First.Principles.AI
Region: DE
Tuesday 06 October 2026 21:15:53 GMT
15336
461
29
38
Music
Download
Comments
Miles Marchiori :
The problem is we can’t see the loss landscape bro. To build a picture with enough resolution to minimise the loss in one step, you’d have to sample the loss across a very high dimensional parameter space at a massive scale, but each sample isn’t meaningfully cheaper than just sampling the gradient in your current state and using that information to inform the next (via gradient descent). So you’re just wasting compute resolving details of the loss landscape that aren’t relevant at all to its minimum. TL;DR gradient descent is a lot better than Montecarlo
2026-10-07 06:29:45
10
Bolvarsdad :
So basically: instead of following the local gradient, train an AI to recognize the surrounding loss landscape and learn how to navigate toward a minimum. The catch is that in high dimensions, you can’t literally visualize the whole landscape, so you’d need a compressed representation of its useful structure.
2026-10-07 08:01:20
4
mobiuspizza :
This require you to evaluate the cost surface enough time to build the image, surely that's equally inefficient as a global particle swarm which evaluates enough points to avoid local minima
2026-10-07 02:38:31
15
Bjsweetz :
It’s just a transformation of the math into geometry, then using a computer vision adjascent algorithm. But it would be cheaper and faster to simply do the calculation. Unless you can compile this very efficiently into a model somehow. But at the end of the day you’re still doing a transform. And once you get to higher dimensional functions your projections become infinitely complex. It’s a cool idea but impractical
2026-10-07 06:38:31
8
Nox. :
Sampling F(x_k + Δ_ij) on a stencil is how finite-difference gradients and Hessians are computed. The heightmap is derivative info, not a bypass. Grids blow up as mⁿ in n dimensions. Every slide is 2D where real models have millions of parameters. That's why gradients win. BFGS already gives implicit inverse curvature, and Nelder–Mead/CMA-ES handle black boxes. Learned optimisers are real, but benchmark them.
2026-10-07 07:48:24
1
Vorname Nachname5855 :
You don't understabd what you are actually saying: Your question is, what if AI just understood math. The meaning is, that it understands it that good, that it just sees. Your pictures already show that.
2026-10-07 04:52:03
12
neri ᗜˬᗜ :
please dont use ai to write your scripts, not only it overlooks important details but also just says blatantly wrong stuff.
2026-10-07 05:52:00
1
Обі-ван Кенобі :
good luck visualizing 1000 dimension surface
2026-10-07 07:33:07
3
Maxorisbad :
The reason we imagine a blind hiker is that to visualise the surface, you need to evaluate it at numerous points, while the blind method only evaluates it at some points, this makes it far more efficient
2026-10-07 09:36:33
0
oXoTHUK_3A_HACKAMU :
bugs also do it when they fly
2026-10-07 04:51:39
2
Artists for Democracy :
Mixed integer optimization
2026-10-06 21:59:38
3
Life42 :
Why a drone? Give him wings
2026-10-07 07:27:12
0
azaral1 :
I made a geometric linear and non linear today
2026-10-07 01:11:31
0
leddy_222 :
thats how my brianbwork
2026-10-07 02:29:45
0
agna mu :
holy cringe. this is essentially just computing a discrete grid and searching for a minimum. infeasible for large scale due to curse of dimensionality.
2026-10-07 01:28:33
13
🜲 마우리 🜐 :
Why all the pop science worshippers got triggered by your post 🤭 is it not Science about free thinking and experimenting?
2026-10-07 08:47:15
1
Aphratuss :
i wonder if you are onto something or on something
2026-10-07 07:50:08
1
Deactive :
Larp ah shi
2026-10-07 03:22:15
1
To see more videos from user @first.principles.ai, please go to the Tikwm
homepage.
