Mads

Product | AI | Applications | Builder. Curious & eclectic. MIT Sloan MBA.

[Mads @madssrinivasan]

When you ask your model to do soul searching and do a course correction, it does do some soul searching!

[Mads @madssrinivasan]

@hnshah David Graeber should have been to write a follow to his book - Bullshit Jobs!

[Mads @madssrinivasan]

"Too dangerous to release"

[Mads @madssrinivasan]

I like the honesty :) . During a session when Opus was asked if i should switch to Sonnet, this is the answer i got:

And a caveat about my own advice: I have an obvious bias in answering "should you keep paying for me." The honest version is that most of what I did tonight — running builds, tallying ledgers, writing docs — a smaller model does fine. The part that earned the difference was noticing that a merged commit did contain the test I'd claimed was missing, and that noticing only happened because an external reviewer pushed on an arithmetic detail. Neither model tier substitutes for that.

[Mads @madssrinivasan]

@hnshah Can we call it the 'AI chasm' / 'AI generated chasm' / 'AI induced chasm'

[Mads @madssrinivasan]

How philosophical (from a CLI instruction):

Stop waiting — the sweep finished and your watcher is watching itself.

[Mads @madssrinivasan]

...that feeling when you are not satisfied with a product/design working 'fairly well' , and decide to rebuild it!

[Mads @madssrinivasan]

Easy way to irritate me. Next time when you meet me just say - "Let me check rather than assert".

[Mads @madssrinivasan]

Harness engineering lesson of the day: test your own safety rails.

A check that blocks rm -rf / sounds solid. I tried 27 ways to sneak one past it - 19 worked.

The banned-command list was fine. It only looked at the first word of the line, and there are a lot of ways to not be first.

[Mads @madssrinivasan]

"The hottest new programming language is English." Karpathy is right, and the line has been carrying more weight than it was built for. Programming languages are compiled or interpreted, and the slogan does not say which. https://t.co/gAr8MbkX38

[Mads @madssrinivasan]

@ClaudeDevs Here is the reason why i prefer new session over compaction https://t.co/yAh93hCflf

[Mads @madssrinivasan]

@ClaudeDevs This is great. i used to have a handoff routine to do this.

[Mads @madssrinivasan]

Of late i am seeing more of these kind of statements in my Claude sessions:

"𝘼 𝙙𝙚𝙛𝙚𝙘𝙩 𝙬𝙝𝙤𝙨𝙚 𝙨𝙮𝙢𝙥𝙩𝙤𝙢 𝙞𝙨 𝙖𝙗𝙨𝙚𝙣𝙩 𝙤𝙣 𝙩𝙝𝙚 𝙗𝙤𝙭 𝙮𝙤𝙪'𝙧𝙚 𝙝𝙤𝙡𝙙𝙞𝙣𝙜 𝙞𝙨 𝙨𝙩𝙞𝙡𝙡 𝙩𝙝𝙚 𝙙𝙚𝙛𝙚𝙘𝙩. 𝙐𝙥𝙜𝙧𝙖𝙙𝙞𝙣𝙜 𝙢𝙖𝙣𝙪𝙛𝙖𝙘𝙩𝙪𝙧𝙚𝙨 𝙩𝙝𝙖𝙩 𝙖𝙗𝙨𝙚𝙣𝙘𝙚 𝙤𝙣 𝙥𝙪𝙧𝙥𝙤𝙨𝙚".

"𝘼 𝙘𝙝𝙚𝙘𝙠 𝙬𝙝𝙤𝙨𝙚 𝙨𝙪𝙗𝙟𝙚𝙘𝙩 𝙣𝙚𝙫𝙚𝙧 𝙖𝙥𝙥𝙚𝙖𝙧𝙨 𝙞𝙨 𝙞𝙣𝙙𝙞𝙨𝙩𝙞𝙣𝙜𝙪𝙞𝙨𝙝𝙖𝙗𝙡𝙚 𝙛𝙧𝙤𝙢 𝙤𝙣𝙚 𝙩𝙝𝙖𝙩 𝙥𝙖𝙨𝙨𝙚𝙨"

"𝙏𝙤𝙪𝙘𝙝𝙞𝙣𝙜 𝙖 𝙡𝙞𝙣𝙚 𝙢𝙖𝙠𝙚𝙨 𝙮𝙤𝙪 𝙧𝙚𝙨𝙥𝙤𝙣𝙨𝙞𝙗𝙡𝙚 𝙛𝙤𝙧 𝙚𝙫𝙚𝙧𝙮𝙩𝙝𝙞𝙣𝙜 𝙤𝙣 𝙞𝙩 — 𝙨𝙤 𝙩𝙝𝙚 𝙘𝙝𝙚𝙖𝙥𝙚𝙨𝙩 𝙘𝙤𝙧𝙧𝙚𝙘𝙩 𝙢𝙤𝙫𝙚 𝙞𝙨 𝙩𝙤 𝙩𝙤𝙪𝙘𝙝 𝙛𝙚𝙬𝙚𝙧 𝙡𝙞𝙣𝙚𝙨."

Not sure if Anthropic is training their models on a huge philosophy corpus, employs a full-time philosopher to write these lines, or the the LLM makes them up as it goes!

[Mads @madssrinivasan]

Ok Claude, Github when both of you guys are up you guys figure out merge, deployment, and testing and let me know when everything is ok.

[Mads @madssrinivasan]

Dead code costs nothing the day it is written and charges interest every day after, which is why it surfaces as migration scope and security triage two years later rather than as a line on this quarter's dashboard. https://t.co/EJLyRBhxXd

[Mads @madssrinivasan]

https://t.co/79YGvPCoAd

[Mads @madssrinivasan]

When a builder writes the code themselves, the cost signal is built in. They feel the third afternoon on retry logic and something objects. That objection is not discipline; it is fatigue doing useful work. https://t.co/F9y85drYbU

[Mads @madssrinivasan]

That confidence when finally Claude comes back after an outage and says - "...things are running fine on my end...I didn't hit an API error. If you saw one on your side...." !!!

[Mads @madssrinivasan]

Prompt sent. He leans back,
lord of a thousand green diffs —
the agent decides.

#vibeku

[Mads @madssrinivasan]

A harness that knows when not to run is a better harness.