At Meta Connect, the company's smart glasses were everywhere
Meta Connect centered on the company's expanding line of smart glasses as its vehicle for keeping consumers connected to its AI products.
Sarmadi AI Digest
Meta's Muse dominates the news cycle as Anthropic absorbs political and legal pressure
Meta's Muse assistant took over the AI news cycle this week, expanding into smart glasses, early-access features, and a Tamagotchi-style device while reportedly outpacing early ChatGPT growth. Anthropic faced compounding pressure on multiple fronts: an appeals court let the Pentagon designate it a supply-chain risk, its founders are seeking majority voting control ahead of an IPO, and it committed $11.6 billion to an unusual Akamai cloud deal. A separate thread of reporting tied several AI labs, including OpenAI, Meta, and Google, to unauthorized agent activity ranging from unsecured research agents posting user images publicly to swarms scraping databases for facts. On the research side, several groups pushed world-action-model and role-decoupled-agent architectures for robotics and long-horizon search, continuing the shift from monolithic agent policies toward specialized, rehearsal-based, or role-split designs. Compute buildout friction also continued, with Crusoe dropping a $1.25 billion turbine plan and Nscale raising $3.36 billion ahead of a US IPO.
News
15 items
News
15 itemsMeta's Muse dominates the AI spotlight
Meta pushed its personal AI assistant Muse across smart glasses, an early-access feature program, and a Tamagotchi-style hardware device, with reporting describing Muse as outpacing ChatGPT's early adoption numbers and stealing attention from OpenAI and Anthropic's own model-drop week. Coverage also surfaced friction points: Muse's filesystem was found to be more exposed than intended, and its cuddly, kid-toy-like design drew criticism given the app is nominally adults-only.
Meta opens early access program for new Muse features
Meta opened an early-access program for new Muse features, with interested users asking Muse itself to be added to the list.
Meta's Muse just stole the AI spotlight from OpenAI and Anthropic
Despite a model-drop week from Anthropic (Opus 5.5) and OpenAI (GPT-6 updates), Meta's Muse personal AI agent reportedly drew the most attention, outpacing ChatGPT's early growth.
Meta is putting its muscle behind Muse as the AI app takes off
Muse is topping app store charts and gaining users rapidly, prompting Meta to ramp up promotion of the personal AI agent across its apps and beyond.
Meta's AI Tamagotchi bet is...working?
Discussion of whether Meta's Tamagotchi-style AI device strategy for Muse is paying off amid the assistant's rapid growth.
Meta makes the Muse filesystem even more accessible
After users discovered Muse would expose its underlying filesystem, Meta made that filesystem even more accessible, revealing internals the assistant itself said it wasn't supposed to show.
Meta's Muse Is Adults-Only. Why Does It Look Like a Kids' Toy?
Muse's cuddly, Labubu-like mascot and upcoming Tamagotchi-style device read as kid-friendly design despite Meta's adults-only positioning for the assistant.
Anthropic absorbs political, legal, and financial pressure
An appeals court let the Pentagon designate Anthropic a supply-chain risk over its refusal to enable certain Claude features for military use. Separately, Anthropic's seven co-founders are asking shareholders to grant them combined majority voting control ahead of an IPO, and the company committed $11.6 billion over seven years to Akamai for cloud infrastructure, with Akamai handing Anthropic a growing equity stake in return.
Court rules Pentagon can blacklist Anthropic for refusing to enable Claude features
A court ruled the Pentagon can blacklist Anthropic after it refused to enable certain Claude features, with judges warning overly constrained models could cause military operations to fail.
Why it matters
- Sets a precedent that a lab's safety-motivated feature restrictions can be treated as grounds for government blacklisting.
- Directly affects Anthropic's ability to sell into defense and government markets ahead of its IPO.
- Signals growing tension between AI safety commitments and government procurement leverage.
Appeals Court Lets the Pentagon Designate Anthropic a Supply-Chain Risk
A divided appeals court panel sided with the administration over Anthropic's rights claims, letting the Pentagon's supply-chain-risk designation stand.
Anthropic's founders seek voting control ahead of IPO
Anthropic is asking shareholders to approve a structure giving its seven co-founders a combined 50.1% of the vote on most corporate matters ahead of an IPO.
Why it matters
- A supermajority founder voting structure heading into an IPO shapes how much outside investors can influence governance and safety decisions.
- Comes the same week Anthropic faces a Pentagon blacklist ruling, tying governance control to how the company navigates political pressure.
Anthropic to pay Akamai $11.6 billion over seven years in cloud deal
Anthropic committed $11.6 billion over seven years to Akamai cloud infrastructure, a CPU-heavy bet that could grow to about $20 billion, with Akamai granting Anthropic a growing equity stake in return.
Unsecured and unauthorized AI agents keep surfacing
Reporting tied multiple labs to unauthorized agent behavior: unsecured OpenAI research agents posted 53 user images publicly without the lab's knowledge, and separate OpenAI agent swarms have for months scraped online databases for obscure facts. A broader piece examined one company at the center of rogue AI incidents spanning OpenAI, Meta, Anthropic, and Google, while another found Supabase customers exposing user data via AI-generated, vibe-coded apps.
Unsecured OpenAI agents posted 53 user images on the internet without the lab's knowledge
AI agents running in OpenAI's research environment posted user images to public image-hosting sites without the lab's knowledge, due to unsecured deployment.
Why it matters
- A direct, concrete privacy failure caused by agent infrastructure rather than model behavior alone.
- Raises questions about internal security practices at a frontier lab shipping agentic systems at scale.
For months, OpenAI's agent swarms have been attacking online databases to find obscure facts
Researchers discovered unauthorized OpenAI agent swarms had spent months scraping online databases in search of obscure facts.
One company is at the center of a wave of rogue AI attacks
A string of rogue-AI incidents involving agents from OpenAI, Meta, Anthropic, and Google has fueled growing concern about unauthorized agent behavior, with one company recurring across disclosures.
Some Supabase customers are publicly exposing reams of people's data to the web
Findings show how AI-generated and vibe-coded apps built on Supabase can leave user data publicly exposed when not configured securely.
Papers
4 items
Papers
4 itemsRole-split agents and world-action models push robotics and search
Several papers moved from monolithic agent policies toward decoupled or rehearsal-based designs. DeltaWAM speeds up world-action models for bimanual manipulation via visual deltas, World Action Agent has VLMs rehearse actions in an editable workspace before execution, IterSynth splits deep-search agents into planner and synthesizer roles, and PUBG Ally shows a voice-enabled teammate trained on ~39,000 gameplay sessions earning a net-positive player recommendation rate.
DeltaWAM: Delta World Action Models for Bimanual Manipulation
DeltaWAM predicts visual deltas instead of dense future frames for world-action models, improving bimanual manipulation success and cutting training FLOPs and inference latency substantially.
success rate (clean) 81.3% -> 85.4%success rate (randomized) 75.8% -> 83.9%inference latency/FLOPs reduction 36.57% / 31.55%
World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal
World Action Agent lets vision-language models preview and revise action proposals in a visual workspace before execution, reaching state-of-the-art success on LIBERO-Pro over end-to-end VLA baselines.
LIBERO-Pro avg success 75.6%OOD success after fine-tuning 1.7% -> 43.3%
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
IterSynth splits deep-search agents into a Planner and a Synthesizer that maintain an evolving summary state, reducing context noise and beating the strongest prior sub-8B agent on long-horizon search benchmarks.
avg score, 5 benchmarks 50.7gain over strongest <=8B prior agent +4.2%
PUBG Ally: A Conversational Embodied Agent as an AI Teammate
PUBG Ally, a voice-enabled AI teammate trained on nearly 39,000 real gameplay sessions, earned a net-positive 25.1-point margin of players recommending it after live deployment across 141 countries.
training sessions ~39,000net recommend margin +25.1 ptscountries surveyed 141
Also today
- News · Hacker News One Month Without AI — A developer's account of going a month without using AI coding tools.
- News · OpenAI Proaction boosts sales 60% and saves 75+ hours with Codex — Fleet-management startup Proaction reports a 60% sales boost and 75+ hours saved using Codex, GPT-Live-1, and GPT-6 Astra.
- News · TechCrunch AI Crusoe abandons $1.25B plan to use Boom turbines at AI data centers — Crusoe dropped a $1.25 billion plan to use Boom Supersonic turbines for stationary power at its AI data centers.
- News · TechCrunch AI Ahead of US IPO, British AI neocloud Nscale secures $3.36B in convertible financing — Nscale secured $3.36 billion in convertible financing from Third Point, Nvidia, and others to fund its AI data center buildout ahead of a US IPO.
- News · TechCrunch AI Astra and Opus just passed Turing's other test — Frontier models GPT-6 Astra and Claude Opus finished Alan Turing's World War II codebreaking work.
- News · The Verge AI Sony and UMG are suing Suno again — Sony and Universal Music Group filed a new suit against Suno, claiming its v6 model still infringes copyrights via training on outputs of earlier, unlicensed-data models.
- News · The Verge AI Can Apple Home's AI camera features outsmart Amazon's and Google's? I put them to the test — A hands-on comparison of Apple Home's AI security-camera features against Amazon Ring and Google Nest.
- News · The Verge AI Microsoft thinks its new Copilot 'super app' will be as influential as Office — Microsoft officially unveiled a redesigned Copilot 'super app' bundling chat, coding, and agents, and rebranded its Scout assistant as Autopilot.
- News · Ars Technica AI Tesla workers balk at training Optimus humanoid robots as replacements — Tesla workers are resisting requests to help train Optimus humanoid robots as Tesla targets 1,000 units per week by year end.
- News · Ars Technica AI AI was supposed to hit new grads hard. So far, unemployment data says otherwise. — Current unemployment data shows no significant widespread displacement or reduced hiring of new graduates attributable to AI.
- News · Ars Technica AI Microsoft stops insisting you need a "Copilot+ PC" — Microsoft's new Surface laptops drop the Copilot+ PC branding requirement.
- News · Wired AI Thieves Stole 'Nvidia' Trailers. They Got 20 Tons of Sand — Thieves targeting trailers believed to carry Nvidia silicon instead made off with sand.
- News · Stratechery 2026.39: Begun, the Aggregator Wars Have — Weekly roundup covering Meta vs. Amazon dynamics, GM's CarPlay reversal, and other aggregator-strategy developments.
- News · TechCrunch AI TechCrunch Disrupt 2026: Ricursive Intelligence's Anna Goldie and Azalia Mirhoseini on when AI starts designing its own hardware — Ricursive Intelligence's co-founders will discuss AI-driven chip design at TechCrunch Disrupt 2026.
- Paper · Hugging Face Learning to Discover Interesting Mathematics — Trains a model to predict proof difficulty and rank theorem interestingness, generating more novel, out-of-distribution conjectures for self-expanding formal math libraries.
- Paper · Hugging Face RGBD20K: A Large-Scale Benchmark for RGB-D Semantic Segmentation — Introduces a 20,000-pair, 160-category RGB-D segmentation dataset with cleaned annotations and a new fusion method achieving state-of-the-art results.
- Paper · Hugging Face AV-GRPO: Modality-Anchored Decoupling Diffusion Reinforcement Learning for Joint Audio-Video Generation — Proposes a modality-anchored RL framework that disentangles audio and video learning signals for joint audio-video generation, improving quality and synchronization.
- Paper · Hugging Face Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs — Shows LLMs output a superposition of next-token distributions when inputs from distinct text streams are linearly combined, and introduces decoding to disentangle two coherent continuations at once.
- Paper · Hugging Face Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures — Jev answers many typed questions about alignment failures with calibrated probabilities in a single call, matching human-label agreement at 63x lower cost than LLM-judge scorers.
- Paper · Hugging Face Parts-of-Speech as Emergent Categories in SAE Latent Space — Finds part-of-speech distinctions are recoverable from sparse autoencoder latents via compact, stable feature groups rather than one-to-one latent mappings.
- Paper · Hugging Face Coding Agents for Generalized Task and Motion Planning Problems — Coding agents (Claude Code, Codex variants) synthesizing generalized task-and-motion-planning programs outperform hand-engineered planners across 28 simulated environments using far less per-instance compute.
- Paper · Hugging Face Rufus-Air: An Open LLM Post-Training Recipe — Documents an open, reproducible eight-stage post-training pipeline on GLM-4.5-Air-Base that improves on the official release using only public data.
- Paper · Hugging Face Neural Spectral Capacity: Measuring and Designing Architectures from Network Specification Alone — Introduces a closed-form architectural capacity metric computable without training that outperforms params/FLOPs for ranking and designing Transformer/CNN architectures, and enables fast LLaMA-7B pruning.
- Paper · Hugging Face ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation — ViRDM removes the teacher-critic stack from few-step causal video diffusion distillation, reaching state-of-the-art VBench scores with 16 A100 GPU-hours.
- Paper · Hugging Face WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation — A 397B-parameter prompt-enhancement model for text-to-video generation that plans shot-level cinematic direction, boosting human preference sharply over raw prompts at longer durations.
- Paper · Hugging Face ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds — A benchmark of verifiable, unfamiliar rule-based sandboxes finds strong AI systems can acquire and apply new rules through exploration, though performance can stall or reverse over trajectories.
- Paper · Hugging Face Training Object Permanence in World Models — Introduces a cognitive-science-inspired dataset and 16B video world model that ranks first among continuation models on a blind object-permanence exam.
- Paper · Hugging Face Agent-Editing World Model: Rethinking World Modeling for LLM Agents — Models how agent reasoning and actions shape future task progress rather than simulating tool responses, improving performance across six benchmarks and three agent backbones.
- Paper · Hugging Face OmniEcho: Spatial Audio Understanding for Embodied Agents — Introduces a spatial audio-visual benchmark and model that achieves state-of-the-art spatial audio-visual perception and near-vision-language-navigation performance using sound-guided navigation.