味深いことでしょう。
これでOpenAIに戻ります。
誰もがOpenAIが何か大きなもの、GPT-5で反撃することを期待しています。
次世代であるGPT-5が次の大きなマイルストーンとなり、誰もが目指そうとするでしょう。
しかし、多くの人々がこれらの高速で安価で強力なモデルが多くのビジネスを破壊し、混乱させることを懸念しています。
OpenAI自体が、OpenAIエコシステムを利用して構築しているさまざまなスタートアップを圧倒することで悪名高いです。
ただし、その質問はサム・アルトマンに投げかけられます。
この速い動きの世界で何を構築すべきか、プロセスを活用し、次の革新の波に押しつぶされないように考えるべきですか?
これは非常に最近のことで、4月15日に彼がこれについて話していることです。
聞いてみてください。
あなたは自分の時間を投資しているとおっしゃいましたが、今日、世界中の多くの大企業と関わっているのを見ます。
私としては投資家として、多くのAI企業を見ていますが、アプリケーションレイヤーのAI企業には投資していません。
ただし、OpenAIが製品を発表し、まるで、その製品が全産業を壊滅させたかのようです。
おそらく、現在AIを活用したりAIを使ったスタートアップを立ち上げるための基本的な戦略は2つあると思います。
モデルが改善されないと仮定するという1つの戦略があります。
そして、それを基にさまざまな小さなものを構築していくということです。
そして、もう1つの戦略は、OpenAIが同じペースで進化し、モデルが同じペースで改善され続けると仮定して構築するというものです。
95%の世界は後者のカテゴリに賭けるべきだと思われますが、多くのスタートアップは前者のカテゴリで構築されています。
そして、私たちが基本的な仕事をするだけで、つまりモデルとそのツールを毎回改善するだけで、OpenAIが私のスタートアップを潰したというメームが生まれます。
GPT-4上に何かを構築している場合、合理的な観察者がGPT-5がGPT-3よりも優れていると言うだろう場合、私たちはあなたが好きでないからではなく、単に使命があるから、あなたを圧倒するでしょう。
GPT-5がはるかに優れていることから利益を得るスタートアップの巨大なセットがあります。
そして、それらを構築し、AIの進歩が私たちが考えるように進む場合、ほとんどの場合、あなたは本当に幸せになると思います。
ほとんどの場合、あなたは本当に幸せになるでしょう。
サム、私が投資できる可能性のある投資テーゼを探している投資家として、蒸し返されないものと、蒸し返される可能性のあるものとは何ですか?
会社に、モデルの100倍の改善が楽しみであるかどうか尋ねてください。
実際、私たちはかなりよくわかります。なぜなら、私たちに次のモデルが欲しいと言ってくる企業を知っているからです。いつ出るのか、いつ出るのか、最初に試したいと思っています。
それは私の会社にとって最高のことになるでしょう。
そして、そのような点で私たちから連絡を受けない企業がたくさんあります。
そして、より良い基盤となる知能がその製品や企業をどのように加速させるかには、明確な道筋がある場合、ほとんどの企業がそのストーリーを非常に明確に説明できます。
クラナはその例ですか?
クラナは良い例です。
なぜなら、クラナにとって、数字は驚くほどです。
次のモデルが期待通りの性能を持つとしたら、それがどれだけ良くなるか考えてみてください。
今朝、医療アドバイザーのようなAIと話しました。彼らはそう呼ぶのかな。
そして、モデルが性能不足である場所を教えてくれました。
これらの種類のことにはまだかなり役立ちます。
しかし、もしモデルがこれだけ指標を改善できたら、他の多くのビジネスが可能になります。だから、もっと早くできるかな?
そうすれば、多くの命を救うことができ、医療ケアにアクセスできなかった人々にもアクセスを提供できるようになります。
そして、いつ実現できるか。遅延するたびに何人が毎日亡くなっているかを教えてください。
実際には効果的なプレゼンテーションでした。
クラナは、個人のショッピング、顧客サービス、従業員の生産性のためのAIであるOpenAIのウェブサイトで最近特集された最初の顧客成功ストーリーです。
GPT-5、LLaMA 4、Cloud 4などに移行するにつれて、AIモデルがどれだけ良くなるか。
会社はこのように改善を続けています。
AIがより良くなるにつれて、その改善の波に乗ることになります。
彼らはより良くなります。
と言っても、必ず購読していることを確認してください。
私たちは、ローカルマシンでいくつかの作業を開始する方法や、クラウドで実行する方法について話す予定です。
今では、費用がますます安くなってきているので、将来的にはホスティングと同じようになると思います。
実際、誰もが自己ホスティングをしていません。
一部の人々はするかもしれませんが、ほとんどの場合、クラウドに外部委託するだけです。それが簡単で、より安全で、かなり安いからです。
しかし、これらのモデルを実行するコストがどれだけ安くなっているか、より小さなモデルがどれだけ効果的であるか、どれだけ速く、安く、より良く、ローカルで実行できるかについての突破口があります。
これにより、信じられないほど新しい機能が開放されるでしょう。
最近、より多くのエージェントが必要な論文を取り上げました。
エージェントを増やせば、サンプルをより多く取ることができますし、複数のエージェントが協力して働く場合、結果がより良くなり、出力も向上し、エージェントも賢くなります。
今、LLaMA 3をGroqで1秒800トークンで実行できるようになったので、それは80億のパラメータのようです。
これにより、多くの機能が開かれるでしょう。
購読していることを確認してください。
もっとたくさんの良いものが近々やってきます。
私の名前はウェス・ロスです。ご視聴ありがとうございました。