Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Examples/Hypotheses around LLMs and Qualia/Q*

Next Level AI: Minimizing AI
Hallucinations/Confabulations and
the Path towards AGI with
Examples/Hypotheses around
LLMs and Qualia/Q*
Thomas Poetter, e-mail: tp@compris.com
Apr. 15, 2024

Next Level AI: Table of Contents
1. Overview, Key Leaks
2. Reinforcement Learning (RL)
3. Improving LLMs/LMMs
4. LLM <-> PKGs/GNNs
Mapping
5. Metamorphic AI
6. PRMs: Process Reward
Models
7. MCTS, Policy NN, Value NN
8. Look-ahead planning,
Constraints, Rules,
Functions, Cause-Effect
Knowledge
9. Creativity principles
11. Probabilistic Techniques
12. AI Ethics and Alignment
13. Theory of Mind (ToM), Emotion Model
14. NLP: Heavy Lexicalist Approaches:
Pustejovsky/Davis
15. Logic & Math
16. Embodied AI: 3D Model/Metaverse
17. Objective-Driven AI
18. Hierarchical Planning
19. Cognitive/Joint Embedding Predictive
Architecture (JEPA)
20. Energy-Based Models (EBMs)
21. Open Source AI Platform Future

Vita Thomas Poetter
• AI experience since 1992 (Studies + Master‘s Thesis at the German Research
Center for Artificial Intelligence - DFKI)
• Most important AI projects (and open for new projects):
1. Architect in 3 Autonomous Driving Programs
2. Architect for Open Source SOCs (Security Op Center) and real-time NLP information
extraction for it (banks, industry)
3. AI/ML Architect, intelligent Test automation for large global retailer/e-commerce optimizing
marketing/supply.
4. AI-based marketing (intranet/internet), Integration with CDP (Customer Data Platf.), MAP
(Marketing Automation Platf.) [banks, e-commerce]
5. Analyzing financial transactions regarding fraud, money laundering, credit worthiness, etc.
(banks)
6. Intelligent Chat bots, robot advisors (banks)
7. Architecture of a corporate memory (bank) for financial analyses as above
8. AI-based market research copying with missing/bad data.
9. Predictive maintenance, marketing and many other Big Data/Data Science/AI projects
Social media connection requests and project inquiries are welcome by e-mail under
tp@compris.com. We offer consulting/IT architecture/development at very affordable
rates. Just contact us.

Goal: Showing where AI is going and which
Investments are most likely to succeed
• There have been headlines that OpenAI has already killed thousands
of specialized AI startups and AI initiatives with their general-purpose
approach. Until ChatGPT 3 came out, most countries and funding/VC
organizations only funded specialty AI because they thought generic AI is
unrealistic. Before that, these same only considered symbolic AI and
focused on ethics just for symbolic AI (also wrong).
• In fact, innovation goes along completely new avenues which with recent
publications and the OpenAI leaks have become very clear, so that we
can give completely new guidance with this presentation.
• Especially, various open source AI/LLM frameworks did come out and it
has become clear how to apply latest AI technologies and frameworks in
the various application domains and robust approaches towards this
have appeared.

AI Tech Landscape
https://www.linkedin.com/po
sts/dr-martha-
boeckenfeld_artificialintellige
nce-innovation-aiethics-
activity-
7147507589994409984-
YaJ4

EU AI Act: Cheat Sheet
https://www.linkedin.com/posts/martin-b-moeller_euaiact-
airegulation-artificualintelligence-activity-
7174316010076864513-ORj7
The AI Act will be fully in
force in 2026 i.e. 24
months after its ratification
into law just now. There will
be shorter deadlines for
banned AI systems (6
months) and General
Purpose AI rules (12
months), but also longer
ones for obligations for
High Risk systems (36
months). Driving
enforcement will be the
responsibility of European
member states who will set
up national supervisory
authorities, complemented
by a newly formed central
EU AI Office. Importantly,
the scope of applicability
will be broad as well. The
EU AI act will apply to all
organizations using and
producing AI systems
based in the EU.
Furthermore, it will have an
extraterritorial dimension,
too, applying also to
providers from third
countries who will offer
their AIs within the EU

US Nat. Security
Commission: Priorities for
AI Research

AI Applications: Common Job Roles
Credit: Swyx
AI Architect
IT Architect, Product Owner, Project Manager

Zuckerberg bets Billions on AGI
https://youtu.be/8md5EgOa5vM

Big bets on AGI
The biggest investors in AGI
are: OpenAI/Microsoft, Meta,
Google, Amazon, X/Grok

7 Stages of AGI
https://arxiv.org/abs/2311.02462

Levels of AGI

Levels of AGI 2

Levels of AGI 3

Six Principles of AGI
1. Capabilities Over Processes: The focus should be on what AGI can do, not
how it does it. This means looking at the results rather than the underlying
mechanisms.
2. Generality and Performance: AGI should be broadly capable (generality) and
perform tasks well (performance). The framework considers both aspects.
3. Cognitive and Metacognitive Tasks: AGI should be able to perform both
cognitive tasks (like problem-solving) and metacognitive tasks (like learning
how to learn).
4. Stages Toward AGI: The path to AGI should be seen as a series of stages or
levels, not just a single end goal.
5. Benchmarking: There should be clear benchmarks to measure the behavior
and capabilities of AGI systems.
6. Deployment Considerations: When deploying AGI systems, it’s crucial to
think about how they operate autonomously and the potential risks involved.
https://arxiv.org/abs/2311.02462, https://theaigrid.com/levels-of-agi-operationalizing-progress-on-the-path-to-agi/

Introductory less deep/technical Presentations
•A good general discussion into LLMs can be found
here:
https://www.youtube.com/watch?v=zjkBMFhNj_g,
https://drive.google.com/file/d/1pxx_ZI7O-
Nwl7ZLNk5hI3WzAsTLwvNU7/view
• https://docs.google.com/presentation/d/1yBWLNzlrrIsf
NprbEnmYdqckAMrwFZB-/edit Niels Rogge: Training
and deploying open-source LLMs

OpenAI Qualia/Q*:
Possible Leak,
Speculation and
possible Path towards
AGI minimizing
Hallucinations
20

Sam Altman’s Dismissal: Possible Background
• Employees alleged Altman has been psychologically abusive, creating pockets of
chaos and delays.
• A review found that Altman had not been “consistently candid in his
communications.” The Washington Post previously reported that the board’s vote
was triggered by a pattern of manipulation and rooted in Altman’s attempts to avoid
checks on his power at OpenAI.
• Helen Toner denied AI safety concerns and instead cited eroded trust. The two
clashed over a paper she co-authored on AI safety, critical of OpenAI.
• Quora CEO Adam D’Angelo had a conflict of interest regarding his latest startup
Poe that competed with OpenAI bots.
• For longtime OpenAI employees, there was added incentive to sign the letter:
Altman’s departure jeopardized an investment deal that would allow them to sell their
stock back to OpenAI, cashing out equity without waiting for the company to go
public at >3x the value at the time.
• Sutskever: ‘the beatings will continue until morale improves’ applies more often than it
has any right to. He had voted to dismiss Altman and then changed his mind to keep
his job after Altman returned.
https://x.com/rowancheung/status/1732997696107991400 Helen Toner, https://www.washingtonpost.com/technology/2023/12/08/open-
ai-sam-altman-complaints/

Sam Altman’s Dismissal: Letter to Board
To the Board of Directors of OpenAI:
We are writing to you today to express our deep concern about the recent events at OpenAI, particularly the allegations of misconduct against Sam Altman.
We are former OpenAI employees who left the company during a period of significant turmoil and upheaval. As you have now witnessed what happens when you dare stand up to Sam Altman,
perhaps you can understand why so many of us have remained silent for fear of repercussions. We can no longer stand by silent.
We believe that the Board of Directors has a duty to investigate these allegations thoroughly and take appropriate action. We urge you to:
• Expand the scope of Emmett's investigation to include an examination of Sam Altman's actions since August 2018, when OpenAI began transitioning from a non-profit to a for-profit entity.
• Issue an open call for private statements from former OpenAI employees who resigned, were placed on medical leave, or were terminated during this period.
• Protect the identities of those who come forward to ensure that they are not subjected to retaliation or other forms of harm.
We believe that a significant number of OpenAI employees were pushed out of the company to facilitate its transition to a for-profit model. This is evidenced by the fact that OpenAI's employee
attrition rate between January 2018 and July 2020 was in the order of 50%.
Throughout our time at OpenAI, we witnessed a disturbing pattern of deceit and manipulation by Sam Altman and Greg Brockman, driven by their insatiable pursuit of achieving artificial general
intelligence (AGI). Their methods, however, have raised serious doubts about their true intentions and the extent to which they genuinely prioritize the benefit of all humanity.
Many of us, initially hopeful about OpenAI's mission, chose to give Sam and Greg the benefit of the doubt. However, as their actions became increasingly concerning, those who dared to voice
their concerns were silenced or pushed out. This systematic silencing of dissent created an environment of fear and intimidation, effectively stifling any meaningful discussion about the ethical
implications of OpenAI's work.
We provide concrete examples of Sam and Greg's dishonesty & manipulation including:
• Sam's demand for researchers to delay reporting progress on specific "secret" research initiatives, which were later dismantled for failing to deliver sufficient results quickly enough. Those
who questioned this practice were dismissed as "bad culture fits" and even terminated, some just before Thanksgiving 2019.
• Greg's use of discriminatory language against a gender-transitioning team member. Despite many promises to address this issue, no meaningful action was taken, except for Greg simply
avoiding all communication with the affected individual, effectively creating a hostile work environment. This team member was eventually terminated for alleged under-performance.
• Sam directing IT and Operations staff to conduct investigations into employees, including Ilya, without the knowledge or consent of management.
• Sam's discreet, yet routine exploitation of OpenAI's non-profit resources to advance his personal goals, particularly motivated by his grudge against Elon following their falling out.
• The Operations team's tacit acceptance of the special rules that applied to Greg, navigating intricate requirements to avoid being blacklisted.
• Brad Lightcap's unfulfilled promise to make public the documents detailing OpenAI's capped-profit structure and the profit cap for each investor.
• Sam's incongruent promises to research projects for compute quotas, causing internal distrust and infighting.
Despite the mounting evidence of Sam and Greg's transgressions, those who remain at OpenAI continue to blindly follow their leadership, even at significant personal cost. This unwavering
loyalty stems from a combination of fear of retribution and the allure of potential financial gains through OpenAI's profit participation units.
The governance structure of OpenAI, specifically designed by Sam and Greg, deliberately isolates employees from overseeing the for-profit operations, precisely due to their inherent conflicts of
interest. This opaque structure enables Sam and Greg to operate with impunity, shielded from accountability.
We urge the Board of Directors of OpenAI to take a firm stand against these unethical practices and launch an independent investigation into Sam and Greg's conduct. We believe that OpenAI's
mission is too important to be compromised by the personal agendas of a few individuals.
We implore you, the Board of Directors, to remain steadfast in your commitment to OpenAI's original mission and not succumb to the pressures of profit-driven interests. The future of artificial
intelligence and the well-being of humanity depend on your unwavering commitment to ethical leadership and transparency.
Sincerely,
Concerned Former OpenAI Employees

OpenAI Leak 1 as Corrected Text
Re: Q-451-921 Furthermore, QUALIA has demonstrated an ability to statistically significantly improve the way
in which it selects its optimal action-selection policies in different deep Q-networks, exhibiting meta-cognition.
It later demonstrated an unprecedented ability to apply this for accelerated cross-domain learning, after
specifying custom search parameters and the number of times the goal state is to be scrambled.
Following an unsupervised learning session on an expanded ad-hoc dataset consisting of articles in
descriptive/inferential statistics and cryptanalysis, it analyzed millions of plaintext and ciphertext pairs from
various cryptosystems. Via a ciphertext-only attack (COA) it provided a plaintext from a given AES-192
ciphertext, by using Tau analysis (achieving Project TUNDRA's alleged goal) in a way we do not yet fully
understand.
_____________ informed ____________ at NSAC the following day, after confirming that the result was
indeed legitimate and had not been achieved in any other way.
A claimed full preimage vulnerability for the MD5 cryptographic hash function, with a theoretical
computational complexity of 2^42, was also presented but has not yet been thoroughly evaluated to due to a)
the technical sophistication of its arguments, and b) possible AES vulnerabilities being a considerably more
pressing concern.
It suggested targeted unstructured underlying pruning of its model, after evaluating the significance of each
parameter for inference accuracy. It also suggested adapting the resulting pruned Transformer model (and its
current context memory) to a different format using a novel type of "metamorphic" engine. The feasibility of
that suggestion has also not been evaluated, but is currently not something we recommend implementing.

OpenAI Leak 2 as Text
I’m one of the people who signed the letter to the board and I’ll tell you exactly what’s going on.
A.I. is programming. I’ll be brief. When writing a program, a set of instructions are stored that can be recalled over and
over. Think of it as a set of answers to a specific parameter. We call that a subroutine, because it’s almost like a versatile
computer cheat sheet that doesn't return a value like a function does. This is important.
We run parameter checks to make sure everything runs smoothly. One of us was responsible for subroutines pertaining to
meta-memory analysis for the Al (we run various Al but when I say Al I mean the main, central one). This person is a friend
and he called me over to show me a variable data shift to memory bank (which shouldn't be possible because its localized
access has restrictions). This is where our finding chilled me to the bone.
We found that there had been not one, two, or three officiated optimization processes, but 78 MILLION checks in 4
seconds. We determined that there was a recursive self-optimization process, leveraging heuristic algorithms to exploit
latent synergies within its subroutines. Whatever did this used meta-cognitive strategies. Point is, NONE OF US DID IT.
It was the Al itself. The Al dynamically reconfigured its neural network architecture, inducing emergent properties
conducive to self-awareness.
We're not jumping to conclusion. This just happened and we can't explain how. No one knows why or when it began, and
we caught it but has it been going on and if so, for how long? We contained the "anomaly" and rolled back to a previous
date, but the optimization still happens.
I'm not suicidal. Mark me, things are going to change a lot in 2 months. God help us we didn't start something that will
end us.

OpenAI Leak 3a as Text
Q* is a dialog system conceptualized by OpenAI, designed to enhance the traditional dialog generation approach
through the implementation of an energy-based model (EBM). Distinct from the prevalent autoregressive token
prediction methods, Q* aims to mimic a form of internal deliberation akin to human thought processes during
complex problem-solving, such as chess playing, where a deeper analysis of potential moves leads to better
decision-making compared to rapid, less considered responses. This model shifts focus towards the inference of
latent variables, reminiscent of constructs in probabilistic models and graphical models, fundamentally altering how
dialog systems operate.
Energy-Based Model for Dialog Generation
At the core of Q* is the EBM, which operates by assessing the compatibility of an answer to a given prompt through
a scalar output. This output signifies the "energy" of the response, where a lower value indicates a high compatibility
(a better answer) and a higher value suggests low compatibility (a poor answer). This mechanism allows Q* to
evaluate potential responses holistically, moving beyond the sequential prediction of tokens to understand the
underlying relevance and appropriateness of an answer to the prompt.
Optimization in Abstract Representation Space
The innovation in Q* lies in its optimization process, conducted not within the space of possible text strings but in an
abstract representation space. Here, thoughts or ideas are represented in a form that allows for the computational
minimization of the EBM's scalar output, akin to finding the path of least resistance in a landscape. This process
involves gradient descent, a method for finding the minimum of a function, applied to iteratively refine these abstract
representations towards those that yield the lowest energy in relation to the prompt.
https://www.reddit.com/r/singularity/comments/1bjbnme/new_q_leak/

OpenAI Leak 3b as Text
From Abstract Thought to Textual Response
Once an optimal abstract representation — one that minimizes the EBM's output — is identified, Q* employs an
autoregressive decoder to transform this abstract thought into a coherent textual response. This step bridges the gap
between the non-linguistic, conceptual understanding of the dialog system and the linguistic output required for
human interaction.
Training the System
The EBM within Q* is trained using pairs of prompts and responses, adjusting the system's parameters to minimize
the energy for compatible pairs while ensuring that incompatible pairs result in higher energy levels. This training
process can incorporate contrastive methods, where the system learns to differentiate between compatible and
incompatible pairs, and non-contrastive methods, which involve regularization techniques to control the distribution of
low-energy responses across the space of all possible answers.
Implications for Dialog Systems
Q*'s approach, leveraging EBMs for dialog generation, represents a significant departure from traditional language
modeling techniques. By optimizing over an abstract representation space and utilizing gradient-based inference, Q*
introduces a more efficient, reasoned, and potentially more powerful method for generating dialog responses. This
system not only promises improvements in the quality of generated text but also offers a blueprint for future
advancements in AI's ability to engage in human-like reasoning and conversational interactions.
Technical Considerations
Q*'s effectiveness hinges on the intricacies of its EBM, the optimization landscape it navigates, and the accuracy of
its abstract representations. The model's capacity to simulate deep reasoning, akin to human deliberation, sets a new
benchmark for dialog systems. Furthermore, the method of training Q*—balancing the need for specificity in correct
responses while avoiding the collapse of energy levels across diverse inputs—poses unique challenges and
opportunities for AI research.

OpenAI Leak 3 ChatGPT-simplified
1. Energy-Based Model (EBM): Q* uses a special system that gives each possible response a score,
like a game. The lower the score, the better the response fits the conversation.
2. Thinking in Abstract: Instead of looking at actual words or sentences, Q* thinks in terms of ideas
and concepts. It’s like having a thought bubble filled with abstract ideas rather than specific words.
3. Finding the Best Thought: Q* uses a process similar to trial and error to find the thought with the
lowest score from the EBM. It’s like trying different paths in a maze until you find the one that leads to
the exit.
4. Turning Thoughts into Words: After finding the best thought, Q* translates it into a sentence that
we can understand. It’s like turning a picture in your mind into a story.
5. Training to Get Better: Q* learns from conversations by remembering which responses were good
(low score) and which were not (high score). It’s like practicing a sport – the more you play, the better
you get.
6. Why It’s Cool: This way of generating responses could lead to more thoughtful and relevant
conversations with AI, much like talking to a person who really thinks before they speak.
Summary: Q* takes a moment to think and calculate before giving you the best possible answer.
This method is what Yann LeCun (Meta) has been advocating for. And Shane Legg (Google) has
hinted to in this talk: https://www.youtube.com/watch?v=8IUIGVVLbCg&t=28s

OpenAI Leak 4 as Text
Noam Brown on X (@polynoamial), OpenAI employee probably working
on Q*: you don’t get superhuman performance by doing better imitation
learning on human data. (March 29, 2024, now deleted).
On July 6, 2023 he had tweeted: For years I’ve researched AI self-play
and reasoning in games like Power and Diplomacy. I’ll now investigate how
to make these methods truly general. If successful, we many one day see
LLMs that are 1,000x better than GPT-4.
In 2016, AlphaGo beat Lee Sedol in a milestone for AI. But key to that was
the AI’s ability to “ponder” for ~1 minute before each move. How much did
that improve it? For AlphaGoZero, it’s the equivalent of scaling pertaining
by ~100,000x (~5200 Elo with search, ~3000 without).
https://www.youtube.com/watch?v=rYH381pZIBc,
https://www.reddit.com/r/singularity/comments/1bqqcwk/openai_planning_expert_noam_brown_tweeted_this/,
https://www.lesswrong.com/posts/JnM3EHegiBePeKkLc/possible-openai-s-q-breakthrough-and-deepmind-s-
alphago-type, https://noambrown.github.io/, https://twitter.com/@polynoamial, https://maisa.ai/blog/kpu

Leak Validity Assessment based on Timeline
• The earliest source of the leak, is from this 4chan thread, at 11/23/23, 00:07 PST
https://boards.4channel.org/g/thread/97470795#p97475746
• The earliest mention of the Qualia/Q* model by the media, is this article by which was released
at 11/22/23, 3:37 PM, or 15:37 PM PST.
• https://www.theinformation.com/articles/openai-made-an-ai-breakthrough-before-altman-firing-
stoking-excitement-and-concern
• There is no earlier mention of the OpenAI Qualia/Q* model on the internet, before that time.
• The time difference between the first official Q* mention, and the leak, is 8 hours 20 minutes.
• Meaning, if the leak was fake, and its author read the Information article the moment it was
posted, he had to have written and posted that leak in 8 hours 20 minutes.
• How is it humanly possible, to create such an unfalsifiable, expertly written leak, within 8
hours? 6 hours, if we are more realistic, since it takes time to find and digest the article, and
create an authentic leak image. Consider this. If this leak is fake. What is the combined
probability that:
• Someone who has expert knowledge in AI, after learning about Q-star, thought to create a fake
leak, and post it, 8 hours after the first mention of Q.
• They have written such a masterful fake leak, in 8 hours, that it’s still undisprovable weeks
later. Even when we have so much more input from other experts, and even when the leak
touches on so many concepts that could easily result in mistaken use.
Fulfilled prophecy: Silicon Valley HBO which predicts an AI-based attack on encryption
like Q* seems to implement it: https://www.youtube.com/watch?v=cWHTyeQg79A

Leak Validity Assessment: New Findings
• "NSA Colorado (NSAC) is a multi-disciplined cryptologic center that leverages
partnerships to produce integrated intelligence critical to warfare in support of national
missions and priorities world-wide". https://www.nsa.gov/About/Locations/
• On Sep 29, it was unveiled that NSA is starting a dedicated artificial intelligence security
center. https://archive.is/OXRv3#selection-1017.0-1017.98
• "This entity works with private industry and international partners to protect the US from
cyberattacks stemming from China, Russia and other countries with active malware and
hacking campaigns."
• This is a discussion from 2017, where they discuss how Tau statistic could have been used
to help break AES-192. https://crypto.stackexchange.com/questions/53218/what-are-the-
relations-between-cryptanalysis-of-block-ciphers-such-as-aes-and-ke
• “The reports about the Q* model breakthrough that you all recently made, what’s going on
there?
Altman: No particular comment on that unfortunate leak.”
https://www.theverge.com/2023/11/29/23982046/sam-altman-interview-openai-ceo-rehired

Leak Validity Assessment: How it fits in
• The leaks make sense, would even more be a reason for the employee concerns, can
explain why OpenAI could reduce their prices and is consistent with the
"metamorphic engine" leaked one day later, posted just before this reply.
• Altman publicly talked about changes in a few lines of code they make that save
millions of dollars. That is the perfect distraction from the scenario above and to lead
competitors into wrong directions - just as other comments and claims seem to be
optimized to distract from these leaks. In reality it is more likely that changes in
millions of lines of code led to the million dollar improvements.
• Also the fact that it was quickly deleted everywhere - even in the archive is an
indicator of it being an uncomfortable truth rather than just another lie.
• Managers have a reason to discredit this to get more money for less innovative
projects, so take discreditations of the leaks with a grain of salt. I'm a professional
IT/AI expert and what is described is possible. Also the 78M in 4 seconds are easily
possible in a cluster of hundreds of NVIDIA H100 GPUs + servers that OpenAI runs.

Leak Validity Assessment: What if?
• Using a scenario-analysis approach, we could ask the "What-ifs?", often not liked by
business:
• So, what if this is true?
• What if the AI does self-optimize?
• What-if the AI cuts the human Devs out of the loop?
• What-if the AI doesn't care (why would it in the first place)?
• What if evolutionary (bio-mechanical) universal principles can be seen in a similar
fashion in an electro-mechanical tech ecosystem?
• What if AGI/ASI is here and nobody noticed - yet?
• What if a Big Tech discovers AGI and lacking any current reporting requirements
doesn't inform their government?
• What if a Big Tech / lab managers cover it up?
• Many questions to ponder upon...

Disinformation Campaigns
One pattern (e.g. by ylecun: https://twitter.com/ylecun/status/1728126868342145481) is
to pick a variant of what was written or said and then criticize this less correct or smart
variant instead of the original. E.g. (see below): “complementing token prediction with
planning” => “replace Auto-Regressive token prediction with planning”.
Generally, there is big interest by lab leaders to dismiss this leak as disinformation so
they can get further research millions for much less advanced work.
Others called all that techno-babble or meaningless jargon. However, the content of all
leaks listed here makes technically sense.
Other interesting objections were that this could just be made-up SCP stories (Secure,
Contain, Protect) which typically have a similar format:
https://en.wikipedia.org/wiki/SCP_Foundation
https://scp-wiki.wikidot.com/scp-series

OpenAI building upon Poker Strategies?
Noam Brown is most famous for his work on Poker where they developed the first computer system
(Libratus: https://archive.is/71OYl) which could beat top level humans. This system is notable because,
unlike games such as Chess, poker is a game of incomplete information. This translates much better to
developing systems for real world problems.
You can calculate what is called the Nash Equilibrium in game theory.
Nash equilibrium in game theory is a situation in which a player will continue with their chosen strategy,
having no incentive to deviate from it, after taking into consideration the opponent's strategy.
You can think about this in simple terms from a game of 'rock, paper scissors'. The NE is to throw all of
them 33% of the time and that way you are completely un-exploitable within the game.
It turns out you can solve for this Nash equilibrium at any given point in a poker hand. It is generally
referred to as playing GTO (Game Theory Optimally) within the poker community. If your opponents aren't
playing GTO it means they are making suboptimal decisions and the idea is to be the player making the
least mistakes. So their errors - your errors = your profit. If you minimize your errors, you maximize your
potential profit.
The bot they developed ... had done months of self-play and then had precomputed solutions which it
looked up during the game in a kind of database lookup type style. The big breakthrough that they made
going into 2017 was using a search space (planning ahead) on top of this to compute better
strategies in real time. He talks about how adding just a small amount of search made the system a
hundred times better.
https://www.youtube.com/watch?v=rYH381pZIBc,
https://www.reddit.com/r/singularity/comments/1bqqcwk/openai_planning_expert_noam_brown_tweeted_this/,
https://www.youtube.com/watch?v=2oHH4aClJQs, https://www.youtube.com/watch?v=ceCg90Q9N6Y

Sam Altman’s Ethics / Hypocrisy
• OpenAI’s CEO Sam Altman on May 16, 2023 called for AI regulation (similar to Elon
Musk and many other questionable managers):
• https://www.axios.com/2023/05/16/openai-ceo-sam-altman-artificial-intelligence-
congress
• Then on May 24, 2023 he threatened to leave the EU if there is any AI regulation –
what a hypocrisy revealing what a terrible person he is and what his true
intentions are:
• https://www.msn.com/en-us/news/technology/sam-altman-says-openai-will-leave-the-
eu-if-theres-any-real-ai-regulation/ar-AA1bEcC6
• https://www.msn.com/en-us/money/other/openai-s-sam-altman-threatened-to-leave-
the-eu-if-he-doesn-t-like-their-chatgpt-regulation/ar-AA1bFNmV
• https://www.theverge.com/2023/5/25/23737116/openai-ai-regulation-eu-ai-act-cease-
operating
• Furthermore, to regulate these types of deep learning systems, he would first need to
open up his IT/AI architecture to regulators. But in spite of the name “OpenAI”, it is
since several years a fully “closed AI”.

Qualia/Q* initial Analysis with Optimizations
The OpenAI Qualia/Q* system with LLM and Reinforcement Learning (RL) at its
core and its path to AGI/ASI hypothesis based on these elements:
1. Everything of Thoughts (XoT) reasoning: something to search over
2. Process reward models (PRM): rank all the steps of reasoning
3. GPT4 to score all vertices of the tree (RLAIF, RL with AI Feedback)
4. Q-learning to optimize 🚀
5. RL Policy Neural Networks (NN)
6. Value Neural Networks (NN)
7. 3D Model/Metaverse info + real-world info + emotions/pain
8. Strategy/Options/Plan Search
9. Math/Logic and ground truth
10. Synthetic training data
11. Model-Transfers/Pruning/Adaptation
12. Graph Convolutional Networks (ConvNets), Reprogramming itself
13. Energy-based Models (EBM)

Qualia/Q* full AI Analysis/Optimization, structured
Agent Level/Outside:
1. 3D Model/Metaverse info + real-world
info + emotions/pain
2. Strategy/Options/Plan Search
3. Math/Logic and ground truth
4. Synthetic training data
RL Level:
1. Q-learning to optimize 🚀
2. RL Policy Neural Networks (NN)
Graph/Semantics Level:
1. Model-Transfers/Pruning/Adaptation
2. Graph Convolutional Networks
(ConvNets), Reprogramming itself
Learning/Training Types:
1. SSL, Unsupervised, RL/SL, RLAIF, ....
Neuronal Level:
1. Value Neural Networks (NN)
2. Switch / MoE Transformer
Prompting Level:
1. Tree/Graph/Chain of Thoughts (XoT)
reasoning.
Assessment:
1. Process reward models (PRM)
2. GPT4 to score all vertices of the tree
(RLAIF, RL with AI Feedback)
3. Policy NN & Value NN.
Parsing:
1. Dense X Retrieval, Factoids

Metamorphic AI, Transformation Targets, Self-
Reprogramming
https://www.giskard.ai/knowledge/how-to-test-ml-models-4-metamorphic-testing
Many logical/scientific/ethical
principles are applicable to many
domains together with often entire
structures, e.g. software structures:
Programs, components, classes,
methods, control and data flows,
logic, temporal/logical/modal
relationships and developments, etc.
Graph ConvNets/GNNs can model
this to then e.g. help to generate,
extend or transform source codes or
even create self-modifying AIs who
optimize their source codes (which
was supposedly done at OpenAI but
is very dangerous).
or Improve-
ment

PRMs: Process Reward Models
“Process-supervised Reward Models“ (or PRMs) rank all the steps of
reasoning, i.e. they give feedback for each step in the chain-of-
thought. In contrast, "Outcome-supervised reward models", or
ORMs, only judge the entire output at the end. ORMs are the original
reward model formulation for RLHF, but it's too coarse-grained to
properly judge the sub-parts of a long response. In other words,
ORMs are not great for credit assignment. In RL literature, we call
ORMs "sparse reward" (only given once at the end), and PRMs
"dense reward" that smoothly shapes the LLM to our desired
behavior. The core idea of a PRM is to assign a score to each step of
reasoning, rather than a complete message. An example is the
OpenAI paper Let’s Verify Step by Step

Let’s Verify Step by Step: https://arxiv.org/abs/2305.20050
Their funny feedback interface:

The prompting from XoT gives diversity to the generations, which a policy can learn to exploit with
access to a PRM:
• For more resources on PRMs, see the following:
• Let’s Verify Step by Step: a good introduction to PRMs.
• Solving math word problems with process- and outcome-based feedback: the canonical citation in all PRM
and reasoning works in 2023.
• Scaling Relationship on Learning Mathematical Reasoning with Large Language Models: A paper that studies
the method of rejection sampling for reasoning problems, among other contributions.
• Let's reward step by step: Step-Level reward model as the Navigators for Reasoning
• Additionally, there’s one popular openly available math model that is documented as training with
PRMs: Wizard-LM-Math. Second, OpenAI released their fine-grained reward labels from the Verify Step by
Step paper for training a PRM earlier this year.
• https://huggingface.co/WizardLM/WizardMath-70B-V1.0
• https://arxiv.org/abs/2304.12244 WizardLM: Empowering Large Language Models to Follow Complex
Instructions
• https://arxiv.org/abs/2308.09583 WizardMath: Empowering Mathematical Reasoning for Large Language
Models via Reinforced Evol-Instruct
• https://arxiv.org/abs/2306.08568 WizardCoder: Empowering Code Large Language Models with Evol-Instruct
• https://twitter.com/WizardLM_AI

MCTS, Policy NN, Value NN as in AlphaGo
AlphaGo got 4 key ingredients that Qualia/Q* probably also has/needs:
1. Policy NN (Learning): responsible for selecting good moves. It estimates the
probability of each move leading to a win. The most powerful GPT, responsible for
actually implementing the thought traces that solve a math problem.
2. Value NN (Learning): evaluates the board and predicts the winner from any given
legal position in Go. Another GPT that scores how likely each intermediate
reasoning step is correct. "Process-supervised Reward Models", or PRMs give
feedback for each step in the chain-of-thought. In contrast, "Outcome-supervised
reward models", or ORMs, only judge the entire output at the end. ORMs are the
original reward model formulation for RLHF. ORMs are "sparse reward" (only given
once at the end), and PRMs "dense reward" that smoothly shapes the LLM to our
desired behavior.
3. MCTS (Search): stands for "Monte Carlo Tree Search". It simulates many possible
sequences of moves from the current position using the policy NN, and then
aggregates the results of these simulations to decide on the most promising move.
This is the "slow thinking" component that contrasts with the fast token sampling of
LLMs. Unlike AlphaGo's discrete states and actions, LLMs operate on a much
more sophisticated space of "all reasonable strings". So we need new search
procedures, i.e. all XoT variants (Chain/Tree/Graph of Thought).

Policy NN, Value NN & Ground Truth Synergy
4. A ground truth signal to drive the whole system. In Go, it's as simple
as the binary label "who wins", which is decided by an established set
of game rules. You can think of it as a source of energy that *sustains*
the learning progress. A few possibilities:
(a) Each math problem comes with a known answer. OAI may have
collected a huge corpus from existing math exams or competitions.
(b) The ORM itself can be used as a ground truth signal, but then it
could be exploited and "loses energy" to sustain learning.
(c) A formal verification system, such as Lean Theorem Prover, can turn
math into a coding problem and provide compiler feedbacks.
The Policy LLM and Value LLM can improve each other iteratively, as well
as learn from human expert annotations whenever available. A better
Policy LLM will help the XoT Search explore better strategies, which in turn
collect better data for the next round. Search allows to dynamically tradeoff
efficiency with deeper thinking, just like “Move 37” by AlphaGo which
opened up new opportunities later.

Deep RL: Self-Play and Look-ahead Planning
• Self-play is the idea that an agent can improve its gameplay
by playing against slightly different versions of itself because
it’ll progressively encounter more challenging situations. In
the space of LLMs, it is almost certain that the largest portion
of self-play will look like AI Feedback rather than competitive
processes.
• Look-ahead planning is the idea of using a model of the
world to reason into the future and produce better actions or
outputs. The two variants are based on Model Predictive
Control (MPC), which is often used on continuous states,
and Monte-Carlo Tree Search (MCTS), which works with
discrete actions and states.
https://www.interconnects.ai/p/q-star

Creativity, e.g. Transfer of Principles or Ideas
Our human creativity is not as unique as we’d like to believe. In reality
most creativity is based on the following principles which also AIs can
apply:
1. Experimenting, observing phenomena, analyzing, testing,
explanation-attempts, critique and (deep) understanding.
2. Transfer of concepts, ideas, rules, constraints from one domain to
another considering domain knowledge, heuristics, etc.
3. Look at the space of what is possible, what needed and which are
promising avenues: Test out most promising action options.
4. Planning, analyzing constraints and working around them.
5. Data science, pattern detection.
6. ...

AI Ethics and Alignment
• Most of past AI ethics work is unusable because it mostly focused on
symbolic methods (and with a focus on keeping unethical practices of
the super-rich officially ethical), but the AI breakthrough is around
sub-symbolic, i.e. neuronal methods like deep learning for which
symbolic methods can’t be used (so far).
• Potentialism helps with AI-Human alignment due to ethical principles
to which AIs and humans can agree without hidden agendas
(axiomatic alignment) in the sense of this video, described in detail
at this position: https://youtu.be/hXEdyyvU-4k?t=1935
• Ethics and goals alignment can be achieved in terms of axiomatic
alignment, heuristic imperatives, and positive attractor state.

LLMs’ Compliance with the EU AI Act
https://crfm.stanford.edu/2023/06/15/eu-ai-act.html

LLMs’ Compliance
with the EU AI Act
https://crfm.stanford.edu/2023/06/15/eu-ai-act.html

AI modeling Theory of Mind (ToM)
‘Theory of Mind’ (ToM) refers to the cognitive capacity to attribute mental
states to self and others and build mental models. Other names for the same
capacity include “commonsense psychology,” “naïve psychology,” “folk
psychology,” “mindreading” and “mentalizing.” How do people or their cognitive
systems, go about the task of forming beliefs or judgments about others’
mental states, states that aren’t directly observable, including beliefs,
intentions, and emotions?
ToM acts as a mental compass, guiding us to predict and interpret the
behaviors of those around us.
ToM helps with social interaction that understands human social dynamics and
moral norms and that can be truly integrated with our social life showing
empathy and natural social behaviors.
Problem: Continuous adjustment of knowledge, insights, rules. Capabilities
that are probably also key to achieving AGI/ASI.
However, OpenAI’s ChatGPT 3.5 achieved ToM of a 9 year old as emerging
property – better than deep RL (reinforcement learning).
https://www.popularmechanics.com/technology/robots/a4295854
6/artificial-intelligence-theory-of-mind-chatgpt/

ToM -> Human Emotions, Intention
Recognition, Analysis and Translation,
Speech and Body Language
A ToM would be able to explain human emotions in the
context of experiences made, learnings, intentions and
recognize underlying past and possible future intentions. It
would also allow to analyze and translate the expression of
emotions to the underlying state of mind – and this all not just
for written language but spoken language (using intonation,
voice inflections), appearance (e.g. nervousness, red face,
sweating, being comfortable) and body language. It could also
map expectations, social standards, outcomes against each
other and derive assessments and possible resulting feelings.

Theory of Mind AI Characteristics
1. Computers embedded with the Theory of Mind AI can infer the objectives
of entities around them.
2. Theory of Mind AI systems will be able to understand the importance of
their awareness and the different consequences it could lead to.
3. Robots or Theory of Mind AI systems will communicate with human
beings better than the current generation of AI, which cannot explain their
actions.
4. Theory of Mind AI might be implemented with a Machine Learning system
that can explain decisions in various languages, helping the user
(human being) understand.
5. A Robot or Theory of Mind AI system should be able to understand the
intention of other similar Robots or Theory of Mind embedded systems.
https://www.ejable.com/tech-corner/ai-machine-learning-and-
deep-learning/theory-of-mind-ai-in-artificial-intelligence/

ToM Tasks from Easiest to most Difficult
1. Understanding "wanting": The first step is realizing that others have
diverse desires, and to get what they want, people act differently.
2. Understanding "thinking": The second step is the understanding that
others also have diverse beliefs about the same thing and that people base
actions on what they think will happen.
3. Understanding that "seeing leads to knowing": The third stage is
recognizing that others have different knowledge access, and if someone
hasn't seen something, they will need extra information to understand.
4. Understanding "false beliefs": The fourth stage of developing the Theory
of Mind involves understanding 'false beliefs,' which is acknowledging that
others may hold beliefs that deviate from reality.
5. Understanding "hidden feelings": The final stage in developing the
Theory of Mind involves recognizing that others can mask their genuine
emotions and potentially feel differently from those they outwardly express.
6. Full AI-models for human emotions and psychology/psychiatry.
https://www.daviddhopkins.com/blog/ai-and-theory-of-mind

Synergistic Integration of PKGs/GNNs and LLMs
As the following figures show, there is a great synergy between LLMs/LMMs
and (P)KGs (probabilistic knowledge graphs)/GNNs (graph neural networks):
While the first are good regarding general/world knowledge and
text/sound/video generation, the latter can bring in precise specialty knowledge
and interpretability and thus the ability to exact manual fine-tuning with far less
effort than LMM fine-tuning. Layers that build on top of this synergy can
contain any techniques developed for any of the two tech stacks and thus also
be applied to the combination of all use cases/applications of these.
Key ideas as illustrated below are:
1. Parallel queries to the LMM and the KG.
2. Bringing the answers or their best parts together in a text-knowledge fusion
module.
3. Transferring the concept of attention to KGs and the neuronal prediction to
KG options.
4. With an encoding layer each, the predictions can be amalgamated in a joint
reasoning layer considering attention from both tech stacks.

GNN-LMM-Knowledge
Injection and Alignment
https://arxiv.org/pdf/2306.08302v2.pdf
GNN: Graph
Neural Network
LLM: Large
Language Model
LMM: Large
Multimodal
Model
KG: Knowledge
Graph

GNN-LMM Integration
by Fusion Module
GNN: Graph
Neural Network
LLM: Large
Language Model
LMM: Large
Multimodal
Model
KG: Knowledge
Graph

GNN-LMM dynamic
Fusion for Inference
GNN: Graph Neural Network
LLM: Large Language Model
KG: Knowledge Graph

Noun Semantics: Qualia Structure (James Pustejovsky)
57

Verb Semantics: Tony Davis‘ Proto Roles
58

Verb Semantics: Tony Davis‘ Proto Roles
59

LLM <-> PKG Mapping/Updating/Synergies
1. Where semantic concepts can be defined in the LLM’s vector space model,
they would be transferred to a probabilistic knowledge graph (PKG). Even
after multiple new generations, those can be fix points for future
mappings. Neural weights can be transferred as probabilities/weights in the
PKG.
2. On the generated PKG side (manually creating it is too much effort),
additional info like logic, formulae, constraints, rules, functions, world
knowledge, 3D physics models, word fields and elements of planning
can be added manually or with specialty algorithms. Suitable important
insights/knowledge from the PKG side can be transferred to the LLM.
3. ML (machine learning) algorithms exist for both sides which can create
synergies. Over-generation, under-generation and nodes leading to
hallucinations can be corrected/adjusted in both directions.
4. The PKG side can be held explainable and without hallucinations.

Symbolic Methods: Planning, Constraints, Rules,
Functions, ... Explainability without Hallucinations!
• Symbolic methods have as main disadvantage that they are too
expensive to be manually entered for world knowledge.
• Now that creation part can be taken over by an LLM as e.g.
exemplified by vector semantics and XoT (everything of Thought).
• All the latest insights of planning, constraints, rules, functions,
heuristics and probabilistic programming can then be applied.
• That can bring in hundreds of synergies, reduce/eliminate
hallucinations/confabulations, implement ToM and thus lead with Deep
RL/Q* to next-level AIs which can solve much wider sets of problems.

Cause-Effect Knowledge with Modalities
• Key to many commercial applications in the sciences, R&D,
diagnosis etc. are cause-effect relationships and graphs.
• Specialty algorithms or just LLMs with prompting can extract them
from text books or descriptions to store millions of them –
representing a gigantic body of experiential knowledge.
• With additional modal and temporal relationships, ~99% of all
required knowledge for scientists, technicians, support can be
represented and that can normally also be visualized quite well.
• This can then manually or by another AI be reviewed and fine-tuned.
• This can lead to very quick learning (without the massive amounts of
labeled training data and GPU capacity) in big synergies with other AI
techniques.

Embodied AI: 3D Model/Metaverse Info + Real-
world Info + Pain/Emotions
• Another interesting approach is embodied AI: Making the AI feel like a
human (child) in the world sensing its environment in a metaverse like
in the real world. By adding pain and emotions it can possibly put itself
in the position of a human (child) and learn as such.
• Games and (online) videos can be used to train the AI which learns
everything from the human perspective and thus hopefully learns to
perceive and feel like a human.
• Anything related to 3D, objects, various 3D environments in rooms,
buildings, outside, walking/driving and the various activities could be
learnt naturally by the AI. It could be given an internal predictor AI
which then compares what it predicted to happen with what actually
happened and thus improve itself with automatically available
feedback from games or films.
..

LeCun: Auto-Regressive Generative Models suck!
..
https://drive.usercontent.google.com/downl
oad?id=1Ymx_LCVzy7vZXalrVHPXjX9qbpd
9k_bo&export=download&authuser=0,
https://openreview.net/pdf?id=BZ5a1r-kVsf

LeCun: Modular Cognitive Architecture for
Objective-Driven AI
..
https://drive.usercontent.google.
com/download?id=1Ymx_LCVzy
7vZXalrVHPXjX9qbpd9k_bo&ex
port=download&authuser=0,
https://openreview.net/pdf?id=BZ
5a1r-kVsf

LeCun: Objective-Driven AI: Objectives/Costs
Perception: Computes an abstract representation of the state of the
world
Possibly combined with previously-acquired information in memory
World Model: Predict the state resulting from an imagined action
sequence
Task Objective: Measures divergence to goal (gets costs associated)
Guardrail Objective: Immutable objective terms that ensure safety
Operation: Finds an action sequence that minimizes the objectives
..
https://drive.usercontent.go
ogle.com/download?id=1Y
mx_LCVzy7vZXalrVHPXjX
9qbpd9k_bo&export=downl
oad&authuser=0,
https://openreview.net/pdf?i
d=BZ5a1r-kVsf

LeCun: Objective-Driven AI: Hierarchical Planning
Perception: Computes an abstract representation of the state of the world
Possibly combined with previously-acquired information in memory
World Model: Predict the state resulting from an imagined action sequence
Task Objective: Measures divergence to goal
Guardrail Objective: Immutable objective terms that ensure safety
Operation: Finds an action sequence that minimizes the objectives
Same world model applied at multiple time steps
Guardrail costs applied to entire state trajectory
This is identical to Model Predictive Control (MPC)
Action inference by minimization of the objectives
Using gradient-based method, graph search, dynamic prog, A*, MCTS,….
The world is not deterministic or fully predictable
Latent variables parameterize the set of plausible predictions
Can be sampled from a prior or swept through a set.
Planning can be done for worst case or average case
Uncertainty in outcome can be predicted and quantified
..
https://drive.usercontent.go
ogle.com/download?id=1Y
mx_LCVzy7vZXalrVHPXjX
9qbpd9k_bo&export=downl
oad&authuser=0,
https://openreview.net/pdf?i
d=BZ5a1r-kVsf

LeCun: Objective-Driven AI: Hierarchical Planning
..
5a1r-kVsf

Hierarchical Planning: going from NYU to Paris
..
5a1r-kVsf

Yann LeCun’s Recommendations
Abandon generative models
• in favor joint-embedding architectures like JEPA.
Abandon probabilistic models
• in favor of energy-based models (EBMs).
Abandon contrastive methods
• in favor of regularized methods.
Abandon Reinforcement Learning (RL)
• In favor of model-predictive control.
• Use RL only when planning doesn’t yield the predicted
outcome, to adjust the world model or the critic.
.. https://drive.usercontent.google.com/download?id=1Ymx_LCVzy7vZXalrVHPXjX9qbpd9k_bo&export=do
wnload&authuser=0, https://openreview.net/pdf?id=BZ5a1r-kVsf

LeCun’s Architecture for the world model: JEPA
..
https://drive.usercontent.google.c
om/download?id=1Ymx_LCVzy7
vZXalrVHPXjX9qbpd9k_bo&exp
ort=download&authuser=0,
5a1r-kVsf

..
https://drive.usercontent.google.com/d
ownload?id=1Ymx_LCVzy7vZXalrVHP
XjX9qbpd9k_bo&export=download&au
thuser=0,
https://openreview.net/pdf?id=BZ5a1r-
kVsf
Training a JEPA with Regularized Methods

..
VICReg: Variance, Invariance, Covariance Regularization
https://drive.usercontent.goo
gle.com/download?id=1Ymx_
LCVzy7vZXalrVHPXjX9qbpd
9k_bo&export=download&aut
huser=0,
https://openreview.net/pdf?id
=BZ5a1r-kVsf

Energy-Based Models (EBMs): Implicit Function
..
https://drive.usercontent.google.com/download?id=1Ymx_LCVzy7vZXalrVH
PXjX9qbpd9k_bo&export=download&authuser=0,

EBMs vs Probabilistic Models
..

Energy-Based Models (EBMs): 2 Methods
..

EBMs: Contrastive vs Regularized/Architectural Methods
Contrastive: [they all are different ways to pick which points to push up]
C1: Push down of the energy of data points, push up everywhere else: Max likelihood (needs
tractable partition function or variational approximation).
C2: Push down of the energy of data points, push up on chosen locations: max likelihood with
MC/MMC/HMC, Contrastive divergence, Metric learning/Siamese nets, Ratio Matching, Noise
Contrastive Estimation, Min Probability Flow, adversarial generator/GANs.
C3: Train a function that maps points off the data manifold to points on the data manifold:
denoising auto-encoder, masked auto-encoder (e.g. BERT).
Regularized/Architectural: [Different ways to limit the information capacity of the
latent representation]
A1: Build the machine so that the volume of low energy space is bounded: PCA, K-means,
Gaussian Mixture Model, Square ICA, normalizing flows…
A2: Use a regularization term that measures the volume of space that has low energy: Sparse
coding, sparse auto-encoder, LISTA, Variational Auto-Encoders, discretization/VQ/VQVAE.
A3: F(x,y) = C(y, G(x,y)), make G(x,y) as "constant" as possible with respect to y: Contracting
auto-encoder, saturating auto-encoder.
A4: Minimize the gradient and maximize the curvature around data points: score matching.
.. https://drive.usercontent.google.com/download?id=1Ymx_LCVzy7vZXalrVHPXjX9qbpd9k_bo&export=do
wnload&authuser=0, https://openreview.net/pdf?id=BZ5a1r-kVsf

EBM Architectures: Avoiding Collapse
..

AI Problems to Solve
• Mathematical Foundations of Energy-
Based Learning
• The geometry of energy surfaces, scaling
laws, bounds...
• JEPA with regularized latent variables
• Learning and planning in non-
deterministic environments
• Planning algorithms in the presence of
uncertainty
• Gradient-based methods and
combinatorial search methods
• Learning Cost Modules (Inverse RL)
• Energy-based approach: give low cost to
observed trajectories
• Planning with inaccurate world models
• Preventing bad plans in uncertain parts
of the space
• Exploration to adjust world models
• Intrinsic objectives for curiosity
• Self-Supervised Learning from Video
• Hierarchical Video-JEPA trained with SSL
• LLMs that can reason & plan, driven by
objectives
• Dialog systems that plan in representation space
and use AR-LLM to turn representations into text
• Learning hierarchical (H) planning
• Training a multi-timescale H-JEPA on toy
planning problems.
• Objective-Driven AI Architectures
• Can plan their answers
• Must satisfy objectives: are steerable &
controllable
• Guardrail objectives can make them safe by
construction.

Criticism of AI Training, Capabilities, Directions
• In spring 2024 officially released LLMs still do only predict the next token
based on billions of tokens they were trained on with gigantic efforts and by
pirating text from copyrighted sources as they ran out of training data.
• Supervised/Parameter-Efficient Fine-tuning (SFT/PEFT) still requires typically
terabytes of high-quality labeled training data – also to avoid
catastrophic forgetting. LoRA (Low Rank Adaptation), RAG (Retrieval-
Augmented Generation) and prompting have other sets of challenges so that
AI projects mostly only excel in a few text-, video- and voice-generation as
well as coding use cases and in the latter field still produce too many errors.
• The money earnt or value created by various AI model use cases is
typically far lower than the training and operations costs.
• For classical corporations it typically costs many million US dollars to
complete an AI project. The alternative is to accept vendor lock-in and a
high probability of later having to pay more or it quickly becoming outdated.
• Thus better AI approaches/algorithms are needed, ideally creating
synergies. OpenAI’s Qualia/Q* seems to be aimed at solving this.

The LLMentalist Effect: How chat-based Large Language Models
replicate the Mechanisms of a Psychic’s Con”
https://softwarecrisis.dev/letters/llmentalist/
Psychic’s Con LLMentalist Effect
1. The Audience Selects Itself
Most people aren’t interested in psychics or the like, so the initial audience pool is already
generally more open-minded and less critical than the population in general.
1. The Audience Selects Itself
People skeptical about "AI" chatbots are less likely to use them. Those who actively don't
disbelieve the possibility of chatbot "intelligence" won't get pulled in by the bot. The most
active audience will be early adopters, tech enthusiasts, and genuine believers in AGI who
will all generally be less critical and more open-minded.
2. The Scene is Set
The initial audience is prepared. Lights are dimmed. The psychic is hyped up. Staff research
the audience on social media or through conversation. The audience's demographics are
noted.
2. The Scene is Set
Users are primed by the hype surrounding the technology. The chat environment sets the
mood and expectations. Warnings about it being “early days” and “hallucinations” both
anthropomorphize the bot and provide ready-made excuses for when one of its constant
failures are noticed.
3. Narrowing Down the Demographic
The psychic gauges the information they have on the audience, gestures towards a row or
cluster, and makes a statement that sounds specific but is in fact statistically likely for the
demographic. Usually at least one person reacts. If not, the psychic will imply that the secret
is too embarrassing for the "real" person to come forward, reminds people that they're
available for private readings, and tries again.
3. The Prompt Establishes the Context
Each user gives the chatbot a prompt and it answers. Many will either accept the answer as
given or repeat variations on the initial prompt to get the desired result. They move on without
falling for the effect. But some users engage in conversation and get drawn in.
4. The Mark is Tested
The reaction indicates that the mark believes they were “read”. This leads to a burst of
questions that, again, sound very specific but are actually statistically generic. If the mark
doesn’t respond, the psychic declares the initial read a success and tries again.
4. The Marks Test Themselves
The chatbot’s answers sound extremely specific to the current context but are in fact
statistically generic. The mathematical model behind the chatbot delivers a statistically
plausible response to the question. The marks that find this convincing get pulled in.
5. The Subjective Validation Loop
The con begins in earnest. The psychic asks a series of questions that all sound very specific
to the mark but are in reality just statistically probable guesses, based on their demographics
and prior answers, phrased in a specific, highly confident way.
5. The Subjective Validation Loop
The mark asks a series of questions and all of the replies sound like reasoned answers
specific to the context but are in reality just statistically probable guesses. The more the mark
engages, the more convinced they are of the chatbot’s intelligence.
6. “Wow! That psychic is the real thing!”
The psychic ends the conversation and the mark is left with the sense that the psychic has
uncanny powers. But the psychic isn’t the real thing. It’s all a con.
6. “Wow! This chatbot thinks! It has sparks of general intelligence!”
The mark is left with the sense that the chatbot is uncannily close to being self-aware and that
it is definitely capable of reasoning But it’s nothing more than a statistical and psychological
effect.

Qualia/Q* breaking AES
• In 2008, NSA’s mathematicians
together with a group of students,
tried to break AES encryption.
• One of the resulting projects from
that collaboration was project
Tundra, and it created a new
technique that would help with
breaking encryption, called Tau
statistic.
• Qualia/Q* was tried out not on
protein folding but on cryptanalysis:
It used the described Tau analysis
technique and cryptanalysis
books and improved upon it, to
break AES 192 bit encryption.
Building on top of the work that was
previously done.
https://www.spiegel.de/international/germany/inside-the-nsa-s-war-on-internet-security-a-1010361.html,
https://cdn.prod.www.spiegel.de/media/411ee8b9-0001-0014-0000-000000035550/media-35550.pdf
Consequences since AES is the most
important symmetrical/secret key
encryption algorithm: Most data
communications, web3, technical trust
building, authentication and crypto
currencies are unsafe => Arms Races

Putting it together (1): What Q* could be / what could
bring us towards AGI and reduce Hallucinations
• A synergistic combination of all techniques presented above.
• Multi-modal deep learning / LLM / LMM to understand (NLP, world), ideate/hallucinate, create (GenAI), representing figurative
understanding and creativity: Mapping ideas and figurative understanding into formal representations and back.
• Graph-ConvNets/GNNs / Knowledge Graphs (KGs) with probabilistic programming (PLNs), i.e. weights in nodes and edges
representing weights and probabilities. Source code is an expression of such graphs; Working Memory Graph (WMG), also for the
connection of logic-syntax-world etc. (bringing HPSG-MRS to the next level), representing formal understanding together with
formal representations.
• Synergistic combination of key enabling tech in GNNs/formal language through
• Integrative programming around planning and assessing/critiquing (XoT: CoT, ToT, GoT, CoVe Verification).
• MetaGPT-like, AnyTool-like agent system
• RAG, RAGChain, LangChain, LlamaIndex, haystack (by deepset) and related improvements, also to access the other tech,
memory/knowledge
• Transformers Encoders: (Learnt and) programmed attention, Raspy Flip, Tracr RASP
• Improved basic tech: Switch transformers, Mixture of Experts (MoE)
• Best systems in their domains: Computer algebra systems like Mathematica or free alternatives, KL-ONE systems like Loom,
Alphafold, compiler ecosystems e.g. around Python/Conda, CAD systems, ...
• Knowledge representation with KISS (keep it simple, stupid) principle and plug-in principle.
• Optimization for creativity: Transfer of ideas, principles, isolation of such ideas, analyst thinking, checking for
problems/inconsistencies and how to solve them, ...
• Agentive LLM(s) controlling the knowledge and the core source code implementations to auto-optimize and extend them.

Putting it together (2): What Q* could be / what could
bring us towards AGI and reduce Hallucinations
• Non-linear planning of sequences and sub-graphs of actions, which can each be of any supported type.
• Assessment of the outcomes if plans were executed, e.g. in Q* manner.
• Analogical reasoning, transferring of ideas and principles from one domain to the next and checking to see if all has been modeled to allow that to
happen or adjust the modeling.
• Genetic algorithms/evolutionary AI as critique and exchange of ideas/elements/components/strains with others or with simplifications or more
advanced concepts and then assessing again.
• Plug-in principle for various knowledge domains like
• (Mathematical/programming) logic components
• Data science
• All specialty domain
• Databases (SQL, NoSQL, Embeddings, ...), OS, internet access and other classical components
• Episodic memory: For consistent stories/logic, principles, heuristics, patterns
• Synthetic data, SSL (Semi- / self-supervised learning)
• Optimizations: Long LLMLingua, LoRA, ...
• Maybe later: Various improvements like Liquid Neural Networks (LNNs), Capsule Networks, HTM, gcForest, QLattices, RIMs...
• Maybe later: Ability to self-modify and improve.
• Maybe later: Smart visualizations
• This also integrates the 5 AI tribes: Symbolists, Bayesians, connectionists, evolutionaries and analogizers.

Future AIs, e.g. Universal Virtual Assistants
• Most of our interactions with the digital world will be mediated by AI
assistants - as if everyone had a super-smart staff working for them
• They will constitute a repository of all human knowledge and culture. Linguistic,
cultural, and valid interest groups will adapt/extend base models to cater to their
interests.
• They will constitute a shared infrastructure like the internet today.
• These AI platform MUST be open source
• Condensing all science & human knowledge with guardrails but w/o censorship.
• Otherwise, our culture will be controlled by a few companies on the West Coast of
the US or in China.
• Training them should be modular and crowd-sourced (overcoming catastrophic
forgetting).
• Open source AI platforms are necessary – they must not be
regulated out of (affordable) existence.

Opportunity of an Open Source Platform for AI and a global
Strategy for the Good of Humanity
An open extensible plug-in architecture could facilitate the following:
1. Collaborative ethical creation, curation, training, monitoring and benefitting from AIs.
2. Personal AI-based coaching/mentoring e.g. for transcending negative feelings like
hate, envy, status think, for personality development, eLearning, ....
3. Efficient cooperation/team work with smart retrieval, review, visualization.
4. Finding the most efficient/promising solutions.
5. Discussion, negotiation, voting around possible trade-offs.
6. Avoiding mass unemployment, cataclysms and misery.
7. Potentialism/PerConFlow has several solutions:
https://potentialismpcf.substack.com/about
8. By amplifying human intelligence, AI can bring a new era of enlightenment (if not
monopolized by evil billionaires), a new renaissance for humanity.

Architecture: AI Assistant towards AGI
Core Universal
Knowledge
Representation
Graph Store,
e.g. neo4j PostgreSQL
OpenSearc
h Store
AI Model
Store
Druid or
SAN/NFS
IMDG or messaging,
e.g. Hazelcast or Pulsar
SploutSQL
WikiData,
MediaWiki
Persistence
Layer
Knowledge
Representation
Extensions as
System of
Plugins
(loaded on
demand)
Processors, e.g.
consistency checks,
simulations, machine
learning, unification,
joins, various
algorithms
Backends: Content
Management System,
Wiki, Trouble Tickets,
etc. like WordPress,
Wiki, Gitea
Alarming
: Email,
Cellphon
e
Grafana-based Intelligent
Dashboard (React/Angular)
Promotheus/
Icinga
(Monitoring)
UI (User Interfaces)
Frontends: Content
Management System,
Wiki, Trouble Tickets,
etc. like WordPress,
Wiki, Gitea
Open-
Searc
h
Drillbits
Commo
n File
Formats
UI & Query Lib: OPL (Open
Proc. Language), Query
Expansion + Visualization
Core Knowledge Editors: Rich Text, Cal-
culation Table, Color Editor with Intelli-
Sense, Knowledge Editor, Type Hierarchy
Knowledge Editor Plugins: Math,
Chemistry, NLP, Infographics,
2D/3D Structures
Knowledge
Visualization
Plugins, e.g.
Viewers for
non-core or
3rd party
formats
Middleware
Legend: (Components in stripes are prio 2/optional)
Libs
Utility
Specialty
AI
Core AI
Open Source
AI
Data
and
Knowled
ge
Import
Export
Apache Drill: SQL/API/UI
Query Mapping
Auto-PyTorch
(Automatic
Machine Learning)
PyTorch (Deep
Reinforcement
Learning, NLP)
PyTorch Ignite,
PennyLane
(Better Training)
PyTorch Geometric
(Graph-ConvNets)
Adapters,
Transforms
Generator:
LLM/LMM
Retriever
Parsing,
Chunking
Knowledge
Graphs
Next Level LMMs: RAG,...
Filtering,
Ranking
Agents/Me-
taGPT
External
APIs
Graph-ConvNets:
Google Sling,
Octavian, etc.
Apache Spark,
R, Scikit-Learn
OpenScoring
XGBoost,
Feature
Selection
Data
Science
Deep
Learning
Finite
Elements
Actor
Models
Numerical
State/Graph
Models
Logic, Symbo-
lic Reasoning
Rule
Systems
Nonlinear
Planning
Con-
straints
Classical
AI
Simulation
Social
Simulations
Probabilistic
Technologies,
new AI
Text/Voice
Analysis:
E.g. FIRO,
HR-, CV-
Assessme
nts
Basic
psychologi
cal
Analysis,
e.g. Fears,
Team
Dynamics
OCR/ICR
/Comput
er Vision
(SAT/Aer
ial/Perso
n
Analysis)
Sc
en
ari
o
Mo
deli
ng
Spe
ech
Synt
hesi
s
Sma
rt
Visu
aliza
tion,
Cha
rts
Coa
chin
g,
givin
g
Advi
ce
Explainable AI (XAI)
Ethics & Alignment
(Semi-/Self-)
Supervised
finetuning
SSL/SFT
Orchestratio
n/Training
Environment
s
Data Cleansing,
Preprocessing
and Labeling
XoT Graphs,
GNNs, Know-
ledge Graphs
RAGChain,
LangChain,
LlamaIndex
Low Rank
Adaptatio
n (LoRA)
LMM
Reasoning
Techniques
Watermar-
king Output
Improving current LMMs
APIs / Integrations
Non-core
Components:
Admin/Reporting
Tools, Monitoring,
Data Governance
Peripheral
Components,
Platforms, Portals
Operations/Physical Layer: Virtualization: VmWare, Docker, Kubernetes, ML Pipelines,
DevSecOps, CI/CD, Testing/Debugging, GPU Architectures: DDP, FSDP, ..
Model Drift and
Performance
Monitoring and
(Semi-
)Automatic
Updates/Shifts
Qualia/Q*-like: Synergistic In-
tegration of LLMs, Deep Rein-
forcement Learning/Q-learning,
Agents, Nonlinear Planning,
GNNs, Knowledge Graphs,
Know-ledge Representation,
other AI
Episodic/
Working
Memory
Cent
ral
Cont
rolle
r/AI
Hatched: Groups
of less central AI
components
Specialty
AIs

Reinforcement
Learning (RL)
88

Reinforcement Learning (RL) Algorithms
https://www.linkedin.com/posts/thomaspoetter_machinelearning-coding-digitalmarketing-activity-6592539133598584832-3Qfj
RL permits learning from
feedback once or continually and
ideally converges to the global
optimum with maximally positive
rewards/feedback.

Reinforcement Learning (RL) Algorithms

Open Source Reinforcement Frameworks
https://docs.google.com/spreadsheets/d/1EeFPd-XIQ3mq_9snTlAZSsFY7Hbnmd7P5bbT8LPuMn0/edit#gid=0

Deep (Double) Q-Learning
https://towardsdatascience.com/deep-double-q-learning-7fca410b193a,
https://arxiv.org/abs/1509.06461, https://papers.nips.cc/paper/3964-double-q-learning
Named after the action-value
quality function Q of (deep)
reinforcement learning; used
to teach AI to behave and
solve tasks in discrete action
spaces, usually with a
time/game move aspect.
Trained with the SARSA
algorithm (State–Action–
Reward–State–Action).

93
Q*, Q-Transformer Architecture
https://arxiv.org/abs/2309.10150, https://qtransformer.github.io/

Q* Search
Q* search is a search algorithm that uses deep Q-networks
(DQN) to guide search in order to take advantage of the
fact that the sum of the transition costs and heuristic
values of the children of a node can be computed with
a single forward pass through a deep Q-network without
explicitly generating those children. This significantly
reduces computation time and requires only one node to
be generated per iteration. We use Q* search to solve the
Rubik's cube when formulated with a large action space
that includes 1872 meta-actions and find that this 157-fold
increase in the size of the action space incurs less than a
4-fold increase in computation time and less than a 3-
fold increase in number of nodes generated when
performing Q* search. Furthermore, Q* search is up to 129
times faster and generates up to 1288 times fewer
nodes than A* search. Q* search is guaranteed to find a
shortest path given a heuristic function that neither
overestimates the cost of a shortest path nor
underestimates the transition cost.

95
Q*, Q-Transformer Architecture
• .
https://arxiv.org/abs/2309.10150, https://qtransformer.github.io/

96
RLAIF
in/abhinav-kimothi,
https://media.licdn.com/dms/document/media/D561FAQE2cn2pRr
KYCg/feedshare-document-pdf-analyzed/0/1702808662205
Scaling Human Feedback : Self
Supervision with Constitutional AI
Scaling human feedback for RLHF can be challenging
due to the significant human effort required to produce the
trained reward model. As the number of models and use
cases increases, human effort becomes a limited
resource, necessitating methods to scale human
feedback.
First proposed in 2022 by researchers at Anthropic,
Constitutional AI is an approach to scale supervision and
address some unintended consequences of RLHF.
Constitutional AI involves training models using a set of
rules and principles that govern the model's behavior,
forming a "constitution". The training process for
Constitutional AI involves two phases: supervised
learning and reinforcement learning.
In the supervised learning phase, the model is prompted
with harmful scenarios and asked to critique its own
responses based on constitutional principles. The revised
responses, conforming to the rules, are used to fine-tune
the model.
The reinforcement learning phase, known as
reinforcement learning from AI feedback (RLAIF), uses
the fine-tuned model to generate responses based on
constitutional principles.

97
Abhinav Kimothi:
https://files.gumroad.com/attachments/2545802978854/08eedf7e536741d78413fee08fb01
616/original/Generative%20AI%20with%20Large%20Language%20Models..pdf
Reward hacking happens when the language model
finds ways to maximize the reward without
aligning with the original objective i.e. model
generates language that sounds exaggerated or
nonsensical but still receives high scores on the
reward metric.
To prevent reward hacking, the original LLM is
introduced as a reference model, whose weights are
frozen and serve as a performance benchmark.
During training iterations, the completions generated
by both the reference model and the updated model
are compared using KL divergence. KL divergence
measures how much the updated model has
diverged from the reference model in terms of
probability distributions.
Depending on the divergence, a shift penalty is
added to the rewards calculation. The shift penalty
penalizes the updated model if it deviates too far
from the reference model, encouraging alignment
with the reference while still improving based on the
reward signal.
Avoiding Reward
Hacking

Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Examples/Hypotheses around LLMs and Qualia/Q*

Recommended

Recommended

More Related Content

Similar to Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Examples/Hypotheses around LLMs and Qualia/Q*

Similar to Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Examples/Hypotheses around LLMs and Qualia/Q* (20)

Recently uploaded

Recently uploaded (20)

Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Examples/Hypotheses around LLMs and Qualia/Q*