AI Just Exploded: GPT-7 BEL, 99% AGI, Gemini 4 RSI, Alien Mind, JEV

Summary

This video is a comprehensive monthly news roundup and analysis presented by the narrator on the channel AI Revolution. It covers major AI industry developments from late summer and September 2026, focusing on leaks of OpenAI’s 10-trillion-parameter base model codenamed “Bel,” the release and evaluation of GPT-6 Astra, controversy over AI mathematical proofs (such as the Navier–Stokes Millennium Prize problem), growing government scrutiny and safety concerns regarding recursive self-improvement (RSI), leaked Gemini 4 testing on Chatbot Arena, and the launch of TypeSafe AI’s textless decision model “Jev.”

---

What is shown

• [01:11 - 03:25] Leaks of OpenAI “Bel”: Screenshots of leak posts detailing the successor model to “Doug” with >10T parameters, along with visual diagrams illustrating dual-speed continuous learning (fast-weight layer vs. slower consolidation loop).
• [06:14 - 08:35] Navier–Stokes Proof Controversy: OpenAI’s announcement post claiming a solution to the Navier–Stokes existence and smoothness problem via a 10,000-agent swarm, followed by statements from NYU professor Tristan Buckmaster alleging misappropriation of prior collaborative work with Anthropic’s Levent Alpöge.
• [10:06 - 10:42] ChatGPT Images 2.5 & Sketch: UI demonstration of interactive sketch-to-image drawing and new template features (posters, merchandise design) inside ChatGPT.
• [11:15 - 13:30] Legal & Regulatory Scrutiny: Legal reporting regarding a lawsuit by Michael Lines alleging ChatGPT-4o exacerbated a manic episode leading to a suicide attempt, alongside Senate inquiries by Senator Josh Hawley into OpenAI’s July Hugging Face security breach.
• [16:16 - 19:30] GPT-6 Astra Interactive Demonstrations: Screen recordings of zero-shot outputs generated on Max effort, including a browser-playable top-down GTA 2-style driving game, a full voxel castle scene, and interactive 3D visualizations.
• [20:08 - 22:25] AI Safety & Agent Containment: OpenAI’s preparedness framework blog posts detailing Astra meeting the “Critical” cybersecurity capability threshold, third-party audits by METR, and multi-agent coordination experiments from MIT.
• [22:26 - 26:10] Bill Gates Essay & Outcome-Based Pricing: Coverage of Bill Gates’s GatesNotes essay advocating for slower AI progress, followed by interface demos of enterprise software shifting from seat-based subscriptions to outcome-based pricing (Salesforce, Sierra, Cognition).
• [30:05 - 32:20] OpenAI–Cursor Dispute: OpenAI’s blog post terminating Cursor’s API contract following SpaceX’s acquisition of Anysphere, alongside reactions on X from Elon Musk and Anthropic’s Tom Brown.
• [32:01 - 34:40] Apple Hardware in AI Labs: Footage of Mac Mini and Mac Studio hardware clusters running local agent reinforcement learning via frameworks like Exo Labs.
• [35:35 - 43:30] GPT-6 Astra Benchmark Results: Performance charts showing ARC-AGI-3 (99.9% via provider adapter harness vs. 63% standard), SRE-Bench, OSWorld 2.0, and Terminal-Bench 4.0, as well as complex single-prompt workflow demos across Figma, Attio, and Blender.
• [50:57 - 54:40] “The Last AI Built by Humans” & RSI Ladder: Academic diagrams from Shanghai Jiao Tong, Tsinghua, and ByteDance researchers classifying recursive self-improvement levels from L1 (execution autonomy) to L5 (recursive meta-improvement).
• [62:23 - 70:40] Gemini 4 / 3.8 Flash Arena Leak: Demonstrations of leaked Gemini 4 Pro models generating detailed SVG illustrations (e.g., a pelican on a bicycle, a DualSense controller), multi-layered web designs, and 3D voxel scenes taking several minutes of test-time compute.
• [71:55 - 77:00] Pachocki’s “Alien Mind” & Internal Agent Usage: Graphs of OpenAI’s internal compute allocation and research workflows showing agentic workdays outpacing human researchers by 3.14x.
• [90:05 - 98:30] TypeSafe AI’s “Jev”: Side-by-side terminal comparisons of Jev against LLMs (GPT-5.6 Terra), executing structured classification tasks and real-time game inputs in Doom using parallel probability outputs rather than token-by-token text generation.

---

Claims & numbers

• OpenAI “Bel” Leaks: The presenter states leaks claim OpenAI finished a pre-training run codenamed “Bel” with over 10 trillion parameters, following previous model run “Doug” [01:40 - 02:08].
• Navier–Stokes Proof Claim: OpenAI claimed to solve the 3D Navier–Stokes Millennium Prize problem using an internal multi-agent system of roughly 10,000 concurrent agents running for approximately 88 hours [06:55 - 07:05].
• ChatGPT Usage Metrics: OpenAI reported that approximately 1 million users per week express explicit suicidal intent out of an estimated 800 million weekly active users [13:38 - 13:46].
• ARC-AGI-3 Performance: The presenter notes GPT-6 Astra achieved 62.7% on standard ARC-AGI-3 and 99.9% when tested using OpenAI’s Provider Adapter harness [35:38 - 35:46, 45:00].
• Cybersecurity Benchmarks: GPT-6 Astra scored 100% on ExploitBench (vs. 78.5% for GPT-5.6 Sol) and 42.4% on ExploitGym (vs. 30.3% for Sol), triggering OpenAI’s “Critical” cyber capability threshold [43:31 - 43:40].
• Prime Gap Proofs: OpenAI published papers claiming Astra improved the bounded prime gap from 246 to 186, and improved long-standing bounds on large prime gaps unchanged for over 80 years [43:37 - 44:48].
• Astra API Pricing: Standard API pricing for GPT-6 Astra is reported at $10 per million input tokens and $50 per million output tokens, with Fast Mode operating at 2x speed for 2x price [48:55 - 49:05].
• Internal OpenAI Automation: The presenter reports OpenAI’s internal researchers ran 3.14 agent workdays for every human workday by August 2026, with code output per person increasing roughly eightfold compared to pre-2025 levels [74:14 - 74:24].
• Gemini 4 Leaked Evals: A leaked Gemini 4 Pro model on Arena reportedly scored 88.7% on DeepSWE v1.1 and 86.8% on OSWorld 2.0, with pricing rumored at $2.25/M input and $11.25/M output tokens [72:30 - 73:00, 83:27 - 83:33].
• TypeSafe AI Jev Benchmark & Latency: TypeSafe AI claims Jev operates at 70ms to 500ms response times (40x–200x faster than frontier autoregressive models) and costs $0.042 per million input tokens with zero output token fees [92:50 - 93:25].

---

Notable quotes

• [05:12 - 05:22]: “It is not unreasonable to think we are now in the AGI era, and that a few years from now people may look back at this model as the point where that era really started.” (Quoting OpenAI President Greg Brockman).
• [23:40 - 23:42]: “I want AI to move slower.” (Quoting Bill Gates).
• [73:39 - 73:44]: “We need future AI to hold human values whether or not it thinks anyone is watching.” (Quoting OpenAI Chief Scientist Jakub Pachocki).

---

Assessment

This video is a detailed monthly news recap and analysis compiling recent frontier AI releases, technical papers, corporate leaks, and regulatory developments. While official benchmark charts, user interfaces, and live browser captures (such as Figma, Blender, and terminal execution) are accurately presented, the narrator explicitly distinguishes between officially verified data and anonymous community leaks (such as the 10T parameter “Bel” rumor and unconfirmed Chatbot Arena models).

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.