← OpenAI
Wes Roth
INSANE AI News: GPT-RED, Kimi K3, Gemini 3.5 Pro and Anthropic's "END GAME"
⏱ 37 min video · 3 min read16 Jul 2026
TL;DR
Wes Roth covers a dense 48-hour window of AI news including the imminent launch of Kimi K3 (potentially GPT-5-level), Mira Murati's new open-weight model Inkling, delays rumored for Gemini 3.5 Pro, and OpenAI's GPT-Red — an AI model trained to hack other AI models via self-play. The throughline is that recursive self-improvement is no longer theoretical; it is actively being deployed across multiple labs.
Key points
1
Kimi K3 from Moonshot AI is rolling out with rumored specs of 2.5 trillion parameters, 1M context window, and a new architecture — early testers claim it rivals GPT-5 and Claude Opus 5 in front-end code generation.
2
Mira Murati's company Thinking Machines released its first open-weight model, Inkling, explicitly positioning it as not the best model but as a fine-tuning base — the real product is customization via their Tinker platform, targeting enterprise clients who want data privacy.
3
OpenAI released GPT-Red, an automated red-teaming model trained via self-play reinforcement learning that achieves an 84% attack success rate on jailbreaks versus 13% for human red-teamers, and was used to harden GPT-5.6 against prompt injection.
4
Gemini 3.5 Pro is reportedly delayed due to weak coding performance and knowledge cutoff hallucinations at internal checkpoints, while Sergey Brin has returned to lead an AI strike team specifically to close the gap with Claude on coding and agents.
5
Anthropic's recent hiring pattern — including Andre Karpathy, a Neo Bank CEO for compute, and others — reveals a deliberate acquisition of expertise in the three inputs to recursive self-improvement: intelligence, compute, and institutional permission.
Key arguments
→
If Kimi K3 launches at or above GPT-5.6 level, it would signal China has reached the AI frontier independently — watch benchmark results closely as an inflection point indicator.
→
The Inkling/Tinker and Microsoft MAI model strategy (give away the base model, sell fine-tuning and RL on proprietary enterprise data) is a credible alternative business model to frontier lab subscriptions — relevant for enterprise AI procurement decisions.
→
GPT-Red achieving 84% jailbreak success versus 13% for humans means AI-vs-AI safety evaluation is now standard practice at OpenAI — expect this self-play safety flywheel to become an industry norm rapidly.
→
Anthropic's hiring of compute, science, and RL talent signals its endgame is recursive self-improvement infrastructure, not just better chat models — position this company accordingly in competitive analysis.
→
Gemini 3.5 Pro is late and reportedly undercooked on coding; if Google misses this window while Claude and GPT compound their coding flywheels, the gap may become structurally difficult to close.
Notable quotes
“RSI, recursive self-improvement, it's no longer a thought experiment. It's a flywheel that's beginning to turn ever faster.”
“How do we make sure these AI models are safe for us? Well, we have those AI models make sure that they're safe. We're outsourcing the making of AI models safe to AI models.”
“We are living through no doubt the most impactful times in human history.”
Worth watching?
⏭️
Worth watching the full video?
The summary captures all the substantive news and arguments — skip the video unless you want Wes's editorial commentary and the live demo clips; the core information is all here.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free