ChatGPT’s intelligent UI has really grown on me

at first it felt like widgets for the sake of widgets but for dense stuff like cox regression it’s markedly easier to follow than a wall of text

can't get over how supposedly ‘1 developer’ spent ~2.5 years building trust as a maintainer of xz-utils (a compression lib thats on basically every linux machine, yet was shockingly maintained by just 1 person), used sock puppet accounts to pressure that burnt-out maintainer into handing over control, and the only reason the backdoor got caught is that a Microsoft engineer noticed SSH logins were taking ~500ms longer than usual and decided to investigate

thankfully it was caught a few weeks before it would have shipped in stable releases of Fedora and Ubuntu, so not many machines ever ran it, but still, that’s a scary close call

had no idea how deep sandboxing Codex and Claude Code in Nono would go

ended up learning landlock, lsms, no_new_privs, kernel ABI versions and how restrictions inherit through fork and exec

I suspect decision models (Jev + the wave of open clones e.g. Amazon's Strands Decider 2B) are going to make applying taste nearly free, since you can encode a rubric once and run it on everything for fractions of a cent

it will push more value to how one defines the rubric though

to put it into perspective, Google’s 2016 monorepo paper reported ~40,000 commits per workday across ~25,000 engineers, and more than half of those were automated

Anthropic is already in that range with a fraction of the engineers

Gemini 4 Argon now generally available on AI Studio

benchmarks

the magic you are looking for is in the work that you are avoiding

MCP & CLI economy

reading code I wrote a year ago and immediately seeing what I'd change (esp. architecturally) is a good indicator that I've actually improved

OTOH, not finding anything to improve would be a tad more worrying

the number of tabs I have open is a pretty reliable indicator of how scattered my thinking is that day

someone needs to create a 'boring' phone with an eInk display

things in AI that didn't make it (RIP):
- the Humane AI Pin ($699)
- Rabbit R1
- AutoGPT-style "fully autonomous" agents from 2023
- prompt engineering as a discipline
- vector DBs as the default retrieval layer for coding agents. turns out grep was plenty

strange how Sonnet 5.5 is essentially on par with Opus 5.5 but there seems to be no official explanation

my guess is that Sonnet 5.5 was trained on distillations of the model after Opus 5.5

I mean, how else do you explain how impressive and cheap it is

if you greatly compress the amount of time that you allocate yourself for a given task, you'll be surprised at what you can pull off

doing a hobby project that you think will take 2 weeks? force yourself to get it done in 2 days

you will naturally get much better at prioritisation

its wild how efficient China is in scaling things in the physical world, and how consistent they have been in doing this across essentially every sector

I am becoming increasingly de-sensitized to model progression, that if a lab were to release AGI, as long as the models preceding it were not noticeably worse, I'd probably write off as a "decent improvement"

if we tried to fit the stable core of SWE it'd probably be something like:
- DSA + big O
- computer architecture
- operating systems / concurrency
- networking
- databases
- distributed systems
- security
- API design

the difference is mechanical engineering's frameworks are 300 year old physics, and ours get replaced every 3-5 years

every model release makes another batch of hand-crafted prompts, skills and scaffolding redundant, if not outright detrimental, as they just eat tokens and context

funnily enough tho, Sutton said exactly 1 year back on Dwarkesh that LLMs are a dead end

was looking into the craziest software failures and found this

in 2012 Knight Capital, one of the biggest market makers in the US, lost $460M+ (~$670M today) in 45 minutes

pretty wild what happened, it was essentially 5 separate failures:
- an old feature called "Power Peg" was n't deleted from prod
- in 2005, part of its code was moved, breaking it, and it was never retested
- in 2012, the flag that used to switch Power Peg on was reused for new code
- a technician didn't copy the new code to 1 of the 8 servers, so on that server the flag switched Power Peg back on
- 97 automated emails flagged the error before market open, but they weren't set up as alerts, so nobody acted on them

turns out that their order router was trying to fill 212 customer orders, but ended up sending more than 4 million orders into the market