Programmer "Productivity": CHART OF THE DAY
Smarter machines—but do they make us dumber drivers and worse coders?...
A possible paradox: “AI” really does seem to be helping programmers spit out code by the truckload. But we are not (yet?) seeing (many?) signs of that helping businesses ship better software or make their users happier. From my over‑eager VW microbus to Anthropic’s IPO roadshow, perhaps we are mistaking frenetic machine activity for genuine progress…
The brand-new VW ElectricMicroBus that replaces the 22-year-old Subaru has:
automatic following distance;
automatic lane-keeping;
speed-control settings that respond to speed-limit traffic signs (alarmingly, not just slowing the car if it is going faster than the limit);
anticipatory pre-curve and pre-intersection braking (if its route-finding is engaged).
But turning on these systems makes driving an immensely more difficult cognitive process. I have to:
do all the normal mental processing to figure out how I would drive;
interrupt the largely non-conscious loop by which my brain decisions then move my feet and hands without much frontal-lobe attention;
assess the differences between what I would be doing and what the car is doing;
decide whether I would be right or whether the machine is right;
take over and react if I decide that the machine is wrong.
as opposed to the normal, relatively boring effort of just driving along. I assure you, when I engage these systems, I AM NOT BORED. I AM THE FURTHEST THING FROM BORED.
Right now it feels to me like our bureaucracies with respect to their adoptions of MAMLM technology in its form of natural language interfaces and very big data, very high dimension, very flexible function classification analyses are in the same position relative to the world’s LLMs and such as I am with respect to the electric brain of the VW.
Now comes Anthropic with the hope of getting a higher price from the rather gullible, most optimistic buyers in the public market for the tranche of its equity it proposes to sell in its rapidly approaching IPO. not being happy with the mere claim that every instantiation of Claude may well be in some sense “conscious”, they are doubling down:
Anthropic <https://twitter.com/AnthropicAI/status/2062568862479208923>: ‘Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor. It’s happening faster than we thought, and the implications deserve greater attention…
Well, yes: Anthropic says (and I see no reason to disbelieve them) that its programmers are writing a lot more code.
But there is an old, possibly (probably?) apocryphal story about the IBM-Microsoft “relationship” back in the day. IBM argued that it was contributing much more to their joint projects. It measured contribution on the number of lines of programming its teams had written. Microsoft pointed out that it was much prouder of how much code they had refactored and simplified, or simply removed as useless. Programs with greater length but no greater functionality were more fragile and error-ridden. Microsoft seems to me to have had the better of this argument.
Do not get me wrong: Claude Code and Claude Cowork are economic engineering and intellectual marvels. As was GPT 3.0. And as was ChatGPT. But no version of the SINGULARITY is nigh.
Probably relevant:
It is from Demirer & al. <https://www.nber.org/papers/w35275>. And it provoked comments:
John Burn-Murdoch: How Much Value Is AI Really Creating? <https://www.ft.com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1>: ‘Coders created or edited almost 300 per cent more files… but… 150 per cent… [in] work submitted for review, and… 30 per cent…in the number of full software releases… [with no] increase in downloads…. Uber CEO Dara Khosrowshahi revealed the company had blown through its entire AI budget for 2026 in one quarter…
And:
Noah Smith: Tokenmaxxxing <https://www.noahpinion.blog/p/roundup-83-i-told-you-so: ‘The things we really want a lot more of may not actually be the things that generative AI is yet equipped to provide…
It is very funny to me that the company with the hubris to call its system "Claude [Shannon]" measures a program by the number of squiggles it would occupy on a screen rather than its Kolmogorov complexity.
Ah, yes! The old "lines of Code" metric ... heh