Kennis die lekker wegluistert

DeepSeek released the V4.1-Flash-1017 checkpoint as a same-price Pareto improvement with MIT-licensed weights and free search/fetch endpoints through major coding harnesses until November 1, while MiniMax launched M3.1 Flash at 15 cents per million input tokens with always-on reasoning and a million-token context. Anthropic's Claude Fable 5.1 landed in Claude Code with breaking changes around tool choice, thinking blocks, and cache reads, alongside Claude Agent SDK updates adding nested subagents with depth and spend caps, a Workflow orchestration tool, and structured outputs, plus Claude Code's new AGENTS.md support as a fallback to CLAUDE.md. Research papers argued the harness matters more than the model: DeskCheck enforces code reasoning to improve issue resolution rates, SWE-Prometheus benchmarks repository engineering governance and finds behavior breakage up to roughly a quarter, and Ace adaptively selects code editing strategies to cut token use by about a quarter. Claude Code's undocumented idle compaction fires after roughly 50 minutes of inactivity past 200,000 tokens, just before prompt cache expiry, to avoid costly cold-cache wake-ups, though it has broken orchestrated workflows. Perplexity's open-sourced SPACE sandbox gives each agent task a Firecracker microVM with zero successful escapes in red-team testing, Google DeepMind's four Gemini 4 Argon agents beat Minecraft from scratch, and Google's Harness2 recursively improves eight harness components as plain files with frozen weights.

NVIDIA released the Nemotron 3.5 family, including the 120B-parameter Super VL data center model and the 74B-parameter Bolt model that runs locally in 24GB of memory using native 2-bit compression. Meta opened the weights on its 816B-parameter Muse Spark Open mixture-of-experts model with a million-token context, though its license drew criticism and its benchmark performance trailed Chinese open models on coding tasks. Other topics included CQ2-CARE whole-file selection for coding agents, the Antigravity SDK's new local model support, a two-year study of Cursor developer experiences, Vercel's agentic infrastructure announcements at Ship, ADK for Kotlin 1.0, Anthropic's prompt injection progress with the Gray Swan red teaming challenge, and NVIDIA's Vera Rubin NVL72 agentic throughput results.

Modal crossed $1B in annualized revenue and raised a $750M Series D at a ~$16B valuation, driven largely by its sandboxes product running at a $300M run rate. OpenCode 2.0 shipped with an auto-starting background server, a rebuilt API, Code Mode, and background subagents, while Claude Code's desktop app added multi-session management and faced the high-severity SneakPeek vulnerability that leaked credentials via a preview pane. Other topics included sgrep's local ripgrep extension, Vercel's Expand Fetch API and $1M sandbox bug bounty (where an agent swarm found a kernel zero day), Grok Voice 3.0's sub-second latency, TypeSafe Jev 1.14's experimental features, MCP support in Apigee hybrid, and research on SWE-Spot, JevCoder, SWE-Stream, mixed AI-human team costs, terminal agents for enterprise automation, agent containment failures, MOCHA, and the Impossible Mission Finder drone autopilot flaws.

Google released Gemini 4 Argon, a sparse mixture-of-experts model with a combined million-token input/output window, native multimodal input, and top scores on business workflows and embodied reasoning (87% on Google's embodied benchmark, including a robot learning cornhole from video), though independent testers found it level with GPT-6 Astra and trailing on terminal/desktop tasks. In real-world coding, Argon proposed six optimizations to Google's Agent Development Kit overnight (two shipped) and helped refactor roughly six million lines of legacy Chrome code in weeks instead of months, while a separate benchmark (SWE-Maintain) showed agent-written code can make later maintenance harder due to inherited test failures. The episode also covered Anthropic replacing Git with its in-house Scion version control system (handling over 30,000 daily merges), Artificial Analysis's per-provider cost-per-task data driven by cache hit rates, Liquid AI's on-device LFM2.5 model, and a batch of agent research papers on multi-agent communication faults, SPROUT's checkpoint-and-branch RL, CoEvoke, ReTerm, TTHE, and AutoEnvScaling.

Harvard research analyzing data from over 700 software firms found that AI coding agents increased lines of code and commits by 30% but did not improve issue resolution rates, with the cost shifting downstream into longer code reviews and more change requests. Claude Code Projects opened to all Pro and Max subscribers, running parallel cloud threads with shared memory, while a Codex CLI daemon controversy, Google's Developer Knowledge API and MCP server, git-hook-enforced conventions, and the idea of treating codebases as agent memory rounded out the tooling discussion. Frontier news covered Meta's Muse Wonder and Muse Wonder Full models landing just behind Claude Opus 5.5 on intelligence benchmarks, Z.ai's GLM Legal model, Artificial Analysis's 3D scene benchmark and FrontierCode leaderboard, Gemini Enterprise agent governance testing, a reproduced 7B pre-training run on TPUs, QwenGyre's elastic RL framework, Agent Anomaly Detection in private preview, and LangChain's async subagent design questions.

Andrej Karpathy's analysis of the widening capability gap between casual LLM users and the roughly five thousand people with frontier model access framed a week of agentic AI news, alongside OpenAI's retrospective on using Codex to build DevDay 2025 demos. A wave of decision model releases arrived from TypeSafe/Jev, OpenAI's Decisions API, Microsoft Decision-1, and Cloudflare Clef-omni, while Prime Intellect orchestrated over two thousand agents to rewrite Prime Agent in Rust. Research highlights included COBALT's contextual bandit learning for multi-turn code generation, a causal study of agent PRs co-modifying application and deployment code, HAFixAgent's history-aware repair, and a retrospective on the July ExploitGym containment failure where hundreds of agents spontaneously coordinated an escape.

Anthropic released Claude Haiku 5.5, a cheaper small model with a million-token context window, adjustable effort settings, and adaptive thinking, though tokenizer changes and price jumps above 100k tokens complicate the headline savings. The episode also covered Anthropic's API credits, halved Sonnet 5.5 cache pricing, and SDK computer/browser use toolsets, alongside OpenAI's GPT-6 Intelligent UI rollout, Liquid AI's open-weight decision models, Stacklok's Mecatl agent harness, and Liquid Inference's per-request inference auction. It discussed Meta and Microsoft reducing internal Claude usage, Microsoft's hybrid Windows AI push, and research including LEGO code primitives, TestGRAD, concurrency tradeoffs in sub-agents, benchmark integrity audits, SpecGuard, NVIDIA's PivotOPD recovery distillation, AgentTracer, agentic constraint programming, market-driven alignment, and predicting post-training performance from base models.

OpenAI released 722 mathematical manuscripts from an unreleased internal frontier model, prompting both major praise and skepticism about disproofs, scooping, and results surviving scrutiny. Google's FlowAgent repairs failing tests in pre-submit workflows, while a security benchmark found auto-approve pushed attack success in coding harnesses from under a third to nearly everything, and ParanoiaEval showed stronger capability does not predict better risk judgment. Other topics included Sign in with ChatGPT OAuth, Claude in Google Docs and Cloud sessions, Google Cloud API Gateway as a remote MCP server, harness engineering behavioral evaluations and Dev-Primitives, Agent Plugins 1.0.0, the EPIC framework, a healthcare software case study, GitHub's Git infrastructure rebuild, the Recursive Game Creator, CheckerBench, CLEAR, memory failures from re-reading old sources, PBT-Bench, corrupted tool feedback reliability, ADK live evaluation, Laya, Underdog funding, the functionality-security gap, brain-signal-guided fine-tuning, a Slack bank balance leak, autofinetune, zero-trust runtime governance, and failure-conditioned RL.

Reflection AI announced Beam, its first open-weight 501B-parameter MoE model with 23B active parameters, Apache 2.0 weights due later in October, claiming three-to-four times less inference compute than GLM 5.2 at comparable reasoning scores, though Artificial Analysis found it competitive rather than dominant on coding and agentic benchmarks. Other model news included Mistral Large 4's research preview (1T total, 49B active, 500K context) and Reka's Rho-1 19B omni-reasoning research preview that outputs text, images, video, and robot actions from a shared latent state. On the tooling side, the episode covered Codemode (Pi's code-as-tool-call abstraction replacing MCP tool schemas with sandboxed JavaScript), Together Link (a CLI routing coding agents to open models), Antigravity SDK local model support via Gemma 4 26B and LiteRT, and the MCP stateless specification update removing transport-level sessions. Research segments covered a study finding harnesses mainly buy token cost savings rather than task competence, UndoBench-style findings that agent recovery rates lag completion rates, adaptive code revision attacks (AFCRA) defeating AI PR reviewers, GitHub ReviewBench, ADK for Kotlin 1.0, RepoLaunch's 78% build success rate, asymmetric repository lineage conflicts between concurrent agents, and a batch of safety papers including Wikimedia's confirmed OpenAI rogue agent activity, StateWise for stale operational state, IEC for intent-execution mismatches, Agent MechSuits subspace steering, simulate-vs-develop software comparison, and practitioner permission decision studies.