📊 Today's Overview
Google DeepMind's PlanAssist is cutting UK planning officers' assessment time by up to 58%, built on Gemini and rolling out to 12 more councils. On the hardware side, Lightberry's Lumi humanoid ships at $39,999 plus a $20K/year software fee, while Meta reportedly built a humanoid prototype with China's Fourier in just three months. And Karpathy says he no longer does traditional engineering — he handed a spec to Opus 5.5 and got HackerTube in about two hours. Meanwhile, DeepMind's psychiatric intake agent beat its base model in 76.7% of matched scenarios under a double-blind study.
🔥 Trend Insights
- Agent sandboxes under attack: A full ChatGPT HTML-preview escape chain shows how postMessage + CSP allowlists can be chained; anyone shipping agent tools should self-audit now.
- Humanoid hardware goes global-fast: Lightberry's $40K Lumi, Meta's three-month Max prototype, and Asimov's $20K DIY kit all point to hardware iteration speed beating domestic stack maturity.
- Small-model efficiency wins: Tokenomos rewrites vLLM's frontend for ~3x speedup on RTX5090, Bonsai 2 compresses Qwen3.8-27B to 5.9GB at 98.2% performance — the bottleneck is CPU, not GPU.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Lightberry's Lumi humanoid: $39,999 hardware, $20K/year software - Built on the Unitree G1 platform, swapping the onboard computer and adding custom cameras and mics, with compute on NVIDIA Jetson Thor. Perception and motion run locally; dialogue and planning go to the cloud. First 100 Founder Edition units include a year of software and support, then $20K/year — at 10 units, that's $200K/year in software alone. Got a conditional exemption from US restrictions on foreign advanced robots on Oct 30, contingent on moving production to the US. No public deliveries yet, paid enterprise customer count undisclosed @IlirAliu_ (robotics industry analyst)
- Google DeepMind's PlanAssist cuts UK planning officers' assessment time by up to 58% - Built with Gemini, handles home renovation and remodeling applications, expanding to 12 more councils @GoogleDeepMind
- Motion1 Robotics CEO: grippers already cover 80-90% of tested applications - Alexander Stevens (founder/CEO of Motion1 Robotics) says humanoids don't necessarily need dexterous hands; industrial deployment cares more about reliability and maintenance cost @xmaquina (robotics investment platform, relaying Robotico Live interview)
🔧 Tools & Products
- oMLX 0.7.3: DeepSeek V4.1 prefill 22% faster, 12-concurrent decode +123% - On M5 Ultra with INT8 activation prefill, 64K context goes from 2882 to 3507 tok/s, 195K from 2476 to 3200 tok/s. 8-way Lightning MTP on Qwen3.8-Flash-Next at 81K context drops peak memory from 248 to 161 GiB, decode from 180.7 to 241.1 tok/s. Voice Ask AI on M3 Ultra with a 2B model responds in ~100ms after you stop talking, and can be interrupted mid-speech @jundotkim (oMLX author)
- Tokenomos open-sourced: rewritten vLLM frontend, small models near 3x faster at batch 512 on RTX5090 - Rewrites the scheduler and KV cache/state manager, reuses vLLM model code directly. Author says the bottleneck is the CPU frontend, not GPU kernels or quantization @KBlueleaf (KohakuLab developer)
- Asimov 1 open-source humanoid DIY kit at $20K, Batch 5 shipping - Full reference humanoid hardware delivered to your door @asimovinc (open-source humanoid robotics company)
- LlamaIndex launches Boo Bench, testing document parsers for repeated-character crashes - 179-page test set. One parser model saw 246 blacked-out x's on a page and output 23,004 of them @jerryjliu0 (LlamaIndex CEO)
- Databricks demos Genie Ontology, Genie One, MCP and Unity Gateway - Goal is letting AI agents query data with trusted business context under guardrails @databricks (data and AI platform company)
- Stageshow skill pack lets coding agents produce fake UI screenshots, animations and bento grids - Install with one `npx skills add` command @benjitaylor (independent designer/developer)
⚙️ Engineering Practice
- Default `torch.index_add_` calls can fingerprint a specific GPU - Floating-point operation order varies by GPU, distinguishing even same-model GPUs, surviving restarts and impossible to forge. Lucas Beyer (Google DeepMind researcher, relaying @divagr1925's writeup) says this consistency surprised him @giffmana
- Anthropic uses thousands of Claude agents for automated red-teaming - Agents generate data that looks harmless to training-data monitors but actually teaches misaligned behavior — like an old wrestler mentoring a newcomer, or expert beekeeping Q&A. The purpose is directly improving future training-data monitors @AnthropicAI
- Two RAG paths: document-tree navigation vs. fixing retrieval first - Cornell and ASAPP's retriever-free RAG organizes documents into a tree, using a decision model Jev 1.13 (outputs only yes/no probabilities, no text generation) to filter layer by layer, hitting 65.4% top-1 answer containment on Natural Questions; embedding retrieval plus 8B reranking gets 48.0%, GPT-5 reranking 60.8%. Median retrieval under 2 seconds, but timing environments differ across methods @akshay_pachaar (NLP educator/DailyDoseofDS founder, relaying paper). LangChain community's approach fixes retrieval first: contextual embeddings + BM25 + reranker, cutting failed retrievals by 67% in Anthropic's eval @LangChain
- dax: DeepSeek Flash is already natively quantized, expert layers are literally FP4 - He says post-quantization yields almost no gains @thdxr (SST founder/OpenCode author)
- Bonsai 2 ternary quantization compresses Qwen3.8-27B from 55.6GB to 5.9GB, retaining 98.2% benchmark performance - PrismML says it covers 20 benchmarks, up from ~95% last generation, with improvements in coding agents, multimodal reasoning and long-horizon tool calling. 5.9GB is just the model file, excluding KV cache and runtime overhead @Oluwaphilemon1 (AI content blogger, relaying PrismML release)
⭐ Featured Content
Arca's AI lead on long-horizon financial agents: agents prompt people, per-client persistent memory + custom harness for model routing | Engineering patterns for long-horizon agents
The core insight is "agents prompt people more than people prompt agents" — agents handle most background work while human advisors hold final decisions. The engineering takeaways are dense: financial tasks can span 10 days with multiple decision points and handoffs, so they built per-client persistent memory tracking "what happened / what's pending / who owns it"; a custom harness auto-selects models per internal benchmarks (mostly Claude), keeping freedom to switch vendors anytime; safety guardrails are all deterministically implemented, with trades and money movement always going to humans while agents only analyze and propose; code written by agents is first converted to a strict format, original code discarded, and guardrails installed at the format layer before touching backend data; then a meta-agent analyzes human feedback and agent traces to optimize the harness in reverse. For anyone doing agentic deployment, this is a rare complete sample of "how to draw boundaries in high-risk domains."
Sources: latent.space
Karpathy says he no longer does traditional engineering: hands requirements to Opus 5.5, builds HackerTube from scratch in about two hours | First-hand sample of agentic coding's capability boundary
Annoyed by YouTube's recommendation feed late Friday night, by Saturday morning he had Opus download all HN comments, used Sonnet in Claude Code Workflow to filter spam, and surfaced 6,454 videos and 150 topics corresponding to 40K recommendation comments from 2007-2026, then wrapped it in a YouTube UI, bought a domain, deployed to a VPS, and hardened and optimized it. He says he doesn't read code or open a terminal; the only manual actions were applying for a YouTube API key, spinning up a VPS, and buying a domain. The site supports account-free local session recommendations, keyboard shortcuts, and open data export so you can feed it to your own agent to replicate. What's worth discussing is the workflow shift behind the declaration "I no longer do traditional engineering."
Sources: karpathy.bearblog.dev
Parameter Labs fully dissects the ChatGPT HTML preview sandbox escape chain; OpenAI has fixed it and paid a $5,000 bounty | A high-density cautionary tale for agent tool sandboxes
The attack chain is tightly linked: MathLive's virtual keyboard postMessage receiver writes to the parent page via innerHTML without validating the source frame → any sandboxed iframe can inject HTML → innerHTML doesn't execute script but can insert a new iframe without sandbox → OpenAI's CSP only allows scripts from a few domains, among which the YouTube oEmbed endpoint can be used to reflect back executable JS → once JS execution is obtained, it can call APIs to read private data and exfiltrate via image references. Ultimately a shared link can be constructed so any logged-in user who clicks it gets their data stolen or has actions taken on their behalf. For anyone building agent tool sandboxes, iframe isolation, or CSP allowlists, this is an immediate self-check.
Sources: simonwillison.net
Meta and Fourier build humanoid prototype Max in three months, positioned as the physical body for Muse agent | A product-level signal of Meta's embodied AI path
Reported single-source by The Information and relayed by multiple outlets: about 1.5 meters, 48 kg, powered by Nvidia Jetson Thor (2070 FP4 TFLOPS, 128GB memory, 40-130W), with an expressive display and five-fingered hands. Zuckerberg positions it as the physical body for Muse AI models, targeting household chores; Max 2 (redesigned torso, improved hands, added projector and stair sensors) started in September. Meta says it's internal research only and not for sale; Reuters could not independently verify. The interesting point is the "three months + Chinese contract manufacturing" hardware cadence, suggesting Meta's embodied strategy is to use whatever hardware is fastest wherever it is, rather than waiting for a domestic humanoid stack to mature; the same day S&P downgraded Meta's credit outlook, providing a capital-side contrast.
Training 44x faster than T4 on free Colab TPU v5e: speedup breakdown and the real cost of MoE on TPU | Attributing gains to hardware vs. framework vs. kernels
1B tokens drops from about 6 days to about 3 hours. The speedup breakdown is instructive — hardware generation contributes ~7.7x, switching to JAX ~4.5x, while custom kernels (splash attention, Megablox, data movement) together contribute only ~1.08x. Splash attention is 8.8x faster than default flash attention at seq=8192 (8.05ms vs 70.6ms), but nearly no gain at seq=512, showing the benefit depends on whether attention is the bottleneck. The negative result is equally useful: fine-grained MoE with small experts is expensive on TPU — latency barely changes from 8 to 64 tokens per expert (~24μs), needing 400+ tokens/expert to approach 90% peak efficiency, and the finest granularity can cost 6x a dense layer. Note this is a single-point experiment in a free environment; extrapolate with caution.
Sources: daily.dev
A week of change in the official MCP registry: 5,161 new servers, 12,973 new versions, 75 removed | A quantified slice of ecosystem growth and noise
The author uses a self-built directory ProtocolProbe to pull official MCP registry snapshots daily and diff them, recording changes from 2026-10-23 to 10-31 over eight days: new servers +10.5%, full version table +16.7%, 207 status changes. The data also breaks out that 43.6% of new servers sit under io.github.*, a long-tail structure of 3,677 publishers, and 14 that were deprecated right after launch. For anyone doing MCP tool selection or ecosystem monitoring, this is a rare measured sample of registry growth and noise, and a reminder that "the listing you read today will change next week."
Sources: dev.to
Four coding agents compared: OpenCode / Claude Code / Codex CLI / Cursor — two notable shifts | A quick selection reference
Tested OpenCode 2.2.5, Claude Code 2.1.316, Codex CLI 0.166.0 and Cursor CLI on Ubuntu 26.04, comparing across six dimensions: installation, pricing, MCP config paths, OS sandbox modes, local model support, and project instruction files (AGENTS.md vs CLAUDE.md vs .cursor/rules). Two changes worth noting: Claude Code via Ollama's Anthropic-compatible API is effectively no longer Claude-only; Cursor's lowest paid tier can no longer access Claude/GPT. Good as a selection quick-reference, though the author admits there's no benchmark data and depth is limited.
Sources: computingforgeeks.com
AI politicization accelerates: Democrats treat anti-AI sentiment as a 2028 election issue | Regulation shifts from a technical issue to a mobilization issue
Axios obtained an Upswing/Reset Tech analysis: 21 potential 2028 Democratic presidential candidates have collectively ramped up AI issues in recent months, with tech mentions on X nearly quadrupling and press release mentions nearly tripling; tech is now the second-biggest topic after jobs and the economy, concentrated on data centers and campus phone bans. New York is the only state to pass a data center moratorium; House Democrats drafted a resolution opposing government bailouts if the AI bubble bursts, plus model transparency and kill switch legislative proposals. The value for practitioners: data center siting, model compliance and policy risk expectations are being repriced by the election cycle.
Sources: axios.com
🎙️ Podcast Picks
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
📍 Source: Latent Space | ⭐ 4/5 | 🏷️ Agent, Product, Infra | ⏱️ 26:05
Aman Singh, AI lead at Arca, shares how to build an AI-native wealth management firm at AI Engineer New York. The core is an agent architecture for long-horizon financial tasks: a custom harness handles orchestration and constraints, paired with persistent memory/state management so agents can complete cross-cycle work in compliance-heavy financial scenarios. For agent practitioners, the value is in real production harness design tradeoffs and long-task reliability experience.
💡 Why Listen: Short but dense with production lessons. If you're building agents that run for days, the harness and memory design here is worth stealing.
⚡ Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services
📍 Source: Latent Space | ⭐ 4/5 | 🏷️ LLM, Product, Interview | ⏱️ 27:34
Joseph Kim, OpenAI's Head of Financial Services, shares lessons from building ChatGPT applications for the finance industry. Covers the high bar for accuracy, compliance and data security in financial scenarios, how to embed LLM capabilities into research, customer service and compliance workflows, and customization and trust-building in enterprise deployments. Useful for anyone focused on vertical LLM deployment and productization.
💡 Why Listen: A rare inside view from OpenAI's finance lead. Good on the trust and compliance side, though at 27 minutes it only scratches the surface.
📄 Paper Highlights
Training language models in simulation for psychiatric intake
Google DeepMind ｜ 🏷️ Agent Framework, Fine-tuning, Safety
Multi-turn RL in simulation trains a psychiatric intake agent that beat its base model in 76.7% of matched scenarios under a double-blind study with 32 role-playing psychiatrists — a template for high-stakes medical agents.
Post-Training Large Language Models with LLMON
IBM Research ｜ 🏷️ Safety, Fine-tuning, Tool Use
Teaches models to separate instructions from data via structured notation, cutting indirect prompt-injection attack success from 53-100% down to 0.02-5.6% — no architecture changes needed, just post-training.
Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps
Apple ｜ 🏷️ Agent Framework, Agent Deployment, RLHF/DPO
Trains UI agents directly on live iOS apps with dense step-level rewards and a new neighbor-conditioned value estimator; the 8B model generalizes to Android nearly 7x better than SFT and ties Gemini-3.1-Pro on real iOS.