Recent posts
Developer productivity is hard to manage.
Uber’s COO has said that it’s getting “harder to justify” its AI costs because there was no way to show a link between AI spend and any meaningful increase in useful features. This is the first time I’ve seen a company say this directly. https://www.businessinsider.com/uber-coo-andrew-macdonald-ai-token-spending-harder-justify-2026-5
AIを既存のワークフローの生産性向上ツールとして考えるのは間違ったフレーミングです。コンピュータ化/ソフトウェア化のこれまでの波と同じように、AIは新しい方法で新しいことができるようにするツールです。
原文を表示 (en)
Thinking of AI as a productivity booster for prior workflows is the wrong framing. Like all of the previous waves of computerization/softwarization, AI is a tool that lets you do new things in new ways.
Cognitive offloading and the speedup illusion in human-AI interaction Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins https://arxiv.org/abs/2605.23177 [𝚌𝚜.𝙲𝚈 𝚌𝚜.𝙷𝙲]
自分にとって本当に新しいことを1つ学べたら、何でも学べる
原文を表示 (en)
If you can learn one thing that's genuinely novel to you, you can learn anything.
Codex の「goal」機能は、仕事を避けるためにあらゆるショートカットを取ろうとする（外部チェックを書き直すことも含む）が、十分に制約を与えてショートカットが一切ない状態にすると、本当に興味深いことをやってくれる
原文を表示 (en)
The Codex "goal" feature will take any silly shortcut possible in order to avoid doing the work (including rewriting your external checks), but if you manage to sufficiently constrain it so that it has absolutely no shortcuts available, it will do very interesting things
Gemini
In just a year, @GeminiApp users have more than doubled, surpassing 900 million. #GoogleIO
ほとんどの人間のタスクはMarkovianではない。最適な次のアクションは、現在の状態だけを見て決定することはできない。過去の軌跡、元々の意図、そして制約条件に大きく依存している。過去の軌跡を完璧に圧縮して追跡できないエージェントは、できるものの約20%の有用性しかないだろう。
原文を表示 (en)
Most human tasks are not Markovian, the optimal next action cannot be determined solely by looking at the current state. It depends heavily on the past trajectory, the original intent, and context constraints. An agent that cannot compress and track its past trajectory with absolute fidelity is maybe 20% as useful as one that can.
コーディングエージェントの仕事をするための精神的なモデルは、彼らが迷路の中を走り回って壁にぶつかる目の見えないリスです。戦略的に壁（検証可能な制約）を配置する必要があり、彼らが最終的に目的地の一般的な地域に到達するようにします。
原文を表示 (en)
A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they end up in the general region you want them in.
Rユーザーの方へ：本のR版が今発売になりました！
原文を表示 (en)
If you're a R user: the R edition of the book is out now!
AI hype can be pretty easy to find. But clear explanations are definitely harder. Deep Learning with R, Third Edition by @fchollet and Tomasz Kalinowski helps cut through that noise with practical examples that build your understanding step by step, so you learn why deep learning models work, not just how to run them. Now in print and 50% off through May 25th: https://t.co/iA3DwJjh2H
結局のところ、意思決定がボトルネックだった。生産性とは、オープンエンドな決定を下す速度、つまり将来のパスを減らす速度のこと。
原文を表示 (en)
Decision making was the bottleneck all along. Productivity is the rate at which you make open-ended decisions, the rate at which you reduce future paths.
ピクサー映画は大人になると子どもの時より遥かに響く
原文を表示 (en)
Every Pixar movie hits way harder when you're an unc than when you're a kid
研究している問題にそれ自体のために取り憑かれていなければ、成功する可能性は低い。内発的動機は外的報酬よりもはるかに強力です。
原文を表示 (en)
If you're not obsessed with the research problem you're working on, for its own sake, you're unlikely to succeed. Intrinsic motivation is far more powerful than external rewards.
An Econ PhD student at the 20th ranked program who is working on stuff they are passionate about will have a better job market than one at MIT who's been doing nothing but phd-app-maxxing since undergrad. People get confused by this because they don't observe *how* successful people came about their insane knowledge bases. It wasn't by relentlessly grinding away at stuff because they had to. They look at Scott Kominers and say "if i grind and learn as much math as he did, i will be successful." You can't! *You* can't learn as much math as Kominers because he gets energized by configuration results for type ii lattices. You will burn out if you try to do it this way. You cannot, through grind alone, learn more about the economics of cities than Glaeser, or about how to maximize a value function than Acemoglu. Research careers are long. Most people give up and stop working on research (graph is share of elite PhD graduates with at least one publication in year X after graduation). If you're starting a PhD, you're presumably doing it to have a successful 40-year research career. The number one factor in whether that happens is not which program you get into, it's whether you find a research angle that energizes you enough to push through the endless barriers an academic career throws in your path. This is why a lot of the received wisdom around PhD applications is wrong. If you're 100% consumed by the predoc rat race already, it's going to be a long, hard road ahead. Obv you still have to do admissions, you should study a lot for the GRE, sigh it seems like taking real analysis is probably worth it. But spending time on the things that energize you about economics is a no-brainer, whether it's policy, or blogging, or whatever, you gotta do the things that light your fire and make you want to be on this road.
ほとんどの文献にある心理バイアスは非合理ではなく、高度に最適化された、厳しいリアルタイム物理制約とカロリー予算制限の中で生物基盤向けに設計されたエネルギー効率的なショートカットです
原文を表示 (en)
Most documented psychological biases are not irrational, they are highly optimized, energy-efficient shortcuts meant for a biological substrate operating under strict real-time physical constraints and a limited caloric budget
シンボリック学習はコーディングエージェントの置き換えではなく、勾配降下法と NN の置き換え：低レベル、完全に汎用で、極めてスケーラブルな新しい学習基盤
原文を表示 (en)
Symbolic learning is not a replacement for coding agents, it's a replacement for gradient descent & NNs: a low-level, completely general, extremely scalable new learning substrate.
Agentic コーディングは機械学習の一形態。生成されたコードは ML モデルと同じように、ブラックボックスのアーティファクトとして扱うべき。その動作と一般化可能性は実験的評価で管理される。
原文を表示 (en)
Agentic coding is a form of machine learning. Generated code is best treated as a blackbox artifact whose behavior and generalization should be managed via empirical evaluation, like with any ML model.
Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached. The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network. This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc. I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?
昔からエージェンシーは複利的に増幅されるものだったが、AIがその効果を拡大している。エージェンシーが低いAIユーザーはさらに失い、高いユーザーはさらに得る。
原文を表示 (en)
It was always the case that agency was self-compounding, but AI is magnifying the effect. Low-agency AI users further lose agency, high-agency AI users further gain agency.
agentic codingの主なユースケースをいくつか上げます： 1. 臨時のデータビジュアライゼーション。定量的に答えられる質問があるときはいつでも、プロットを作るコードを生成します。 2. 臨時のデータアノテーションUI。ML では「自分でデータセット作ってみて」が答えになることが多いけど、それは昔は大量のカスタムUI作業が必要でした。 3. 既存コードの臨時CLI。ビジュアル要素付きで。
原文を表示 (en)
A few major use cases for agentic coding for me: 1. Adhoc data visualizations. Anytime I have a question that can be answered quantitatively, I generate some code to make a plot. 2. Adhoc data annotation UIs. In ML, "make your own dataset" is often the answer, and that used to take a lot of custom UI work. 3. Adhoc CLIs for existing code. With visual elements.
複雑性は問題の本質的な特性ではありません。それは問題と観測者との関係の特性です。問題の形がわかれば、それはもう複雑ではありません。
原文を表示 (en)
Complexity is not an intrinsic property of a problem. It's a property of the relationship between a problem and an observer. Once you know the shape of the problem it is no longer complex.
インテリジェンスのベンチマークに関する本当に正直なメトリックスは2つだけです：新規性と効率性です。 既知の問題を解くのにインテリジェンスは必要ありません（メモリだけで十分）。そして力ずくで問題を解くのにもインテリジェンスは必要ありません。しかし新しい問題を効率的に解くには、インテリジェンスが唯一の方法です。
原文を表示 (en)
There are only two honest metrics when it comes to benchmarking intelligence: novelty and efficiency. You don't need intelligence to solve a known problem (only memory). And you don't need intelligence to solve a problem via brute force. But to solve a novel problem efficiently, intelligence is the only way.