2,575 karma · joined March 30, 2013
https://x.com/chrisbarber
See also: https://x.com/wangzjeff/status/1983914310738047291
And also Nick Gray's 2 hour cocktail party book
My personal thoughts on events:
(These don't really apply to parties, but they do apply to non-party events)
1. Do intro circles: If it's a 5-25 person event with a handful of people that don't know each other, do an intro circle about 15-20 mins after the start time. Turns it from something where people show up and might meet 1-3 random people that they happen to walk up to, vs something where everyone gets 1 point of contact with anyone else. Works well up to about 25 people, haven't tested it beyond that. Go round say name, and then pick a few questions depending on the audience (eg could be something you'd like help with, something you're reading about, etc). For non-parties (eg meetups, work mixers, things that don't have alcohol or aren't late), the easiest way to improve any event is for the host to do a brief intro circle.
2. The best events to host are the ones you wish you could attend but that don't exist
3. Minimize uncertainty for attendees: Clear parking info/photos and a photo of the space is always helpful too.
4. Host more events: Very positive sum. Even can be simple discussion groups. Anything that you enjoy doing where it'd be more fun with a few other people. Playing video games together, reading papers together, discussing how you're using AI coding tools, whatever. Workshops, mixers, talks, parties, peer groups, etc. If you enjoy reading about it on HN or twitter, you'd probably also enjoy discussing it with people directly. The world is undersupplied for events.
Therefore: If you want lots of users, design for the median user; if you don't, this doesn't apply to you
TBD on if it ships on time, how good it is, etc, but fuck, this is pretty cool.
It seems like Microsoft stock is then the most straightforward way to invest in OpenAI pre-IPO.
This also confirms the $500 billion valuation making OpenAI the most valuable private startup in the world.
Now many of the main AI companies have decent ownership by public companies or are already public.
- OpenAI -> Microsoft (27%)
- Anthropic -> Amazon (15-19% est), Alphabet/Google (14%)
Then the chip layer is largely already public: Nvidia. Plus AMD and Broadcom.
Clouds too: Oracle, Alphabet/GCP, Microsoft/Azure, CoreWeave.
I wonder what criteria that panel will use to define/resolve this.
Is it about jobs/tasks, or cognitive capabilities? The majority of the AI-valley seems to focus on the former, TFA focuses on the latter.
Can it do tasks, or jobs? Jobs are bundles of tasks. AI might be able to do 90% of tasks for a given job, but not the whole job.
If tasks, what counts as a task: Is it only specific things with clear success criteria? That's easier.
Is scaffolding allowed: Does it need to be able to do the tasks/jobs without scaffolding and human-written few-shot prompts?
Today's tasks/jobs only, or does it include future ones too? As tasks and jobs get automated, jobs evolve and get re-defined. So, being able to do the future jobs too is much harder.
Remote only, or in-person too: In-person too is a much higher bar.
What threshold of tasks/jobs: "most" is apparently typically understood to mean 80-95% (Mira Ariel). Automating 80% of tasks is different to 90% and 95% and 99%. diminishing returns. And how are the tasks counted - by frequency, by dollar-weighted, by unique count of tasks?
Only economically valuable tasks/jobs, or does it include anything a human can do?
A high-order bit on many people's AGI timelines is which definition of AGI they're using, so clarifying the definition is nice.
“For us to truly understand what the policy solutions look like, we need to spend a bit less time talking about the specifics of the technology and trying to convince people of our particular views of how it might go wrong – self-improving AI, autonomous systems, cyberweapons, bioweapons, etc. – and more time listening to people and understanding their concerns about the technology. There must be more listening to labor groups, social groups, and religious leaders. The rest of the world which will surely want—and deserves—a vote over this.”
“Right now, I feel that our best shot at getting this right is to go and tell far more people beyond these venues what we’re worried about. And then ask them how they feel, listen, and compose some policy solution out of it.”
Most people don't care about open source for its own sake.
Find the things they already care about where open source gives them more of what they already want, with less effort.
Don't try and convince them to care about it, just find things they already care about where the open source version is genuinely better (from their perspective, not yours). Promote it on the merits they already care about (speed, cost, UX), not on the fact that it's open source (only people who already care about open source care about that).
This tweet recapping this paper https://x.com/lugaricano/status/1969159707693891972
This tweet with recaps of various papers presented at "The Economics of Transformative AI" by NBER in Palo Alto a few weeks ago https://x.com/lugaricano/status/1968704695381156142
What makes you not see that potential here? Seems like the obvious answer, at least if this is aimed at being a nice small profitable side project.
My recap of the PDF is something like:
1. There are good books about the near-term economics as AI.
2. There aren't many good books about "what if the AI researchers are right" (e.g. rapid scientific acceleration) and the economic and political impacts of those cases.
3. The Second Machine Age: Digital progress boosts the bounty and widens the spread, more relative inequality. Wrong on speed (e.g. self driving tech vs regulatory change).
4. Prediction Machines: AI = cheaper prediction. Which raises the value of human judgement, because that's a complement.
5. Power and Prediction: Value comes when the whole system is improved not just from smaller fixes. Electrification's benefits arrived when factories reorganized, not just when they added electricity to existing layouts. Diffusion is slow because things need to be rebuilt.
6. The Data Economy: Data is a nonrivalrous asset. As models get stronger and cheaper, unique private data grows in relative value.
7. The Skill Code: Apprenticeship pathways may disappear. E.g. survival robots prevent juniors getting practice reps.
8. Co-Intelligence: Diffusion is slowed by the jagged frontier (AI is spiky). Superhuman at one thing, subhuman at another.
9. Situational Awareness: By ~2027, $1T/yr AI capex spend, big power demand, and hundreds of millions of AI researchers getting a decade of algo progress in less than a year. (Author doesn't say he agrees, but says economists should analyze what happens if it does)
10. Questions: What if the AGI-pilled AI researchers are right, what will the economic and policy implications be?
1) reward/incentive/expected good feelings
2) effort/displeasure of doing the thing and the result
One way to increase #1 is to make it more socially involved. If you're working on a project solitarily, start going to events and talking about it with people, or write about it online. Humans are massively socially motivated.
For #2, one way to address this is with emotional processing. Often something is unpleasant because it reminds of something we didn't like from the past. So really digesting those emotions can allow the expected displeasure to fade because we kind of integrate it into our brains/bodies. But the key for this is that it has to be emotional processing, not intellectual processing.
Easier to start, harder to stand out. More competition, a more effective "sort" (a la patio11).
RLHF providers:
1. Surge. $1b+ revenue bootstrapped. DataAnnotation is the worker-side (you might've seen their ads), also TaskUp and Gethybrid.
2. Scale. The most well known. Remotasks and Outlier are the worker-side
3. Invisible. Started as a kind of managed VA service.
4. Mercor. Started mostly as a way to hire remote devs I think.
5. Handshake AI. Handshake is a college hiring network. This is a spinout
6. Pareto
7. Prolific
8. Toloka
9. Turing
10. Sepal AI. The team is ex-Turing
11. Datacurve. Coding data.
12. Snorkel. Started as a software platform for data labeling. Offers some data as a service now.
13. Micro1. Also started as a way to hire remote contractor devs
OpenAI projects 50% gross margins for 2025
The other companies don't include free users in their GM calculations which makes it hard to compare
2. It's probably not very well represented in the training data. Humans don't often ask other humans to count to a million. Asking why an AI can't count to a million seems similar to asking a human why they can't easily recite the alphabet backwards (unless they've tried it before, or they really think it through ie count forwards to get each prior letter, or they use a 'tool' e.g. writing it down).