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: IBM buys HRL, gaining spin-qubit and sensing research, built on electron spin in silicon, closing as a pending research deal in Q3 2026.
IBM signed a definitive agreement on July 23 to acquire HRL Laboratories from Boeing and GM, adding silicon-spin qubit engineering and quantum sensing to a stack that has been superconducting-only for a decade; HRL's chips move into IBM's New York wafer foundry, terms undisclosed, close expected by Q3 2026. The signal is not capacity, it is hedging — silicon spin qubits ride CMOS-compatible fabrication, so the bet being placed is manufacturability at scale, not more qubits next year, and buying a second modality outright says IBM does not consider the architecture question settled on the way to Starling in 2029. Note what the announcement does not contain: a demonstrated result. Starling and Blue Jay are still roadmap, so if you own any quantum-adjacent planning, keep algorithm work at the abstraction layer instead of encoding transmon assumptions, and read every 2029-and-later date in a vendor deck as intent rather than schedule.
HOW TO READ THIS Read top to bottom: WISER and E.ON set up the benchmark, pit a quantum model against a classical one, watch the quantum forecast hug the real load curve, then see the resulting error cuts.
WISER and E.ON benchmarked two hybrid quantum-classical models for electricity demand forecasting against a classical multi-output Gaussian Process baseline. The projected-quantum-kernel GP cut mean absolute error 62% on simulator and 40% on real quantum hardware, and the same write-up reports a 100-qubit utility-scale run that held predictive accuracy in only 76% of evaluations. That third number is why this one is worth your time: most QML results lead with the simulator figure and let you assume it survives contact with a real device. If you are evaluating a quantum ML pitch, ask for the hardware number and the failure rate — a vendor showing only simulation is answering a different question than the one you asked.
HOW TO READ THIS Read downward from the photon pair to separate wavelength bands sharing live fiber, then to preserved entanglement between Evanston and Chicago, as reported by [Northwestern](https://news.northwestern.edu/stories/2026/07/quantum-internet-leaves-the-lab?fj=1).
Northwestern researchers distributed entangled photons across 24.4 km of live fiber between Evanston and downtown Chicago at above 94% fidelity, while that same strand carried two 800 Gbps classical data channels. The trick was placing the quantum signal in the O-band, away from C-band production traffic. Dedicated dark fiber has been the single largest cost objection to metro quantum networking, and this says you may not need it. For infrastructure planning, that reframes the question from 'can we get dark fiber' to 'what does coexistence cost us in the existing optical budget.'
HOW TO READ THIS Read top to bottom: DARPA funds PsiQuantum, PsiQuantum clears stage gates A and B to reach C, that gate is an evaluation not a delivered machine, and $125M unlocks at Stage C while status stays research.
PsiQuantum signed a $125 million performance-based agreement with DARPA on July 22 under the Quantum Benchmarking Initiative — its largest U.S. government contract to date, funding hardware, software, and system-level testing across Milpitas and Chicago. Read carefully what the money buys: QBI Stage C is an evaluation program that exists to verify whether a utility-scale design actually works as designed, and PsiQuantum is one of only two companies still in it. Government money for adversarial verification is a considerably healthier signal than government money for a machine. Treat 'selected for QBI Stage C' as a claim under test, not a claim confirmed.
Self-hosted agent + enterprise search platform with a built-in agent builder — the reference point when pricing a private alternative to a hosted assistant.
Open-source AI email assistant; worth reading as an agent-permissions design study as much as an inbox tool.
MCP servers for UniFi Network, Protect, Access, and Drive — MCP reaching into physical infrastructure, with the blast radius that implies.
Desktop app that generates 3D models from images entirely on your local GPU — no cloud round-trip, no upload.
An agent skill that strips the tells out of AI-written text — the detection arms race now ships as an installable package.