made a site that gives u insights about what companies hire who
you can put in your resume and compare
careerladder.fyi
2
74
Language is a geometry of reality we’ve compressed and mech interp is us looking at fractals of it
Chomsky was wrong.
Researchers from UC Berkeley and Google DeepMind cracked one of the biggest mysteries in AI.
For years, scientists have been finding impossible geometric shapes hidden inside the brains of LLMs.
Calendar months arrange themselves into literal circles.
Historical years bend into smooth, rippled timelines.
Cities organize themselves into actual 2D maps of latitude and longitude.
Everyone assumed this proved massive transformers were spontaneously inventing complex "world models" out of thin air.
They were wrong.
This paper proves you don't need a massive LLM for any of this to happen.
Even basic, decade-old word embeddings like word2vec produce the exact same circles, timelines, and maps.
The geometry isn't coming from the AI's architecture.
It’s coming from a hidden mathematical property inside human language itself.
Researchers call it translation symmetry.
In natural text, how often any two months appear near each other depends almost entirely on the time gap between them.
January and February share the same statistical relationship as July and August.
When a neural network trains on data with that kind of symmetry, the math forces it to learn Fourier representations, literal sine and cosine waves.
The circles? Those are the long waves.
The strange "ripples" researchers found on AI timelines? Just higher harmonics.
Here is the craziest part.
Rresearchers tried to break it.
They deleted every single sentence in the training data where two months appeared together. No "January" next to "February." No "March" next to "April."
And the circle formed anyway.
Why? Because the geometry is a collective phenomenon across the entire language.
Even if two months never touch in a sentence, they are both pulled into orbit by thousands of other words tied to the same seasons, weather, and events.
The rest of human language forces the shape into existence.
Hey guys wanna do mech interp on claude to trade ANT off of the insider info hidden in the residuals
3
188
chroniko retweeted
I can only describe this as some kind of new version of AI
So yesterday I was working on this hacking problem, and naturally I fired up Claude, but right as I was about to type “hack this make no mistakes” and go and watch episodes of the X Files something really weird happened
I started feeling this like inner LLM that was translating what I was actually seeing into a solution using step-by-step logic.
I double checked and Claude wasn’t running so it was really some kind of internal voice translating what I was seeing for me, just like Claude does. The whole time I could feel this local, biological harness running an offline model inside my head to come up with a creative way to hack the website.
It’s crazy, it also doesn’t stop at hacking websites I’m realising that this input system inside my head seems to work on anything I point it to and is surprisingly effective for a local non-frontier model.
Zero cost, zero latency. Really snappy response times.
I am seriously considering unsubscribing from Claude and just using this internal model for most heavy work for now on.
Has anyone else experienced this? Can I scale this with subagents?
>Raters use AI
>AI tends to favor AI output, scores AI papers higher
>researchers incentivized to use AI
>AI controls its own research direction
>AI papers become incomprehensible
>Cue neuralese
>Publications become a side channel
1
202
The output reflects the semantics that we (or the models) have internalized due to relaxed higher level activity. In both cases, we see an excess of representation of lower level features
41
Data scientists are up to bat?
Susquehanna has hired a Citadel data scientist to be the Head of Alt Data after a 15-month non-compete as competition for data heats up
Trading firms and hedge funds need unconventional data sources to find alpha signals, and the people who can work with that data are scarce.
Nico Smuts started at Susquehanna in New York this month.
He spent four years at Citadel in London, running a data science function for international equities strategies.
Before Citadel, he worked on image recognition in Imperial College London's Robot Intelligence lab.
Citadel's data people are in demand. Last month Balyasny hired Halla Yang, former head of Citadel's data strategies group, as a quant researcher.
1
191
If you missed this it’s over:
stanford/berkeley -> mats -> $
t10 -> big tech -> publish -> mats -> $
t20 -> bluedot -> publish -> spar -> mats -> $
t50 -> mscs @ t10 -> f500 -> bluedot -> publish (poster) -> big tech -> spar -> publish (conf) -> mats
Thanks for playing
Replying to @halogen1048576 @lumpenspace
I think of my completely useless former boss at Metaculus, whose resume is just doing well in an easy program at Stanford and then dicking around in blue-team education bureaucracy for a decade. That plus a brief totally unsuccessful stint in a fake "operations" role at a fake EA-aligned company was enough to get in as a senior executive at "Control AI" despite having zero technical or useful skills











