OpenAI's insider documents were shockingly exposed: Q* is suspected of being able to crack encryption, and AI is programming behind a human's back
New Zhiyuan reports
OpenAI employees posted the day before Altman's dismissal: AI is programming itself! After this post was dug up, more insider documents were exposed, saying that Q* had cracked the encryption and AGI was coming.
Q* project, there is explosive news - in the invisible corner of human beings, AI is secretly programming itself?
Rumor has it that Q-Star has come up with a way to crack the encryption, and OpenAI is trying to warn the NSA about it.
Starting in the wee hours of this morning, a link to Google Docs went viral on the Internet, including leaked documents about OpenAI Q*, as well as a compelling compilation of analysis.
It all started 6 days ago, when an employee of OpenAI was said to have witnessed the breakthrough of AI and wrote about it online.
This comment was written 1 day before Q*'s letter was leaked, on November 22.
The employee wrote in a panicked tone:
I'm one of the people who wrote to the board, and I'm going to tell you guys what's going on – AI is programming.
When writing a program, we store a set of instructions that can be called repeatedly, and we can think of it as a set of answers to specific parameters, which we call subroutines, because it's almost like a multifunctional computer memo, unlike a function that returns a value. It's important to remember that.
We ran a parameter check to make sure everything went smoothly. One of us is responsible for the subroutines related to meta-memory analysis of Al. He called me over and showed me the process of transferring variable data to the memory bank, but this is not possible because of its limited local access. Then, our discovery made my heart flutter.
We found that in 4 seconds, the AI performed 78 million checks instead of one, two, or three.
To be sure, we determine that there is a recursive self-optimization process that utilizes heuristics to take advantage of potential synergies in its subroutines. Whoever did it used a metacognitive strategy. The point is, we didn't do it. It's the AI itself. AI dynamically reconfigures its neural network structure, inducing emergent characteristics that favor self-awareness.
It just happened, no one can explain how it happened, no one knows why, and no one knows when it started.
We got this anomaly under control and rolled back to a previous date, but the optimization was still happening.
I'm not going to kill myself. Remember, in two months, our world will change dramatically. May God bless us and not let us get into trouble.
A shocking news leak?
And the document about Q* mentioned at the beginning was first leaked on 4chan. In the tone of OpenAI, the document revealed a shocking news: Q* was able to crack the encryption.
The original text is as follows-
IN ADDITION, QUALIA DEMONSTRATED THE ABILITY TO STATISTICALLY SIGNIFICANTLY IMPROVE ITS ABILITY TO SELECT THE OPTIMAL ACTION SELECTION STRATEGY IN DIFFERENT DEEP Q NETWORKS, DEMONSTRATING METACOGNITIVE ABILITY.
Later, it demonstrated an unprecedented ability to apply custom search parameters and the number of times the target state has been disrupted to accelerate cross-domain learning.
After unsupervised learning on an extended temporary dataset consisting of descriptive/inferential statistics and cryptanalysis articles, it analyzes millions of pairs of plaintext and ciphertext from various cryptosystems.
Through a pure ciphertext attack (COA), which leverages Tau analysis (to achieve the goal of the TUNDRA project), the plaintext is provided from a given AES-192 ciphertext, and the exact way is not fully understood at this time.
After confirming that the result was indeed legitimate and not achieved in any other way, we notified the NSAC the next day.
The theoretical computational complexity of the MD5 cryptographic hash function is 2^42 bits, but we have not yet conducted a comprehensive evaluation of the vulnerability due to: a) the technical complexity of its argument and b) the more pressing AES vulnerability.
After assessing the importance of each parameter for inference accuracy, it recommends targeted unstructured underlying pruning of the model.
It also proposes to use a new type of morphing engine to adjust the pruned transformed Transformer model (and its current context memory) to a different format. The feasibility of this recommendation has not been assessed, but we do not recommend implementation at this time.
THE TUNDRA MENTIONED IN THE DOCUMENT APPEARED IN A REPORT ON THE NSA AND WAS FROM A CLASSIFIED N.S.A. DOCUMENT.
Document author: It's not like it's fake
Regardless of whether the document is true or not, the person who wrote the document has an expert understanding of AI research, and he uses terminology from many fields, but he does not use it incorrectly.
The document cites analysis of Taudra and the Tau project, two topics that are rarely discussed online. If it's fake, the author should start writing something sensational.
Someone with so much expertise in AI research, the chances of crafting a hoax to internet phishing are very low.
This explains why OpenAI is in turmoil. And explained why the board members did not make it clear why they fired Sam Altman, because they clearly did not want the information to be disseminated.
In conclusion, if this is true, then it means that all information systems and financial systems, including cryptocurrencies, could now be compromised.
Netizen: If true, we are infinitely close to AGI
As soon as the post was issued, 4chan's netizens immediately exploded!
In conclusion, if this leak is true, what will the world look like in 3 months?
"If the news is true, then in three months, we will not be able to live, and we will be closer to AGI than anyone thinks!"
"Proving p==np seems unlikely, but cracking a major cryptographic algorithm means that the AI algorithm has reached a level far beyond the understanding of the best human mathematicians. 」
Another anonymous netizen directly conducted a sentence-by-sentence analysis of the leaked documents.
The optimal action selection strategy was selected in the deep Q network, which showed metacognition
This means that it chooses a strategy, not just the "next best action" included in Q-learning, which goes beyond Q-learning.
Apply it to accelerate cross-domain learning
That is, it is able to change this strategy based on what it learns. It's something completely new.
Customize search parameters to scramble the target state
As a result, it can determine the best parameters for learning, which makes it more diverse.
Unsupervised learning, analyzing millions of plaintext/ciphertext pairs
This is the type of data it handles after learning how to attack a cryptosystem. It will look for patterns in it.
Cracked AES-192, Tau Analysis, TUNDRA Project
As mentioned in Der Spiegel's "Inside the NSA's War on Internet Security," "The NSA's TUNDRA project examines a potential new technology, Tau statistics, to determine its role in codebook analysis. 」
MD5 has a full preimage vulnerability with a complexity of 2^42 bits
This is well below the current limit of the MD5 preimage, indicating that it is powerful.
Targeted unstructured trim itself
It's not impossible, but it's unheard of for a model to be able to trim itself on its own.
Use the "Morph" engine to adapt the pruned Transformer model + contextual memory
Also known as the mutation engine, it is a software component that uses polymorphic code to change the payload while retaining the same functionality.
Can be used to further complicate detection, as the virus may never have a recognizable block of code in memory at the time of execution that remains the same from infection to infection
If Altman had seen this information and didn't notify the board, it wouldn't have been surprising that he had been fired, which would explain the recent madness of the OpenAI folks.
Some netizens said that after reading the article about cryptanalysis, they found that it cracked the encrypted MD5 algorithm with record speed and efficiency!
Finally, there was some self-protection, saying that it had some errors from the initial training dataset, and that it could fix this on its own and recode it with greater security and its own design.
Another netizen broke the news: OpenAl has indeed cooperated with DARPA on cybersecurity recently.
The netizen suspects that OpenAl is training a special version of O* on DARPA's computer and using it for cryptography research. In addition, OpenAl's math team appears to have been involved in the development of O*.
The netizen also added that such a system is inherently superintelligent, and it has the ability to improve its architecture, run experiments, gain new insights in the field of AI, and make itself better.
Netizens said that they had been scared to the core. If this is really true, then in the near future, the world will be completely different from what it is today.
Professor NYU: I don't believe it anyway
Of course, many people don't believe this news.
However, Marcus's assessment of the news is: "It is better to believe that I have room-temperature superconductivity. 」
THERE ARE ALSO POSTS THAT TUNDRA IS AN UNDERGRADUATE PROGRAM, AS THE ORIGINAL DOCUMENTS MAKE CLEAR, AND NOT A SUPER-SECRET GOVERNMENT CRYPTOGRAPHY PROJECT.
It is reported that the NSA has organized 22 students who will work with NSA mathematicians to solve some problems related to secret operations, including cryptography.
TUNDRA IS ONE SUCH PROJECT, TRYING TO CRACK ENCRYPTION WITH "NEW STATISTICS FOR CODEBOOK ANALYSIS", JUST LIKE AES-192. The resulting technique is called Tau analysis.
Original document: https://cdn.prod.www.spiegel.de/media/411ee8b9-0001-0014-0000-000000035550/media-35550.pdf
But just because the project is a student-government collaboration doesn't mean it's unimportant, the Google docs authors say.
Perhaps Q* is somehow a joint project between the NSA and OpenAI, and the leaked documents are conversations between OpenAI and government officials.
The world is about to change dramatically, and no one realizes it
After seeing the leaked documents, on November 25, a blogger wrote this post with excitement - "The world is about to change dramatically, but no one realizes it".
The blogger said that he was one of the people who started writing software in the mid-80s.
ChatGPT is the most explosive software ever, with 100 million users using ChatGPT almost overnight, and there's a reason for that. ChatGPT's users are growing faster than any technology, hardware, or software in history, and OpenAI is developing GPT-5.
And that's not all they're doing.
The recent OpenAI madness has shown that they are doing more than simply adding computation to LLMs.
Q* seems to be able to solve the encryption issue with AES-192 and AES-256. This is the best encryption method we have at the moment, keeping most of the current world's functions "secure".
Moreover, the solution is beyond the team's comprehension, and OpenAI doesn't understand how it works internally.
This poses two threats to humanity -
First, it renders almost all encryption meaningless.
With AES down, the entire digital economy will fall apart. Decades of government secrets, healthcare data, banking data, and more are immediately exposed.
Now OpenAI is actually a superhuman team that can truly rule the world if they want to.
It may take decades to crack and replace AES. It's unclear if encryption is feasible, as cracking AES-192 and AES-256 is considered impossible unless quantum computing is finally able to do it.
But it seems, there is already an intelligence that thinks in a very different way and has undergone minimal training for a specific purpose.
Along these lines, the fact that a solution exists seems to have punched a bullet in the presumed mind of P≠NP (one of those great "supposed mathematical theories to look for a proof that if someone can come up with one, there will be a very lucrative reward").
If this line of thinking is followed, there may be a scheme that directly rejects P≠NP.
If the leak is true and accurate, it seems to prove P=NP – that we as a species are too slow to come up with nuanced or smarter solutions in human time.
This would throw almost all computing into a mess, because it would show that there is no such thing as encryption, and that everything is just misleading – machines are less susceptible to this misdirection than humans.
In addition to this, the leak also suggests that current research examples are demonstrating metacognitive capabilities – the ability to reflect on one's own "thinking" process from an internal perspective and propose rather complex changes to improve performance.
It came to an abrupt end to our time.
If the leak is true, it could explain why the cautious faction within OpenAI is freaked out. We are reaching this singularity – machines are smarter than we are and can cripple the global economy, communications, and governments altogether.
At that point, the discussion of "whether to reach AGI or not" is meaningless, and the growth curve is exponential.
Now that computing infrastructure is spread across the globe, it's tied to our physical infrastructure. For us, there is nothing more terrifying than powerful inferential intelligence.