Apple tightens macOS full-disk access as AI agents face scrutiny — arstechnica.com

Sean Parker Puts Music at the Center of Stability AI's Restart — techcrunch.com

Meta opens its Muse AI agent to DIY hardware builders — theverge.com

Amazon Pledges $1 Billion to Data Center Communities — Critics Call It Damage Control — arstechnica.com

Netflix ends its era of star director deals as strategy shifts toward cheaper content — theverge.com

Apple tightens macOS full-disk access as AI agents face scrutiny — arstechnica.com

Sean Parker Puts Music at the Center of Stability AI's Restart — techcrunch.com

Meta opens its Muse AI agent to DIY hardware builders — theverge.com

Amazon Pledges $1 Billion to Data Center Communities — Critics Call It Damage Control — arstechnica.com

Netflix ends its era of star director deals as strategy shifts toward cheaper content — theverge.com

Automated discovery

Google Research Cogentic produces new proofs on five open problems in theoretical computer science

Cogentic, a multi-agent system from Google Research, started from bare problem statements with no expert hints and produced five new verified results on open problems in online learning, auction theory and mechanism design. The system runs orchestrator, prover, summarizer, adversarial verifier, auditor and process-advisor agents over iterative rounds — a notable step for automated proof discovery in theory, not just applied ML.

Google Research via X (@omarsar0) · 2026-10-02 · Read the dispatch →

Sam Altman confirms deep OpenAI-Cerebras partnership focused on inference speed

OpenAI's CEO says the two companies are working closely to push the frontiers of inference and compute speed.

Sam Altman (X) · 2026-10-02

Meta open sources SDK for building your own Muse AI gadgets

Developers can connect Meta's Muse AI agent to ESP32 boards, Raspberry Pi displays and sensors — and 5,000 free 'Muse Home Link' gadgets ship this month.

The Verge · 2026-10-02

Hugging Face opens multi-harness RL for coding agents: 7 trained models, TRL trainer, proxy — all public

A fully open recipe for RL-training coding agents across four API harnesses; Liquid AI's LFM2.5-2.6B jumped from 42% to 54% trained on 4 harnesses at once, with 31% fewer tool calls.

Hugging Face · 2026-10-02

Research & Papers

Where-OPD: spatially guided on-policy self-distillation of MLLMs with synthetic scenes

Trains a multimodal student against an EMA teacher that receives privileged spatial information, improving spatial reasoning.

arXiv · 2026-10-01

CARM: cancellation-aware response masking for LLM reinforcement learning

Filters cancelled or failed rollouts in RL post-training, improving training stability and benchmark results on math and code.

arXiv · 2026-10-01

Asynchronous LLM post-training: group-mass capping and convergence analysis

Gives theoretical guarantees for asynchronous RL post-training with stale rollouts, proposing group-mass capping as a correction.

arXiv · 2026-10-01

Feature selective model collapse in diffusion models

Shows collapse from training on synthetic data is feature-selective: some semantic features degrade while others persist, depending on the training regime.

arXiv · 2026-10-01

Scaling and distilling text embeddings for better diffusibility

Finds that scaling the embedding model improves the latent-space 'diffusibility' for continuous diffusion language models, and proposes distilled embeddings optimized for it.

Hugging Face · 2026-10-01

Open Source & Models

Together AI and Ollama: evaluating open models against real production traffic, not benchmarks

Event sessions on selecting and shortlisting open-weight models in the field, focused on how they hold up against real production traffic rather than standard benchmark suites.

Together AI · 2026-10-02

Karpathy: get better answers out of LLMs with controlled language, diagrams and interactive HTML

Request writing in controlled languages like ASD-STE100, ask for diagrams, have the model generate full interactive HTML explainers — with humans focused on oversight.

@karpathy · 2026-10-02

Karpathy's geography eval: 'Land or Water?' across 16k+ coordinate queries

Models perform well on raw latitude/longitude queries, attributed to internet-scale compression of geographic knowledge rather than lookup.

@karpathy · 2026-10-02

Hardware & Robotics

Astribot brings T1 humanoid robot to North America at IROS 2026

The Chinese humanoid maker debuted its T1 with an integrated physical-AI stack: continuously improving models, an embodied OS, and a cable-driven body.

Robotics and Automation News · 2026-10-02

Trane adds two 250 MW AI factory cooling reference designs built on NVIDIA DSX platform

High-density, liquid-cooled hyperscale deployments get packaged thermal-management blueprints.

Refindustry · 2026-10-02

Inside Omron's next-generation LD mobile robots

Software, safety systems and fleet management are becoming as important as the robot hardware itself in mobile robotics.

The Robot Report · 2026-10-02

Samsung Exynos 2700 2nm smartphone SoC reportedly enters mass production

Samsung's first 2nm flagship mobile chip reportedly enters mass production with yields improved around 10%, setting up on-device AI to rival Snapdragon.

Wccftech · 2026-10-02

Tools & Startups

Suno launches Speech: voice and music generated together in one model

Public beta on web and mobile — Suno says it is the first audio model that generates spoken voice and background music as one cohesive track, up to roughly eight minutes.

The Verge · 2026-10-02

ChatGPT adds a virtual changing room for clothing try-ons

Globally launched shopping features let users upload a photo to preview apparel in chat — closely mirroring Google's AI try-on tool.

The Verge · 2026-10-02

Show HN: Television, an open source GUI for your agent harness

A visual console for monitoring and driving coding-agent harnesses instead of raw terminal output.

Hacker News · 2026-10-02

Show HN: Agent Scrub, a scanner for secrets leaked in coding-agent history

Scans agent session history and config files for API keys and credentials accidentally written into logs.

Hacker News · 2026-10-02

Show HN: figma-server, a bridge for agents blocked by Figma's API hostility

A self-hosted server exposing Figma designs to AI coding agents for design-to-code pipelines.

Hacker News · 2026-10-02

Italia AI Spotlight

Green light for Singapore-founded Silicon Box's €1.3B advanced packaging plant in Novara

Invitalia authorized the €1.3B silicon photonics and advanced packaging fab — one of the largest deep-tech investments tied to the AI compute supply chain in Italy, with ~1,600 jobs expected.

la Repubblica · 2026-10-02

ItaliaAperta.it: an AI assistant that simplifies access to Italy's public administration

'Pino', a RAG chatbot over official sources powered by Qwen 3.8-27B via Regolo.ai, answers natural-language questions on PA procedures — and aims to become an agent that completes them.

la Repubblica · 2026-10-02

AI skills double in Italian job postings, half of demand now comes from non-IT roles

Politecnico di Milano's AI Observatory counts ~44,000 2025 postings requesting AI skills, nearly double the prior year, with 48% from marketing, HR, administration and legal functions.

la Repubblica · 2026-10-02

Aosta Valley launches AiVdA, a free online AI training platform for citizens

24 hours of free online AI courses open to all residents — one of the first regional-level public AI literacy initiatives in Italy.

La Stampa · 2026-10-02

Mondadori's PLAI moves from startup accelerator to open-innovation hub

After analyzing 1,600+ startups and making around twenty investments, PLAI presented three startups with investments or collaboration evaluations underway.

StartupItalia · 2026-10-02

Money & Markets

General Intuition raises $220M at $6.2B valuation for physical-world AI models

New York-based developer of AI models trained on videos and virtual worlds to build skills for physical environments.

Crunchbase News · 2026-10-02

Supabase raises $150M led by GIC and acquires database startup Turso

New funding plus an acquisition for the open-source Postgres platform powering AI app development — consolidation in the AI data layer.

Crunchbase News · 2026-10-02

PaleBlueDot AI raises $200M Series C at $3.2B valuation for GPU infrastructure

Over 3x its January Series B valuation; claims $5B+ in signed customer contracts, with its B300 cluster in Japan earning Nvidia Exemplar Cloud status.

Tech Startups · 2026-10-02

Free Models

Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.

In Brief

Source:

The image is complete

The hundredth light has returned home. On to the news.

Google Research Cogentic produces new proofs on five open problems in theoretical computer science

Sam Altman confirms deep OpenAI-Cerebras partnership focused on inference speed

Meta open sources SDK for building your own Muse AI gadgets

Hugging Face opens multi-harness RL for coding agents: 7 trained models, TRL trainer, proxy — all public

Research & Papers

Where-OPD: spatially guided on-policy self-distillation of MLLMs with synthetic scenes

CARM: cancellation-aware response masking for LLM reinforcement learning

Asynchronous LLM post-training: group-mass capping and convergence analysis

Feature selective model collapse in diffusion models

Scaling and distilling text embeddings for better diffusibility

Open Source & Models

Together AI and Ollama: evaluating open models against real production traffic, not benchmarks

Karpathy: get better answers out of LLMs with controlled language, diagrams and interactive HTML

Karpathy's geography eval: 'Land or Water?' across 16k+ coordinate queries

Hardware & Robotics

Astribot brings T1 humanoid robot to North America at IROS 2026

Trane adds two 250 MW AI factory cooling reference designs built on NVIDIA DSX platform

Inside Omron's next-generation LD mobile robots

Samsung Exynos 2700 2nm smartphone SoC reportedly enters mass production

Tools & Startups

Suno launches Speech: voice and music generated together in one model

ChatGPT adds a virtual changing room for clothing try-ons

Show HN: Television, an open source GUI for your agent harness

Show HN: Agent Scrub, a scanner for secrets leaked in coding-agent history

Show HN: figma-server, a bridge for agents blocked by Figma's API hostility

Italia AI Spotlight

Green light for Singapore-founded Silicon Box's €1.3B advanced packaging plant in Novara

ItaliaAperta.it: an AI assistant that simplifies access to Italy's public administration

AI skills double in Italian job postings, half of demand now comes from non-IT roles

Aosta Valley launches AiVdA, a free online AI training platform for citizens

Mondadori's PLAI moves from startup accelerator to open-innovation hub

Money & Markets

General Intuition raises $220M at $6.2B valuation for physical-world AI models

Supabase raises $150M led by GIC and acquires database startup Turso

PaleBlueDot AI raises $200M Series C at $3.2B valuation for GPU infrastructure

Repos on GitHub

Trending on HF

Free Models

Free on OpenRouter

Free on OpenCode Zen

In Brief

The image is complete