DICTWF

Disruptive innovation invester / $TSLA Invester / Innovation is Key to Growth

[めろーね @melone3710]

【Anthropicの件を受けてのAI投資シナリオは】
悲観楽観が混在していていますが、AI投資全体の現在地、現状の雑感です。

自分が個人的にほぼ確信持ってるスタンスは概ね、
【AI投資熱が冷めるまで、半導体恩恵企業は儲かり続け、欲深な市場は決算の度に見直して株価があがる可能性が高い】
です。

来年にかけて大手ITのFCFもマイナスになっていき、ここからは赤字のグロース企業が「２，３年後には凄い未来になります！大儲けです！」とか言って粘りながら投資され続けるのと似た同じ構図
その先で、いよいよ赤字が続きすぎ、収益化も緩やかな伸びになり・・・曲線的・放物線的に雲行きが怪しくなった時が天井になる可能性は濃厚
だけど、例えば油断していた時にとんでもないロボがでてきたり技術的なブレイクスルーがくれば風向きも変わるし、そんなことは現場の人間ですら現段階では誰にもわからない。

■雰囲気の楽観論で「AI・ロボ・FAIでバラ色未来は確定！！」だから株価は上がり続ける！
も
■極端な悲観論で「規制されたからAIバブルは終わり！株価は下落へ！」
どっちも両方バイアスだと思う。

現状を客観的に評価すれば、
【AI相場は最終的にどうなるかは分からない】
ってのが本質
「そう言っておけば外さないからズルイ！」とかそういう話じゃなくてｗ
「サイコロ振ったら偶数と奇数それぞれ出る可能性は５割」というのが単に事実であるのと同じで、その%が刻一刻と動いてるってだけ。
この「AIの最終的な未来は誰にも分からない」という主張をずっと自分はしてるけど、ちゃんとロジックはある。

実際に情報を朝から晩まで舐め回しても、「強弱要素が折り重なって」いて、楽観も悲観も「80%とか90%みたいな確度で断言できる要素」が本当にない。

「毎年〇兆円投資して減価償却して、〇年後に自動運転で〇兆円、ロボで〇兆円、APIで〇兆円利益あげて・・」とか、ある程度の「計画」はあって試算はできたとしても・・・
【人類が経験したことがほぼない、半導体先進国を中心とした全世界的規模の未曾有の投資】
であり、人類が受け入れた事のない新技術であるがゆえに、開発者や投資側ですら以下のことが恐らく本当に朧気にしか把握できてないのが根本的な理由になってる。
＝＝＝
「事業ごとに細分化してどこでいくら売上げるのか」
「そのためにコストはいくらかかるのか」
「その目標達成までに何年かかるのか」
「技術的な堀は何年有効なのか」
「各技術的ブレイクスルーはいつくるのか」
本来、単体のグロース企業なら仮にでも当然想定すべき事業計画的なものを算定するなどまず不可能で、細かい算段なんて全くできるわけがない状況。

ここが本質。
＝＝＝
過去にも鉄道、航空、ITネット網、色々な大規模投資はあったけど、毎日情報を見れば見るほど、
【過去最大の不確実性と規模が大きい投資】
なのではないかと思う。

・人間の数か月分の長期記憶ですら無理な中で、社会実装の為のハルシネーションや遅延・不具合・停止などへの対策
・最終責任は人間であることによる制約で完全自動化できない領域＝法的な問題
・今回のAnthropicpのように広義の意味でのAIの地政学リスク
・フィジカルAIの費用対コストの正当性
・自動運転の法的な問題や、そもそも利用者としての真の需要はどの程度か
などなど・・
今回のように、今後起こりえる不確実性が多すぎる。

「正確な未来を予測する」のは困難ではある中で、「ハイテク企業としてはやらざるを得ないチキンレースで突っ走ってるのが真実

だから「断言できるような未来なんてない」というのが現在地。

決してAIバブルが崩壊するとも、AIツルハシ企業が延々と儲かり何年も株価が上がり続けるとも強く主張することは難しい、現段階ではできない。

一つポジティブなことをあえていうとすれば、
【24年～25年のアッシェンブレナー、革ジャンやアモデイ達が想定してた世界と比較して、今のところはほぼ順調にその通り推移している】
ということ。
私が確認できる情報の範囲内ではわからなくても、
【公には言えないような彼らには見えている（知ってる）技術】を元にして投資の確信としている可能性も高い。
例えば、自動応答、擬似人格やロボや自動運転が2,3年後にここまでできる、だから社会もここまで変わる
という確信的な予測が既にあるということ。

もしそれがあるとするなら、この投資にも納得がいく話になる。

今のＡＩ相場はそういうものだと思ってます。

[Nic Cruz Patane @niccruzpatane]

Lars Moravy, VP of Tesla Vehicle Engineering, says the Tesla Cybercab is officially the most efficient EV ever certified at 165 Wh/mi!

That’s around 40% more efficient than a Model 3. WOW.

[みくかぶ🦋専業投資家👼 @miku919191]

今日の革ジャンのインタビュー内容をまとめました☺️

よく分からない自称AI有識者の意見を読むより、AIのトップの革ジャン本人のインタビューを読む方が絶対いいです‼️
NVIDIAが何を見ていて、次にどこへ向かうのかわかります。
AIがバブルなんて考え吹っ飛びますw

$NVDA

[がす | 月面マスドライバー × イーロンAOC公認 @lunarmassdriver]

イーロン・マスク：破壊的革新者か、それとも？

「イーロン・マスクは科学者ではない。エンジニアでもない」という批判に対し、彼を間近で知る人々はこう語っています。スレッドでまとめました。↓ $TSLA

[みくかぶ🦋専業投資家👼 @miku919191]

革ジャンのインタビュー
（めちゃ長いです）

☑️1 AI投資はなぜ「過剰」に見えるのか

市場では現在のAI投資について、「支出が多すぎる」「設備投資が過剰だ」という見方が強い。
特にAmazon、Microsoft、Google、Metaといったハイパースケーラーが行っている巨額の設備投資を見ると、短期的にはキャッシュフローが悪化しており、非合理的に映る。
しかし、この見方自体が根本的に問いを間違えている。
これは通常のIT投資ではないし、単なるデータセンターの拡張でもない。
人類の計算インフラそのものを、ゼロベースで作り直す段階に入っている。

☑️2 　AIは「計算のやり方」そのものを変える
AIは単なるアプリケーションではない。
検索、広告、研究、設計、意思決定、ソフトウェア開発など、あらゆる分野において「計算のやり方」そのものを変えてしまう存在だ。
従来のコンピューティングは、人間が事前にロジックを定義し、それを機械が実行するという形だった。
AIはそうではない。
文脈を理解し、状況に応じて判断し、必要な計算をその場で選び取る。
これは段階的な改善ではなく、明確な転換点を伴う変化だ。

☑️3 　AIは「使えないデモ」から実用フェーズに入った
数年前まで、AIは好奇心の対象だった。
デモとしては面白いが、実務で使えるレベルではなかった。
幻覚が多く、信頼できないという評価も正しかった。
しかしここ1〜2年で状況は完全に変わった。
推論能力が大きく改善され、文脈を保持し、ツールを使い、複数の工程をまたいだ作業が可能になった。
その結果、多くの企業がAIを試験導入ではなく、本格的に業務へ組み込み始めている。

☑️4 需要は存在するのか
需要が本当に存在するのか、という問いに対する答えはすでに出ている。需要は存在する。
しかも、異常なほど強い。
OpenAIやAnthropicのような企業は、創業からそれほど時間が経っていないにもかかわらず、すでに高い収益性を示している。
もし計算能力が2倍になれば、収益は2倍ではなく、4倍になる可能性がある。
制約条件は需要ではなく、供給側にある。

☑️5 市場は「需要待ち」ではなく「供給不足」の段階にある
AI市場はいま、「需要が立ち上がるのを待っている段階」ではない。
すでに需要は存在しており、「供給が追いついていない段階」にある。
計算能力がボトルネックになっており、そこが解消されれば、利用は一気に拡大する。

☑️6 ダークファイバーとダークGPUは同じではない
2000年代初頭、光ファイバーへの過剰投資が行われた。
その結果、大量の未使用回線、いわゆるダークファイバーが生まれた。
現在のAIインフラをそれと同列に語る人がいるが、前提条件がまったく違う。
ダークGPUというものは存在しない。
世界中のGPUはフル稼働している。

☑️7 GPUは余っていない
6年前に販売されたGPUですら、現在では価格が上昇している。
これは供給過剰ではなく、深刻な供給不足を示している。
もしこれがバブルであれば、未使用の計算資源が大量に存在するはずだ。現実はその逆だ。

☑️8 次世代AIは10倍の計算を要求する

次世代AIモデルは、前世代と比べて10倍の計算量を必要とする。
さらに、扱うデータ量も10倍になる。
AIは単に答えを返す存在ではない。
自ら「次に何を計算すべきか」を判断し、その判断がさらに計算需要を生む。

☑️9 計算需要は指数関数的に増える
この構造によって、計算需要は線形ではなく、指数関数的に増えていく。
一度この軌道に入ると、自然に止まることはない。

☑️10 巨額投資は異常ではない
したがって、現在見えている巨額のAI投資は、過剰でも異常でもない。
むしろ、この規模でなければ追いつけない段階に入っている。

☑️11 AWSの歴史が示すもの

現在のAI投資を見て「やり過ぎだ」と感じる人は、過去にAWSをどう見ていただろうか。
2008年から2009年にかけて、Amazonは莫大な設備投資を行い、ほとんど利益を出さなかった。
当時の市場では、「なぜここまでキャッシュを燃やすのか」「いつ利益が出るのか」という批判が絶えなかった。
しかし結果として、AWSは現在、年間およそ1,400億ドル規模の売上を生み、約300億ドルの利益を稼ぐ事業になっている。
当時「過剰」と言われた投資は、後から見れば極めて合理的だった。
AI投資も、まったく同じ構図にある。
初期段階ではキャッシュフローが悪く見えるが、それは将来の巨大なキャッシュフローを前倒しで支払っているにすぎない。

☑️12 AI企業はすでに「異常な存在」になっている
OpenAIやAnthropicのような企業は、人類の産業史の中でも極めて異例だ。
創業からそれほど時間が経っていないにもかかわらず、すでに大規模な売上ポテンシャルを持ち、かつ高い成長率を示している。
これは「いつかそうなるかもしれない」という話ではない。
すでに数字として現れている現実だ。
もし計算能力の制約がなければ、これらの企業はさらに急成長する。
ここでもボトルネックは需要ではなく、計算インフラにある。

☑️13 インフラ構築は一度で終わらない
AIインフラは、道路や橋のように一度作って終わるものではない。
モデルは進化し続け、それに伴って必要な計算量も変わり続ける。
5年から7年ごとに、インフラ全体の世代交代が必要になる。
安定した状態に到達するまでには、7〜8年、あるいはそれ以上の構築期間が必要になる。
したがって、「どこかで投資が止まる」という前提自体が誤っている。

☑️14 ソフトウェアの定義が根本的に変わった
従来のソフトウェアは、事前に定義されたロジックを実行する存在だった。
ExcelやPowerPointは、あらかじめ用意された機能を人間が操作する道具にすぎない。
AI時代のソフトウェアは違う。
文脈を理解し、状況に応じて判断し、必要に応じてツールを使い、結果を生成する。
これは「エージェント」という新しい形態だ。
その結果、ソフトウェアは静的な存在ではなく、常に計算を必要とする動的な存在になった。

☑️15 なぜ計算需要は爆発するのか
将来のアプリケーションでは、テキストだけでなく、画像、音声、動画、3Dなど、あらゆる要素がリアルタイムで生成される。
すべてのピクセル、すべての音が計算結果になる。
これは従来のITインフラでは想定されていなかった負荷だ。
そのため、GPU需要は量だけでなく質の面でも別次元に入る。

☑️16 「バブル論」への最大の反論
2000年代初頭の光ファイバー投資では、利用率が一時的に7%程度まで低下した。
それに対し、現在のGPUは常にフル稼働している。
未使用の計算資源は存在しない。
この一点が、「バブル」と決定的に異なる。

☑️17 NVIDIAの競争優位はどこにあるのか
競合企業はしばしば「推論用チップが40%安い」と主張する。
しかし重要なのは、チップ単体の価格ではない。
NVIDIAは、性能、ソフトウェア、ネットワーク、開発環境を含めた全体最適を行っている。
その結果、1トークンあたりの実効コストが最も低くなる。
性能が10倍になれば、コストは5分の1になる。
この差は単純な価格競争では埋まらない。

☑️18 スケールとスピードという二重の壁
NVIDIAは史上最大規模のコンピューティング供給網を持っている。
同時に、技術進化のスピードも最速だ。
競合が同じことをしようとすれば、まず規模で負け、次に速度で負ける。
この二重の壁が、追随を極めて困難にしている。

☑️19 クラウド企業との特殊な関係
AWS、Azure、Google Cloudは、NVIDIAの最大顧客であると同時に競合でもある。
それでもNVIDIAはロードマップを共有し、顧客が自社チップを開発することも妨げない。
これは弱さではなく、自信の表れだ。
どの選択をされても、最終的にNVIDIAが必要になる構
造を理解している。

☑️20 NVIDIAは唯一の共通プラットフォーム
世界中のスタートアップ、研究機関、大企業が、NVIDIAの上でAIを構築している。
すべての主要クラウドに存在する、唯一の共通アーキテクチャだ。
このエコシステム自体が、さらなる需要を生み出している。

続く

[Ming @tslaming]

$TSLA 🇨🇭 During his fireside chat at Davos, Elon Musk highlighted several key developments and future goals for Tesla 🔮

Focusing heavily on robotics and autonomous software, the following breakdown captures the key insights and strategic updates he shared regarding the company’s future 🎙️

🫘 Mission evolution and "Sustainable Abundance"

Musk clarified that Tesla’s mission has expanded beyond its original focus on sustainable energy to include the concept of "sustainable abundance".

He argued that the only way to truly solve global poverty and ensure a high standard of living for every person on Earth is through the widespread deployment of AI and robotics.

By creating a world with ubiquitous, low-cost AI and millions of humanoid robots, Musk envisions a global economic expansion that is "beyond all precedent".

He believes this shift is necessary because there simply aren't enough young people to perform the labor required to care for an aging global population.

🚘 The "Robot within the car" and transferable tech

Musk emphasized that Tesla's progress across different sectors is driven by a unified engineering discipline.

He explicitly referred to the software in Tesla vehicles as an upgradable "robot within the car", noting that frequent software updates—sometimes as often as once a week—constantly improve the vehicle's capabilities.

He explained that the real-world artificial intelligence and engineering developed for Tesla's automotive fleet are directly transferable to the Optimus humanoid robotics program.

This synergy allows Tesla to apply its existing expertise in navigation, vision, and actuators to create functional humanoid robots more rapidly.

🤖 Tesla Optimus and the future of labor

Musk revealed that Tesla Optimus prototypes are already performing simple tasks in factories. He expects that by the end of this year, they will be capable of more complex industrial work, eventually reaching a level of reliability where they can be sold to the public by the end of next year.

He envisions these robots eventually outnumbering humans, providing a "benign scenario" where robots saturate all human needs.

Musk stressed that these machines must be designed with extreme safety in mind so they can be trusted to watch over children or care for elderly parents in a home environment.

🚖 FSD and the Robo-taxi rollout

Musk asserted that autonomous driving is now "essentially a solved problem". He highlighted that the safety of FSD is becoming so evident that some insurance companies are offering customers half-price premiums if they use the software while driving.

The goal for the near future is a massive expansion of Tesla’s robo-taxi service, which Musk expects to be "very widespread" across the U.S. by the end of 2026.

Furthermore, he mentioned plans to seek regulatory approval for supervised FSD in both Europe and China, potentially as early as next month.

🔋 Energy infrastructure and Solar manufacturing

To support the energy-heavy future of AI and robotics, Musk stated that Tesla is working to manufacture 100 gigawatts of solar power per year within the United States. He estimates it will take approximately three years to reach this scale of production.

This initiative is aimed at overcoming high tariff barriers that currently make solar deployment artificially expensive.

Musk's goal is to provide the steady-state power and battery storage needed to meet the exponentially increasing electrical demands of modern technology.

[Ming @tslaming]

BOOOM 🚨 TESLA INVENTED A FILE FORMAT CALLED ".SMOL" BUT IT IS HUGE BECAUSE THIS "TINY" INNOVATION ACCELERATES FSD TRAINING BY 400% ⚡️

You can build the world’s fastest AI training cluster, but it is useless if your storage system buckles under the pressure. For Tesla, that pressure has become existential: they are no longer just processing video; they are processing reality, and standard digital containers can simply no longer hold the weight.

Realizing that legacy formats like MP4 and CSV were acting as a "restrictor plate" on their AI development, limiting the engine's performance no matter how hard they pressed the accelerator, Tesla decided to reinvent the file system from scratch.

Patent WO 2024/073080 introduces a proprietary data format designed to slash Input/Output Operations (IOPS) by 400%.

By abandoning traditional storage methods for a novel "header-first" architecture, Tesla has effectively unlocked a 4x speed multiplier for their FSD training pipeline without adding a single new microchip.

They call it ".SMOL", which sounds like "small", but its impact on the future of autonomy is massive.

However, to understand the magnitude of this innovation, we first have to look at the invisible wall Tesla hit.

⏳ The "data loading" bottleneck in AI supercomputing

The training of complex machine learning models, specifically the "Occupancy Networks" used by Tesla to predict the physical space around a vehicle, is an incredibly resource-intensive process. These models consume millions of video clips, sensor readings, and "ground truth" data points. These are verified, real-world examples that serve as the answer key for the AI's test.

To understand why this is so difficult, imagine playing a 3D version of Tetris where the game board is the real world. The car must instantly predict which empty spaces are safe to drive in and which are "occupied" by solid objects like curbs, pedestrians, or trees. The "Occupancy Network" is the brain that solves this puzzle, but it needs to watch millions of hours of driving footage to learn the rules.

In a traditional computing environment, the Graphics Processing Unit (GPU) is the fastest component, capable of crunching numbers at lightning speeds. However, the GPU often sits idle, wasting expensive time and electricity, simply waiting for the storage system to find, retrieve, and load the next batch of training data from the hard drive.

This bottleneck creates a phenomenon known as "Data Starvation". Think of it like owning a Ferrari (the GPU) but fueling it with a coffee straw (the storage system). No matter how fast the engine can spin, it can only go as fast as the fuel arrives.

Traditional file formats are not optimized for the specific, randomized access patterns required by deep learning. When a training system needs a specific frame or tensor (a block of mathematical data) associated with a specific timestamp, standard formats often require the computer to parse through large portions of a file sequentially to find the relevant data. This is akin to trying to find a specific song on a cassette tape where you have to fast-forward past everything else to get to it.

Furthermore, these formats are often bloated with redundant information. For example, if a file contains ten minutes of driving data, static attributes like the car's model, VIN, or map region might be repeated in every single row of data.

Across petabytes (millions of gigabytes) of storage, this redundancy creates massive wasted space and forces the storage system to read gigabytes of useless information just to get to the actual pixel data needed for training.

Facing this mountain of wasted effort, Tesla engineers realized they couldn't just optimize existing tools. They had to build a new one.

📑 Tesla's solution: A "header-first" indexed architecture

Tesla’s solution is a bespoke file format designed specifically for the high-performance needs of the Dojo supercomputer and their GPU clusters. At the core of this invention is a sophisticated "header-first" architecture that turns the data file into a random-access database.

Think of standard video formats like an ancient scroll. If you want to find a specific sentence in the middle, you are forced to unroll the entire thing from the beginning until you find it. This "sequential access" is agonizingly slow for a computer.

Tesla’s format, by comparison, functions like a reference book with a detailed master index at the very front. When the system generates a file, it creates a master index in the header that maps every single timestamp to a specific memory offset. This is a precise byte location on the disk.

This allows the training system to perform "random access" reads. Instead of scanning through a file to find a specific moment in time, the processor reads the header, calculates exactly where that data lives on the disk, and "skips" directly to that memory position. It effectively allows the processor to "teleport" through the data, skipping over gigabytes of irrelevant information.

To solve the redundancy issue, the format segregates data into dynamic and static segments. Data that does not change over the course of the recording, such as the vehicle's dimensions, the camera calibration settings, or the map data, is written only once at the beginning of the file immediately following the header.

The dynamic data, such as velocity, video tensors, and steering inputs, are then stored in subsequent rows. This separation ensures that the system never wastes bandwidth reading the same static variable twice.

The patent documentation suggests that this specific architectural change results in an approximately 11% decrease in file size and, more importantly, reduces the number of Input/Output operations (IOPS) by a factor of four.

But to truly appreciate why this custom architecture is necessary, we must compare it against the standard tools that failed to keep up.

🆚 Comparison: Why standard formats fail at scale

In the world of standard data storage, formats like CSV, JSON, or standard MP4 containers are designed for compatibility and linear playback, not for the massive parallel processing required by AI. These legacy formats impose a hidden "tax" on performance that becomes crippling at Tesla's scale.

Standard video files, for instance, are optimized for Netflix, not Neural Networks. They utilize a compression technique known as inter-frame compression, where the file stores one complete "Keyframe" followed by a series of partial frames that only describe changes.

This creates a massive inefficiency for AI training: if a supercomputer needs to grab a random batch of 50 frames to train a model, it cannot simply view "Frame 50" in isolation. It often has to find "Frame 1" and mathematically reconstruct frames 2 through 49 just to figure out what Frame 50 looks like. This is wasted computational effort, spending 98% of the GPU's power decoding data it will immediately throw away.

Text-based formats like CSV suffer from a similar "parsing penalty" due to variable row widths. Because the number "100" takes up more characters than "1", every row in a dataset has a different physical length on the disk.

If you tell a computer to "go to Row 1,000,000", it cannot simply teleport there. It is forced to start at Row 1 and scan every single comma and newline character for the first 999,999 rows just to locate where the millionth row begins. This turns a simple retrieval task into a heavy processing job.

Tesla's format eliminates both problems by treating video and sensor data like a "Look-Up Table", or a cheat sheet. Because the header contains the exact "Byte Offset" (address) of every data row, the read operation is deterministic. The system doesn't need to scan for newlines or reconstruct previous video frames; it reads the address from the header, jumps instantly to that specific byte on the hard drive, and grabs a perfectly self-contained "bundle" of data.

Ideally, you can think of standard formats like a cassette mixtape: if you want to hear the fifth song, you have to physically fast-forward through the first four songs to get there. The act of "seeking" takes time.

Tesla’s format is more like a vinyl record. You can see the grooves (the index), lift the needle, and drop it instantly on the exact second of the song you want to hear. It is instant, precise, and random-access.

This shift from sequential to random access doesn't just look good on paper; it fundamentally changes the physical relationship between the computer's components.

🏎️ The hardware bridge: Optimizing the "data loader"

The patent explicitly highlights how this file format serves the physical architecture of high-performance compute clusters, specifically a component called the "Data Loader". In these systems, there is often a distinct struggle between the Central Processing Unit (CPU) and the Graphics Processing Unit (GPU) that functions much like a high-speed relay race.

In this analogy, the CPU acts as the "Feeder" that runs ahead to grab raw files, unpack them, and prepare them, while the GPU acts as the "Runner" that takes that package and sprints through the complex mathematical calculations.

The problem in modern AI is that the Runner (GPU) has become thousands of times faster than the Feeder (CPU). The GPU finishes its lap in milliseconds and reaches back for the next baton, but the CPU is often still struggling to open the packaging of the next file.

With standard formats like MP4 or CSV, the CPU is bogged down by "administrative" tasks: it must open the container, scan to find the right timestamp, decode the compression, and mathematically convert that data into a format the GPU can understand. This heavy lifting forces the CPU to drop the baton, causing the GPU to stop running and sit idle. This is a phenomenon known as "Data Starvation".

In a cluster with thousands of H100 GPUs, this idle time burns millions of dollars in electricity without producing any intelligence.

Tesla’s .smol format is designed to eliminate this administrative burden entirely. Because the data is stored in raw, pre-transposed tensors with a precise byte-offset index, the CPU no longer needs to "decode" or "convert" anything. It simply looks up the address, grabs the specific byte range, and hands it directly to the GPU.

This shifts the CPU's role from being a "Translator" (which is slow and computationally expensive) to a "Logistics Manager" (which is fast and efficient), ensuring the Feeder is always one step ahead of the Runner and keeping the training cluster running at 100% capacity.

Once the data is successfully handed off to the processor, the system employs another trick to ensure it isn't choked by unnecessary volume.

📰 Memory layout: The "columnar" optimization strategy

The patent details a clever "columnar" organization strategy within the data rows to further optimize read speeds. Within the data file, different types of data (columns) are not just thrown in randomly; they are strictly arranged based on their data size, from smallest to largest.

This architecture mimics how humans read a newspaper. When you scan a physical paper, you skim the headlines first to see if the story is relevant before you commit the mental energy to reading the long, dense paragraphs below.

Standard file systems often force the computer to do the opposite—loading the massive "body text" (video data) just to find out what the "headline" (metadata) says. It is akin to packing your passport at the very bottom of a full suitcase; just to check your flight number, you have to unpack your entire wardrobe.

Tesla organizes the data to prevent this inefficiency. The system places smaller, lightweight data elements, like simple integer velocity values or steering angles, at the very beginning of the data row and places the massive, heavy data elements, like complex 4D video tensors, at the very end.

This allows for a technique known as "early rejection". If a specific training job is looking only for "high-speed highway driving", the reader can grab the first few bytes of the row, check the velocity, and if the car is moving slowly, stop reading immediately. It never touches the massive video chunks at the end of the row, saving the system from clogging its limited bandwidth with gigabytes of useless pixel data.

Efficiency, however, is only half the battle. In the high-stakes world of autonomous driving, speed is worthless without absolute precision in how the machine perceives time.

📸 Syncing the senses: Multi-camera "bundling"

A critical detail revealed in the patent is the concept of "bundling" within data rows. In the context of Tesla’s "Occupancy Networks", which aim to build a real-time 3D volumetric map of the world, synchronization is everything. If the data from the left camera is even a few milliseconds out of sync with the right camera, the AI cannot accurately calculate depth or velocity.

Without precise synchronization, the system essentially behaves like a "dizzy" AI. To understand why this matters, imagine walking down the street with a unique condition where your left eye sees the world as it is now, but your right eye sees the world as it was half a second ago.

If a cyclist rides past you, your left eye might see them ten feet away while your right eye sees them twenty feet away. Your brain would fail to merge these two images, resulting in double vision and a complete inability to judge the cyclist's speed. This is exactly what happens to an AI if its cameras are not perfectly synchronized: a fast-moving car will appear in two different places at once, leading to "ghosting", phantom braking, or dangerous swerving.

Tesla’s file format solves this by treating a "time instance" as a unified, unbreakable container. It is essentially a "time capsule" of reality. When the system requests data for timestamp 1337.37, it doesn't just grab a loose collection of images; it retrieves a perfectly aligned "bundle" of tensors representing the entire 360-degree view (front, side, repeater, and B-pillar cameras) frozen at that exact millisecond.

This ensures that when the Occupancy Network stitches these eight camera feeds together, the edges align perfectly. The curb on the left camera flows seamlessly into the curb on the front camera, creating a cohesive, solid 3D representation of the world rather than a glitchy, fragmented mess. To store these synchronized realities efficiently, Tesla had to abandon the language of traditional computing and adopt the native language of AI.

📦 Data types: Native tensor support and encryption

Unlike generic file formats, this architecture is built natively for Machine Learning. The data rows don't just store text or numbers; they are designed to store "tensors", which are the multi-dimensional arrays that are the fundamental language of neural networks.

What is a Tensor? A tensor is simply a container for numbers that preserves their relationship to each other.
- A shopping list is a 1D Tensor (a line of items).
- A spreadsheet is a 2D Tensor (rows and columns).
- A color video is a 4D Tensor (Height x Width x Color x Time).

The patent describes the ability to store these tensors in transposed forms directly in the file. This is a significant optimization because neural networks often require data to be "flipped" or transposed before processing. By storing the data in the format the model expects, Tesla removes the need for the CPU to perform these costly transformation operations during the training loop.

Standard formats are like buying "flat-pack" furniture from IKEA that you have to spend hours assembling before you can use it. Tesla's format stores the data "fully assembled", meaning it is pre-rotated and pre-packaged exactly how the GPU likes it, so it can be used the instant it is loaded.

Additionally, the format supports encryption at the tensor level, ensuring that specific sensitive data streams can be locked down without making the rest of the file unreadable. This shift to a native tensor-based architecture unlocks a capability far beyond just better self-driving cars.

🤖 Universal "ego": Built for Optimus, not just cars

A fascinating, easily missed detail is the broad definition of the "ego", which is the technical term for the device collecting the data. The patent explicitly states that the system is not confined to vehicles but includes "a general purpose, bipedal, autonomous humanoid robot".

This confirms that this file format is the foundational data infrastructure for the Tesla Bot (Optimus) as well, serving as a sort of "Rosetta Stone" that allows distinct machine species to speak the same language.

The format's ability to handle heterogeneous sensor data is critical here because a robot experiences reality differently than a car. A vehicle primarily focuses on perception (lanes, obstacles, signs), whereas a humanoid robot relies heavily on proprioception (balance, joint torque, finger pressure). If Tesla used standard, rigid file schemas, they would need entirely separate data pipelines for the car and the robot.

However, because this format uses "tensors", or generic mathematical containers, it effectively decouples the AI's "mind" from the machine's "body". To the training cluster, it does not matter if a specific tensor represents the radar return from a Model Y bumper or the tactile pressure from Optimus’s pinky finger; it is all just math.

This abstraction allows Tesla to ingest any kind of data into the same unified training pipeline without needing to invent a new file format for every new hardware product, creating a single, scalable "brain training" engine for every machine Tesla builds. But supporting such a diverse range of machines requires a file structure that is rock-solid and unbreakable.

🔒 Efficiency: The "read-only" enforcement

A unique characteristic of this file format is its strict "read-only" nature (immutability). The patent explains that once these files are generated, they are effectively sealed—never to be edited, appended, or modified again.

While this might sound like a limitation in a world accustomed to dynamic documents, it is actually a deliberate performance feature. You can think of standard editable files like a loose-leaf binder: you can add pages anywhere, but the page numbers eventually stop making sense, and finding things becomes a chore.

Tesla’s format is more like a printed encyclopedia. Because the text is permanently "etched in stone", the index at the front is guaranteed to be 100% accurate forever. Page 50 will always be Page 50.

If the system allowed engineers to insert new data into the middle of a file, it would break the rigid byte-offset mapping stored in the header, forcing the computer to waste precious milliseconds re-calculating where the data moved to.

By enforcing this immutable structure, Tesla ensures that the memory layout remains perfectly contiguous and predictable. This predictability is the secret sauce behind the reported 4x reduction in IOPS; the reading hardware never has to "double-check" the file structure. It trusts the index blindly, allowing it to fetch data at the theoretical speed limit of the drive.

🚀 The future is Exabyte: Why this matters for the next decade

When you combine the architectural choices of the header index, the columnar layout, and the read-only enforcement, the result is a strategic advantage that scales linearly with Tesla's ambitions.

This patent serves as a critical enabler for Tesla's ambition to solve Full Self-Driving (FSD) and general-purpose robotics. As Tesla's fleet grows, the amount of training data flowing back to their data centers is shifting from petabytes to exabytes. Without this specialized file format, the "data loading" phase would become the hard ceiling on how fast they can improve their AI.

In the immediate term, the impact of this invention is mathematical and massive. Training a single version of FSD requires processing billions of video frames, and by reducing file size by 11%, Tesla effectively saves over 100 petabytes of storage at the exabyte scale. This represents tens of millions of dollars in hard drives and electricity that simply do not need to be purchased.

Furthermore, increasing data throughput speed by 400% directly reduces the "Epoch Time", or the time it takes the AI to learn from the entire dataset once. If an epoch drops from four days to one day, engineers can test four times as many ideas per week, accelerating the critical path between FSD versions.

Looking beyond these immediate efficiency gains, this format provides the unified "digital nervous system" for Tesla's entire robotics roadmap. Since it is agnostic, caring only about mathematical "tensors" rather than specific car parts, it allows the Dojo supercomputer to train a Model Y in the morning and an Optimus robot in the afternoon using the exact same data pipeline.

While competitors are still struggling to integrate off-the-shelf software, Tesla has reinvented the most basic unit of computing, the file itself, to build a vertically integrated machine learning engine. They aren't just building better AI models; they are building a faster assembly line for intelligence.

[Elon Musk @elonmusk]

Now that the AI5 chip design is in good shape, Tesla will restart work on Dojo3.

If you’re interested in working on what will be the highest volume chips in the world, send a note to [email protected] with 3 bullet points on the toughest technical problems you’ve solved.

[Elon Musk @elonmusk]

@imPenny2x AI4 by itself will achieve self-driving safety levels very far above human.

AI5 will make the cars almost perfect and greatly enhance Optimus.

AI6 will be for Optimus and data centers.

AI7/Dojo3 will be space-based AI compute.

[米国株インサイト🚀テーマトラッカー @invest_us_jp]

🚀【米国株テーマトラッカー】リリース！

本気の投資家向けに開発しました。
✅ テーマの盛り上がりを追跡
✅ 次の高騰テーマを逃さない
✅ 今なら全機能無料

2026年のトレンドテーマは？ 新年初日の勝ち組・負け組を分析👇
一位は...是非、アプリをご覧ください！
https://t.co/4DzzDqmKWb

[HUTCH ハッチ (岡元兵八郎) @heihachiro888]

「とても裕福になるために、20回も正しい判断をする必要はありません。おそらく4回か5回で十分でしょう……」
「すべてのことについて意見を持たなければならない、と思い込むのはひどい間違いです。」―ウォーレン・バフェット

[Startup Archive @StartupArchive_]

Peter Thiel on Elon Musk’s success: “Maybe the rest of us are too risk averse”

“When Elon was building both Tesla and SpaceX in the 2000s, people thought he was just really, really crazy — even those of us who’d worked with him at PayPal. There was this PayPal book that David Sacks and I thought of writing, and the Elon chapter was I think entitled something like ‘The Man Who Knew Nothing About Risk’ or something like this. There are all these crazy Elon stories I could tell.”

Thiel continues:

“If one of the two companies had succeeded, you would say, well, maybe he got really lucky. But when two out of two companies that people thought were completely harebrained in the 2000s, when they both succeed, you have to reassess it. Maybe the rest of us are somehow too risk averse or there’s something about risk he knows that we don’t.”

Video source: @AspenInstitute (2024)

[やす 🇺🇸 ベンチャーキャピタル @YasLovesTech]

Noteを書きました🗒️
2026年のS&P500リターンは「14％前後」と予想しています。

1万字超とかなり長めですが、マクロ・EPS・PERの視点から、その根拠をできるだけ丁寧に整理しました。

参考になりましたらRT・いいねしてもらえると励みになります🙏

https://t.co/spJ3DwVDlE

[Sawyer Merritt @SawyerMerritt]

In a new interview today, Ron Baron explains what drew him to invest in @elonmusk and his companies:

"When it became obvious his cars were meeting very strong demand, we invested $400 million between 2014 and 2016. We’ve made about $8 billion on that so far.

In 2017, we began to invest in SpaceX. We’ve invested $1.5 billion so far, and that’s now worth about $10 billion.

What separates him apart? Young people are dying to work there. He goes into these meetings and, in any one discipline they’re working on, takes them into a room for a couple of hours.

He goes around and talks to everyone, asking them questions about what they’re doing. He keeps getting deeper and deeper, and then at the end of the meeting, people walk out of that room—especially first-time engineers—shaking their heads and saying, “Can you believe what we’ve just heard? I’m so glad I survived that meeting.”

The people who work there do so because they want to work with this person—Elon. What you've invested in is a company with a culture. He gives responsibility to talented people. He's not afraid to fail."

[Elon Musk @elonmusk]

The Sun is an enormous, free fusion reactor in the sky. It is super dumb to make tiny fusion reactors on Earth.

Even if you burned 4 Jupiters, the Sun would still round up to 100% of all power that will ever be produced in the solar system!!

Stop wasting money on puny little reactors, unless actively acknowledging that they are just there for your pet science project jfc.

[*Walter Bloomberg @DeItaone]

MUSK: THE SUN IS AN ENORMOUS, FREE FUSION REACTOR IN THE SKY. IT IS SUPER DUMB TO MAKE TINY REACTORS ON EARTH

[Sawyer Merritt @SawyerMerritt]

Elon Musk's ecosystem of companies is evolving into an incredibly interconnected network. Advancements in one fuel progress in others.

SpaceX will deploy AI-focused satellites equipped with Tesla-designed chips and powered by Tesla solar panels. Those satellites support xAI, which trains Grok, the same AI that ultimately runs inside Tesla vehicles and Optimus robots. Your Tesla will stay connected through Starlink.

As SpaceX begins generating massive revenue from space-based datacenters, and further growth with Starlink, that funding accelerates the push to send Optimus robots, and ultimately humans and Tesla vehicles to Mars. SpaceX will build a dedicated satellite constellation around Mars called "Marslink," providing global internet connectivity just like Starlink does on Earth.

The synergy between Elon's companies is increasing rapidly, most people just don't realize it yet. I think Elon feels that AI accelerates the mission of all of his companies, including SpaceX’s goal of making life multi-planetary.

A rapid pace of innovation and progress is central to how he operates, which often results in him being a first mover. That’s why some of his decisions aren’t immediately understood, but over time, they tend to make sense. In my opinion, space-based AI data centers will be no different.

[Elon Musk @elonmusk]

A major additional factor should be considered.

Satellites with localized AI compute, where just the results are beamed back from low-latency, sun-synchronous orbit, will be the lowest cost way to generate AI bitstreams in <3 years.

And by far the fastest way to scale within 4 years, because easy sources of electrical power are already hard to find on Earth. 1 megaton/year of satellites with 100kW per satellite yields 100GW of AI added per year with no operating or maintenance cost, connecting via high-bandwidth lasers to the Starlink constellation.

The level beyond that is constructing satellite factories on the Moon and using a mass driver (electromagnetic railgun) to accelerate AI satellites to lunar escape velocity without the need for rockets. That scales to >100TW/year of AI and enables non-trivial progress towards becoming a Kardashev II civilization.

[phil beisel @pbeisel]

Autonomy, the future of auto

Let's predict some potential future states in an autonomous world.

Cities:
Robotaxis will dominate urban mobility to such a degree that many cities will eventually prohibit user-driven vehicles (except for commercial or specialty use) from entering downtown cores. Street parking will vanish, replaced by dynamic pickup and drop-off zones optimized for autonomous fleets. Parking garages will become obsolete and repurposed. The city will become more green.

Highways:
On highways, the shift will be equally transformative. “HOV lanes” will evolve into “autonomy lanes,” where driverless vehicles can travel at higher speeds and tighter spacing with network-level coordination. Vehicle “trains” will form automatically to boost aerodynamics and efficiency (could cut energy consumption per mile by 25%). It will eventually feel weird to drive alongside a sea of robots.

Ownership and Economics:
Personal car ownership will decline sharply. Cost per mile of Robotaxi travel could fall below $0.25 (cheaper than public transit today) making private ownership economically irrational for most people. Vehicles will shift from being depreciating assets to productive, revenue-generating units.

Infrastructure and Industry:
Gas stations will fade away, replaced by high-throughput charging depots designed for fleet vehicles. Auto insurance, traffic enforcement, and even car dealerships will shrink dramatically (yea!). Governments will adapt with new taxation models (boo!) perhaps a per-mile “autonomy tax” to replace fuel taxes, parking revenues, etc.

Culture:
The automobile, once the symbol of freedom, will evolve into a mobility utility. Yet, paradoxically, autonomy could expand freedom for millions: the elderly, disabled, and those unable or unwilling to drive.

What are your thoughts?