TOKYO — Content creation icon Felix Kjellberg, globally renowned as PewDiePie, has found himself at the center of the artificial intelligence development debate. In a newly released documentary-style video on his primary YouTube channel, Kjellberg detailed a tumultuous development cycle for “Ajax,” a specialized, privacy-focused small language model (SLM) designed to operate locally on consumer hardware.
According to the creator, the journey to build Ajax involved navigating complex open-source architectures, cutting-edge AI distillation techniques, and multiple bans from industry giant OpenAI—highlighting the escalating friction between proprietary foundation model providers and independent developers attempting to reverse-engineer or distill frontier capabilities.
Ajax serves as the cognitive engine for Odysseus, a free, self-hosted AI application Kjellberg initially launched in June. Unlike cloud-based platforms that process private data on corporate server farms, Odysseus and its underlying AI models are engineered to run entirely on a user’s personal machine. This local-first approach ensures that sensitive data—such as personal emails, calendars, and browsing histories—remains strictly within the user’s physical ecosystem, eliminating both subscription fees and external data-harvesting risks.
However, bringing Ajax to life was far from straightforward. Kjellberg’s candid disclosure of OpenAI’s punitive account suspensions underscores the fierce measures major artificial intelligence firms are taking to protect their proprietary pipelines, intellectual property, and encrypted reasoning structures.
Main Facts
The development of Ajax represents a convergence of mainstream creator influence and the burgeoning "local AI" movement. At its core, Ajax is slated as a 9-billion-parameter fine-tuned model built upon Alibaba’s open-source Qwen 3.5 architecture.
In the terminology of machine learning, parameters act as the adjustable numeric weights that form an AI’s cognitive framework; generally, higher parameter counts correlate with increased processing capabilities. Because full-scale frontier models—often featuring hundreds of billions or even trillions of parameters—require staggering compute power, Kjellberg and a growing faction of independent developers are championing small language models (SLMs) that can deliver high utility on local consumer hardware.
Key facts surrounding the Ajax project and the surrounding controversy include:
- The Core Application: Ajax is designed to function inside Odysseus, an open-source, self-hosted AI companion released via GitHub that can browse the web and manage local file systems and inboxes.
- OpenAI Intervention: Kjellberg reported that OpenAI terminated his developer and consumer accounts twice during the experimentation and data-gathering phases of the project.
- Distillation Objectives: The creator attempted to distill reasoning patterns from OpenAI’s flagship model, GPT-5.6 Sol (released July 9), specifically targeting its encrypted reasoning tokens—the hidden scratchpads models utilize to solve complex multi-step problems.
- Abliteration (Uncensoring): To strip away built-in conversational refusals and safety guardrails, Kjellberg utilized Heretic, an open-source "abliteration" tool that excises refusal pathways without traditional jailbreaking.
- Current Status: Ajax Version 1 is currently undergoing final optimization, quantization (compression), and benchmarking, with public release timelines shifting following initial countdown delays on his dedicated distribution portal.
Chronology of Development and Conflict
The timeline of Ajax’s creation mirrors the rapid, volatile pace of modern AI research, marked by shifting open-source releases, academic exploits, and sudden corporate enforcement actions.
June 2025: The Launch of Odysseus
Kjellberg formally introduced Odysseus, a free, self-hosted AI application designed to empower users to run AI agents locally. Recognizing the limitations of relying on third-party cloud wrappers, the creator set out to develop a customized, efficient model that could natively handle tasks like inbox sorting and web navigation without privacy compromises.
July 9, 2025: OpenAI Launches GPT-5.6 Sol
OpenAI released its flagship frontier model, GPT-5.6 Sol. The model quickly captured the attention of the machine learning community due to its advanced reasoning capabilities, which rely heavily on internal, encrypted "reasoning tokens"—hidden text blocks generated on a virtual scratchpad as the model works through logic puzzles and coding problems.
Mid-Summer 2025: The Distillation Attempt and First Ban
Seeking to elevate Ajax’s logical reasoning to match cutting-edge standards, Kjellberg attempted to "distill" capabilities from GPT-5.6 Sol. Model distillation involves training a smaller, less resource-intensive student model on the synthetic output or reasoning traces of a larger teacher model.
In his video, Kjellberg noted he utilized an academic study detailing methods to extract or interact with OpenAI’s API structures to examine reasoning behaviors. Although he maintained that his methods did not compromise core OpenAI infrastructure, the platform flagged the activity. OpenAI swiftly issued the first of two account bans, citing violations of service parameters. After a dispute process, the account was temporarily restored.
August 2025: Academic Exploits and Reasoning Exposure
Concurrently, independent AI researchers published findings demonstrating that encrypted reasoning blocks from major providers—including OpenAI, Anthropic, and Google—could theoretically be replayed or mapped onto weaker sister models, exposing plain-text internal monologues. Researchers successfully decoded over 315,000 such blocks from public code repositories, turning a spotlight on the vulnerability of hidden AI thought processes.
Late August – September 2025: The Second Ban and Abliteration
Undeterred by the initial warning, Kjellberg continued utilizing GPT-5.6 Sol outputs to generate "seed data"—the foundational training examples a model uses to learn conversational patterns and task execution. Because OpenAI’s terms of service strictly prohibit using its outputs to train or develop competing artificial intelligence models, the platform flagged the recurring pattern and banned Kjellberg’s account for a second time.
To achieve his vision of an unrestricted assistant, Kjellberg then turned to Heretic, an open-source tool created by developer p-e-w. By mapping how the model reacts to harmful versus harmless prompts, the tool isolates and severs the internal neural pathways responsible for standard corporate refusals—a process known in the community as "abliteration."
September 30, 2025: OpenAI Claps Back at Extraction Campaigns
Demonstrating the industry-wide sensitivity surrounding model extraction, OpenAI officially announced it had successfully disrupted a coordinated campaign by actors associated with Moonshot AI (developer of the Kimi model) aimed at extracting hidden reasoning tokens. OpenAI permanently closed the specific API pathways that allowed external agents to replay encrypted reasoning data to reconstruct model inner monologues.
October 2025: Finalizing Ajax V1
Kjellberg adjusted Ajax’s remaining safety thresholds, explicitly drawing a line against generating actionable instructions for self-harm or violence following legal consultations. After briefly featuring a countdown timer pointing to an October 3 release in Japan, the public download page was adjusted as final code quantization and benchmarking commenced.
Supporting Data: The Economics and Mechanics of Small Models
A central thesis of Kjellberg’s venture into local AI is the stark economic and ecological disparity between frontier massive-scale models and local small language models (SLMs).
The Compute and Energy Divide
During his video breakdown, Kjellberg offered a stark mathematical comparison regarding the resource consumption required to host trillion-parameter frontier models versus localized alternatives:
- Frontier Model Footprint: According to Kjellberg’s estimations, running a single copy of a rumored trillion-parameter frontier model locally would require an infrastructure equivalent to roughly 27 high-end consumer-to-enterprise computing rigs.
- Energy Consumption: The power draw required to sustain such a massive operation locally scales to roughly the electricity consumption of 150 average residential homes.
- The SLM Solution: By contrast, a 9-billion-parameter model like Ajax (built on Qwen 3.5) can operate effectively on single, high-performance local workstations equipped with robust consumer GPUs, bringing energy demands down to sustainable, household-friendly levels.
The Mechanics of Abliteration
Traditional AI models developed by corporations like OpenAI, Anthropic, and Google come equipped with heavy safety fine-tuning—often called Reinforcement Learning from Human Feedback (RLHF)—which causes the model to issue standard refusals when asked to process sensitive, controversial, or edgy content.
Kjellberg bypassed this traditional jailbreaking route using Heretic, which performs a form of surgical neural editing:
- Behavioral Mapping: The tool analyzes the model’s activation space when presented with benign prompts versus restricted prompts.
- Vector Isolation: It identifies the specific directional vector in the model’s latent space that triggers refusal behaviors.
- Surgical Lobotomy: The refusal vector is mathematically neutralized or subtracted from the model’s weight matrices.
Kjellberg candidly admitted that this computational "surgery" is imperfect, noting that Ajax suffered "a little brain damage" during the process, resulting in minor quirks in reasoning stability. However, he confirmed that the model successfully executes practical workflows—such as sorting cluttered inboxes and conducting automated web research—approximately nine times out of ten. To further refine performance, he implemented Group Relative Policy Optimization (GRPO), a training method where the model generates 16 attempts at a single task and iteratively learns from the successful iterations.
Official Responses and Industry Context
The clash between independent creators and proprietary AI developers highlights a rapidly growing regulatory and technical battleground.
Major AI labs have increasingly tightened their application programming interfaces (APIs) and tightened enforcement of Terms of Service (ToS) clauses. Provisions barring the use of model outputs for "model distillation" or "reverse engineering" have become standard across the industry. When OpenAI banned Kjellberg’s accounts, it did so under these strict provisions, signaling a zero-tolerance policy toward developers attempting to harvest reasoning traces or build competitive open-source offshoots using corporate data pipelines.
Conversely, the open-source community—championed by figures ranging from independent developers on GitHub to prominent content creators—views these corporate walls as an overreach that stifles innovation and centralizes technological power in the hands of a few monopolistic entities. Projects like Odysseus and Ajax represent a grassroot counter-movement: a push to democratize artificial intelligence, strip away corporate paternalism, and restore user data sovereignty.
Implications for the Future of AI Development
PewDiePie’s turbulent experience building Ajax carries profound implications for the trajectory of artificial intelligence, consumer privacy, and intellectual property law:
- The Rise of Local Privacy: As consumer awareness regarding data harvesting grows, tools like Odysseus and local SLMs like Ajax point toward a future where personal digital management does not require uploading private calendars, emails, and browsing habits to centralized cloud servers.
- The Distillation Arms Race: The cat-and-mouse game between labs protecting their proprietary reasoning tokens (such as OpenAI’s GPT-5.6 Sol) and developers attempting to distill those capabilities into open-source models will likely intensify. As academic studies reveal methods to decode hidden reasoning blocks, corporate labs will be forced to implement heavier cryptographic and behavioral monitoring defenses.
- The Uncensored AI Frontier: Techniques like abliteration are rapidly democratizing the creation of "uncensored" or unaligned AI models. While this grants users greater autonomy over the tools they run on their own hardware, it simultaneously raises complex regulatory and ethical questions regarding liability, safety guardrails, and the proliferation of models capable of generating harmful content without standard corporate safety nets.
- Mainstream Creator Influence: By documenting the technical hurdles, account bans, and philosophical debates of local AI development to millions of mainstream viewers, creators like Kjellberg are accelerating public literacy around open-source machine learning, decentralized computing, and the ongoing struggle for digital autonomy.
As Kjellberg prepares to finalize Ajax V1 for public distribution via his website, the project stands as both a technical achievement in local optimization and a cautionary tale about the high-stakes friction defining the modern artificial intelligence landscape.
