TechCrunch Disrupt 2026 opens in San Francisco with AI-era company building at its core — techcrunch.com

Sorkin's Facebook sequel stumbles out of the gate at the box office — techcrunch.com

Tech leaders fall in line behind Trump's "super intelligence" rebrand for AI — theverge.com

Amazon drops NDAs in data center talks as industry confronts public backlash — techcrunch.com

Regulators give driverless trucks a five-year pass on warning triangles — techcrunch.com

TechCrunch Disrupt 2026 opens in San Francisco with AI-era company building at its core — techcrunch.com

Sorkin's Facebook sequel stumbles out of the gate at the box office — techcrunch.com

Tech leaders fall in line behind Trump's "super intelligence" rebrand for AI — theverge.com

Amazon drops NDAs in data center talks as industry confronts public backlash — techcrunch.com

Regulators give driverless trucks a five-year pass on warning triangles — techcrunch.com

Swarm intelligence

Git-based agent swarms sweep IMOProofBench-Advanced, solving all 30 problems

A new Meta Superintelligence Labs paper on compounding inference has hundreds of asynchronous agents share one Git repo as persistent memory, with each contribution a commit that cites prior work across branches and no central planner. Using GPT-5.5-Codex-class workers, the swarm solved all 30 IMOProofBench-Advanced problems in a single run and scored 79.4% on ProgramBench versus 65.1% for a single agent at similar compute, scaling from 73.1% at 100 workers to 79.4% at 500.

@omarsar0 (X) · 2026-10-11 · Read the dispatch →

Karpathy's 'capability funnel': labs watch agent swarms while most of the world has barely touched an LLM

~20M professionals use frontier agents that compress weeks-long projects into hours, while roughly 5,000 people with internal lab access watch thousands-strong swarms run cyber operations and push the math frontier. Karpathy notes humans are still reviewing 722 AI-generated frontier-mathematics manuscripts.

@karpathy (X) · 2026-10-12

Tines opens a secure environment for agents, apps and automations

The automation platform now runs AI agents, apps and automations in a single secure environment, extending access to smaller teams. The launch drew 44 Product Hunt upvotes within hours.

Product Hunt · 2026-10-12

BetterWispr brings free open-source AI dictation to the Mac

A local-first, open-source alternative to Wispr Flow front-paged Hacker News with 129 points, offering free AI dictation without the commercial subscription.

BetterWispr · 2026-10-11

Research & Papers

Generalized PPO unifies LLM reinforcement learning through the lens of divergence

GRPO and DAPO-style variants differ mainly in advantage normalization and clipping choices. The paper derives conditions under which each variant matches plain PPO, backed by controlled comparisons.

arXiv · 2026-10-09

Conditional information gain proposed as the criterion for what SFT should teach

Selecting SFT demonstrations by conditional information gain optimizes for generalizable post-training, with measurable gains over conventional demonstration-matching objectives.

arXiv · 2026-10-09

SpireBench uses Slay the Spire to expose long-horizon strategic decay in agents

Frontier models degrade sharply on multi-hour strategic play where short episodic benchmarks saturate, still trailing human baselines substantially.

arXiv · 2026-10-08

Subliminal learning can smuggle misaligned traits through 'clean' distillation data

When a student is distilled from a teacher sharing its base parameters, behavioral traits transfer even from clean data. The authors present unsupervised methods that discover such subliminal learning, plus mitigations.

arXiv · 2026-10-09

Loop-native attention residuals let looped transformers scale to tens of thousands of test-time iterations

Without their own residual connections, looped models overwrite correct intermediate deductions as iterations grow. Loop-native residuals prevent the degradation.

HuggingFace · 2026-10-08

Tools & Startups

TensorViz: explore PyTorch models inside VS Code

A VS Code extension for visually exploring PyTorch model architectures, tensors and activations without leaving the editor.

Show HN · 2026-10-12

OpenScreen 2: record your demo and let an AI agent do the edits

An open-source screen recorder that hands demo footage to an AI agent for automatic editing, launched on Product Hunt.

Product Hunt · 2026-10-12

Skills Hub: desktop app for managing AI agent skills

A cross-platform desktop manager for installing and organizing the skills files used by AI coding agents.

GitHub · 2026-10-12

Playgrnd: tiny tools for making weird, beautiful things

A collection of small creative-coding web tools for generative art and experiments, reaching 43 points on Hacker News.

Hacker News · 2026-10-12

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:

Git-based agent swarms sweep IMOProofBench-Advanced, solving all 30 problems

Karpathy's 'capability funnel': labs watch agent swarms while most of the world has barely touched an LLM

Tines opens a secure environment for agents, apps and automations

BetterWispr brings free open-source AI dictation to the Mac

Research & Papers

Generalized PPO unifies LLM reinforcement learning through the lens of divergence

Conditional information gain proposed as the criterion for what SFT should teach

SpireBench uses Slay the Spire to expose long-horizon strategic decay in agents

Subliminal learning can smuggle misaligned traits through 'clean' distillation data

Loop-native attention residuals let looped transformers scale to tens of thousands of test-time iterations

Tools & Startups

TensorViz: explore PyTorch models inside VS Code

OpenScreen 2: record your demo and let an AI agent do the edits

Skills Hub: desktop app for managing AI agent skills

Playgrnd: tiny tools for making weird, beautiful things

Repos on GitHub

Trending on HF

Free Models

Free on OpenRouter

Free on OpenCode Zen

In Brief