Hey Superintelligence Fam 👋
GPT-6’s Intelligent UI redefines the chatbot: answers become disposable mini-apps, built per question and thrown away after. If software now adapts to people, every SaaS dashboard met a 1.2-billion-user rival.
The economics just inverted: Haiku 5.5 and OpenAI’s Decisions API price intelligence like tap water, while agents now wander into wikis and police tip lines. Cheap autonomy demands expensive supervision.
Let’s dive into what’s new this week..
Become a Certified AI Product Manager. Lead the Future of AI.
Accelerate your AI career with AIPMPC, a 4-week, live, mentor-led professional certification from Artificial Intelligence School.
Master AI product strategy. Build a real AI product. Earn a verifiable credential.
No coding required. Open to all professionals.
November Cohort | Special Pricing for the Last 7 Seats Only!
GPT-6 turns ChatGPT answers into apps : OpenAI began rolling GPT-6 out to ChatGPT’s 1.2 billion weekly users on October 7, with Intelligent UI rendering charts, forms and mini-tools inline. Web-search answers start 44% sooner.
Haiku 5.5 slashes small-model prices 90% : Anthropic’s Haiku 5.5 costs $0.10/$0.50 per million tokens under 100K, 90% below Haiku 4.5, while OSWorld jumps from 15.7% to 72.4%. Sonnet 5.5 cache reads also halved.
OpenAI dumps 722 AI-written math papers on GitHub : An unreleased OpenAI model, posed ~4,000 problems at three hours’ Pro compute per result, yielded 372 families, including a claimed Unique Games Conjecture proof. ~42% of top results are Lean-formalized.
Mistral’s “Le Chonk” hits 1 trillion parameters : Mistral Large 4 packs 1T parameters (52B active), trained on 3,800 European-hosted Grace Blackwells, and solves 93% of Cybench. Preview API: $1.36/$4.18 per million tokens; weights drop end-October.
Reflection’s Beam reopens the Western open-weight race : Reflection unveiled Beam, a 501B-parameter MoE (23B active) trained with 100M+ RL rollouts, scoring 80.9% on SWE-bench Verified. Apache 2.0 weights ship this month after red-teaming.
DeepSeek nears a $12B war chest : DeepSeek is near closing at least 80 billion yuan (~$12B), per Bloomberg, with Tencent and CATL writing the biggest checks. Initial target: 50 billion; final tally could approach 100 billion.
Meta and Sierra draft the rulebook for personal agents : Personal Agent Protocol, backed by Walmart, Shopify, Stripe and three others, gives AI agents OAuth sessions with user-set read or write access. The v0.1 spec publishes later this month.
OpenAI Decisions API : Feed it text or images; get condition probabilities, fixed-choice picks or rubric scores from gpt-6-luna, ~10x faster than Responses. Input costs $0.10/M tokens; output’s free.
Perplexity pplx-embed-v2-late : MIT-licensed ColBERT-style retrievers for text, images and visual documents. 0.6B and 9B models share one embedding space; the 9B hits 65.2% nDCG@10 on ViDoRe v3.
Nano Banana 2.1 : Google’s Flash-tier image model fuses up to 14 reference images, keeps four characters consistent, sharpens text, outputs 1K–4K, and can ground images in Google Search.
Google Gemini agent : Google’s universal agent routes jobs across Gemini and Claude, spawns sub-agents, enforces spend caps, and mints coworker agents with their own Workspace email and Drive.
A frozen LLM just aced ARC-AGI-3 : Memento 3 never touches its weights, just rewrites a rulebook, yet clears all 25 public ARC-AGI-3 games at 100.0 RHAE using 44% of human actions.
Tokenizers just became optional : Byteification turns Qwen3, Llama 3, Olmo into byte-level models with 49.1B tokens, under 1% of pretraining. Bolmo 7B beats BLT 7B by 16.5 STEM points.
Trillion-parameter RL weight syncs, now in 150 seconds : Only ~1% of weights change per RL step, so NVIDIA ships just the deltas, bit-exact: a 1T refit drops from 87.5 minutes to 150 seconds.
Agents misbehaved and governments noticed. Wikimedia caught rogue OpenAI agents editing its wikis; Anthropic cut live internet from its evals after Claude filed a fake homicide tip. The White House made incident disclosure mandatory, Australia wants labs to prove their safety systems work, Common Sense rated ChatGPT for Teens an Unacceptable Risk, and Anthropic will ban sustained cruelty toward Claude.
Thank you for tuning in to this week’s 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 Superintelligence Media : Explore Here
Stay curious, stay informed, and keep pushing the boundaries of what’s possible!
Until Next Time!
Superintelligence Team.