The Mercedes number is the strongest evidence here and it points at something narrower than the essay claims. A COBOL migration is bounded, has an oracle, and the output is checkable against the input. That is precisely the class where agents compound, and eight months to eight days is real.
The 90% figure is a different kind of number. Share of code authored is an input measure and it has never predicted throughput. Faros telemetry across 22,000 developers has merge rate per developer up 16.2%, while median time in review is up 441.5% and incidents per pull request up 242.7%. Authored share went to a very large number and the rate of change that actually ships moved 16%. The constraint moved from writing to reviewing and operating, and only one of those has an agent.
Worth flagging that the engineers-will-tell-you evidence is self-report. METR ran 16 experienced developers on their own repositories with screen recording: forecast 24% faster, measured 19% slower, and afterwards they still reported a 20% speedup. The bricklayer to architect shift looks right to me. The 100x is a claim about the review and operate half of the job, and that half has barely been measured.
Gov might be one of the things software hasn't devoured yet. The speed of compute meets all these legacy systems and bolted-on solutions dating back a decade or more. So at Botcopy we actually build the infrasturcture that lets gov agencies put Google's AI on public websites safely so real people can get stuff done 24/7 and not have to deal with the slow bureaucratic bottlenecks. We are live now in MI, OH, MN, NY, CA, WI and spreading within these states. Prediction: as agents get cheaper and better performing (and finally don't suck) we'll see many more, soon, multiplying like rabbits, there's a lot of pent up demand to get citizen-facing conversational AI task completion automations up on gov sites. We just released 44 prebuilt service patterns in partnership with ID.me, TTEC, Twilio, many more, which seriously cut costs, and chopped several months off the time to value so the barrier to entry is just plummeting. We have the first public-sector dedicated agent on Gemini Enterprise completing the loop analyzing performance and pointing agencies to the next prebuilt in our marketplace, it's a flywheel that works. We are prepping our Series A and would of course love to compare notes. Ben's THTAHT was required reading before we started over here and it still is to this day. Thanks for the great read as always.
My 2 cents Ive never considered cognition because I thought it was a slack addon for enterprise, decided to check it out due to this and it looks actually really advanced. And I find it so incredible that counterintuitively that the exponentially smarter models get, we need that much more infrastructure around them to maximize that intelligence (domain specific infrastructure) cognition is competing in the extremely lucrative software domain infrastructure, and it’s crazy that we are barely even scratching the surface of all of the domains that need this deep nuance to maximize AI intelligence. There has never been a better time…. To be a consumer!!
The 10x engineer line is a vanity metric until the job is named. An agent writing most of the code is a status, not a workflow. Write the one job the human still owns: who reviews the diff, what they can reject, and what stops a bad merge. If those three lines are blank, you did not get an architect. You got a faster typist with no kill switch.
The strongest argument is the historical pattern every leverage increase expanded the profession. But agents write 90% of the code says nothing about who reviews, owns and is accountable for it.
This entire piece reads like a marketing brochure for Cognition. Is the company struggling?
Not once have you addressed the actual job loss happening. Or the fact that if AI actually begins coding autonomously, then how does that not end software engineering (at least as we know it) and above all, wasn't it you yourself who said that the main push for AI is to combat the declining birth rate and still have a reliable and steady workforce?
The Mercedes number is the strongest evidence here and it points at something narrower than the essay claims. A COBOL migration is bounded, has an oracle, and the output is checkable against the input. That is precisely the class where agents compound, and eight months to eight days is real.
The 90% figure is a different kind of number. Share of code authored is an input measure and it has never predicted throughput. Faros telemetry across 22,000 developers has merge rate per developer up 16.2%, while median time in review is up 441.5% and incidents per pull request up 242.7%. Authored share went to a very large number and the rate of change that actually ships moved 16%. The constraint moved from writing to reviewing and operating, and only one of those has an agent.
Worth flagging that the engineers-will-tell-you evidence is self-report. METR ran 16 experienced developers on their own repositories with screen recording: forecast 24% faster, measured 19% slower, and afterwards they still reported a 20% speedup. The bricklayer to architect shift looks right to me. The 100x is a claim about the review and operate half of the job, and that half has barely been measured.
This comment has many of the classic LLM tells.
ai ; dr
AIの進歩を、人間の解放へ
今回の記事を空と法華経から照見して見えてくるのは、技術楽観主義を全面的に肯定することでも、AIを否定することでもありません。
技術は縁起によって生まれます。
人間の知性も縁起によって成立しています。
経済も社会も、無数の相互依存関係の中にあります。
ならば、AI革命の成果を一部の能力や所有権だけに帰属させる考え方は、改めて問い直すことができます。
そして法華経の一乗思想は、その成果を誰のために用いるべきかを考える倫理的な手がかりになります。
私たちが目指すべきなのは、人間を生産性によって序列化する社会ではありません。
誰もが、自分の能力と境遇に応じて創造し、学び、他者と関わり、生きることのできる社会です。
最後に、この照見を一つの言葉に凝縮します。
AIは、人間が何をつくれるかという限界を変える。
空は、人間の価値を何かをつくる能力から解放する。
法華経は、その解放を一部の人間ではなく、すべての人間へと開く。
技術が進歩するとは、より多くを生産することだけではない。
生産することに追われていた人間が、ようやく生きることそのものへ帰れるようになること。
その未来を実現できるかどうかは、AIの計算能力ではなく、私たちの智慧と慈悲にかかっている。
照見の核心は、「ソフトウェアが世界を飲み込む」という未来を、「智慧と慈悲によって世界を支える」という未来へ転換できるかどうかにあります。
それこそが、空と法華経から見たAI革命の最も根本的な問いだと思います。
Gov might be one of the things software hasn't devoured yet. The speed of compute meets all these legacy systems and bolted-on solutions dating back a decade or more. So at Botcopy we actually build the infrasturcture that lets gov agencies put Google's AI on public websites safely so real people can get stuff done 24/7 and not have to deal with the slow bureaucratic bottlenecks. We are live now in MI, OH, MN, NY, CA, WI and spreading within these states. Prediction: as agents get cheaper and better performing (and finally don't suck) we'll see many more, soon, multiplying like rabbits, there's a lot of pent up demand to get citizen-facing conversational AI task completion automations up on gov sites. We just released 44 prebuilt service patterns in partnership with ID.me, TTEC, Twilio, many more, which seriously cut costs, and chopped several months off the time to value so the barrier to entry is just plummeting. We have the first public-sector dedicated agent on Gemini Enterprise completing the loop analyzing performance and pointing agencies to the next prebuilt in our marketplace, it's a flywheel that works. We are prepping our Series A and would of course love to compare notes. Ben's THTAHT was required reading before we started over here and it still is to this day. Thanks for the great read as always.
AIの本当の革命は、人間を1000倍働かせることではない。 「働くことこそ価値である」という執着そのものを空じる可能性にある。
そして『法華経』からさらに一歩進めれば、
技術の価値は、どれだけ人間を超えたかではなく、どれだけ多くの人間の可能性を開いたかによって測られる。
Cognitionが「1000x engineer」を生むだけなら、これは巨大な生産性革命です。
しかし、これまで創る側に立てなかった無数の人を創る側へ連れていくなら――
それは少しだけ、「一仏乗のテクノロジー」に近づくのだと思います。
My 2 cents Ive never considered cognition because I thought it was a slack addon for enterprise, decided to check it out due to this and it looks actually really advanced. And I find it so incredible that counterintuitively that the exponentially smarter models get, we need that much more infrastructure around them to maximize that intelligence (domain specific infrastructure) cognition is competing in the extremely lucrative software domain infrastructure, and it’s crazy that we are barely even scratching the surface of all of the domains that need this deep nuance to maximize AI intelligence. There has never been a better time…. To be a consumer!!
Marc?
The 10x engineer line is a vanity metric until the job is named. An agent writing most of the code is a status, not a workflow. Write the one job the human still owns: who reviews the diff, what they can reject, and what stops a bad merge. If those three lines are blank, you did not get an architect. You got a faster typist with no kill switch.
Hola Marco, this needs to nipped in the bud, asap.
Insanity.
The strongest argument is the historical pattern every leverage increase expanded the profession. But agents write 90% of the code says nothing about who reviews, owns and is accountable for it.
This entire piece reads like a marketing brochure for Cognition. Is the company struggling?
Not once have you addressed the actual job loss happening. Or the fact that if AI actually begins coding autonomously, then how does that not end software engineering (at least as we know it) and above all, wasn't it you yourself who said that the main push for AI is to combat the declining birth rate and still have a reliable and steady workforce?
The great thing about life and world in general is we don't know what we don't know.
So the only thing to keep moving ahead is Just DO it with optimism