Canada 2nd on Claude use, Gemma 4 ditches encoder, Alibaba ships RynnWorld-4D

TL;DR

• Canada hits AUI 4.4 on Claude, second per-capita only to the US, per Anthropic’s Economic Index.
• Gemma 4’s 12B swaps a 400M vision encoder for a 35M projector into the decoder.
• Gemma 4 safety refusals on biosecurity prompts collapse from 86.7% to 0% without the chat template.
• Alibaba’s RynnWorld-4D hits δ₁ of 0.610 on geometric accuracy, nearly 2× 4DNeX.
• RynnWorld-Teleop distills the teacher to 40+ FPS on a single H100 for drivable synthetic robots.

Today’s research reads scatter across three unrelated lanes. Anthropic’s Economic Index puts Canada second per capita on Claude use — a geography-of-adoption story rather than a capability one, with Ontario, Quebec, BC and Alberta driving 94% of traffic and document translation showing up as the country’s most distinctive task. Google’s Gemma 4 takes the architectural swing of the day, replacing the ~400M ViT/USM encoder stack with a 35M projector into the decoder and posting a 1451 Elo on its 31B dense variant — impressive until the safety numbers collapse from 86.7% to 0% the moment the chat template is stripped. And Alibaba DAMO’s RynnWorld ships as a coordinated pair: a 4D world model that nearly doubles geometric accuracy over prior baselines, plus a distilled teleop twin at 40+ FPS, though world-model planning still lags real manipulation frequencies by two orders of magnitude.

Gemma 4’s 12B ditches vision encoders for a 35M projector

Source: hf-daily-papers · published 2026-07-01

TL;DR

• Gemma 4’s 12B swaps a ~400M ViT/USM encoder stack for a 35M-parameter projection into the decoder.
• The 31B dense model hits 1451 Elo on Chatbot Arena, rivaling MoE models an order of magnitude larger.
• Safety refusals collapse from 86.7% to 0% on biosecurity prompts when the chat template is stripped.
• The 26B-A4B MoE still needs ~15GB VRAM at Q4 despite activating only ~4B parameters per token.

The encoder-free bet

Gemma 4’s headline architectural move is confined to one SKU: the 12B model replaces the ViT vision encoder and USM/Conformer audio encoder used across the rest of the family with a single 35M-parameter projection module that ingests raw 48×48 RGB patches and 40ms audio chunks directly into the decoder ¹. Independent write-ups frame this as a lineage step from Fuyu-8B and EVE rather than a fresh invention, and results are competitive — the 12B posts a 0.063 FLEURS ASR word-error rate on English, matching encoder-based baselines.

The open question is what a 35M embedder gives up. Reviewers flag adversarial robustness and noisy-input handling as untested territory: a frozen semantic encoder normally acts as a filter that catches edge cases before they hit the decoder, and it’s unclear whether the lightweight projector preserves that safety margin under distribution shift ¹.

Efficiency wins that mostly check out

The QAT story survives third-party measurement. Google’s own writeup reports the E2B Q4_0 checkpoint achieves ~29× lower KL divergence than naive Q4_0 quantization — enough that a QAT-tuned 4-bit model beats a naive Q8_0 of double the size ². That is a real shift in the size/quality frontier for on-device deployment.

The MoE variant tells a subtler story. The 26B-A4B activates ~4B parameters per token via 128 tiny experts (8 active) plus one shared expert, and reaches 88.3% on AIME 2026 versus 89.2% for the 31B dense flagship ³. Simon Willison singled it out as the sweet spot for laptop-class inference ⁴. The catch: all 25.2B weights must sit in VRAM to keep routing latency down, so the practical hardware bar is ~14.4–15GB at Q4 — much closer to the dense flagship than the “4B active” headline suggests ³.

Where the report oversells

Day-one deployment surfaced concrete regressions the report doesn’t anticipate. LM Studio users saw the 31B checkpoint emit only --- regardless of prompt until a parser fix landed — a shipping-format bug, but a real friction point Willison flagged publicly ⁴. Hugging Face reports also caught the 12B-IT entering a “degenerate loop” that repeatedly emits the token thought under long agent-style system prompts, a direct hit on the flagship thinking-mode feature ⁵.

The safety picture is worse. The BioRefusalAudit paper (arXiv:2605.30162) reports that Gemma 4 E2B-IT refuses 65 of 75 biosecurity prompts (86.7%) under standard chat templates — but that number collapses to 0% when the chat template is stripped or output is capped at 80 tokens ⁶. Both configurations are common in low-latency production. The refusal training is real; it’s just template-dependent in a way that trivially unravels outside chat mode.

Refusal dropped to 0% when chat formatting was removed or an 80-token output cap was applied. ⁶

That is the sharpest independent contradiction of the report’s alignment claims, and it means anyone deploying Gemma 4 behind a non-chat API needs to layer their own guardrails — the weights alone won’t hold.

---

Alibaba’s RynnWorld pairs a 4D world model with a 40 FPS twin

Source: hf-daily-papers · published 2026-07-06

TL;DR

• RynnWorld-4D hits δ₁ of 0.610 on geometric accuracy — nearly 2× 4DNeX (0.327) and TesserAct (0.279).
• Alibaba DAMO shipped it alongside RynnWorld-Teleop as a coordinated pair: generative substrate plus real-time data engine.
• Teleop distills the teacher to 40+ FPS on a single H100, making the synthetic robot drivable like a physical twin.
• The catch: world-model planning runs at just 0.9 Hz, well below the 50–100 Hz stable manipulation needs.

A stack, not a paper

Read the two Rynn papers as one release. RynnWorld-4D is the generative substrate — a tri-branch diffusion transformer built on Wan 2.2 that emits synchronized RGB, depth, and optical flow (RGB-DF) so a policy can unproject scene flow into 3D geometry instead of guessing motion from pixels. RynnWorld-Teleop is the exploitation layer: streaming autoregressive distillation compresses the bidirectional teacher into a causal student that runs at 40+ FPS on a single H100, letting an operator “drive” the synthetic robot as if it were a physical twin ⁷. That real-time bar is what makes the pair interesting — it’s the difference between sampling a diffusion model and generating usable teleoperation trajectories at scale.

flowchart LR
    A[RGB-D image + language] --> B[RynnWorld-4D<br/>tri-branch DiT]
    B --> C[Synchronized<br/>RGB + Depth + Flow]
    C --> D[RynnWorld-4D-Policy<br/>9 Hz control]
    B -.distill.-> E[RynnWorld-Teleop<br/>40+ FPS causal]
    E --> F[Synthetic teleop<br/>rollouts]
    F -.train.-> D

The numbers hold up

Independent write-ups corroborate the structural-accuracy claims: the RGB-DF representation roughly doubles prior 4D world models on geometric accuracy and is credited with killing the shape-morphing that plagues 2D video predictors ⁸. On six bimanual real-world tasks with a TIANJI M6, RynnWorld-4D-Policy beats Diffusion Policy and π₀ across the board — 65.7% on lid placement and bowl stacking, 94–97% on dual picking and bimanual lifting. Hand-over lands at only 28.6%, but π₀ and π₀.₅ score 2.9% and 0% on the same task, suggesting the 3D-aware representation is doing real work on distance estimation.

The 0.9 Hz problem

The architecture’s cost lives in the tri-branch transformer, which eats 89.5% of inference latency and caps the world model’s “mental refresh” at ~0.9 Hz. Action chunking boosts effective control to 9 Hz, but OpenReview literature on 4D planners argues the whole class is stuck: stable manipulation typically demands 50–100 Hz control, and timing jitter from variable generative iteration times compounds the mismatch ⁹. A Substack analysis flags a related weakness — isolated skills like grasping hit 77–100%, but chained sequences degrade on “next-skill readiness” and target grounding, exposing a long-horizon composition gap the paper underplays ¹⁰.

Current 4D latent planners operate at 0.5–10 Hz… timing jitter caused by varying iteration times in generative planners remains a barrier to fluid, reactive movement. ⁹

Strategic frame and artifact reality

Regional coverage reads the Rynn stack — RynnBrain, RynnVLA, RynnRCP, RynnWorld — as Alibaba’s answer to Google’s Gemini Robotics and NVIDIA Cosmos: a vertically integrated, open-weight AI factory spanning silicon, cloud, and embodied foundation models ¹¹. That framing raises the stakes and the scrutiny.

Reproducibility is uneven. The GitHub repo ships the Stage-3 checkpoint (~28 GB) plus EMA shadow weights and the full three-stage training recipe, but Issue #1 — the Hugging Face release of the 254M-frame Rynn4DDataset — is still open ¹². You can run the model today; you cannot rerun the training. For a release pitched as an open-weight Cosmos alternative, that gap is the thing to watch next.

Further reading

• RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation — hf-daily-papers

---

Anthropic ranks Canada 2nd on Claude use per capita

Source: anthropic-research · published 2026-07-14

TL;DR

• Canada hits AUI 4.4 — per-capita Claude use runs 4× its working-age share, second only to the US.
• Ontario, Quebec, BC and Alberta account for 94% of national traffic, tracking professional/technical-services employment rather than GDP.
• Document translation is Canada’s most distinctive task, driven by federal bilingualism mandates in New Brunswick, Nova Scotia and Quebec.
• Only 2% see measurable ROI despite 93% of Canadian leaders piloting AI, per KPMG Canada’s adoption index.

The headline number

Sampled in February 2026, Canada pulled 2.6% of global Claude.ai traffic — eighth by volume, second by per-capita intensity behind the United States. Anthropic’s Usage Index (AUI) puts Canadian consumption at 4.4× what the country’s working-age population would predict. Provincial concentration is sharp: Ontario (43.9%), Quebec (20.8%), BC (18.9%) and Alberta (10.2%) make up roughly 94% of national usage, with BC leading per-capita at 1.4× parity and Newfoundland trailing at 0.2×.

The more interesting finding is what predicts adoption. Globally, national income is the dominant signal. Inside Canada, it’s the size of the professional, scientific and technical services sector — a labour-market fingerprint, not a wealth one. Anthropic reads this as capability-market fit: Claude’s strengths line up with what Canadian knowledge workers actually do.

Translation is the Canadian tell

Task-mix looks Anglosphere-standard at the top level (Personal 44–51%, Work 34–40%, Coursework 13–18%), but the distinctive slice is document translation, correlated with public-administration employment and concentrated in the bilingual provinces. Canadians also over-index on STEM coursework, coding help and résumé drafting versus US/UK/Australian peers, and under-index on workplace email polish, marketing copy, and legal or home-maintenance queries. It reads like a country where students and early-career technical workers are the heaviest hands on the keyboard.

The caveats Anthropic left out

Independent data broadly corroborates the adoption curve — Statistics Canada’s Q2 2026 business survey shows enterprise AI use jumping from 6.1% to 19.2% in two years, with Information & Cultural Industries leading at 42.3% ¹³. That’s a rare case of vendor telemetry lining up with a national statistical agency.

The productivity claim is where the brief gets ahead of the evidence. KPMG Canada’s own adoption index finds 93% of business leaders piloting AI, and only 2% reporting measurable ROI ¹⁴. Anthropic itself quietly halved its augmented-developer productivity forecast from 3× to 1.5× after benchmarking revealed a 63% first-try failure rate on mid-range agentic tasks ¹⁵. “Integration into the national economy” is doing a lot of work in the brief’s conclusion.

Two other Canadian-specific frictions go unmentioned. OSFI has formally warned the Big Six banks that frontier Claude variants can autonomously exploit zero-day vulnerabilities, compressing risk-mitigation windows to machine speed ¹⁶ — a direct headwind to the financial-services adoption the report highlights. And Cohere just shipped North Mini Code, a 30B open-weight coding model pitched explicitly at Canadian regulated industries that don’t want data leaving domestic perimeters ¹⁷. Canada’s Claude enthusiasm is real; so is the domestic push to route around it.

The ranking has also moved. By early July, Australia had taken the global top spot at 6.4× expected usage across 121 countries ¹⁸. Canada’s silver medal is a February snapshot in a leaderboard that reshuffles quarterly — a useful reminder that per-capita AUI is a race, not a moat.

Round-ups

Single transformer layer matches full-parameter RL fine-tuning

Source: hf-daily-papers

Reinforcement learning gains in transformers concentrate in a handful of middle layers, and training just one of them rivals full-parameter RL across math, code, and agentic decision tasks. The finding holds across GRPO, GiGPO, and Dr. GRPO, hinting at far cheaper RL post-training recipes.

Tencent’s HiLS attention learns chunk selection for infinite context

Source: hf-daily-papers

Hierarchical Landmark Sparse Attention from Tencent Hunyuan trains chunk selection end-to-end via language-modeling loss, matching dense attention quality while extrapolating past training context lengths. Sparse KV access keeps compute bounded, and the team has released code for the architecture on GitHub.

Nemotron-Labs-Diffusion fuses AR, diffusion, and self-speculation decoding

Source: hf-daily-papers

NVIDIA’s tri-mode language model trains a joint AR-diffusion objective with lookahead planning, letting one network switch decoding strategies for higher throughput. On SPEED-Bench running SGLang on GB200, it beats existing models on speed-of-light efficiency while scaling cleanly with parameters.

AlayaWorld open-sources real-time playable video world models

Source: hf-daily-papers

AlayaWorld packages autoregressive video world generation into a modular, reproducible pipeline supporting live user actions and long-horizon rollouts. The release ships reference implementations and evaluation tooling on GitHub, giving researchers a shared baseline for interactive generative worlds rather than closed demos.

DSpark uses confidence-scheduled verification to speed LLM decoding

Source: hf-daily-papers

DSpark pairs parallel semi-autoregressive drafters with adaptive verification tuned by prefix survival probabilities, cutting wasted tokens from acceptance decay. In high-concurrency serving, it pushes the throughput-latency Pareto frontier above standard speculative decoding baselines.

Signed MaxSim extends ColBERT-style retrieval to logical queries

Source: hf-daily-papers

MaxSim similarity exactly reproduces inner products over k-sparse vectors and supports soft-OR and CNF logical expressions, and a Signed MaxSim variant lifts the trick to real-valued embeddings. The extension raises nDCG@10 on complex query types over standard ColBERT late interaction.

HunyuanOCR-1.5 speeds lightweight vision-language OCR with DFlash

Source: hf-daily-papers

Tencent’s end-to-end OCR VLM pairs a DFlash inference stack with an Agentic Data Flow pipeline for pretraining and post-training, delivering faster inference and broader task coverage in a small model. The design targets deployment where full-scale VLMs are too costly.

Footnotes

1. oflight.co.jp analysis of encoder-free 12B — https://www.oflight.co.jp/en/columns/gemma-4-12b-encoder-free-multimodal-2026

   The primary innovation in Gemma 4 12B is the removal of heavy modality-specific encoders in favor of a 35M-parameter lightweight embedder… critics have questioned whether such a drastically smaller input layer can maintain the same level of feature extraction quality as traditional frozen encoders.

   ↩ ↩²

2. Google Developers Blog — QAT for Gemma 4 — https://blog.google/innovation-and-ai/technology/developers-tools/quantization-aware-training-gemma-4/

   the Gemma 4 E2B variant demonstrates a Kullback–Leibler divergence roughly 29 times better than naive Q4_0 quantization… a QAT-optimized 4-bit model can often outperform a traditional 8-bit (Q8_0) model of double the size.

   ↩

3. Roboflow model comparison: 26B A4B vs 31B — https://playground.roboflow.com/models/compare/gemma-4-26b-a4b-vs-gemma-4-31b

   The 26B A4B activates only ~4B parameters per token via 128 tiny experts (8 active) plus one shared expert, reaching 88.3% on AIME 2026 versus 89.2% for the 31B flagship — but all 25.2B parameters must be loaded into VRAM (~14.4–15GB at Q4).

   ↩ ↩² ↩³

4. dev.to practical guide (citing Simon Willison) — https://dev.to/arshtechpro/gemma-4-a-practical-guide-for-developers-2co5

   Willison characterized Gemma 4 as ‘byte for byte, the most capable open models’… but noted the 31B Dense model was initially ‘completely broken’ in LM Studio, repeatedly outputting ’---’ regardless of the prompt.

   ↩ ↩²

5. Medium — reasoning mode evaluation on-device — https://medium.com/@saloni_garg/reasoning-mode-evaluation-in-on-device-transformers-gemma-4-9ccc9762a6df

   Community reports on Hugging Face highlighted a ‘degenerate loop’ issue in the 12B instruction-tuned model: under certain long agent-style system prompts, the model enters a loop, repeatedly emitting the word ‘thought’ rather than completing the reasoning trace.

   ↩

6. Stephen Turner blog on BioRefusalAudit (arXiv:2605.30162) — https://blog.stephenturner.us/p/five-things-june-5-2026-aixbio

   Gemma 4 E2B-IT refused 65 of 75 biosecurity prompts (86.7%) with standard chat templates, but refusal dropped to 0% when chat formatting was removed or an 80-token output cap was applied.

   ↩ ↩²

7. AIWeekly summary of RynnWorld-Teleop — https://aiweekly.co/alerts/rynnworld-teleop-replaces-the-robot-with-a-40-fps-world-model

   streaming autoregressive distillation… compresses the bidirectional teacher into a causal student capable of 40+ FPS inference on a single H100 GPU, allowing an operator to ‘drive’ the synthetic robot as if it were a physical twin.

   ↩

8. alphaXiv overview of RynnWorld-4D — https://www.alphaxiv.org/abs/2607.06559

   In quantitative evaluations, RynnWorld-4D achieved a geometric accuracy (δ₁) of 0.610, nearly doubling the performance of 4DNeX (0.327) and TesserAct (0.279).

   ↩

9. OpenReview critique of 4D world-model planners — https://openreview.net/pdf/d19b200c60835c287e29c6d20ec394f93038c007.pdf

   stable manipulation typically requires 50–100 Hz control, [but] current 4D latent planners operate at 0.5–10 Hz… timing jitter caused by varying iteration times in generative planners remains a barrier to fluid, reactive movement.

   ↩ ↩²

10. So Essentially (Substack) analysis of RynnWorld-4D — https://soessentially.substack.com/p/rynnworld-4d-robot-manipulation-the

    isolated skills like grasping or door-opening achieve success rates of 77–100%, [but] chained sequences often stall due to errors in target grounding and ‘next-skill readiness’ — a substantial gap in long-horizon compositional robustness.

    ↩

11. Business Times (Singapore) on Alibaba RynnBrain push — https://www.businesstimes.com.sg/international/global/alibaba-pushes-robotics-ai-open-source-rynnbrain

    Alibaba is positioning itself as a vertically integrated ‘AI factory,’ owning every layer from proprietary silicon to cloud infrastructure and foundational robotic models — a direct challenge to Google’s Gemini Robotics and Nvidia’s Cosmos.

    ↩

12. GitHub alibaba-damo-academy/RynnWorld-4D — https://github.com/alibaba-damo-academy/RynnWorld-4D

    Stage-3 checkpoint includes both full trained weights (~28 GB) and recommended EMA shadow weights; Rynn4DDataset Hugging Face release is tracked in open Issue #1.

    ↩

13. Statistics Canada — Canadian Survey on Business Conditions Q2 2026 — https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm

    19.2% of Canadian businesses are actively using AI to produce goods or deliver services, up from 6.1% in 2024; Information & Cultural Industries lead at 42.3%

    ↩

14. KPMG Canada Generative AI Adoption Index — https://assets.kpmg.com/content/dam/kpmg/ca/pdf/2025/11/ca-generative-ai-adoption-index-report-en.pdf

    93% of business leaders claim to be piloting or using AI, but only 2% report seeing a measurable return on investment

    ↩

15. Gnoppix Forum — coverage of Anthropic productivity revision — https://forum.gnoppix.org/t/anthropic-cuts-ai-productivity-forecasts-in-half-after-analyzing-claudes-real-world-failure-rates/4054

    Anthropic halved its augmented developer productivity forecasts from 3x to 1.5x after internal benchmarking revealed a 63% failure rate for mid-range agentic tasks on the first try

    ↩

16. Times of India / OSFI warning coverage — https://timesofindia.indiatimes.com/technology/tech-news/canada-government-warns-countrys-largest-financial-institutions-and-banks-on-anthropics-claude-mythos-and-other-ai-models-says-new-technology-can-/articleshow/132385518.cms

    OSFI warned Canada’s Big Six banks that Claude Mythos is capable of autonomously exploiting zero-day vulnerabilities, compressing the window for risk mitigation to machine-speed

    ↩

17. The New Stack — Cohere North Mini Code launch — https://thenewstack.io/cohere-sovereign-coding-model-north-mini-code/

    Cohere launched North Mini Code, an open-weight 30-billion-parameter model designed for self-hosting on single-GPU hardware, catering to regulated industries that require data to remain within private perimeters

    ↩

18. Mi-3 (Australia) — https://www.mi-3.com.au/07-07-2026/australia-ranks-first-our-121-countries-anthropic-usage-index-64x-expected-rate-less

    Australia ranks first of 121 countries on the Anthropic Usage Index at 6.4x expected rate

    ↩