Not true. Predicting the correct answer can require learning an algorithm that produces this prediction, so yes it can learn to calculate.
Prediction creates pressure to understand, that's why they become this intelligent, how they can reason, emergent behavior and so forth.
Doesn't mean they calculate exactly like a calculator though.
If someone says to a LLM "what happens if the glass falls from the table onto a stone floor?" it has to have learned what gravity is, what a stone floor is, glass density etc etc etc this to give the most accurate answer possible.
We also predict what happens before something happens because we have an understanding of what an object is, that glass breaks, what gravity is etc etc, so in that way quite similar to how we think / reason. We are also awarded (dopamine) when we predict right. LLM's are "awarded" too, though not very similar to humans.
@DX I intentionally framed the posting in a way that people can make their own conclusions and linked to Tao's blog post with the comments instead of directly to the AHM statement.
But the thing is, even if it is just a minority of 900 in this case, there are still many in the math community agreeing with them. It looks more like a 50-50 split at the moment.
“A minority,” “many,” and “50–50” could be compatible, but only if we’re clear about which groups we’re counting.
But “uneasy about how OpenAI released it” and “agreeing with the AHM” aren’t the same thing. AGMAI wants responsible release, funding for understanding the results, and open access to the models.
Most of the critics seem to be in the first camp. 50-50 between what, exactly? I haven’t seen any actual poll, just blog comments and X, which skew toward whoever’s angriest.
Not saying there isn't concerned mathematicians. But using X as a gauge of the response to something is self-selection bias. Of course people who use X and care enough to reply aren’t representative of everyone. And they are the most publicly vocal. No one posts "I don't care either way" or "it is what it is".
Social media naturally attracts those loud enough, and strong of enough opinion, they feel the need to stop and post it to the world.
Thats why I dislike social media. People have a hard time grasping numbers of people as a percentage, 5000 comments on a post could be a sliver of the whole in the world. You only see the ones that feel enough about a thing, to post about it.
If you go view Kroger's social media, you could conclude a lot of people have strong opinions, for or against a specific grocery store chain. But its only a large group (to a humans concept of 'large group'), that is a fraction of all people.
You want the true collective opinion on it, don't use social media. Use a site the collective goes to like mathoverflow.
These help establish what mathematicians think, not just mathematicians that visit and post on social media think.
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
Not true. Predicting the correct answer can require learning an algorithm that produces this prediction, so yes it can learn to calculate.
Prediction creates pressure to understand, that's why they become this intelligent, how they can reason, emergent behavior and so forth.
Doesn't mean they calculate exactly like a calculator though.
If someone says to a LLM "what happens if the glass falls from the table onto a stone floor?" it has to have learned what gravity is, what a stone floor is, glass density etc etc etc this to give the most accurate answer possible.
We also predict what happens before something happens because we have an understanding of what an object is, that glass breaks, what gravity is etc etc, so in that way quite similar to how we think / reason. We are also awarded (dopamine) when we predict right. LLM's are "awarded" too, though not very similar to humans.
This is all anthropomorphism. It doesn’t know what gravity is, even if it has all the definitions of gravity. Even we don’t know what gravity really is, but we at least experience something we call gravity.
Anyway, I don’t know what the issue is. It doesn’t need to be able to calculate or experience gravity to be a helpful tool.
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
This is all anthropomorphism. It doesn’t know what gravity is, even if it has all the definitions of gravity. Even we don’t know what gravity really is, but we at least experience something we call gravity.
Anyway, I don’t know what the issue is. It doesn’t need to be able to calculate or experience gravity to be a helpful tool.
Never said it's not helpful just because it might not understand it better or similar to us though. I 100% agree.
It infers it. thats the difference.
Consuming billions of pages of human writing. It can reference: glass, table, falling, floor and know the most probable cloud of words that form the sentence of that question.
If you ask the physics of it, and rate of acceleration, and such of the glass and G forces experienced. It can infer the math from those millions/billions of math examples.
It doesnt learn/conceptualize the 'act' of a glass falling, and what a floor, table or glass 'is'. Nor the idea of gravity as a 'thing' it needs to understand as we do. To do any of that.
Download your own models, and log the mechanistic interpretability (which pegs it hits) with a thing like circuit-tracer, and none of them is it 'grasping' the concept as an action in the way we do. Nor does it invent new data it has 'learned' to do so.
Weights are frozen at inference. The model doesn’t learn or update its world model from experience.
if it did such things, we wouldnt need things like Temperature, seed, Top-k, Min-p etc.
Without those, the answer is always exactly the same, for the exact same question. No new information it 'formed' exists to guide it different the next time.
If you want to test reproducible output locally (that no new concepts or grasping is produced), fix the seed, use temperature 0 and a fixed seed with your sampler settings, keep the same quant, backend, and hardware. And ask it the same copy/paste question 100 times, you will get the same word for word reply.
Edit: Well depends on your model loader and wrapper. Some use the host machine clock for FP calculation randomness. Freeze time in local system time if so.
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
Also doesn't mean they are right even if so. Just means thier minds was changed.
We can say the same about a lot of things people changed thier opinions on that was wrong (and later changed again): Cigarettes aren't dangerous, Thalidomide, Eugenics, Eggs was healthy, then not healthy, Lobotomies.
Just because the general public now thinks a thing. In no way makes it any more true.
I myself don't care if one average joe, or all of them think a thing, about a thing. It doesnt weigh into my decision on if it is to me.
Everyone could be persuaded to think the moon landing was fake. Doesn't mean I will go "most people think this now. Will just let them do the work for me, and lean towards accepting that too".
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
It doesnt learn/conceptualize the 'act' of a glass falling, and what a floor, table or glass 'is'. Nor the idea of gravity as a 'thing' it needs to understand as we do.
of course not, i am not of the opinion that these advanced AI's are like humans at all or would NEED to be like us either. It doesnt exist in the physical world so there's a lot it doesnt need but also can't really grasp as we would always.
probably repost..but ah well.. this is really scary (even though I have to admit there are quite a few games from the 90's i would love to see updated...from companies that does not exist anymore that is)
shitloads of new stuff in my pc. Cant keep track of it all.
When (not if) AI solves prime factorization - finds a quick way to do it - then security is majorly fucked
I can only advise everyone to have offline or local backups of everything critical
When (not if) AI solves prime factorization - finds a quick way to do it - then security is majorly fucked
I can only advise everyone to have offline or local backups of everything critical
AI or not, today there's only one known way to solve prime factorization: Shor's algorithm using powerful enough quantum computer.
Gemini search wrote:
We are roughly three orders of magnitude and a few hardware generations away from a quantum computer capable of breaking standard 2048-bit RSA encryption.
Current Hardware vs. Required Scale
• Where we stand today: Leading state-of-the-art systems like IBM’s Flamingo-class processors have scaled to roughly 1,000 physical qubits, and Google’s Willow chip operates around 105 physical qubits, but they lack deep error correction.
• What is required: Thanks to recent algorithmic breakthroughs (such as optimized circuit designs and quantum low-density parity-check or qLDPC codes), estimates have dropped sharply. Breaking RSA-2048 now theoretically requires between 100,000 and 500,000 physical qubits configured into reliable error-corrected logical qubits, down from past estimates of 20 million.
Expert Timeline Expectations
• Early-to-mid 2030s: Most independent consensus and risk assessments (such as the Global Risk Institute) point to a 50% cumulative probability of a cryptographically relevant machine arriving in this window
lol. 3 years ago this shit was state of the art, i am glad i have kept some of this garbage on my channel, so fun to go back
4 years ago this video went viral on Facebook and on Instagram (people contacted me, wanted me to make music videos, i made 2 and earned a little bit of money lol):
probably one of the first viral AI videos, not too many made them, you had to stitch together the frames yourself.
total game changer this tool, currently using it to reverse engineer fallout 4 so that the agent can understand enemy behaviors better. it (Codex, Astra high) got it from source code to doing something useful with it in under 4 minutes. I suspected it would take far longer.
Will be so useful for modding overall, people building 3D importer/exporters, tools like Nifskope etc (i think they have figured out most stuff for Bethesda models but unlikely everything).
Crazy mods will come for sure.. sorry, i meant InSanE mods.
Edit: oh yes indeed. Fixed our issue with enemies not falling back and going into cover when hurt badly. Damn, this will be so useful.
I just need it to make it full screen now. Took around 45 min - 1 hour maybe, unsure, went to sleep but woke up to a Friday13th.exe Controller works.
Just figuring out what to do with it.. make it more like a 16 bit game maybe, add day and night schedules? No! I know, i need to make you able to bUiLd yoUR oWn hOuSe and gather all the REsoUrcEs
Signature/Avatar nuking: none (can be changed in your profile)
You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum
Controller works.
