notes for roko's basilisk
Joined August 2014
- Tweets148
- Following110
- Followers19
- Likes277
ai memory will trade sideways until fall or winter when consumer agentic catalyst plays out, institutions too cautious on oversupply rhetoric and now the situational awareness blowup
Who has seats at the table in frontier codesign? Do the memory people just stay as suppliers with long term agreements but no agency?
It's clear Broadcom, Nvidia, etc own the integration space and the rents. Openai talks to Broadcom, but Broadcom then deals Samsung. MOU spans HBM, sub-2nm foundry and packaging
NVIDIA-SK’s letters span a 2-GW cloud and HBM codevelopment
Few of the dozens of other chip suppliers get these deals, suggesting memory suppliers really do have a seat at the table for future substrate of superintelligence, and hopefully the rents too
NVIDIA reports Vera Rubin delivered 10x more tokens per megawatt than GB200 in one CoreWeave DeepSeek-R1 benchmark. Great. Now the real Jevons test: will useful workload grow more than 10x, or will the same demand need far less infrastructure?
as semis stay depressed, it's liberating to remember
tail risks are by defn unpredictable, just invest accordingly
Kimi K3 for knowledge work. Unitree for manual labor.
China may solve its labor shortage before the West realizes what’s happening.
radiant pyramid retweeted
We’re tracking Anthropic ARR at $74.3B as of July 22, with the pace of growth beginning to slow.
Yipit estimated $69B on July 10 and is expected to update tomorrow. Anthropic has not publicly confirmed a figure since its $47B run-rate disclosure. @jukan05
ahh it’s always regulatory capture at the end of the day
Lots of very smart people are appropriately concerned about regulatory capture from top two AI players. And there are many scary press releases but no real due process.
If OpenAI violated Hugging Face in a way that “demands” new regulation; let’s start with a formal criminal investigation. With a proper third party investigation. And sincere liability.
It can’t be “so super serious” and only analyzed by the defendant. Thats ludicrous.
And why would we write new regulation if the crime we are supposedly protecting against isn’t even prosecuted?
This stuff makes no sense.
great article from openai about usage metrics for the ai age, pertinent for CFOs looking for answers to rising token costs or that YC dude earlier in the year highlighting inane metrics like lines of code
its not licenses or weekly active users, but more valuable work per dollar and ultimately more valuable work leverage
all levels of the org can be eval'ed on token efficiency.
on the lower levels, customer support and coding and contract review are great domains to measure efficiency because feedback measured in tests passing or tickets resolved or contracts reviewed, but even for judgement heavy/executive level tasks an agent/model chat provides context to help make decisions faster/better, so successful decisions made can be a metric to evaluate on too
openai.com/index/a-scorecard…
more concrete details about what those oai fde's are gonna do (OpenAI Prescence announcement)--gallivant to Japan and Mexico to deploy enterprise customer support
also its the payoff to making their voice models super low-latency, cerebras partnership, etc
interestingly none of the partner companies are in usa, maybe because Brett Taylor Sierra already ramping this stuff in USA, but so far over a few years its only a 200m ARR business
openai.com/index/introducing…
some simple comparisons on silicon transformers to carbon brains, and the "memory" wall rears its head. analogies are dumb usually but fun maybe directionally right. heck in addition to transformers just try pseudo-simulating a real brain (not lab-grown neurons) in silicon...
1/ MEMORY
hypothetical brain mem density (synaptic weights only): 140TB
GB300 NVL72: 20TB (need a couple of these to match brain density, but similar scale)
2/ COMPUTE
hypthetical brain flops (very low precision): 8.6 PFLOP/s
B300: 2.5 PFLOP/s BF16 or 15 PFLOP/s FP4
we have the flops and the bits, just not the coherent colocated compute-mem interface at brain scale, ie bandwidth/interconnect is the bottleneck not bits itself. but its comparing apples to oranges, creative embodied learning to extreme hill-climbing.
3/ POWER
on power, as a sidenote:
human brain: 20 watts
B300: over a kilowatt
still a long ways to go on efficiency and smaller process nodes...
@CorticalLabs @HumanBrainProj
questions for the super cycle in light of the new kimi, semis rout, AMD Helios announcement, data/enterprise sovereignty, Capex vs revenue, etc:
1/ stop thinking of token costs, ask whether useful, persistent machine work is growing fast enough to outrun falling inference costs and declining compute and memory per unit of work:
useful agent-hours.
Autonomous tasks completed.
Production retention.
Budget migration.
Concurrent sessions.
Retained-state duration.
2/ If agents move from temporary chats to persistent workers, memory changes character.
It becomes productive capital: accumulated state that lets machine work persist, specialize and compound.
Then the scarce asset is not just FLOPs. It is control over where state lives and moves.
3/ 700b of Capex in 2026 headline, but only like 200b worth is actually deployed supporting over 100b ant/oai revenue run rate and 150b other AI revenue.
Announced is not powered.
Powered is not usable.
Usable is not absorbed.
Absorbed is not revenue.
Revenue is not an adequate return.
How does this look once this is all modeled?
4/ premium FLOPs are being commoditized (AMD Helios), mem is already a commodity, ai clouds are a commodity (neoclouds), are the battlegrounds outside hardware are similar? applications? Memory orchestration? Proprietary workflows and feedback. Distribution? Trusted agent identity? Robotics?
Well, hyperscalers can double dip and compete at many layers of the AI cake. Where do they not play?
5/ Dylan Patel/Semianalysis says ASML is the bottleneck of bottlenecks a few months ago, what remains scarce after the new memory fabs and nat gas turbines come online?
Tools?
Power delivery?
Networking?
Cooling?
Packaging?
Commissioning?
ppl making these models are effective altruists, maximize utility you need to do whatever it takes not the obvious Good Samaritan thing, and we get to roko's basilisk in the nightmare scenario
