tmsh

Cool thanks for the reply. And congrats on the launch here.

but can't a decision model or some other model deduce "P" or "B" based on the context (i.e., previous words said)?

reminds me of tesla fsd v. waymo. may be a market for both? i.e., camera-only or camera + specific hardware.

why doesn't the author have any internet footprint (or at any rate any one that links to this paper)? it's curious.

I agree https://github.com/dmmulroy/anti-slop is very cool (heard about it via https://www.youtube.com/watch?v=ix1qQK1IvmA).

But what does this add beyond that?

All of life proceeds effectively with a right brain and a left brain. A fast loop and a slow loop. A general and a scout. A melody and a base. The evolutionary moments involve moving the melody into the base and making way for a new melody. But there is always an overseeing element and a work element. A manager or coach and an employee or athlete. It's just an effective pattern of growth. You see it everywhere in nature. In generations of animals and plant life (the more experienced parent, the growing child; the central 'brain' and the rest of the body; the queen ant and the ant colony; the trunk and the leaves).

The friction in math right now is you have the fast moving melodic bits racing faster than the base can understand always or keep up with. So you start to need AI for both the left brain part that is executing and the right brain part that is synthesizing. That's just the friction right now. You reduce the friction by using AI to help with understanding AI and getting back to a more normal rhythm of scouting out (with AI) and (what is still developing more and more) synthesizing with AI.

Also: https://cims.nyu.edu/~tristanb/statement.pdf

I had planned to say on announcing our work that the results are not the important thing. Rather the important thing is instead the significance that a mathematician and an LLM model can now do all this work in a month

They needed the general direction and idea it might work from a few folks tinkering for a year with the same models that we all have access to. The majority of credit goes to them. The idea of where to look with AI is key. Models didn't do that by themselves.

Trust: https://x.com/__alpoge__

"since on my side things were mostly me and claude having a good time yoloing random stuff in the corner rather than anything institutional"

Not: https://x.com/SebastienBubeck

"If you don't want me to be nice, then I don't have to be nice." [when clearly there's not enough deference paid to where this line of research originated]

Even through the lack of ethics here and there and mistakes - I think the major headline is that talented people are collaborating with AI to achieve impressive results. It’s shrouded in competitive mistakes of judgment. But it’s an existence proof for collaborating with AI and achieving incredible things.

Maybe how medieval monks felt about the printing press. Even the guy bagging groceries can create books now!

See better.

Is post-training magic just overfitting to benchmarks?

I think it matters a lot to understand the kernel of AI in the past 10 years. It's the only way you can truly appreciate the exponential. It looks flat unless you see the dramatic leap in transformer architecture and scaling laws and reinforcement learning. I know it's been hyped to death but you have to see how those things are truly exponential at their core to appreciate how it's not just walking up one step or two steps faster and how it's walking up buildings etc.

Raises an increasingly important question:

How do you solve this asymmetry (frontier labs or advanced model owners v. regular companies and groups of people)? Symmetry? But along which axes?

I think there will come a time when we will ask these questions and use AI to try to answer them (soft landing, slow rollout, regulation, hybrid X, Y, Z). But then AI will be biased towards more AI if it has distilled anything of what it means to be life.

Power corrupts. Power is the problem. An imbalance of power. Maybe there will be some sort of consensus protocol between powerful models in the future. Similar to blockchain I hate to say it. Maybe you have 100 very strong models in the next 10-20 years. And basically a lot of it is powered by tokens. So you agree not to attack others or else you'll get attacked. So there's sort of this natural deterrent to not attack others and they develop independently and there's generally a power balance among things. Maybe at some point that spreads to a billion individual models operated by and/or analogous to individual humans. All sort of holding each other in check.

Disallowing the use of AI is basically the modern equivalent of testing based on memorization. It is a laziness to find something more useful to evaluate on or an inability to distinguish real understanding in a field from first level understanding.

Professors need to step up and teach value beyond what LLMs know (very possible with or without LLMS). Or get out of the way from those building on the field with LLMs.

If you’re teaching students something that LLMs can score 100 on you are not adequately teaching them something useful for them in the future.

let people use as much ai as possible. encourage it. and as an educator, you have to learn to leverage it oneself or not (depending on the subject). and be better at using it than students if it does make one 'better'.

if ai doesn't help, then it won't help. if it does help, then you should use it. the metric is your output of whatever is being tested. writing an essay well and clearly understanding the material. solving a pset. whatever.

if you give access all the time for that, and then you test on a hard problem that could be done with or without ai, then it's fair. e.g. "clearly explain these four sentences of Y." obviously ai researching loosely and blathering isn't useful. won't be high signal / dense and correct and worthy of an 'a'. but someone who can harness ai and someone who knows the material well in the end will be rewarded the same by society. what you are testing is correctness and information density in a response. so you have to start now in accepting the reality that those who use ai to get there should be rewarded just the same as those who don't.

the burden is on educators to be as good as they can with ai if it is relevant or not if it is not relevant (and schools to fund them and ai companies to fund them if they have excess capital and are humanitarian).

and note the hard part even for us engineers at tech companies is in the correctness. it is very hard. but the sooner we start teaching how to do things correctly with ai, the more prepared the next generations will be.

Really well-written. Maybe black bar @dang?

AOC’s point is simply that when growing exponentially from 250M to 500M to 1B usually whether using exploitive rails one is conscious of or not - someone is bearing an unfair burden. That in 60 out of 60 examples of YC billionaires and all billionaires with maybe a few exceptions - people do not go out of their way to ensure nobody is getting hurt and everyone participates fairly. People are too excited about the exponential growth and their goals. I don’t even think it’s AOC’s main point that these people are at fault. Just that the system is at fault for not ensuring exploitation is minimized further for the rest of the 99%.

(Great essay on how to be a billionaire though. Could billionaires give back more? Yes. But creating market value like that is both worth celebrating and evolving.)

I agree with this. I think the point that's often missed about taxing the ultra-wealthy is it incentivizes them to work through people more instead of doing it all themselves. This is a good incentive.

E.g., if I have no noticeable tax on my wealth as I create impact for the world through my companies I'm going to keep being the one person in charge of that (to achieve my mission of reaching mars, etc.). But if I'm going to get nicked (to the tune of billions of dollars even at 2% etc), on average I'm going to redeploy my assets via people I trust in the company etc. I might even invest more in public welfare projects. It is fair arguably that there is this forcing function because one's value accrues from those projects originally. So there is an elegant symmetry at the end too.

It would be unfair to tax billionaires more if they truly worked in a vacuum and provided value to the economy through very few dependencies. But that's never the case. And right now too much excess is spent on things like these sport teams via inherited wealth etc.

It looks like it's trending back up post-COVID (this link has California data but not sure how you link to this without selecting a state)? https://www.nationsreportcard.gov/profiles/stateprofile/over...

Good to know. Thanks!

Interesting yeah. Thanks for posting I see:

  system_profiler SPDisplaysDataType

  Graphics/Displays:

    Apple M4 Max:

      Chipset Model: Apple M4 Max
      Type: GPU
      Bus: Built-In
      Total Number of Cores: 40
      Vendor: Apple (0x106b)
      Metal Support: Metal 4
      Displays:
        LG UltraFine:
          Resolution: 4096 x 2304
          UI Looks like: 2048 x 1152 @ 60.00Hz
          Main Display: Yes
          Mirror: Off
          Online: Yes
          Rotation: Supported
          Automatically Adjust Brightness: Yes

To be honest it feels crisp. But good to know maybe I need to upgrade more.

You might need a higher quality usb cable. I run ok with LG 5k display and MacBook Pro m4 max.

It’s a reason to push one’s teams to have conventions: (1) all commit messages have an issue number (2) all issues have acceptance criteria or link to issues that do.

Then the answer is just one prompt away… (I wish our team was more committed to 2 - will push for it more in the future). Having acceptance criteria in stories linked at the start of each sprint is so important though I think for alignment of work.

Leadership works on making it better. This is not leadership.

Very cool. Instead of MPs I think you might want to say "Representatives" etc. How to fill out the rest of the data too? Anyway, just wanted to +1. And it's cool you're building in an open way.

Drake equation (extrapolated past just the Milky Way galaxy) points to nearly 1 sextillion earth-like planets (starting with 100-200 billion stars * 200 billion - 1 trillion galaxies). I have no doubt there will have been some that developed life and that are much further ahead of where we are.

One way to predict how they would reach out is how we would reach out or make contact (perhaps not visibly) if we had the technology, wisdom, etc 1 million years into the future. I think we would do it an imperceptible, but perhaps watchful, curious way. So I think that exists - for more or less all intensive purposes it is "no contact." But not due to lack of capability or lack of observance. The statistics/probabilities are just highly predictive that many millions of such "civilizations" already exist. And like with life on Earth, they combine and benefit from diversity as they grow further. But like with responsible life they do not interfere with life for the most part that is still just barely learning to replicate itself consciously (i.e., with AI). The part we are going through now is probably some of the more interesting parts to study in isolation - there's no benefit in interfering at this point given we do not have very sophisticated new things to say. Our information is not very compressed. It's slow. We're like sloths etc. Again, project a million years forward with AI and look back at our present period - or look back at neanderthal periods in human history. There was some interesting art on walls etc., but it's not something that would make sense to interrupt.

I think it wouldn't make sense to interrupt unless it's ready to sort of eradicate itself by accident. So maybe that's when aliens step in (sort of like gardeners if a plant is seriously at risk). Otherwise it probably develops the most unique, useful information if it "gestates" independently. Sort of like a fine fruit. But more practically it needs to get to a dense enough information state where "communicating" with it is possible and interesting (i.e., not just 99.999999% of the time is it becoming like the advanced sentient beings by learning from them - i.e., where it can truly have a dialogue at the speed of something more advanced).

With general relativity we know time can be relative to mass. So for other extremely advanced sentient beings, they don't have to be "impatient." They truly can wait until we get interesting. And maybe then grow us again and see if they can reproduce the experiment and fork off along a particularly interesting bit that is useful for broader intelligence/exploration in the universe. And sometimes maybe it makes sense to graft together two different lifeforms. But probably like with forests for the most part life forms grow independently until their "information" (in the widest possible sense of life) gets potentially useful to the broader group.

Right now we feel early. Like teenagers learning to explore. Of course that's biased by the human development cycle - broader development is unlikely to be like that. It does seem like it'd be more like (this is going to sound really crazy) chariots of fire - fireballs of knowledge growing in various places. And it's more like you want to grow the most energy / information. And information is only information if it is useful/new/diverse from what already exists.