Every iPhone since the 6S has been running a background speech detection system every second of every day. Battery barely affected. Here's the two-pass architecture Apple built to make that possible:

This has been running on every Apple device since the iPhone 6S. 2.2 billion devices. Always on. Battery barely affected. Apple documented this in their ML research. Hiding in plain sight. machinelearning.apple.com/re…

Auth0 starts at $35/month. Okta charges $2/user/month for SSO, then got breached in October 2023. The breach exposed names and email addresses of every Okta customer support user. HAR files for 134 customers. There is a self-hosted alternative. It is called Authentik. 21,763 stars on GitHub. Here is what it does: → SAML 2.0 and OAuth2 / OpenID Connect identity provider → LDAP outpost for legacy apps → SCIM user provisioning → MFA: TOTP, WebAuthn, passkeys, SMS, static tokens → Drag-and-drop authentication flows with custom stages → Social login: Google, GitHub, Azure AD, and more → Forward-auth outpost to protect any app without modifying it → Role-based access control with self-service enrollment Here's the part that matters for small teams: SSO, SAML, and OIDC are not paywalled in the community edition. Okta charges $2/user/month for SSO at the minimum. For a 50-person company, that's $1,200/year for SSO alone. Auth0: $35/month and up. Okta SSO: $2/user/month, breach track record. Authentik: $0. Self-hosted. MIT core license. 21,763 GitHub stars. Community edition actively maintained. (Link in the comments)

A new paper formally proved that AI agents using imperfect world models can be systematically exploited by adversaries. Not "might be exploitable." Essentially unavoidable on large policy sets. The paper is called "Imperfect World Models are Exploitable." The definition: a world model is exploitable if it implies that one policy should be strictly preferred over another, while the environment's true transition model implies the reverse. The adversary doesn't break the model. The adversary exploits the gap between what the model believes and what the environment actually does. Key findings: → Exploitation is essentially unavoidable on large policy sets (formally proved) → The conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation → A "safe horizon" exists: a bounded time window within which exploitation can be avoided → Bridges the formal connection between reward hacking and model exploitation The implication for teams building agentic systems with world models: the gap between your model and reality is not just a performance issue. It's a security surface.

Figma Make just connected directly to your local codebase. Visually edit a design in Figma, ship the change straight to production code. No export, no handoff, no gap between design and dev. The design file and the codebase are turning into the same surface. x.com/figma/status/205998717…

Just found a structured AI learning hub that covers the full path from beginner to researcher. It's called ai-learning-roadmaps. 516 stars on GitHub. Covers everything, end to end: → Math foundations (linear algebra, calculus, statistics) → Python and framework foundations → Machine learning and deep learning → Transformers and large language models → RAG and agentic AI → MLOps (deployment, monitoring, scaling) → Research paper reading and methodology → Career-oriented learning paths for different roles → Curated communities and discussion platforms Resources are tagged by quality. High-impact content widely used in academia, research, and industry is starred separately. Follow modules sequentially or in parallel depending on your goals. Getting started guide included: Python setup, virtual environments, Colab/Jupyter, essential libraries. No subscription. No course fee. 516 stars. 68 forks. MIT licensed. (Link in the comments)