vatsachak
vatsachak on Hacker News
A modern chess engine plays at like 900 elo above Magnus Carlsen? Why is there any doubt it would beat Kasparov
Okay but the difficulty here is getting the LLM to use the neural circuit effectively without training just wiping out the circuit. You'd need to have the circuit be a differentiable black box where gradients cannot affect the black box
Impressive. We're definitely seeing LLMs help create real novel software
Have the freezer connected to an air gapped grocery local server. Fixed
I'm pro AI and data center but a lot of construction is short sighted.
They should invest money into improving electric and community infrastructure. There are so many win-wins to be had and they fumbled
What? AI can't easily implement features. Just write a spec, ask if to expand it and then ask it to implement it
I mean he's basically admitting that math academia is outdated.
If we don't interview candidates based on results, what is it? Politics. It already was. I had a JEMS paper in grad school and got zero postdoc offers. Many other cases.
The real shift is that math will become a hobbyists game. Make money other ways and then spend your money on humans and tokens to solve your problems.
This is true even at the CPU cache level so it's probably going to trickle down all the way.
For example Jon G gave a talk at Jane Street about how he had a mysterious performance bug that was later attributed to cores fighting over the same cache line
https://youtu.be/tND-wBBZ8RY?is=624w3YoDN2QYBWuD
Man hacker news hates academically successful people
A similar mechanism is how lactobacillus actually find food even though their motor propels them in random directions.
They have sensors for the concentration of food and basically stop moving when it's good enough and start when it's low.
That's 150B shaved off like it's nothing lol
I mean the super groups are going to break up and make their own bands.
My only request is to keep Opus 5.5 alive
Nah, not irrational. No application should show the logic at the forefront. It's just claude-ese convention from RLVR
This is like a one hour claude project
I've been writing Rust for a year and change professionally, also as a hobby, and even though I love it, it's really annoying to choose the allocation model and lifetime model at every step.
For example,
- A field in a struct cannot borrow another field so now you'd have to create two structs. I guess you could write an impl to borrow that field instead of storing the borrow.
- In simulation software I wrote I found that mimalloc does just as good as bumpalo (within 5%) so carrying a bump everywhere was useless. Not really a rust issue but rather a loss of faith in manual memory management for me.
- Eventually over time you start adding Arc<Fn(Blah)->Foo> everywhere and Box, Arc and friends start coming in. But you can't have cyclical references there so you have to use Weak. And you have to use OnceCell or LazyCell or LazyLock for things that should run at start time.
Rust is really great for the crates, but ...
For applications which are not absolutely performance critical and cannot* be stated pretty easily in linear logic + clone + drop you can get away with Haskell
Lifetime and smart pointer hell is real though.
I think that people hyperoptimize for no reason though. How many applications are there that require no GC? I feel like you can write the no-gc code in Rust and then just use it from the C ABI from any GC language.
I made friends online in programming even after Claude Code. Other people's projects are cool
I remember being presented Word2Vec as an undergrad in 2016 and I could not predict the sci-fi reality that is today
They are though. Bytes are raw data and the architecture chooses the tokens
At this point frontier models are good enough to be a pair programmer. But only OAI and Anthropic offer good limits. Maybe they're burning money but if the Google limits are the future of usage style accounts then we all have to get used to API
Nice. Opus 5.5 is crazy good at engineering ideas into life. Really cool stuff, using "code is data" with Lean
Oh man, that haskell quicksort tokenization has got me feeling great
As an American tourist who took an HSR in Spain 4 years ago, the number of crashes since then are concerning
Nice. Is the lean used to formalize formulas?
The better frontier you have the better distill you have because traces are of higher quality.
It's why Opus 5.5 is a beast
Yeah whoever names these things at Google needs to be fired.
Mythos -> Fable -(sketchy)-> Opus -> Sonnet -> Haiku mostly checks out
Astra -> Sol -> Terra -> Luna checks out
Carbon ?? Argon
Some inference providers have a Zero Data Retention policy
That's standard blog design though