eigenspace

@hassankhan Intel principal engineer pulls maybe 400k and they wonder why they've fallen behind

TSMC’s total compensation for a veteran 10 year engineer is ~college grad compensation at FAANG

@cis_female container ships are measured in "gross registered tons" (sounds like a mass unit, is actually a volume unit) and "displacement" (sounds like a volume unit, is actually a mass unit)

Weird commodities units: Oil measured in barrels Methane measured in MMbtu tungsten in "Metric Ton Units" = 10kg of WO3 = 8kg of tungsten gold in troy ounces diamonds in carats wheat in bushels milk in CWT=100 lbs cotton in bales coffee in bags wood in cords

@DanielleFong the bulk of water usage comes from off-site power generation

people say that data centers use water. this is mostly incorrect. even wet cooling towers recycle most of their water, though like 10% a cycle evaporates to come down later as rain. it might even seed clouds to have humidity that high. however, a data center / desalination plant will generate net water. a natural gas plant that condenses exhaust will make net water as well. what are you going to complain about then?

boi about to go over the falls

@signulll the 20m packages are real tho 🤫

this is a high tier psyop from sam. he’s talking directly to the market. by floating a number like $100m: - he sets an anchor so absurd that anything meta offers now feels cheap. - he forces meta to either deny or match, both of which hurt them. - he gives his own employees a loyalty premium, they feel like they’re sitting on a $100m chip. - he blurs price discovery for everyone else. no recruiter knows the real going rate anymore. - & most importantly, he reframes the narrative from “openai is losing people” to “they’re so valuable it takes $100m to steal them” zero marginal cost, massive narrative leverage. & meta’s cto just confirmed it’s working by responding at all. textbook information asymmetry + narrative warfare.

>no vector DB >look inside >faiss

@LauraDeming gotta separate kinetics and thermodynamics. thermo/stat mech only looks at conditions of equilibrium - an "attractor trajectory" is a kinetic process of one state transitioning to another, while thermo only looks at the end equilibrium state

I don't...buy the ergodic hypothesis in stat mech. Any recs for how to get comfortable (even empirical?) Seems just obviously not something that follows from first principles - unless it's like every system is noisy and that's why? But aren't there attractor trajectories that would kinda violate it? Stat mech would be so much easier if I believed in statistics, lol

@typedfemale back in the good old days before they realized each well was pulling from the same source

@KiwiSoggy hit shawty with the skip connection

@kalomaze their recruiters are thirsty af rn

getting word that like ~80% of the llama4 team at Meta has resigned

@willmenaker it's pronoun shit. can't use "soldier" because only the army has soldiers - don't want to be misgendering our marines, airmen and seamen. "warfighter" is gender neutral

When did this shift to calling soldiers "warfighters" start? I remember during the bush/war on terror years there was brief flirtation with our "warriors" but it never stuck. This current iteration feels like a soy-right mix of HR culture and marketing hokum.

@dystopiabreaker alphafold can't tell you if the molecule is going to wreck your liver

on the other website everyone is dunking on this thinking that it means that scientists are going to be using a chat bot to do this, zero knowledge of alphafold or the modern paradigm of computational biology

@CL207 weakness comes from necking which is (functionally) irreversible

how many times can you bend metallic glasses frames before it becomes weak ? can you heat it up and reorganize the way these atoms are aligned or something and its strong again or what, someone from crystallography and metal mechanical properties send help

@astridwilde1 it's a confirmation bias machine. do with that what you will

sincere question, is DeepResearch actually useful yet? am i just doing it wrong? feels like throwing money into a pit and nothing useful or interesting ever comes back up out of the pit

@torchcompiled also real heads remember figure 8.1 from the Goodfellow book

@torchcompiled high dim critical points literally don't exist. "find a point where billions of parameters all have zero gradient" yeah right plus grad norm is proportional to param norm - param norm can increase grad norm independent of loss. this is one reason you decrease lr over time

why does grad norm continually grow with training? this is mad unintuitive. absolutely destroying my mental model of 2D convex bowl optimization.

@Jiankui_He w2c drug approval poster

Gene editing should be regulated by international laws.

@bennpeifert scenario 3 but we implement at scale anyway because it generates shareholder value

one scenario is we get AGI or whatever, good luck with that another scenario is we get substantial productivity benefits as AI takes over a lot of lower skill tasks in engineering a third scenario is AI keeps hallucinating and making overconfident mistakes and remains a toy

My guy ⁦@bigblackjacobin⁩ pushing back against AI hype. The idea of a "subprime AI crisis" driven by massive speculative overinvestment in data centers and training models that are unable to deliver on their wild promises is a risk scenario we should be considering

@ghosttyped they served their purpose. extracted like oil and distilled into VC carry. discard the husks and harvest the next crop of youth

At 28 almost all my formerly cracked 19-24 year old friends have burned out, even the highly successful ones

@kitten_beloved >find the source >case study of a single child >main result is measure of fungal metabolites in urine, "autism reversal" is simply asserted with no data lol. lmao.

If autism is caused by a mold named Aspergillus I'm gonna be so mad