The Lyceum: AI Daily — Apr 03, 2026

Google shipped Gemma 4 under Apache 2.0 — the most permissive license it's ever attached to an open model — and Netflix released its first-ever public AI model, a physics-aware video eraser that's already running on hobbyist GPUs. Meanwhile, Anthropic filed to create a PAC, tightened third-party Claude access, and is fighting the Justice Department in court simultaneously. The theme today isn't capability — it's positioning. Every major lab is making bets about whether openness, control, or political influence is the better long-term moat.

• Gemma 4 (Google DeepMind): Four open models for agentic reasoning under Apache 2.0, including edge variants that run offline on phones and Raspberry Pi.
• VOID (Netflix): Physics-aware video object removal model — Netflix's first public release — open-sourced on Hugging Face and GitHub under Apache 2.0.
• Qwen3.5 2B (Alibaba): Non-reasoning open model available via DeepInfra at $0.02/$0.10 per million tokens.
• GLM-5 Turbo (Z-ai): 203K context window model via AtlasCloud at $1.20/$4.00 per million tokens.
• MiniMax Speech 2.6 Turbo (MiniMax): New speech model available via Together.

The most important thing Google shipped today isn't a benchmark — it's a legal document.

Gemma 4 drops under Apache 2.0, the same permissive license used by Qwen 3.5 and more open than Llama 4's community license. No monthly active user limits, no acceptable-use enforcement, full freedom for sovereign and commercial deployments. Previous Gemma licenses reserved Google's right to terminate access; that clause is gone. Per The Register, the launch comes amid an onslaught of open-weights Chinese models from Moonshot AI, Alibaba, and Z.AI — Google is offering enterprises a domestic alternative that won't route sensitive data through foreign infrastructure.

The flagship 31B dense model ranks third on the Arena AI open-model leaderboard with a 256K-token context window, per 9to5Google. Edge variants (E2B, E4B) run fully offline on phones and single-board computers with 128K context. Per WaveSpeedAI's analysis, this is a calculated commoditization play: give developers commercially friendly weights, and some will pay for Google Cloud to scale them.

What changes if this works: The mid-tier API business shrinks. Startups build products on free weights instead of renting inference. Cloud providers compete on managed services, not raw model access. The signal to watch: If paid products built on Gemma 4 start shipping this quarter, the commoditization strategy is landing. If enterprises still default to proprietary APIs despite the license, the switching costs are higher than Google thinks.

Most companies release AI models to recruit engineers. Netflix released one amid a need for it in production.

VOID — Video Object and Interaction Deletion — doesn't just paint over objects. It infers what the world would look like if the object had never been there, including physical consequences: remove a person holding a guitar, and the guitar falls naturally. The model is fine-tuned on CogVideoX with interaction-aware mask conditioning, per Netflix's GitHub repo. In a 25-person human preference study, VOID was chosen 64.8% of the time versus 18.4% for Runway in the study reported by the authors' preprint — directional numbers from the authors' own preprint, not independent evaluation.

It's Apache 2.0 licensed, per Veo4's writeup, meaning any VFX house or indie filmmaker can run it today. The arXiv preprint details the full method. Early Reddit threads report roughly 40GB GPU requirements and careful per-frame masking for best results — research-grade, not consumer-ready.

What changes: A streaming company just open-sourced production-grade video AI rather than keeping it proprietary. That's a different posture than Hollywood has taken toward AI tools. Watch for: Whether Runway, Adobe, or other VFX tools respond with their own open releases, or whether Netflix remains an outlier.

Anthropic filed to create AnthroPAC, funded through voluntary employee contributions capped at $5,000 annually, with a bipartisan board and FEC disclosure, per TechCrunch. Per Axios, it plans contributions to both parties during the midterms.

The context makes this loud. Per the Washington Examiner, Anthropic is fighting the Pentagon over a scrapped $200 million contract — the company wanted input on how the military used Claude; the department disagreed. Per Bloomberg, the DOJ is appealing a federal judge's order blocking Trump's ban on government use of Anthropic's AI. Per The Hill, Trump-aligned figures are already skeptical the PAC will actually donate bipartisanly.

What changes: Anthropic is no longer content to let policy happen to it. This is a research lab becoming a political actor in real time. The signal: Watch the first FEC disclosure. If contributions skew heavily toward one party, the bipartisan framing collapses — and Anthropic's government relationships get more complicated, not less.

Anthropic announced that Claude subscriptions will no longer cover usage on third-party tools like OpenClaw starting April 4 at 12pm PT, per LLM Stats. The framing is capacity management. The effect is platform consolidation.

This follows this week's Claude Code usage-cap complaints, where developers burned through daily quotas in hours on complex agentic tasks. Anthropic is approaching $19 billion in annualized revenue, per Crescendo AI — these aren't decisions made from weakness. They're decisions about where the margin is.

What changes: Developers who built workflows around Claude through third-party clients now pay separately via API or switch to Anthropic's own tools. Watch for: If this triggers visible migration toward Gemma 4 or OpenAI's API, the Claude ecosystem is more fragile than revenue numbers suggest.

OpenAI acquired TBPN, a tech-business talk show, per the Wall Street Journal. According to Tech Startups, TBPN was profitable with roughly $5 million in 2025 ad revenue and on track for $30 million in 2026.

TBPN wasn't a mass-market property — it was a show founders, investors, and journalists actually listened to. Owning that conversation pipeline is a different kind of asset than a model or an API. OpenAI is now in the media business, which raises real questions about editorial independence and coverage of competitors. Watch for: Whether Anthropic, Google, or xAI respond by building their own narrative infrastructure. If one lab owns a podcast, the others may feel the pressure.

Chevron is in talks to build a natural-gas plant dedicated to powering a Microsoft data center in Texas, per Axios. Clean Intelligence data cited by Axios suggests around 30% of planned data-center capacity now targets isolated "energy island" setups, up from near zero last year, as of the Axios report.

This lets hyperscalers scale without waiting on utilities — but locks in fossil-fuel infrastructure as grids push greener. Google is pursuing a similar arrangement, per The Guardian, with a Texas plant that would emit 4.5 million tons of CO₂ annually — more than San Francisco.

What changes: AI infrastructure is decoupling from the public grid. That solves the capacity bottleneck but creates a climate accountability problem no one has a good answer for yet. The signal: If utilities start losing their biggest potential customers to on-site generation, expect regulatory fights over stranded grid costs.

Georgia's legislature adjourns Monday, April 6, and three AI bills sit on Governor Kemp's desk awaiting his signature as of April 3, 2026: SB 540 (chatbot disclosure and child safety), SR 789 (AI study committee), and SB 444, which prohibits insurance coverage decisions based solely on AI, per the Transparency Coalition.

That last one is the one to watch. Similar legislation is moving in Alabama and Kentucky simultaneously — Kentucky's HB 227 (chatbot safety) passed the House 96-0. Tennessee signed SB 1580 this week, making it illegal for AI to represent itself as a qualified mental health professional; it passed both chambers unanimously (Senate 32-0, House 94-0).

What changes: A coordinated red-state push to regulate AI in healthcare and child safety is building faster than any federal framework. The cumulative compliance surface for consumer AI companies is growing weekly. Watch for: If Kemp signs all three before Monday's adjournment, compliance teams need to start treating state law as the primary constraint, not a footnote.

Benchmarks from live freelance work show agents are useful but far from autonomous replacements. Per Allwork.Space, the Remote Labor Index found the best agent fully automated only 2.5% of end-to-end freelance projects across coding, design, and analysis in the study; GPT-5 managed 1.7% in the same study.

This aligns with Anthropic's own research showing Claude covers about a third of computer/math tasks in workplace contexts, and NBER work finding AI raises support-agent productivity roughly 14% in that study, with larger gains for newer hires.

What changes: The "AI replaces everyone" narrative gets a reality check with actual numbers. Agents are raising the floor for juniors and speeding routine work — not bulldozing paid gigs. Watch for: If these numbers move meaningfully by Q3, the displacement timeline is real. If they stay flat, the bottleneck is task complexity, not model capability.

A new preprint introduces ARC-AGI-3, an interactive benchmark where agents are dropped into abstract puzzle environments with no instructions and must explore, infer rules, and plan — more like learning a strange board game than answering trivia, per the arXiv paper.

The results are humbling: humans eventually solve 100% of these tasks in the benchmark; leading AI systems score under 1% in the benchmark. This isn't about bigger context windows or better training data — it's about whether models can build internal world models and adapt from experience.

What changes: ARC-AGI-3 is likely to become the go-to yardstick for "agentic intelligence" claims, the way ImageNet once anchored vision progress. Watch for: If any lab reports meaningful gains on ARC-AGI-3 this year, it signals genuinely new reasoning capabilities — not just better prompting.

A streaming company teaching AI physics, a safety lab filing PAC paperwork, and a search giant giving away its best models like free samples at Costco. The 2.5% automation number is the kind of statistic that will age either very well or very badly — and nobody knows which yet. Back Monday.

If someone you know is making AI decisions without reading this, that's on you. Forward it.