Large neural networks may be beginning to realize: the chief scientist of openai leads to controversy, and everyone quarrels

Frontier strategy Xiao Xiao From the Aofei temple
qubits | official account QbitAI

“ Today's large neural networks may have begun to take shape .”

OpenAI Chief scientist Ilya Sutskever stay Twitter Put down this sentence , It can be said that the words are not surprising and die endlessly .

Now let's AI The circle fried the pot , Artificial intelligence 、 Replies from experts in the field of neuroscience are rolling in .

The most fierce counterattack is the Turing prize winner 、Meta AI Chief scientist Yann LeCun.

LeCun Fight back on the spot , Academic leaders have come to an end

LeCun The first reply to this is ：

Disagree ！（ current AI） Don't reach “ Slightly conscious ” The lower limit of , Not even reached “ Large neural networks ” Upper limit .

He also said that , If you have to give it to the present stage AI define , that ：

I think you need a specific type of macro architecture that the current network does not have .

See their own employees and their original entrepreneurial partners being hated ,OpenAI cofounder Sam Altman Also personally comment on the end ：

OpenAI The chief scientist expressed his curiosity and openness to a mysterious idea , It's using “ Probably ” One word is the premise .

and Meta Chief AI But scientists say categorically “ No ”.

The tone of the following sentence is slightly ironic , It seems to imply OpenAI More results have been achieved due to an open attitude .

This may explain the past 5 Many situations in .

At the end of the tweet ,Sam Altman Don't forget to poach competitors foundation ：

honorific Meta AI The researchers ： My email address is [email protected] We are looking for ！

next ,LeCun And I pulled it up OpenAI Another co-founder musk , Compare the differences between the latter's aviation and aerospace technologies .

One can build faster planes , And break the height record .

But if one's goal is to enter space orbit , We must study cryogenic tanks 、 Turbine pump, etc .

Don't be so flashy . You can ask musk .

except LeCun Outside , other AI People in the industry are also right OpenAI This speech was written and spoken out .

well-known AI Experts 、 Professor at the University of New South Wales Toby Walsh Cut in ：

Every time such speculative remarks are released , We all need to spend months of effort , Only in this way can people dispel their doubts .

DeepMind Senior research scientists scoff ：

If this view holds , There may also be a little spaghetti in the large wheat fields .

Although in Twitter The criticism occupies the majority , But there is still some support OpenAI In the speech .

One from MIT CSAIL Research Fellow Tamay Besiroglu Express ：

See so many famous machine learners laughing at this idea , What a disappointment .

This makes me less hopeful about their ability to solve some important problems in the coming decades .

He thinks that , image OpenAI instead of Meta Such a laboratory is more likely to solve the profound problems in this field in the near future 、 Strange and important questions .

in addition , There is also an unexpected gain from this quarrel , That's it Sam Altman Revealed GPT-4 The latest news of .

Judging from his statement ,GPT-4 Probably GPT-3 The continuation of .

last year , by OpenAI Provide super large AI chip WSE-2 The company Cerebras According to ,GPT-4 There are about 100 Trillions of parameters , And it will take years .

Back to the storm of controversy “ The core ” Look at , What is it that makes Ilya Sutskever Send out such feelings ？

Perhaps we can see from his experience .

From the initial AlexNet, To the later AlphaGo, To witness GPT-3 and Codex Wait for the model to be born , Almost every time AI In the field of “ Breaking circle ” technology , He is involved .

stay AGI On the road “ The farther you go ”

Ilya Sutskever Ben graduated from the University of Toronto , Used to be Geoffrey Hinton Of the students .

in fact ,Sutskever It is AlexNet One of the authors of .

2012 year , stay Hinton Under the guidance of the old man , He and Alex Krizhevsky Jointly designed this architecture , In those days ImageNet Win the challenge , Lower error rate than the second place 10.8%.

Hinton Founded by DNNResearch After being acquired by Google ,Sutskever Join Google brain as a research scientist .

He once participated in the development of famous AlphaGo, Become one of the authors of many papers , The latter in 2016 In the go game in 4 Than 1 Defeated lishihu .

During Google , He also worked with two other scientists at Google brain , Put forward NLP One of the classic frameworks seq2seq Algorithm .

2015 end of the year ,Sutskever Leave Google , With musk 、Sam Altman And others co founded OpenAI.

2018 year , I worked as a research director for more than two years （research director） after , He became OpenAI Chief scientist of （chief scientist）.

so to speak OpenAI from GPT-2 Developed to GPT-3, From defeating DOTA2 Champion team Rerun To be able to write code and play games Codex, He is one of the witnesses who personally participated in it .

Under his guidance ,OpenAI Is gradually moving towards AGI（ Universal Artificial Intelligence ） The path of .

2021 At the beginning of year , Multimodal model DALL·E and CLIP appear , It opens the connection between words and images .

at that time OpenAI Put forward the research that ,CLIP It is very similar to the human way of thinking , Let netizens call directly AGI The arrival of is faster than expected .

It was also last year ,GitHub and OpenAI Cooperate to launch an automatic code completion tool GitHub Copilot,AI Start get To the skills of some programmers .

Earlier this year ,OpenAI Mathematics of research AI Model Lean Further , After adding the neural theorem prover, we successfully solved two International Olympiad topic .

Now OpenAI We are studying a large model with billions of parameters GPT-4, Maybe this is also related to Sutskever This feeling is related to .

“200 Million parameters AI I can't even add or subtract ”

But in fact ,AGI The road of seems not so clear .

At least recently , Assistant professor at New York University Brendan Dolan-Gavitt Find out ,200 A big model with 100 million parameters GPT-NeoX Even the most basic integer arithmetic problem is wrong .

GPT-NeoX Is not OpenAI Official model , It's a name EleutherAI The big open source model developed by the machine learning group of （ because GPT-3 No open source ）.

△ Part of the results show

stay Brendan Out of 100 Topic integer addition 、 Subtraction and multiplication ,AI It's only right 10 Problem , Although other answers and correct answers are also “ almost ”, But after all, it's all wrong . so AI Did not really understand the four operations .

as for “ Whether neural network has self-consciousness ” This discussion , Netizens also expressed their views .

Some netizens think that , These big guys are not so much talking about AI Are you conscious , Rather, they are discussing what the definition is ：

There are also people who make fun of , Deep neural networks enter “ metaphysics ” Deep water ：

Some netizens put a magic expression bag of robots in the comment area of this discussion ,“AI A frown , Realize that things are not simple ”：

about “ Neural networks are beginning to take shape ” This view , Do you think? ？

Reference link ：
[1]https://futurism.com/conscious-ai-backlash
[2]https://weibo.com/5124470749/LfhTasFiP
[3]https://www.linkedin.com/in/ilya-sutskever/
[4]https://twitter.com/ilyasut/status/1491554478243258368
[5]https://twitter.com/sama/status/1492645047585570816
[6]https://en.wikipedia.org/wiki/Ilya_Sutskever
[7]https://openai.com/blog/authors/ilya/
[8]https://gist.github.com/moyix/267d122f9d92268e3432b05f08976816