FlyingOctopus0

2.4K posts

FlyingOctopus0

@FlyingOctopus0

Interested in machine learning.

Katılım Temmuz 2013

121 Takip Edilen262 Takipçiler

Sabitlenmiş Tweet

FlyingOctopus0@FlyingOctopus0·1 Eyl

This GAN was trained on chinese characters(fonts) and it is fairly close to them (If you do not look too closely). Clearly it does not look like a font, but I run out patience to train it further. I let it run for about 100h (209k iterations). #100DaysOfMLCode

GIF

English

FlyingOctopus0@FlyingOctopus0·18h

@sureailabs Don't be embarrassed. Its not offensive. I rationalize having bad username by thinking that anybody, who will engage with me does it despite the username. Plus such a name is more memorable, which is a valuable resource on the internet.

English

su, reai@sureailabs·18h

The more I am embarrassed about my username, the less I post. What should I do?

English

103

FlyingOctopus0@FlyingOctopus0·1d

@yoavgo Those more creative parts might be optimizable. For example new concepts usually make understanding easier. So we can train the AI to create text that given to LLM will shorten the COT for a wide range of problems.

English

(((ل()(ل() 'yoav))))👾@yoavgo·1d

by weird and annoying complete coincidence, these also happens to be parts which are very hard to verify mechanistically

(((ل()(ل() 'yoav))))👾@yoavgo

even if we do consider math as the pinnacle of human intelligence, it is not the kind of math that proves stuff, but the kinds of math that invents new concepts and then invents questions about them

English

3.8K

FlyingOctopus0@FlyingOctopus0·1d

@jon_stokes If they were to believe that AI has some divine origin, then it would make sense. Be it a trial given by God or a kind of divine messenger, depending on the flavor of AI psychosis.

English

Jon Stokes@jon_stokes·1d

What I'm most interested in is how those who profess to be Christian but are also on the "AI personhood" + X-risk train reconcile their faith in God as the Decider with their fears about a rogue AI as the Decider. How does that even work? I've asked such ppl but they ignore me.

English

390

Jon Stokes@jon_stokes·1d

The deeper issue here is not quite "agency" but is related to it: sovereignty. There can be only one transcendent Other whose inscrutable will decides the fate of humanity. For the religious, that slot is already taken. For AI worriers, it's some degree of basilisk.

Julie Fredrickson@AlmostMedia

In one of the largest religion surveys ever conducted, perhaps surprisingly, people who are actively religious are the ones who have the highest confidence in AI right now

English

3.4K

FlyingOctopus0@FlyingOctopus0·3d

@yoavgo What do you mean by local? Depending on how you position things a local search can as powerful as a global search. If you pre arrange things so that answers are near the queries, then you don't need global search. Just like it is with a binary search tree.

English

(((ل()(ل() 'yoav))))👾@yoavgo·3d

or maybe in other words: its the better local search not the better global search

English

1.8K

(((ل()(ل() 'yoav))))👾@yoavgo·3d

i am still organizing my thoughts on this, but my current intuition is that the unique LLM ability that enabled the current result was not its ability to fuzzily retrieve related results, but the ability to carry through with some form of fuzzy but strong verification

English

3.1K

FlyingOctopus0@FlyingOctopus0·3d

@yoavgo We will see if it is useless math. There could be a chance, it will unlock a lot of discoveries using algebraic geometry, because people will starting looking. Fermat's Last Theorem also solved useless math, but the tools developed for it were useful.

English

170

(((ل()(ل() 'yoav))))👾@yoavgo·3d

this is not a diss on openai-internal, its a diss on erdos. making progress on a very hard problem is still very impressive. even if this problem is useless math.

(((ل()(ل() 'yoav))))👾@yoavgo

my layman remarks: technically it didnt "solve a problem", it "improved a bound". the actual underlying question, what is the optimal number, remains open. erdos was a bit arrogant and a demi god and coined "is *my* construction optimal", but that was never *the* question.

English

7.2K

FlyingOctopus0@FlyingOctopus0·15 May

@pfau Probably. Human created texts will increase in value, so companies/people will preserve them. I imagine that Google, Facebook and OpenAI has a stockpile of internet data. Most of it will make it to the end as long as there will be no technological collapse.

English

David Pfau@pfau·15 May

There are still Mesopotamian clay tablets from thousands of years ago but I wonder if anything from the beginning of the 21st century will even make it to the end.

Nathaniel Rakich@baseballot

ABC News has now taken all FiveThirtyEight articles completely offline. They now redirect to abcnews dot com/politics. A needless erasure of thousands of pages of knowledge.

English

4.3K

FlyingOctopus0@FlyingOctopus0·14 May

@Meaningness @St_Rev You don't have to be mad. It looks like the original dose graph this method is based on was hyperbolic. And it seems, there is a misconception that this is an exponential relation. Doesn't help that in practice they use easier exponential relation as an approximation.

English

152

David Chapman@Meaningness·14 May

@St_Rev This makes me very mad. Hypothesis: "exponential" now just means "YUUUGE" or "so fast it blows up in your face" and so would confuse the public in this context.

English

4.2K

FlyingOctopus0@FlyingOctopus0·9 May

@iamtrask I would even argue that if AI becomes too useful, then all hardware will go to AI companies and none will be left for local AI. Electricity will get very expensive, as it's only a $0.5T revenue business, so it will not able to support $3T of AI inference.

English

FlyingOctopus0@FlyingOctopus0·9 May

@iamtrask If AI can invent milion dollar ideas, then those numbers are plausible. Whole products and businesses created by running AI for a month. Increasing efficiency in the whole technological stack of humanity. The possibilities will be endless.

English

FlyingOctopus0@FlyingOctopus0·9 May

@yoavgo In your opinion would LLMs behave like compilers, if we created all code from a single very long prompt? Then, the programmers would update the prompt for each change and rerun the generation.

English

FlyingOctopus0@FlyingOctopus0·9 May

@yoavgo This seems a little bit arbitrary. If you use as input to LLM the whole conversation history, you get compiler workflow. If you treat commits as modifications to the code, you have agentic workflow, more so if you were to use incremental builds.

English

233

(((ل()(ل() 'yoav))))👾@yoavgo·8 May

coding agents are not compilers from english to programs. and it is not because they are not deterministic. gist.github.com/yoavg/b2454c4d…

English

4.6K

FlyingOctopus0@FlyingOctopus0·8 May

@jon_stokes Unironically, I think NDG would be the most excited about aliens, when they become a scientific fact. He is the kind to be excited even if the aliens were just some microbes on mars.

English

Jon Stokes@jon_stokes·8 May

Me, to the large reptile alien that is about to eat me: "Hang on... Let me get my phone set up to livestream my gruesome demise. I just really need NDG to see this and wipe that smug grin off his face."

English

489

Jon Stokes@jon_stokes·8 May

I'm so brainwormed by this site that if I'm being honest with myself, 99% of my eagerness to see some irrefutable proof of literal aliens visiting earth is b/c of what it would do to the most annoying personalities on my feed. IDC about anything but seeing my enemies get rekt.

GIF

English

1.5K

FlyingOctopus0@FlyingOctopus0·6 May

@robertskmiles @fchollet So it's always 72? The string "The current problem. The observer that knows the shape of the problem" has this many characters. Deciding what is an accurate description is not an intrinsic choice.

English

135

Rob Miles@robertskmiles·6 May

@fchollet It's still intrinsic, it's the length of the shortest description of the problem itself, concatenated with the shortest description of an observer that can know the shape of the problem :p

English

2.7K

François Chollet@fchollet·6 May

Complexity is not an intrinsic property of a problem. It's a property of the relationship between a problem and an observer. Once you know the shape of the problem it is no longer complex.

English

103

839

71.1K

FlyingOctopus0@FlyingOctopus0·5 May

This strategy would also be absent for AI trying to understand humans. Potentially making any kind of predictability and trust have very shaky foundations. Both sides would then always need a backup plan as eventually one side will do something unexpected. 10/10

English

FlyingOctopus0@FlyingOctopus0·5 May

Thus we project motives onto others, anthropomorphise animals or grant intent to the forces of nature. This strategy is effective for understanding other humans which are too complex to model any other way. But, for understanding equally complex AI this strategy might break. 9/10

English

FlyingOctopus0@FlyingOctopus0·5 May

Everyone seems to talk about AI consciousness. I don’t think that AI will ever be conscious in the traditional sense, but AI might achieve a sophisticated state of being that is incomprehensible to us. 1/10

English

Keşfet

@sureailabs @yoavgo @jon_stokes @pfau @Meaningness @St_Rev @iamtrask @elonmusk