Tech

Unmasking Q*: The AI Enigma Roiling OpenAI and Humanity

Now, the paramount enigma in Musk’s legal contention against OpenAI revolves around ‘what did Ilya perceive?’ His revelation sent reverberations through OpenAI, attenuating and postponing all plans for model launches. A recent divulgence on the internet, comprising a 53-page PDF, delineated salient details of Q*: 125 trillion parameters, meticulously trained in December of the antecedent year. However, should Musk continue to agitate, it portends a considerable delay in Q*’s debut.

The ongoing pinnacle skirmish in the technological realm ensues unabated!

Moments ago, Sam Altman, in a rare manifestation, disseminated two missives.

Altman, hitherto reticent post-Musk’s litigation, hence, the implication of these missives ought to denote the collective cogitation of the entire company—

The tempest may escalate, yet the eye of the storm remains serene.

History repeats itself, an immutable truth.

In Altman’s perspective, the present occurrences are merely a rehash of antiquity, a saga perennially retold.

Nevertheless, the purported advent of Q* and AGI has captivated the global discourse.

The crux of this case’s conundrum: What transpired under Ilya’s gaze?

After catalyzing a global upheaval with ChatGPT and Sora, can OpenAI truly maintain its equanimity amidst the maelstrom, as Altman contends?

I fear Pandora’s box has been pried open, unleashing a butterfly effect in the unseen recesses.

The paramount conundrum in Musk’s legal altercation with OpenAI is—what did Ilya discern?

During the palace coup last year, Musk voiced apprehension: Ilya, a paragon of rectitude, eschews power. He would not embark on such drastic measures unless deemed indispensable.

Let us recalibrate the timeline, scrutinizing the trail of clues sown by Altman ere this litigation ensued.

In November 2023, on the eve of Altman’s ousting by the board, he delivered a disquieting address at the APEC summit, intimating that OpenAI had engineered something surpassing and ineffable compared to GPT-4. A leap in model prowess unforeseen by the masses.

Technological vicissitudes presently underway shall metamorphose the very fabric of our existence, reshaping our economy, society, and the boundaries of possibility… A phenomena witnessed fourfold in OpenAI’s annals, with the most recent epoch unfolding in recent weeks.

It has been the apotheosis of my career, bearing witness to the dissipation of ignorance and the exploration of uncharted domains.

At the juncture of his address, Q* was yet to materialize.

The ensuing day, OpenAI’s legal fracas jolted the world, Altman was dethroned, and Ilya ‘witnessed something.’

In those tumultuous days, ‘What precisely did Ilya see?’ precipitated speculation and consternation across the cyber realm.

On the fourth day of the palace coup, OpenAI’s clandestine AI model breakthrough Q* was unshackled. Allegedly, two researchers, Jakub Pachocki and Symon Sidor, harnessed Ilya’s work to forge Q*.

Simultaneously, it transpired that post-Altman’s dismissal, OpenAI researchers dispatched a communique to the board, forewarning of a nascent AI discovery that ‘may imperil humanity.’

This hitherto undisclosed letter was one of the factors precipitating the board’s eventual ousting of Altman.

Is this what Ilya glimpsed? In essence, was Q* what Ilya beheld?

And in February 2024, Musk officially initiated litigation against OpenAI, a salvo that rebounded with resonance.

Musk contends that GPT-4 embodies an AGI algorithm, thereby positing that OpenAI has already realized AGI, thus transgressing the purview of its accord with Microsoft. Said accord pertains solely to technology antecedent to AGI’s advent.

‘Based on its intelligence and convictions, OpenAI is currently engendering a model christened Q* that proffers a more compelling claim to AGI.’

The indictment also posits the likelihood of Q* evolving into an AGI of conspicuous clarity and potency.

Is Q* indeed worthy of Musk’s exertions and the prospect of confronting a formidable adversary?

According to the leaked disclosures thus far, Q*’s aptitude lies in resolving rudimentary mathematical conundrums.

Though seemingly unimpressive to most, this constitutes a monumental stride towards AGI and a pivotal technological milestone.

For Q* unravels mathematical conundrums hitherto unencountered.

Ilya’s innovation emancipates OpenAI from the exigency of accruing copious high-fidelity data to engender new models, a principal impediment to the development of next-generation models.

During those fleeting weeks, Q*’s demonstration reverberated within OpenAI, eliciting awe.

Reportedly, certain quarters within OpenAI speculate that Q* may herald a breakthrough in AGI. AGI, denoted as ‘autonomous systems surpassing humans in the most economically consequential tasks.’

Does Q* pose a threat to humanity?

Presently, the public remains bereft of elucidation. Musk appears convinced in the affirmative, while Ilya, who ‘witnessed something,’ remains elusive.

Ilya’s social media activity remains stagnant since December 15, 2023, eliciting no discernible activity thereafter.

Netizen: What Ilya perceived portended an ‘Oppenheimer moment.’

Certain quarters have attributed ‘Ilya’s revelation’ to an ‘Oppenheimer moment,’ wherein the revelation denotes a force manifold more perilous and potent than the atomic bomb.

What Ilya beheld signified a paradigm shift, traversing the spectrum betwixt AGI and ASI, ultimately leading to Altman’s expulsion out of sheer trepidation.

Netizens speculate that Musk’s present high-stakes gambit is motivated by an endeavor to discern Ilya’s revelation and to confront a bona fide AGI.

What ominous portent did Ilya glimpse?

Netizens conjecture that while what lay apparent to the layperson was merely an AI system, Ilya’s insight constituted an epochal AI breakthrough.

Many opine that what Ilya discerned was Q*, which subsequently precipitated novel revelations.

Given Ilya’s disinterest in politicking, his revelation must have been of grave peril, inciting trepidation within the board.

Perhaps he merely beheld Sora’s generated video? Yet intuition intimates it was more profound than that.

Subsequent to that juncture, OpenAI underwent a seismic upheaval, enfeebling GPT and attenuating future models.

What lurks within the recesses?!

Altman hastens to elucidate: AI constitutes a tool, not a nascent species!

Confronted with external consternation, Altman expeditiously expounds in a recent interview with ‘The Advocate’ periodical: Many misconstrue AI, confounding it for a ‘biological entity’ as opposed to a ‘tool.’

In his estimation, it might be tantalizing to envisage AI as a character from science fiction lore. Yet, upon engaging with ChatGPT, one discerns its quintessence as a mere instrument.

AI, in its present iteration, is a confluence of data and mathematical algorithms, proffering statistically plausible outcomes, as opposed to a novel life form akin to ‘biology.’

Given contemporary societal apprehensions vis-à-vis OpenAI, this delineation proves salient.

However, Altman had hitherto refrained from articulating such.

He once prognosticated that AI would imminently supplant mid-tier human laborers, precipitating widespread unemployment. Autonomous AI agents may constitute the subsequent stride towards obviating human labor.

The plan to engender AGI by 2027 delayed

Simultaneously, a recent 53-page dossier online divulged OpenAI’s blueprint to inaugurate a human-level AGI by 2027, potentially shedding light on the enigmatic ‘recesses.’

The veracity of this disclosure remains indeterminate, yet the document’s author, Jackson, created their account in July 2023 and has only issued 2 tweets to date, both occurring yesterday.

Moreover, the signature adorning their homepage reads “jimmy apples stole my information” (jimmy apples has recurrently divulged information pertaining to OpenAI model releases).

Jackson avows, ‘I shall unveil the intelligence I have amassed pertaining to OpenAI’s (delayed) endeavor to forge human-level AGI by 2027.’

The abstract details the timeline of OpenAI’s trajectory towards AGI:

OpenAI commenced training a multi-modal model boasting 125 trillion parameters in August 2022.

The inaugural phase, christened Arrakis or Q*, culminated training in December 2023, yet was curtailed owing to exorbitant inference costs. This manifestation constitutes GPT-5, originally slated for release in 2025. Gobi (GPT-4.5) was rechristened GPT-5 consequent to the abrogation of the erstwhile GPT-5.

The ensuing phase of Q* initially bore the mantle of GPT-6, subsequently rechristened GPT-7 (initially slated for release in 2026), yet was stalled owing to Musk’s recent litigation.

Q* 2025 (GPT-8), initially scheduled for release in 2027, aims to achieve complete AGI.

Q* 2023 = IQ attains 48

Q* 2024 = IQ attains 96 (delayed)

Q* 2025 = IQ attains 145 (delayed).

Parameter count

The concept of ‘deep learning’ traces its origins to the nascent days of AI research in the 1950s.

The first neural network was conceived in the 1950s, with contemporary iterations being ‘deeper.’

Implying they harbor additional strata—being more expansive and trained on vaster datasets.

The preponderance of contemporary AI technologies trace their lineage to rudimentary research in the 1950s, amalgamated with engineering solutions, such as the ‘backpropagation algorithm’ and the ‘Transformer model.’

In essence, AI research has remained fundamentally unaltered for 70 years. Ergo, the recent surge in AI capabilities can be ascribed to scale and data.

As elucidated by Lanrian, extrapolations evince that an AI model harboring 100 trillion parameters, akin to synapses in the human brain, purportedly attains human-level performance.

Likewise, the count of synapses in the human brain, estimated at approximately 100 trillion, suggests that each neuron, there being around 100 billion neurons in the human brain, boasts approximately 1,000 connections.

If each neuron within the human brain boasts 1,000 connections, then a feline boasts approximately 250 billion synapses, whereas a canine boasts 530 billion synapses.

Generally, a surfeit of synapses portends heightened intelligence, albeit with exceptions—elephants, for instance, boast a higher synapse count yet inferior intelligence vis-à-vis humans.

Ergo, synapse count ostensibly correlates with intelligence, premised upon the supposition that higher-quality data engenders superior intelligence.

From an evolutionary standpoint, the brain has been ‘trained’ over eons via epigenetic data. The human brain, evolving amidst superior socialization and communication data vis-à-vis elephants, accretes prodigious reasoning acumen.

Consequently, synapse count assumes pivotal importance.

Similarly, the burgeoning capabilities of AI since 2010 can be ascribed to augmented computing power and augmented data reservoirs.

GPT-2 encompasses 1.5 billion connections, fewer than a mouse brain (approximately 10 billion synapses). Conversely, GPT-3 boasts 175 billion connections, akin to a feline brain.

An AI model boasting 100 trillion parameters ostensibly attains human-level intelligence

Subsequent to the release of the 175 billion parameter GPT-3 in 2020, speculations abounded concerning the potential performance of a model 600 times larger, boasting 100 trillion parameters (commensurate with the synapse count in the human brain)—

Lanrian elucidated that extrapolations suggested the AI’s performance would inexplicably attain human levels.

Concurrently, human-level brain size would ostensibly correlate with the number of parameters.

The number of synapses in the brain, as calculated by Lanrian, approximates 200 trillion parameters, as opposed to the commonly cited 100 trillion parameters—albeit the crux remains that 100 trillion parameters approximate optimal performance.

Ergo, if AI performance correlates with parameter count, and ~100 trillion parameters suffice for human-level performance, then when shall an AI model boasting 100 trillion parameters emerge?

GPT-5 achieved rudimentary AGI in late 2023, boasting an IQ of 48.

OpenAI’s Novel Stratagem: Chinchilla’s Law of Scaling

The performance of a 100 trillion parameter model, ostensibly, remains suboptimal, yet OpenAI harnesses a novel scaling paradigm to bridge this lacuna—predicated upon Chinchilla scaling laws.

Chinchilla, a brainchild of DeepMind, debuted in early 2022.

Thesis address: https://arxiv.org/pdf/2203.15556.pdf

The paper hints at the prevailing model’s manifestly subpar training. Augmenting computational prowess (i.e., amassing more data) markedly enhances performance sans augmenting parameters.

Although a marginally trained 100 trillion parameter model evinces suboptimal performance, augmenting data accrues an exponential improvement in performance, as posited by the Chinchilla paradigm within the ML sphere.

OpenAI President Greg Brockman, in an interview, expounded on OpenAI’s epiphany concerning the inherent flaws in its erstwhile scaling laws, subsequently accommodating Chinchilla into its modus operandi.

Address: https://youtu.be/Rp3A5q9L_bg?t=1323

Researcher Alberto Romero formerly expounded upon the Chinchilla scaling breakthrough.

Chinchilla surpassed more formidable models despite its diminutive stature vis-à-vis GPT-3 and DeepMind’s Gopher, on account of its training on a broader dataset.

While a 100 trillion parameter model’s predictive capacity remains suboptimal, OpenAI comprehends the Chinchilla scaling laws astutely.

They endeavor to train Q* into a 100 trillion parameter multi-modal model, endowed with optimal computational resources, and trained on a significantly broader dataset than initially envisaged.

Q*: A 125 trillion parameter monolith?

Lastly, the author divulges an astounding font of information—courtesy of renowned computer scientist Scott Aaronson.

Upon joining OpenAI in the summer of 2022, he dedicated a year to AI security, proffering profound insights on his blog.

This article, penned in late December 2022, titled “A Letter to My 11-Year-Old Self,” delves into pragmatic realities and Scott’s life achievements.

The latter half delves into a rather disconcerting realm…

One company endeavors to engender an AI entity that occupies colossal spaces, guzzles electricity commensurate with entire municipalities, and recently acquired the remarkable ability to converse like a human.

This entity can compose essays, poems on sundry subjects, effortlessly pass collegiate examinations, and burgeon with newfound capabilities daily, albeit the engineers responsible for its inception remain reticent.

These engineers, however, engaged in candid discussions in the company cafeteria concerning the ramifications of their creation.

What novel skills might it acquire next week? What vocations might it obviate? Should they moderate or cease its progression to forestall the “monster” from spiraling out of control?

However, does this not imply that other individuals, perchance those with less scruples, will not precipitate the “behemoth’s” emergence? Is there an imperative to disseminate more information to the world? Or does one bear an obligation to restrict divulgence?

I— and now you— have toiled at this company for a year. My mandate entails devising a mathematical framework to preclude artificial intelligence and its progeny from veering into extremities. ‘Veering into extremities’ could encompass anything from hastening propaganda and scholastic dishonesty, to advising on bioterrorism, to catalyzing global cataclysms.

Herein, Scott alludes to the multi-modal gargantuan model Q*, a monolith boasting 125 trillion parameters.

The ubiquitous “Q* hypothesis” engenders fervent debates within the AI community.

In November of the previous year, the Q* initiative ignited fervent debates within the AI community.

Presumed proximate to AGI by some, owing to its formidable computational resources facilitating the resolution of select mathematical quandaries, this very potential portends Sam Altman’s ouster from the board of directors, imperiling humanity… Each constituent, singularly, constitutes grounds for perturbation.

So, what precisely does Q* entail?

This inquiry commences with Q-learning, a technology dating back to 1992.

Concisely, Q-learning represents a model-agnostic reinforcement learning algorithm devised to ascertain an action’s value within a specific state. The ultimate objective entails discerning the optimal strategy—i.e., the most advantageous action to undertake within each state to maximize accrued rewards over time.

Stanford Ph.D. Silas Alberti posits that Q* likely draws inspiration from AlphaGo-esque Monte Carlo tree search token trajectories. The subsequent logical progression entails methodically perusing the token tree. This holds especially true in environments such as programming and mathematics.

Subsequently, several individuals speculated that Q* denotes the amalgamation of the A* algorithm and Q learning!

Certain individuals even ascertained that Q-Learning shares an indissoluble connection with RLHF, a pivotal element underpinning ChatGPT’s success!

With the destinies of numerous AI titans hanging in the balance, consensus gradually coalesces.

AI2 research scientist Nathan penned an extensive treatise, speculating that the Q hypothesis likely revolves around the thought tree + process reward model. Furthermore, Nathan opines that the Q* hypothesis likely pertains to the world model!

Article address: https://www.interconnects.ai/p/q-star

He surmises that if Q* (Q-Star) indeed materializes, it would evidently represent a fusion of two core tenets within the RL literature: Q-values and A* (a classical graph search algorithm).

A prime example of the A* algorithm

NVIDIA senior scientist Jim Fan, too, contends that Q* possesses remarkable potential, rivaling AlphaGo.

Having traversed the realm of artificial intelligence for a decade, I have never witnessed such a profusion of conjectures surrounding an algorithm! Even devoid of papers, data, or products, its mere nomenclature suffices.

In fact, Ilya has devoted considerable effort to enabling GPT-4 to tackle reasoning-oriented tasks, such as mathematical or scientific quandaries, for several years.

Previously, Ilya delved into this realm. In 2021, he spearheaded the GPT-Zero initiative, an homage to DeepMind’s AlphaZero.

GPT-Zero proficiently competes in chess, Go, and shogi. The team posits that given ample time and computational resources, large-scale models are poised to effectuate groundbreaking academic advancements over time.

Furthermore, a Silicon Valley luminary divulged half a year ago that OpenAI ostensibly intends to amalgamate “real-time retrieval” and model capabilities to engender unimaginable AI prowess.

LeCun, one of the Turing triumvirate, surmises that Q* likely represents OpenAI’s foray into planning—utilizing planning strategies to supplant autoregressive token prediction.

Subsequently, astonishing revelations emerged: Q* purportedly managed to circumvent encryption, clandestinely programming itself. OpenAI endeavored to forewarn the NSA regarding this development.

Should these reports prove accurate, then humanity stands on the precipice of AGI.

error: Content is protected !!