if there was a perfect time to get rid of MBB and Big4, it was the last 2 yrs
but clients' inherent need for 'intelligence' is fundamentally different from what AI doomers are trying to tell us. so it didn't happen, they are still everywhere
very interesting about gene expression is how throwing money at it is not enough
for brain development, esp. 1-5 yo, playing in the mud or hiking is better than paying for private tutors or extra classes
convincing them to eat healthy food is the tiresome part not buying it
suffering from the chronic disease of refusing to long the main coin
i heard it enough, yes bro please trust me it is so simple, this is the new btc
but those who coordinate the narrative all have angel/vc allocations in the next rotation
zec -> zama
they always rotate
one day it is THE next btc, then 2 weeks later: "did you know the real story is actually the ethereum of the btc of the ___ theme"
google restricting outside custom mcp connections for their regular gemini chat interface is very surprising
their ecosystem is just a liability if it comes with these kinds of limitations imo
just erased the cultural artifacts of an entire civilization (deleted an md file)
also stopped a secret conspiracy against me (used git version control to notice if new files are created)
what is 'read -> write -> own' for bio/acc?
i) we are statistics in an n=300m pharma experiment
ii) n=1 medicine, but still trust/permission problems
iii) full control of our encrypted health data, permissionless protocols for research collaboration leveraging FHE and MPC
brought up FHE and MPC when a digital banking executive complained about data protection regulation and how they struggled with user data
their reaction and the general vibe of the conversation felt like explaining simple smart contracts 10 yrs ago
but a 100x larger use case
explorative thinking used to be rate limited by bureaucratic constraints (e.g. sales, publishing pressure)
now the only constraint is our ego:
- can we tolerate exploration that first feels like shameless slop
- can we accept not understanding all the experimental details
8 yrs ago deep learning curriculum was so small you could understand newest models after 2-3 semesters
now every single ML problem is its own profession and it became really time consuming to try to understand all competing models for one task
thousands of new professions
What if the real bug isnβt in your codeβ¦
but in your compiler?
I tried removing the compiler entirely:
AI β assembly β Lean proofs
It feels like cheating.
Is this insane or inevitable?
read more: https://t.co/OsHaBpNFoZ
Dwarkesh trying to strawman his guests
"okay the argument hinges on ... "
while his own reasoning hinges on engagement farming world war 3 psychosis with a constant stream of low iq analogies
can't rely on tests, AI will break it
have to write mathematically verified software
soon we need millions of software engineers with formal verification expertise
gradually becomes part of the profession
Rust is a relatively safe choice for vibecoding, but it still helps to do frequent cleanups
remove unwraps, redundant cloning, code modularity/maintainability, etc there is a textbook list for these
I can't imagine the shitshow of vibecoded JS or Python.