Self-Made AI & Million-Token Monsters
9th Feb 2026 | Superintelligence Newsletter
Hey Superintelligence Fam đ
AI just crossed a mind-bending threshold. OpenAIâs GPT-5.3 Codex - the first model instrumental in building itself - and Anthropicâs million-token Opus 4.6 are locked in a coding rivalry thatâs pushing machine intelligence into uncharted exponential territory.
When AI models help create their own successors, the growth curve stops being linear. Meanwhile, this intelligence explosion rattled Wall Street hard - an $830 billion SaaS sell-off proving that autonomous code-writing agents arenât future talk, theyâre here now.
Letâs dive into whatâs new this week..
Exclusive free datasets drops on Mozilla Data Collective
100+ organisations are open sourcing and sharing their datasets exclusively through Mozilla Data Collective. Get your models to work for users all over the world. TTS, ASR and LLM coverage, from Khmer, Tolaki and Betawi, to Chinese, English and Spanish. Join MDC for trustworthy human-source data with a clear consent trail.
OpenAI Unveils GPT-5.3: The First AI Model to Build Itself : OpenAI has launched GPT-5.3-Codex, a landmark model that assisted in its own development and debugging. Operating 25% faster than previous versions, this release signals a move toward self-improving autonomous agents.
Anthropic Debuts Claude Opus 4.6: Entering the âVibe Workingâ Era : Anthropicâs latest flagship, Opus 4.6, introduces a massive 1-million-token context window and âagent teams.â The model is so capable at enterprise tasks that executives describe a shift toward frictionless, autonomous âvibe working.â
The SaaSpocalypse: Anthropicâs AI Breakthrough Wipes Out Billions : The stock market faced a âSaaSpocalypseâ as Anthropicâs new legal and automation tools sent software stocks tumbling. Investors fear that AI agents are now directly substituting for traditional subscription-based software services.
OpenAI Frontier : OpenAI Frontier formalizes enterprise AI agents with shared memory, permissions, and auditability, signaling a shift from standalone models to governed, multi-agent systems inside organizations.
Claude Opus 4.6 : Claude Opus 4.6 strengthens long-context reasoning and parallel task execution, reinforcing Anthropicâs strategy around agentic work patterns rather than single-prompt AI usage.
OpenAI GPT-5.3 Code : GPT-5.3 Codex expands autonomous coding loops, reflecting how AI is moving beyond assistance toward handling full development cycles with minimal human intervention.
VibeTensor: System Software for Deep Learning, Fully Generated by AI Agents : Researchers released VibeTensor, an open deep learning runtime stack entirely generated by LLM coding agents, spanning Python APIs to CUDA memory management, showing AI can build complex systems. It includes a C++20 core, autograd, and caching allocator, and passes basic training workloads on H100 and Blackwell GPUs.
Reinforcement Learning via Self-Distillation (SDPO) : This paper introduces SDPO, a reinforcement learning method that turns rich textual feedback into dense token-level learning signals using the model itself as teacher, improving sample efficiency and accuracy on reasoning and programming benchmarks.
Self-Improving Pretraining : Self-Improving Pretraining from Meta FAIR replaces next-token prediction with sequence-level generation guided by teacher rewrites and a judge model, boosting generation quality, factuality, and safety early in training on Llama2 and RedPajama data.
Global AI ethics focused on structural oversight as the UN Secretary-General formally launched the Independent International Scientific Panel on AI to close global knowledge gaps. Simultaneously, the 2026 International AI Safety Report warned of accelerating biological misuse and deepfake fraud, while the European Commission faced scrutiny for missing critical EU AI Act guidance deadlines for high-risk systems.
Partner spotlight: Open, multilingual, training-ready data that helps AI teams spend less time wrangling datasets and more time building models that actually ship.
Thank you for tuning in to this weekâs edition of Superintelligence Newsletter! Stay connected for more groundbreaking insights and updates on the latest in AI and superintelligence.
For more in-depth articles and expert perspectives, visit our website | Have feedback? Provide feedback.
To explore Advertising : Explore Here
Stay curious, stay informed, and keep pushing the boundaries of whatâs possible!
Until Next Time!
Superintelligence Team.