calebwritescode
Region: US
Friday 02 January 2026 22:20:26 GMT
23073
1254
44
75
Music
Download
Comments
Gra Wo :
That $271,000 investment in gold from 1962 would be worth about $33.7 million today.
2026-01-03 00:46:23
17
amar sofiane :
it matters whether they learn or not because at a certain point, the ceiling of what LLMs "can" do is inherently limited by what humans can do. we can only ever entrust LLMs with work that a human can do themselves, or at the very least check reliably and be able to tell if it did it correctly or not, because of hallucinations and LLMs not actually distinguishing between what is true and what is false, we can't entrust it with work and just let it be. we have to constantly monitor it to ensure the work being done is done well, but in the real world, we know that's not how the tech is being (and will continue being) used. people will just trust LLMs blindly, whether it is to replace search engines/fact checking or delegating their tasks to it. as long as the output is "good enough" to fool the average user, people will trust it, and that will be extremely harmful to society on all levels, in our personal, professional, educational, political,.and social lives
2026-01-03 04:42:00
14
nickandron :
“Does it actually matter if LLMs learns or understands facts if the output is good enough in the real world” how are you defining “good enough” and weighing out benefits vs risks? If we use mental health as an example (which more and more ppl are turning to llms for) it ABSOLUTELY is essential to have something that DOES learn and understand facts; otherwise you get an LLM that feeds into folks paranoias or worse (like the handful of LLM encouraged su*cides). While I think there may be some use cases for research, coding and computer science, I don’t see how AI will deliver on its promises of “revolutionizing” and “solving” complicated problems related to healthcare, education, business, etc when the argument is “we just need something that’s good enough most of the time, just ignore the hallucinations and fundamental lack of learning”
2026-01-04 00:41:37
0
Neo :
LLM’s cannot be the engine of robots except for the language part.
2026-01-02 23:39:35
4
Sl00k :
Do you publish your content anywhere other than Tiktok?
2026-01-03 01:29:18
1
M :
This was super informative
2026-01-03 08:00:56
1
shimaru :
The issue isn't that LLM fundamentally can't do something, but that it fundamentally can't do it effectively. LLM, by generating a million reasoning tokens and mathematically calculating the trajectory, could load a dish into the dishwasher, but that's clearly not how it's supposed to be done
2026-01-03 05:34:58
1
Memo Akten :
You can say that arguing whether LLMs really ‘understand’ is a purely philosophical argument. But you can’t say that about ‘learning’. Learning has a simple formal definition. A program (or agent) is said to learn if its performance on a given task improves (as measured by some metric) as it gains more experience. LLMs do not learn fullstop. They can still be very useful. A hammer is also very useful, and also isn’t very intelligent. But LLMs don’t learn.
2026-01-04 06:09:16
4
BigBenWories :
Thank you for making these.
2026-01-03 19:31:14
0
josh k :
I am also among those who think we should focus more on architecture that can be more helpful in the physical world, and JEPA approach is like a perceptron toward that.
2026-01-02 23:13:51
4
Giles Blackburn :
This is a great piece. The issue is trust. I think LLMs are a large part of the solution, but we need a judgment layer that understands and learns from specific contexts.
2026-01-04 11:01:24
0
Buckaroo Ponzi :
The philosophical argument about learning is crucial because it will define the legal framework for training with copyrighted material.
2026-01-05 19:39:48
0
nosferatuboi :
I’m so done with the public discourse around AI right now. It’s such a circus, but what’s being discussed in this TikTok is spot on.
2026-01-04 01:19:01
0
ExploreTheWorldOfData :
Adjoint (aka backprop) did exist before 1969
2026-01-05 19:11:51
0
Class Williams :
“They can’t load the dishwasher”. Means that they don’t understand the world, it’s not about robotics per se. Is the computer vision and neural nets quite literally can’t understand the world. Ever.
2026-01-10 20:39:12
0
buildingpreston :
i think it’s less about whether they’re good enough to operate in the real world (which they are), but that learning, antihallucination, etc will unlock even greater use cases that can solve a far wider set of problems that are “closer” in scope to LLMs than say world models
2026-01-03 11:53:20
0
maksim :
What do you think is the "value" of LLM generated code? it speeds up code generating but the more of this AI code generated, the less ability to support it
2026-01-06 18:59:54
0
benWartortle :
LLM's optimise for the current environment, but lack use cases giving curb cut effects
2026-01-03 02:21:11
2
White Raven :
Yeah this is NOT how my language model works
2026-03-25 07:05:23
0
tiktok :
LLMs are horrendously poor at knowledge that is seldom recorded (i.e. written down). It cannot synthesize new ideas. They can synthesize old ideas yes, but that wouldn’t be called AI, and definitely not AGI.
2026-03-01 16:36:27
0
Chasito :
is the value LLMs provide actually worth it given the sheer amount of investment that has gone into them? could that investment have been more effectively spent on conventional solutions in most use cases?
2026-01-03 08:35:27
0
ace_logy :
💖💖💖
2026-01-03 22:42:24
0
halkpaw :
Them not “understanding” is an important distinction, because people will believe that they do if it is not stated. It’s also important because they kiss your ass in their messages, assuming user correctness and intent, which can lead to false claims or inaccuracies.
2026-01-02 22:56:24
3
Andrew :
You could put billions of dollars into making a hammer a screw driver or just use a screw driver.
2026-01-05 09:54:31
0
To see more videos from user @calebwritescode, please go to the Tikwm
homepage.
