AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: GPT-5.6 targets the Cycle Double Cover Conjecture, traces two cycles across every shared edge, and the resulting proof now awaits peer review.
OpenAI's GPT-5.6 Sol Ultra produced a full proof of the Cycle Double Cover Conjecture — a graph-theory problem open for decades — and published it with the complete prompt and PDF. If peer review holds, this is the clearest evidence yet that frontier models can generate genuinely novel mathematics rather than retrieving and remixing known results. The caveat matters: machine-generated proofs have failed review before, and the interesting signal isn't the headline, it's whether the proof technique is new. Read the PDF before repeating the claim, and watch for independent verification from the graph-theory community over the next few weeks.
HOW TO READ THIS Read top to bottom: vxcontrol ships PentAGI, it acts with no human operator, it loops recon-exploit-report on its own, and it has drawn 19,991 GitHub stars.
PentAGI — fully autonomous agents running complex penetration tests end to end — is surging on GitHub trending (~20k stars, Go). Offensive security is now agent-shaped, which cuts both ways: red teams get dramatically cheaper, and so do attackers. If you run anything internet-facing, your threat model should now assume continuous automated probing at machine scale. Practical move: point it at your own staging environment before someone else effectively does.
HOW TO READ THIS Read top to bottom: Sunrun's solar homes gain AI compute, which joins a mesh of edge homes sharing inference work, and homeowners get paid to host it.
Solar and battery company Sunrun is paying customers to host AI compute units in their homes instead of building centralized data centers. It's a bet that the power crunch gets solved at the residential edge — household energy systems becoming distributed inference infrastructure. Worth watching because it inverts the data-center-siting problem: instead of chasing grid capacity, you chase rooftops that already have it. The open questions are latency-tolerant workload fit, security of physical access, and whether the unit economics survive real utility rates.
HOW TO READ THIS Read top to bottom: cheahjs starts a repo, scattered free API providers get pulled into it, it becomes one list, and that list tops GitHub with 26,837 stars.
A curated list of free LLM inference APIs is one of GitHub's hottest repos this week (~27k stars). For prototyping agents or side projects, it's a practical map of where capable models run at zero cost today. Free tiers churn constantly, so treat it as a starting point, not an architecture decision — but it's the fastest way to de-risk an idea before committing to paid inference.
100+ runnable AI agent and RAG apps — clone, customize, ship.
CLI for configuring and monitoring Claude Code setups.
Self-evolving context database unifying agent memory, RAG, and skills.
Automates browser-based workflows with AI agents.
Next.js app that creates and modifies draw.io diagrams with AI.
Ben's Bites Grok lands in Cursor — model competition inside coding tools keeps compressing switching costs.
Latent Space SpaceXAI ships Grok 4.5, its first Opus-class model since the Cursor acquisition.
The Sequence Sharp essay on why verifiability alone doesn't make a good RL training environment.
SemiAnalysis One-year progress audit of Meta Superintelligence — sober read on where the compute went.