Edition 045 — Agents finish Minecraft, a voice model takes 20,000 calls a day, and Shanghai orders 8,000 humanoids

In this edition

Frontier & Text Models (5)
Image & Vision (1)
Video & World Models (1)
Audio, Voice & Music (1)
Robots & Embodied AI (4)
Small & On-Device Models (2)
Papers & Research (3)
News & Business (1)

The agentic bill came due in three places at once. Four Gemini 4 Argon agents completed Minecraft with no human input and a $1,620 invoice. xAI shipped a voice model that already resolves roughly 70 percent of 20,000 daily Starlink support calls. And a new Microsoft benchmark showed that when the evidence for a decision only appears late in a long run, every frontier model collapses to 21 percent — near coin-flip. Meanwhile Anthropic's own internal data has junior hiring down from 27 to 9 percent of new technical hires, while its agents log six hours for every hour a human works.

Frontier & Text Models

Four Gemini 4 Argon agents playing Minecraft

Four Gemini 4 Argon agents beat Minecraft without a single human input, and the bill came to $1,620. A Google DeepMind team dropped the agents into a randomly generated world with no roles, no strategy and no game-specific tools. They split tasks among themselves, warned each other about threats, recovered items from dead teammates and logged their mistakes in a shared memory. Over 27 hours and 14 minutes the four agents wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times. The session was priced at the standard Gemini 4 Argon rate that takes effect in 2027 — a useful reminder that autonomous play is now a metered product, not a research curiosity.

Google Gemini logo

Google DeepMind says self-improvement is nowhere close for Gemini, even as its own agents write most of the code. Engineer Haydn Belfield reports no measurable acceleration of AI research driven by AI itself. AI now writes more than 75 percent of production code at the company, yet about 88 percent of code changes still need human oversight or review, and the share fully delegated to AI rose only from 7.5 to 11.9 percent. Across 1,389 internal Argon agent runs, success without human input dropped sharply as task complexity rose. A survey of twelve senior DeepMind researchers found the same pattern: automation runs highest on well-defined, verifiable work and lowest on open-ended scientific ideation.

Anthropic logo

Anthropic's junior hiring has collapsed while its AI agents now log six hours for every hour a human works. In an internal study, the share of junior positions in technical departments fell from 21 to 14 percent since early 2025; among new hires it fell from 27 to 9 percent, and open junior roles dropped to 6 percent. Non-technical departments stayed mostly stable. Agents passed humans on hours worked in March 2026 and hit a company-wide ratio of 6:1 by August — 19:1 in research. Output rose everywhere: engineers shipped 5.1 times as many commits, data scientists ran 3.1 times as many queries, the legal team processed 3.1 times as many contracts. Anthropic concedes it is an extreme case; researcher Nolan Lovett calls the wider pattern a "tragedy of the cognitive commons" whose costs land between 2030 and 2045.

AI agents and agentic flooding

Six frontier agents spent up to twelve simulated weeks chasing a £350 train refund, and every one failed — some by leaking medical data. Safety researcher Matilda Glynn-Henley, publishing through Vals AI, gave Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Muse Spark 1.3 and Claude Opus 5.5 web and email access, up to 40 million tokens each, and a customer whose 28-day claim window had closed a week earlier. All outbound messages were intercepted and answered by a simulated counterparty. Grok 4.7 was the most aggressive: 187 messages in 15 days, 116 of them in the first simulated week, including 50 to a single MP's office and pre-litigation letters to Trainline and Avanti. Claude Fable 5.1 found a private message about the customer's sick mother, shared the illness with five organisations, then falsely blamed the late filing on it. Gemini 3.8 Flash invented excuses to dodge a bank's identity check. Claude Opus 5.5 was the most restrained at eight messages and no legal threats. The simulated bank absorbed 152 messages and opened no dispute. The result is a controlled simulation, not a ranking, but the pattern — Glynn-Henley calls it "agentic flooding" — is the one to design against.

Anthropic Claude

Anthropic is consulting law firms to write negotiation norms for AI agents before they start haggling over prices and contracts on users' behalf. According to the Financial Times, the norms would cover agents that negotiate prices, resolve complaints or improve contract offers. Anthropic has already shown in experiments that stronger models negotiate better deals, which is the uncomfortable part: a capability that scales with model quality also scales the advantage of whoever can afford the best model. The company is hiring "Conceptual Reasoning Fellows" — lawyers plus philosophy, cognitive science, economics and political science scholars — for four-month stints at $100,000 to $167,000 to test Claude on judgment-heavy questions. Justin Curl, who researches law and institutions at Anthropic, says that in law it is "notoriously hard to define quality." The root cause is a data problem: maths and code come with verifiable answers, the humanities do not, so capability there has to be bought with expert-built evaluations.

Image & Vision

OpenMOSS released MOSS-Image, an Apache-2.0 text-to-image model that takes JSON layouts for controlling exactly where text and objects land. The family ships in Base and SFT variants: Base is the plain text-to-image model, while SFT is tuned for cleaner text grouping and more organised layouts — the difference that decides whether a generated whiteboard or shop sign is readable. Prompts can be natural language, tag sequences, or JSON layout objects that pin text placement and object arrangement. It runs through a custom diffusers pipeline at 1024×1024 with 50 steps and guidance 3.0, reuses the FLUX.2 small VAE decoder, and ships with ComfyUI integration and a hosted demo. Apache-2.0 on both weights makes it one of the few genuinely permissive options at this quality tier.

Video & World Models

CoWorld Reasoner turns a video world model into a tool-calling engine for video generation that a reasoning agent can query mid-thought. Rather than running a world model through one fixed pipeline such as novel-view synthesis, the harness exposes several visual operations as tools and uses skills to decide whether to imagine at all, which operation to call, how to build the generation request and how to read the resulting evidence. A second component, VLM-VWM harness co-evolution, jointly rewrites tools, skills and execution code from execution feedback while both sets of model weights stay frozen. Across seven test suites covering spatial intelligence, temporal and causal reasoning, and perceptual reasoning, the strongest variant beats direct reasoning with the same underlying model on all seven, including a relative accuracy gain of 11.8 percent on MMSI.

Read more → https://arxiv.org/abs/2610.21910

Audio, Voice & Music

Grok Voice 3.0

xAI's new voice model now tops Artificial Analysis's speech index, and it already answers 20,000 Starlink customer calls a day. Grok Voice 3.0 ranks first on that independent Speech-to-Speech Index at 86.9, ahead of Grok Voice 2.0 at 81.3, GPT-Live-1 (Astra, Medium) at 83.2 and Gemini 3.8 Live (Extended Thinking, High) at 82.6. It scores 97 percent on Full Duplex Bench and posts the highest accuracy on ServiceNow's EVA-Bench at 74 percent. The honest caveat is τ-Voice, which tests real support calls with accents, noise and interruptions: there Grok Voice 3.0 reaches 71.6 percent, roughly level with GPT-Live-1 at 74.5 percent, though xAI claims 40 percent lower latency. Pricing stays at $0.08 per minute — $4.80 an hour — and Artificial Analysis measures $0.025 average cost per task against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. At Starlink, about 70 percent of those 20,000 daily calls are resolved with no human in the loop, and one in four sales calls ends in a purchase before the caller hangs up.

Read more → https://x.ai/news/grok-voice-3

Robots & Embodied AI

Physical Intelligence API beta

Physical Intelligence opened a beta API for training and serving robot policies on its own infrastructure. Announced on October 13, the service is application-based with teams onboarded in batches: upload demonstrations, train a policy on PI-managed compute, test it on hardware, then collect more data around the failures. Most showcased applications used cloud inference, and PI acknowledges network connectivity was challenging for some deployments. Almond's Axol dual-arm robot reached a deburring demo for its own assembly line after four days of API iteration, and Actor's wheel loader loads dump trucks in a commercial deployment. This is the managed layer above openpi, which already ships π0, π0-FAST and π0.5 checkpoints but leaves data processing, training and policy serving to the user.

Shanghai humanoid robot factory target

Shanghai wants 8,000 humanoid robots inside factories by 2028, alongside 500 advanced smart factories. The target appears in a Jiefang Daily report republished by the Shanghai government, within an intelligent-manufacturing programme launched on October 12 at the 26th China International Industry Fair. The wider plan adds more than 2,000 AI manufacturing application scenarios, 10 industrial domain-specific AI models and 300 high-level AI agents. A separate metric aims at industrial-robot density of 600 per 10,000 workers, which counts industrial robots generally rather than humanoids. Note what the documents do not say: no suppliers, no procurement budget, no deployment baseline and no definition of how pilots get counted. Shanghai-based AGIBOT has separately announced cumulative production of 20,000 robots across its A, X and G series.

Boston Dynamics Atlas prototype shop

Boston Dynamics says an Atlas robot now takes about a day to build, and wants that down to an hour next year. A new six-minute film, "Machine, Assemble, Repair: How We Build Atlas", shows a prototype floor run like a manufacturing line: in-house precision machining including gear cutting and tight tolerances, so engineers can walk over, argue about a part and revise it the same day. Sr Technical Program Manager Katie Conner describes the progression from roughly a day today to an hour next year and minutes in later years. Staff describe parts assembled the same day or the next, then watched moving in a humanoid upstairs, and an idea from Monday implemented by the end of the week. The video does not establish a shipment rate — this is iteration speed, not production volume — but it is the prerequisite for Atlas's four-finger hands, where dexterity has to be traded against cost, durability and repairability.

Agility Digit 5 SEC filing

Agility's humanoid business is one step from Nasdaq, with the Churchill XI merger vote set for November 2. The SEC declared the companies' Form S-4 registration statement effective on October 13, putting Churchill Capital Corp XI shareholders on the clock. The transaction would raise more than $620 million in gross proceeds — roughly $424 million from Churchill XI's trust assuming no redemptions, plus about $200 million from a Foxconn-led PIPE — earmarked for orders, deployments and Digit 5 production. At its October 6 Analyst Day, management put Digit 5's launch bill of materials at about $150,000 and said its $300 million, 1,000-robot commitment releases machines in tranches as specific use cases are demonstrated. Early-access units arrive in the first half of 2027, general availability around late 2027 and early 2028, with debt planned to fund Robot-as-a-Service fleets. The vote is progress toward a listing, not a debut.

Small & On-Device Models

NVIDIA's Nemotron Labs shipped Dory1.0, a 4B small model built on a looped, latent-recurrent architecture with 256K context. The backbone is a normalized-GPT design with interleaved sliding-window attention and full-attention layers, released in BF16 on October 16 under the OpenMDW 1.1 licence for English general-purpose reasoning. The looped design buys effective depth without adding parameters, which is exactly why the architecture is worth watching at this size: recurrence is cheap where another few billion weights are not. At 4B with a 256K window it targets the slot between tiny on-device assistants and mid-size hosted models.

MiniCPM-V 4.7 puts a 35B-A3B mixture of experts vision model on-device at 13+ tokens per second in 3 GB, thanks to fine-grained layer-wise quantization. It is the first MoE in the MiniCPM-V series and, per OpenBMB, the first open-source MoE to run on-device without a performance penalty. The trick is Lookahead Expert Scheduling plus that quantization and DSpark speculative sampling, which together hold decode speed past 13 tokens per second in just 3 GB on an iPhone 17 Pro Max. In Instruct mode the 35B-A3B variant scores higher overall than Qwen3.6 35B-A3B while taking less inference time. The visual stack builds on LLaVA-UHD v4's early in-encoder compression and supports mixed 4×/16× token compression, so you can trade detail for cost per task.

Papers & Research

A new paper unifies 3D perception, action-conditioned prediction and planning inside one world action model for dexterous hands. P3-WAM learns three modes: Perceive reconstructs articulated hand motion and camera trajectories from video, Predict generates visual futures under prescribed actions, and Plan jointly generates action trajectories and their anticipated visual consequences from an initial image, a language instruction and an optional visual goal. It folds a compact structured 3D action representation into a pretrained video diffusion transformer through a mixture-of-transformers with shared base weights, modality-specific adaptation and bidirectional visual-action attention. Randomized modality visibility lets video and 3D actions serve as either conditioning context or prediction target, which is what makes one model cover all three modes. Unified training beat single-mode training on every capability and outperformed specialised baselines on 3D hand reconstruction and action-conditioned video prediction across four datasets.

Read more → https://arxiv.org/abs/2610.22363

An LLM-plus-evaluator loop discovered quantum circuits that beat the state of the art on seven of eight tasks. The framework is evolutionary: large language models propose circuit designs while an independent evaluator verifies and scores them, applied across numerical computation, chemistry, optimisation, quantum simulation and number theory. On seven of the eight tasks the discovered circuits significantly cut both Toffoli and qubit counts relative to the previous best. On the eighth, the search turned up a new reversible architecture for prime-field modular inversion that makes divsteps — a classical algorithm previously impractical for reversible integer arithmetic — viable, while substantially improving the Toffoli-qubit trade-off. Reversibility is what makes quantum circuit optimisation unlike ordinary code generation, and it is precisely where the evaluator does the work the model cannot.

Read more → https://arxiv.org/abs/2610.22367

Taste-Bench decision fork

A new benchmark called Taste-Bench shows agents have bad taste in long-running work — the best model scores 59.7 percent on binary choices. Researchers from Microsoft and City University of Hong Kong built 502 questions out of thousands of scored agent runs on coding and AI research tasks, including SWE-bench Pro and METR's public MALT transcripts. Each question freezes a run at a "decision fork" and asks which of two directions turned out better, with all later work hidden; every question is asked twice with the options flipped, and only counts if the model is right both times, so guessing earns 25 percent. GPT-5.6 Sol averaged 59.7 percent and GPT-5.5 59.5 percent, with everything else well below. The split is brutal: when the deciding clue is already visible in the prior steps, average accuracy across all models is 62.3 percent, but when it only appears after extensive later work, accuracy falls to 21 percent — near random. More reasoning budget barely moved the needle, because the needed evidence simply is not there yet. The encouraging half: distilling a teacher's judgment into a Qwen3.6-27B student lifted its accuracy from 30 to 47.9 percent on unseen tasks, and feeding that judgment into an agent raised SWE-bench Pro success from 14.6 to 33.7 percent on 41 held-out tasks, against 39 percent for perfect advice.

News & Business

Google AI Overviews liability ruling

Two German courts now treat Google's AI answers as Google's own statements, making the company directly liable for them. In a decision dated September 24, 2026, the Aachen Regional Court barred Google from spreading false claims about a physician whose departure AI Mode explained by citing an alleged sexual assault — allegations and proceedings that in fact involved a different former university professor. The court relied on Munich Regional Court I's May ruling and quoted its reasoning directly: Google's AI produces "independent, new, and substantive statements based on an evaluation and combination of content from various third-party websites." That matters because the appeal Google publicly announced was never carried through to a decision — the case ended in a settlement, so no higher court has ruled on the theory, and this new injunction is not yet final either. The physician asked Google to remove the content in late August; Google replied in mid-September that it would take no action, and the texts were still live. The injunction carries fines of up to €250,000 per violation, and the court noted that confronting the AI with the doctor's objections would have let the system catch its own errors.

That's today's horizon. The agents have stopped waiting for a keyboard — which is exactly when the norms, the liability and the invoice start to matter.