Open Source Catches Fire

K2.6 seamlessly coordinates heterogeneous agents to combine complementary skills: broad search layered with deep research, large-scale document analysis fused with long-form writing

• Open-source model achieving 58.6% on SWE-Bench Pro vs GPT-5.4's 57.7%
• Agent swarm architecture scaling to 300 sub-agents across 4,000 coordinated steps
• 66.7% on Terminal-Bench 2.0 beating GPT-5.4's 65.4%
• Long-horizon coding with 12+ hour autonomous sessions
• Claw Groups enabling heterogeneous agent collaboration

OVERRATED: running tons of agents in parallel; working on too many things at once. UNDERRATED: using one or two agents at a time; focusing on the task in front of you; thinking deeply; finishing

• Agent parallelization often creates more noise than value
• Context-switching between multiple agents reduces quality
• Deep focus with fewer agents produces better results
• Finishing tasks completely beats opening multiple PRs

Unless you opt out by August 17th 2026, data from Jira and Confluence will automatically be used for AI training

• Automatic opt-in for AI training starting August 17, 2026
• Some data cannot be opted out at all on certain plans
• Follows similar moves by Google, Zoom, Dropbox, Adobe, Slack, Figma, GitHub
• Default assumption: every SaaS will train on non-E2E encrypted data

AI-generated tracks now represent 44% of all new music uploaded to its platform... receiving almost 75,000 AI-generated tracks per day

• 44% of new music uploads are AI-generated
• 75,000 AI tracks uploaded daily, over 2 million monthly
• Only 1-3% of total streams are AI music
• 85% of AI music streams detected as fraudulent
• 97% of listeners can't distinguish AI from human-made music

Grok Imagine

• Grok can generate Word docs, PDFs, PowerPoints, Excel spreadsheets
• Native LaTeX compilation in Grok Files
• Custom image/video templates for Grok Imagine
• 300M+ monthly sessions on App Store

Agent gateways feel like we are racing backwards into the MS-DOS era... One process, one token, with the LLM holding the line

• Current agent architectures mirror MS-DOS's security failures
• Single-token, single-process design creates systemic vulnerabilities
• Wirken demonstrates process separation and Ed25519 per-channel identity
• Shell commands run in hardened containers with capability drops
• History teaches that wrapper-based security always fails

Designers have been granted prototyping super powers with AI, but the models are still not good enough to consistently one-shot implementations you'd want to merge to master on large, critical applications like Basecamp without programmer review

• AI excels at prototyping but not production-ready code
• Large critical applications still require human review
• Gap between demo-quality and merge-ready code remains significant

We're expanding our collaboration with Amazon to secure up to 5 gigawatts of compute for training and deploying Claude.

• 5 gigawatts of compute capacity for Claude training
• Nearly 1 gigawatt expected online by end of 2026
• Amazon investing additional $5 billion with up to $20 billion more
• Roughly equivalent to 5 nuclear reactors worth of compute
• Marks shift from talent/data bottlenecks to power contracts

John Ternus, senior vice president of Hardware Engineering, will become Apple's next chief executive officer effective on September 1, 2026.

• Tim Cook transitions to Executive Chairman after 15 years as CEO
• John Ternus, hardware engineering lead, becomes new CEO
• Effective September 1, 2026
• Apple grew from $350B to $4T market cap under Cook
• Transition follows long-term succession planning

Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7)

• State-of-the-art open source coding model
• 4,000+ tool calls over 12 hours of continuous execution
• 300 parallel sub-agents vs K2.5's 100
• Beats closed models on multiple coding benchmarks
• Available via API, weights on HuggingFace

Claude Code & Codex land near the human median, but with far tighter dispersion & no extremes.

• 146 economist teams got wildly different answers from same dataset
• AI agents landed near human median with minimal variance
• No extreme outliers in AI results unlike human teams
• AI reviewers consistently ranked outputs: Codex GPT-5.4 > GPT-5.3 > Opus 4.6 > humans
• Suggests AI is ready for scalable research work

Now, Codex can help with what you've been working on without you restating context.

• Chronicle feature uses recent screen context to improve memories
• Codex remembers what you've been working on automatically
• Research preview for PRO subscribers on Mac
• Part of broader push into persistent agent memory
• Builds on last week's memory feature release

LeWM didn't need a massive, centralized supercomputer. It has just 15 million parameters. It trains on a single, standard GPU in a few hours.

• Yann LeCun's JEPA approach vindicated by new LeWorldModel paper
• 15M parameters vs billions in LLMs
• Trains on single GPU in hours
• Plans 48x faster than foundation models
• Solves representation collapse with elegant math instead of hacks

Quarterly reminder Krugman was absolutely dead to rights with this quote people try to dunk on him with: since 1998 total factor productivity grew 0.6% per year and prior to 1998 it had been growing 0.68% per year.

• Total Factor Productivity grew 0.6% annually post-1998 vs 0.68% before
• The fax machine genuinely mattered more than people admit
• Consumer surplus has no price so doesn't show up in productivity metrics
• Efficiency of light production improved 900-1600x over 200 years

Kimi K2.6 dramatically improved throughput from ~15 to ~193 tokens/sec, ultimately achieving speeds ~20% faster than LM Studio.

• 12-hour continuous execution with 4,000+ tool calls
• Optimized 8-year-old financial matching engine for 185% throughput gain
• 300 parallel sub-agents executing 4,000 coordinated steps
• Supports proactive 24/7 autonomous agents
• Motion-rich frontend with WebGL shaders and Three.js