golol

Where does the angular momentum go?

that's probably dalle

Consider n real numbers a_0, ..., a_{n-1}. That's not very elegant.

I think another way to think about this is that subtly trying to consider AI in your AI-unrelated research is just respecting the bitter lesson. You need to at least consider how a data-driven approach might work for your problem. It could totally wipe you out - make your approach pointless. That's the bitter lesson.

Nice article. I didn't read every section in detail but I think it makes a good point that AI researchers maybe focus too much on the thought of creating new mathematics while being able to repdroduce, index or formalize existing mathematics is really they key goal imo. This will then also lead to new mathematics. I think the more you advance in mathematical maturity the bigger the "brush" becomes with which you make your strokes. As an undergrad a stroke can be a single argument in a proof, or a simple Lemma. As a professor it can be a good guess for a well-posedness strategy for a PDE. I think AI will help humans find new mathematics with much bigger brush strokes. If you need to generalize a specific inequality on the whole space to Lipschitz domains, perhaps AI will give you a dozen pages, perhaps even of formalized Lean, in a single stroke. If you are a scientist and consider an ODE model, perhaps AI can give you formally verified error and convergence bounds using your specific constants. You switch to a probabilistic setting? Do not worry. All of these are examples of not very deep but tedious and non-trivial mathematical busywork that can take days or weeks. The mathematical ability necessary to do this has in my opinion already been demonstrated by o3 in rare cases. It can not piece things together yet though. But GPT-4 could not piece together proofs to undergrad homework problems while o3 now can. So I believe improvement is quite possible.

>Each individual function eventually does zero things.

Lambda calculus, basically :)

You are considering issues of like 0.25x - 4x efficiency gain or loss while ignoring that the real issue is possible vs impossible. If I have a household tobit that can vacuum the floor taking 4x as time as me I don't care. It's still extremely useful. This is why we human-imitatinf approached are used everywhere. It won't be the most efficient solution but the data contains information that helps make it possible at all.

Chatbot Arena is not really a great benchmark imo

Deepseek is not on par with o1.

I am a math PhD student and I already draw some value from recent reasoning models. I strongky believe than in 1-2 years LLMs will become established tool for scientists to help with coding and math. I really don't think you can call this a con...

powerful general BCI, FDVR, cyborg intelligence and so on. Elon Musk clearly stated this many times.

b...but stochastic parrot!

Modern operating systems should include by default a very simple private/public key system to sign arbitrary files. I think it should not be very complicated? We badly need this in the age of AI.

oh thanks I think Im blind :)

I found the image at the bottom very surprising and enlightening. I wonder where it is from. It's really a new thought for me.

o3-mini is not really "a version of the o3 model", it is a different model (less parameters). So their language strongly suggests, imo, that Deep Research is powered by a model with the same number of parameters as o3.

This mentality is why teenagers can't use a file system. Why do tech people love to hide as much state as possible. Does it really help anyone?

I don't find OpenAIs naming conventions confusing, except that the o for omni and the o for reasoning have nothing to do with eachother. That's a crime.

The things that are missing are what stops us from having useful agents so far: Agency, judgement, sense of time, long horizon planning, not being gullible. I kinda feel like some amount of ego is necessary to get a model to behave like that.

Doesn't writing out the trailing zeroes mean that you clain 4 significant digits? At least in the physics context.

This is kind if true. I feel like the reasoning power if O1 is really only truly available on the kinds of math/coding tasks it was trained on so much.

If they stick the landing first try I'm really impressed abd SpaceX has real competition.

Hmmm without a human control it is not all that clear to me that the variation problems are not more difficult.

In order to replace actual humans doing their job I think LLMs are lacking in judgement, sense of time and agenticism.

The shape doesn't matter! Non-humanoid shapes give minir advantages on specific tasks but for a general robot you'll have a hard time finding a shape much more optimal than humanoid. And if you go with humanoid you have so much data available! Videos contain the information of which movements a robot should execude. Teleoperation is easy. This is the bitter lesson! The shape doesn't matter, any shape will work with the right architecture, data and training!

If OpenAI training on youtube videos violates copyright then so does Google training on them.

It just doesn't make sense. So it is immoral for engineers to design weapons? That's like saying it's immoral to butcher an animal. Sure go ahead, ban butchering animals in your society so everyone can feel good about themselves while they import their meat and thereby outsource the butchering instead. Or you say that everyone just stops eating meat? Fine that works but here the analogy breaks, because everyone won't just stop needing weapons.

What I can agree with is that placing a moral cost on engineering designing weapons in some way increases the actual labor cost of weapon R&D, I guess.

Lambda calculus consists pretty much just of function compositions, which in principle have a tree structure. Is this just lambda calculus in a dress?

All I see is trivialities.

But why would the dataset be worse, they can just use the same one as before.