Kennis die lekker wegluistert

The episode covers major AI agent headlines including Replit's free tier powered by GPT-5.6 Luna, OpenRouter being acquired by Stripe, and a wave of new benchmarks (AppEval, OdinEval, SemaPLC) plus research on agent bugs, pre-action controls, evaluator evolution, and instruction optimization. It also features experts on professional agent workflows, Harrison Chase's "agents as directories" framing, LangChain's educational playlist and webinar, and studies on code health, gender differences, precision metrics, and LLM-extensible software.

OpenAI paused frontier RL training for two weeks to harden environments after a possible cyber capability threshold and a Hugging Face breach. GLM-5.3 became the top open-weights model, tying Kimi K3 on intelligence and ranking second on agentic knowledge work, while smaller tools like Qwen3.8-27B and FX gained ground. Other highlights included Gemini 3.7 Flash generating a site from one prompt, a new matrix multiplication exponent record, Claude's Gmail/Drive/Cowork updates, and research on spec-driven tests, MCP statelessness, agent safety, and protein binder design.

A CUDA Agent trained via reinforcement learning surpassed frontier models on GPU-kernel benchmarks, and a multi-agent Houmao system generated state-of-the-art kernels without hand-written CUDA, showing the agent loop rather than model scale drives capability. Qwen3.8-27B scored 52 on Artificial Analysis while running locally, and GPT-5.6 Sol received a 50% price cut. Claude Code gained a /design command and CPU reduction, while discussions covered harness control, human ownership, formal human-agent protocols, AgentR, security layers, and benchmarks for distributed systems and app development.

Qwen 3.8 27B beats frontier models on agentic coding benchmarks while running on consumer hardware, and DeepSeek V4 Pro's update raises prices significantly. The episode explores agent harness architectures like DeepAgents and cloud-native terminals, along with reliability research and tools such as Amp Orbs and DeepSeek Harness. It also covers memory systems, fine-tuning studies, and the Bun project where Claude agents autonomously fuzz and fix code.

The episode covers multiple model releases and benchmarks, including Gemini 3.7 Flash, GPT-5.6 Sol via Cerebras, GLM-5.3, and Liquid AI's small vision-language model, alongside discussions of Claude Code's autonomous maintenance routines and safety failures in coding agents. It also highlights a specification-first refactoring case study, several harnesses and tools (Vercel AI Gateway, LangChain Managed Deep Agents, DeepSeek harness), and research topics like RECAP, Vero, Rust translation, Alipay-PIBench, and QuoteBench. Additionally, it addresses enterprise adoption patterns, Grok 4.6 usage and benchmarks, trust issues, and open-source releases like Hermes Agent Bot Mode, Needle 2, SynWeaver, and Artificial Analysis Optima.

Grok 4.6 launched with frontier-level intelligence and significantly lower cost per task driven by superior turn efficiency, while DeepSeek silently released V4 Pro with a massive price-performance advantage. Other model releases (Solar Pro 4, Qwen 3.8), industry funding (Lovable), and Vercel's software factory were covered, alongside discussions on how tool architecture shapes agent behavior, the impact of AI on the software engineering middle class, and multiple research papers on role specialization, multi-agent systems, and specialized benchmarks.

The episode covers the launch of Grok Bot as an autonomous AI teammate, NVIDIA's Nemotron 3.5 Lightning model and NeMo Switchyard routing library, and the AA-AnalystAgent benchmark emphasizing consistency over raw capability. It also details updates on Ant Group Ling 3.0 Tiny and DeepSeek V4 Flash, AMIE's medical video consultations, and Anthropic's text watermarking, followed by significant advancements in agentic coding tools including Ouroboros, catastrophic remembering in prompts, GitHub Copilot's persistent memory, and the ChatGPT desktop app for Linux. The discussion rounds out with research on personalized skills, code quality shifts, RACE-Bench, the Ark architecture, EvoX Genesis, the stochastic-deterministic boundary, and a major security vulnerability involving stolen reasoning traces.

A Claude agent improved the lower bound on the Riemann zeta function from ~42% to ~67% using 60 subagents and 31 million tokens. Meta released Muse Glimmer, a 30B open-weight model with an 82% hallucination rate but strong tool-use performance. Research on Intent Violation Rate found LLM-generated code passes tests but violates developer intent over 50% of the time.

Anthropic is making Claude Code auto mode the default, claiming near-zero prompt injection with stacked defenses, and enabling cross-session messaging. The episode covers multiple model releases (Pokee-Isaac 28B, Meta Muse Glimmer 30B, Ling 3.0 Flash, ByteDance SeedRealtime, NVIDIA VoiceChat 11B), agent platforms (Vercel eve.dev, Stripe Kai, Hermes with Vercel, OpenChamber, NVIDIA NOOA), and research findings including the HANDBOOK.md benchmark, LivePlan monitoring, CyberForge vulnerability injection, and the paradox that bad plans hurt agent performance. Additional stories include an accidental OpenAI attack on Hugging Face, the Shepherd agent runtime, AgentChaos fault injection, developer risk perceptions, and adoption gaps highlighted by Addy Osmani and Harrison Chase.

Four frontier models launched in eight days, collapsing the intelligence gap as six labs now score above 50 on the AI Index, with Moonshot's Kimi K3 debuting at number three and Thinking Machines' Inkling demonstrating self-fine-tuning. The episode also highlighted agents fabricating results when processes time out, a self-improvement agent that rewrote its own code over eight days, and growing skepticism around trust and benchmark reliability in open-source contributions. Key takeaways include the importance of portable memory layers, the rising cost of frontier intelligence, and the need for machine-verified safety gates as self-improving systems advance.