Filip Piekniewski
@filippie509.bsky.social
Hype-free rational view on AI.
Saw this posted on LinkedIn. I guess I'd probably agree.
April 10, 2025 at 4:52 AM
Saw this posted on LinkedIn. I guess I'd probably agree.
Full deepseek is <1TB download, so technically speaking you can fit a compressed version of the _entire_ Internet on one of these. In fact twice over. Let that sink in - compressed knowledge of the whole Internet on something the size of a fingernail and costing less than $200.
February 3, 2025 at 5:38 PM
Full deepseek is <1TB download, so technically speaking you can fit a compressed version of the _entire_ Internet on one of these. In fact twice over. Let that sink in - compressed knowledge of the whole Internet on something the size of a fingernail and costing less than $200.
Found a full version of the photo!
January 29, 2025 at 6:27 PM
Found a full version of the photo!
FrontierMath my ass. Marching towards AGI one fraud at a time.
January 19, 2025 at 8:44 PM
FrontierMath my ass. Marching towards AGI one fraud at a time.
Seventh flight of "Starship" vs seventh flight of Saturn V some 55 years ago...
But the booster landing was👌
But the booster landing was👌
January 16, 2025 at 11:55 PM
Seventh flight of "Starship" vs seventh flight of Saturn V some 55 years ago...
But the booster landing was👌
But the booster landing was👌
Seems like a good day to post this.
January 15, 2025 at 1:49 AM
Seems like a good day to post this.
Internet in 2025
January 3, 2025 at 5:11 PM
Internet in 2025
LOL. I guess self-driving could be seen as a kind of a weird digital entertainment...
December 19, 2024 at 1:41 AM
LOL. I guess self-driving could be seen as a kind of a weird digital entertainment...
Must have become conscious and escaped. Clear sign of singularity.
December 12, 2024 at 12:45 AM
Must have become conscious and escaped. Clear sign of singularity.
Benedict sees the writing on the wall. At some point everyone will.
December 4, 2024 at 9:14 PM
Benedict sees the writing on the wall. At some point everyone will.
One of my top spots for camping in SW US hidden close to Las Vegas.
December 1, 2024 at 3:25 PM
One of my top spots for camping in SW US hidden close to Las Vegas.
Hmm I wonder how will this work out. Let’s look at a similar company TuSimple… ohhh
November 27, 2024 at 3:46 PM
Hmm I wonder how will this work out. Let’s look at a similar company TuSimple… ohhh
Captcha is always a good and objective test for the actual capabilities of AI. We are currently here. According to some folks in 1-3 years we will be at a level of AGI supertintelligence. From these two statements I can construct another captcha: can you see a problem?
November 23, 2024 at 5:17 PM
Captcha is always a good and objective test for the actual capabilities of AI. We are currently here. According to some folks in 1-3 years we will be at a level of AGI supertintelligence. From these two statements I can construct another captcha: can you see a problem?
When we look back at this in mid 2030, and then 2040, nobody will acknowledge that anyone had actually believed this. In fact nobody will believe anyone could have believed it. Such are popular delusions and madness of crowds.
November 22, 2024 at 4:07 AM
When we look back at this in mid 2030, and then 2040, nobody will acknowledge that anyone had actually believed this. In fact nobody will believe anyone could have believed it. Such are popular delusions and madness of crowds.
My followers on here quadrupled in the last week. I had an account for many months now, so there is no question the trend is real. Even the bros on twitter are noticing…
November 19, 2024 at 7:04 PM
My followers on here quadrupled in the last week. I had an account for many months now, so there is no question the trend is real. Even the bros on twitter are noticing…
Which suggests the spiral like problems could perhaps be simpler if periodic/trigonometric functions were allowed inside the MLP components. So why should we care about spirals, isn't this just purely academic nonsense? Well spirals are baby fractals.
November 19, 2024 at 7:04 PM
Which suggests the spiral like problems could perhaps be simpler if periodic/trigonometric functions were allowed inside the MLP components. So why should we care about spirals, isn't this just purely academic nonsense? Well spirals are baby fractals.
Which means that if we knew that transformation up front and applied it to the input pixels, the problem turns from horrific to trivial, solvable by just one neuron. We just need to know the inverse of the initial transformation.
November 19, 2024 at 7:04 PM
Which means that if we knew that transformation up front and applied it to the input pixels, the problem turns from horrific to trivial, solvable by just one neuron. We just need to know the inverse of the initial transformation.
But note that the spiral gets generated from a super trivial linearly separable set by a very simple and differentiable transformation, which is basically a rotation matrix scaled by the absolute distance from the origin:
November 19, 2024 at 7:04 PM
But note that the spiral gets generated from a super trivial linearly separable set by a very simple and differentiable transformation, which is basically a rotation matrix scaled by the absolute distance from the origin:
Why is this so difficult? Well simply speaking the number of hyperplanes necessary to carve out the spiral shape is huge (infinite in the limit) and the error landscape to align them is absolutely horrifying (remember we need to sample this set randomly).
November 19, 2024 at 7:04 PM
Why is this so difficult? Well simply speaking the number of hyperplanes necessary to carve out the spiral shape is huge (infinite in the limit) and the error landscape to align them is absolutely horrifying (remember we need to sample this set randomly).
This is a two spiral problem. It's a well known example of a dataset in 2d that is very very hard to learn by a simple multilayer perceptron that gets (x, y) as input (and is supposed to learn image intensity. Thread (some thinking out loud)👇
November 19, 2024 at 7:04 PM
This is a two spiral problem. It's a well known example of a dataset in 2d that is very very hard to learn by a simple multilayer perceptron that gets (x, y) as input (and is supposed to learn image intensity. Thread (some thinking out loud)👇