Your Lumis briefing — Monday, October 5, 2026
Geometric analysis proves global 'unsafe direction' defenses for text-to-image models are fundamentally flawed
T2I safety teams relying on universal toxic subspace filters (e.g., ESD, UCE) should treat them as broken. Patch or replace before deployment; adversarial bypass is now formally characterized.
→ arXiv cs.AI125B 'Qwen 3.8 Flash Next' runs at 100 tokens/sec on a single RTX 4090 via Strata
Operators can now run frontier-class 125B models on consumer GPUs. Re-evaluate self-hosted inference costs and latency benchmarks—prior assumptions about hardware requirements are obsolete.
→ Hacker News / GitHubImproper redaction exposes Google data center's true water and electricity consumption in Nebraska
Actual infrastructure footprints of hyperscalers are now on record. Researchers modeling AI energy/water costs and policymakers setting data center regulations have new ground-truth figures to work from.
→ Hacker News / 1011nowWant this every morning,
tailored to your sources?
Lumis synthesizes Hacker News, arXiv, The Batch, Latent Space — and any RSS feed you choose — into three sharp items before your day starts.
No spam. Invite when your slot opens.