Sam Altman Draws a Line Against AI as Authority

Sam Altman Draws a Line Against AI as Authority

In a terse post today, Sam Altman wrote: "I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue" [4]. Coming from the CEO whose product people increasingly treat as oracle, confidant, and therapist, this is a notable public correction — not a safety paper, a direct statement.

It echoes prior interviews where Altman has been explicit that AI isn't divine and shouldn't replace personal judgment or spiritual authority [5]. The replies are mostly agreement — treating a model's output as infallible is exactly the failure mode builders should be designing against, not encouraging.

This matters beyond philosophy. If your product surfaces AI recommendations as decisions rather than inputs, you're building the exact dynamic Altman is warning about. Good UX in this era means making the model's uncertainty visible, not hiding it behind confident prose.

Tavus Claims a Video Turing Test Win With Griffin

Tavus unveiled Griffin on October 1 — a full-duplex, video-to-video "Human Interaction Model" — and the number that's driving 15 million+ views is this: 48% of 54 blind-study participants believed they were talking to a real human after a one-minute call, up from 2.4% on Tavus's prior stack [6][7]. Griffin also tops NVIDIA's VideoFDB benchmark at 3.83/5, startlingly close to the human baseline of 3.92 [8].

It's research preview only — Tavus is explicitly delaying broad access for safety and disclosure work, which is the right call given what this technology enables [6]. The use cases being pitched are tutors, health coaches, and companions — exactly the categories where "is this a person" matters most.

The discourse online is appropriately focused on disclosure norms rather than raw capability hype. That's progress. A model good enough to pass as human on video isn't a demo anymore, it's a deployment decision with real consent implications attached.

Aleph Alpha Launches Kolibri, a Sovereign European Open-Weights Model

On the Day of German Reunification, Aleph Alpha released Kolibri: 78.1B total parameters (3.46B active, MoE architecture), a 1M context window, full weights on Hugging Face under Apache 2.0 [9][10]. Trained on 20 trillion tokens across data centers in Germany and Finland, it's explicitly positioned as the sovereign alternative for regulated European sectors that can't or won't route sensitive workloads through US labs.

This is a meaningful European statement, even if Kolibri won't beat Opus 5.5 on raw benchmarks. Apache 2.0 licensing plus EU-based training infrastructure directly answers the sovereignty question that's been looming over every European enterprise AI procurement conversation since the AI Act enforcement ramp began. The Hugging Face community reaction has been warm, with calls for ecosystem support to match the US and Chinese open-weight momentum.

EU AI Act Enforcement Begins to Bite

The AI Office activated enforcement powers on August 2, 2026, and the first wave of RFIs went out to GPAI providers in late August covering safety, transparency, and copyright [11][12]. Fines for prohibited practices run up to €35M or 7% global turnover — numbers big enough to change board-level risk appetite, not just compliance checklists. High-risk obligations got partial breathing room, pushed to December 2027 via the Digital Omnibus, but the GPAI transparency rules are live now.

The timing is almost too on-the-nose: the same week Europe ships a genuinely strong sovereign model, it's also tightening the regulatory vice on the startups that would build on top of it. The criticism writing itself on X is that SMEs face compliance costs that US and Chinese competitors simply don't carry [12].