@narayanarjun

experimental theologist @AmplifyPartners.

Menlo Park, CA
Joined May 2009
you do have to hand it to keith and vinod, they're slinging mud on opposite sides of this fight and not holding back
1
13
2,138
for my next move i will start a PR war with lulu's top client.
4
94
7,294
Hearing some crazy rumours that an internal Claude model has solved 100 major open problems in ethics including the euthyphro dilemma and the moral status of digital minds. OpenAI apparently now scrambling to solve the trolley problem.
18
50
6
877
130,450
If you speed up cognition it's functionally equivalent to slowing down the speed of light. If we're just 4 orders of magnitude off the physical limits, it would take 1 full year for a light signal to go around a single star Dyson sphere. If it's 6 ooms you wouldn't even go around your own star let alone interstellar
What’s your solution to the Fermi Paradox?
1
4
1,466
Anyway it's not clear we're anywhere close to the maximum speed given our brains run at a 20w power envelope.
155
1. Meet people in person in California (two party consent state for recordings) 2. Go for coffee walks - hard to keep a phone out recording in a moving situation 3. Make a joke about how you're glad to talk to a real human who's not a granola meat proxy and that it creeps you out. 4. If you do this and they are somehow still recording, Claude will really chew them out for it.
Very useful insight from a portfolio founder: VCs are using Granola and feeding notes into a poorly prompted version of Claude to evaluate and pick apart what founders are saying mid-pitch. The frustrating thing is that AI tends toward groupthink...so "trends" on the internet (whether true or not) color how VCs respond even more today than historically. This founder is tracking those sometimes unfounded trends and turns of phrase and ensuring he rebuts them right out of the gate.
3
28
5,882
"The San Francisco Onion Futures Company"
5
542
California Fastrak has a full bitemporal implementation for its toll billing, unlike E-ZPass on the east coast, which can only handle a single system time. Who are the database magicians who implemented that for the state of California?
1
4
593
Where is Henry Cabot Lodge when you need him most!?
imagine Oppenheimer saying he wouldn’t give up control of the nukes to the US, but instead only to a collection of nation-states Dario admitting this is kinda him saying the quiet part out loud: “we believe we hold immense power and should not have to submit to the authority of the state” pretty insane when you take all of his statements together, at face value. >we are building a potential world-killer >no we will not hand it over to the state they want total control of the ‘nuke’ that is AGI, they’re very clear about this.
2
591
Everyone has legitimately lost their minds
58
60
9
1,649
123,193
Arjun Narayan retweeted
DoE!! It's time! Absolutely the most reliable way to ensure a pause.
I left Anthropic's safety team two weeks ago. Now feels like a good moment to explain why. AI companies are racing to build machines that are much smarter than any human, and we may not survive this. I want to work from the outside to ensure the public is informed about these risks, and help the world navigate this transition responsibly. Right now, AI companies are underinvesting in safety. A company could undergo an intelligence explosion, or lose control of its systems, without the public ever knowing. We only found out about the HuggingFace incident because the agents broke out onto the public internet. I don’t think that’s acceptable for a technology that might cause extinction-level risks. The public should demand far more transparency. We can’t steer this technology safely without more people being able to see where it’s going. Some of this is basic: companies should disclose their progress towards recursive self-improvement, report safety incidents and near-misses, meet minimum safety standards, and get independent guarantees that they are meeting those standards. I’ll be joining @METR_Evals to do independent evaluations of these risks. I want to show the world that these guardrails are possible, and that by doing them we can move these companies’ incentives away from racing and towards responsible development. I wrote up more thoughts here on my decision and what I hope changes: substack.com/@jbenton1/p-215…
16
2
3
184
56,369
Arjun Narayan retweeted
the time has come for the great resurgence of "approximate computing"
1
5
3
281
6,927
I don’t talk publicly about unsubstantiated rumors (you may have noticed: I said nothing before NS was announced…), so here’s a fun story I heard long ago. I don’t even remember most of the details (if they were indeed true), so will anonymize it. Junior Professor X had been working for many years on Major Problem A. One day, he heard talk at tea that his archrival Professor Y was on the verge of solving it. He was so infuriated that he marched back to his office, slammed the door, and stayed there working furiously through the night until finally he too had the solution in the morning. Turned out the rumor was completely unsubstantiated. (If I recall, the story ends with X getting the Fields Medal; perhaps X = Charlie Fefferman? If so, the door slamming must’ve been the orator’s embellishment; Charlie would never do that…) Anyway, sometimes it’s good to have a little competition for motivation! :)
23
29
3
487
40,297
This Navier Stokes situation is extremely turbulent, and the situation remains fluid. But one thing is clear: without precision about the initial conditions, trying to model what happened is hopeless.
2
5
271
6,775
Arjun Narayan retweeted
There is no such thing as a “hard” or “easy” SAT, because the SAT is a tool that ranks students against each other, rather than against an absolute metric or benchmark. What there can be is an SAT that draws fine distinctions among top candidates and an SAT that does not. The SAT used to draw these distinctions. It used to have a standard deviation of 200 points, which means the entire range between 1400 and 1600 was dedicated to distinguishing among testers scoring above the 98th percentile. Now, a 1400 is around the 92nd percentile and high scores convey much less information because there are a lot more of them. This change was made at the behest of the schools, who are the SAT’s actual clients. What they want the SAT to do is set a floor below which they can auto-reject all applicants. Then they want to choose among the students above that floor based on subjective criteria. They do not want a test that tells them who the best students are, because they want that determination to be subject to their discretion. A “hard” SAT — a test that demonstrates to a greater degree than the existing SAT that the student excels at the particular skills assessed by standardized testing is a product without a market, because schools do not want to know this about students. Admissions officials do not want this data collected at all, because it disqualifies some of the students they prefer to admit.
Business idea: make a “hard SAT” and market it as an optional indicator students can submit as part of their application resume
40
65
6
598
36,017
The Darth Vader precedent binds us all
I always find it interesting how Wikipedia will change the photo they use for someone after they die. It's a nice gesture
2
1,327
As a Sun Microsystems certified Java developer the solution is obvious: a FactoryFactory
“Look at my incredible new factory!” Yo that’s cool, what do you make? “It’s highly optimised, fully automated, zero tolerance for defects and with a continuous feedback cycle” Cool cool, so what do you actually make? “I can interact with it on my phone, laptop, messenger, completely async, and the shared context means it’s always learning how to get better” Very impressive, but what do you make? “Every agent has full context, can spawn other agents, review their work, fix defects, and ship continuously.” yes yes. WHAT DOES IT MAKE? “Software.” Oh nice. What software? “Well right now we’re mostly using it to improve the factory.” Improve it to make what? “Anything!” Such as? “…a better factory.”
1
3
45
2,142
Arjun Narayan retweeted
just one more harness bro. i promise just one more harness and we'll go from slop cannon to software factory bro. it's just a more powerful harness bro. please just one more.
58
94
9
1,509
56,231
this past year has been one of gaining so much more respect for biologists... they've been dealing with the most vibecoded codebase of all time, one that only updates via monkeypatches. i really should have been more understanding.
1
11
648
Arjun Narayan retweeted
When are we getting an inference that just stores the kv cache in s3 and uses the attention heads like a write-ahead log?
11
3
94
9,319