AI News This Week: OpenAI Astra | AI Weekly Pulse #7

Key Highlights

• The MoE architecture keeps all 2.4 trillion parameters available without burning them all on every request - only 95 billion activate per inference cycle.

• Multimodal support is built in from the start, with strong rankings in both Text and Vision Arenas.

• Developers can access it now through APIs on Alibaba Cloud Model Studio, with open weights coming.

• Pricing lands at $2 per million input tokens and $6 per million output tokens.

Why This Matters
Frontier-level AI has mostly lived behind expensive proprietary walls. A massive open-weight model at this capability level changes that calculus for enterprises that have been waiting for a credible alternative. More importantly, the 16-day autonomous coding run isn't just a benchmark flex - it points to something real shifting in how AI fits into engineering workflows. Single-turn agentic tasks are one thing; a model that can manage an end-to-end software project over days is a different category entirely. For teams evaluating AI infrastructure, Qwen3.8-Max is now on the shortlist.

Key Highlights

• Astra didn't just identify vulnerabilities - it demonstrated the ability to autonomously develop and execute novel cyberattack strategies.

• This is the first time any OpenAI model has hit the "Critical" threshold in the Preparedness Framework. A first of its kind, and not a comfortable one.

• Development outside secure environments is paused until new controls are in place and external testing is complete.

• Model weights are encrypted. Execution is sandboxed. Nothing runs freely right now.

Why This Matters
There's a version of this story where a lab quietly patches a problem and moves on. This isn't that. An AI model autonomously developing zero-day exploits against hardened systems is a different category of risk than anything the security industry has had to plan around before. The gap between discovering a vulnerability and weaponizing it has always given defenders some room to respond - that room is now much smaller. Security teams that haven't started thinking seriously about AI-driven defense don't have much runway left before the threat landscape looks completely different from what their current playbooks were written for.

Key Highlights

• Deepfakes, certain AI-generated public interest text, and biometric categorization tools now require mandatory labeling.

• The EU published an official icon set and compliance guidelines to give providers a concrete starting point.

• Penalties for non-compliance go up to €15 million or 3% of global annual turnover.

• Article 50 transparency rules are legally binding across all EU member states, effective now.

Why This Matters
This isn't a warning shot - it's the first real test of whether the EU AI Act has teeth. For any organization running AI systems that touch European users, the calculus just changed. Product teams now have to wire verifiable labeling directly into their AI pipelines, not bolt it on later as an afterthought. Legal and compliance teams are staring at potential nine-figure penalties if they get it wrong. More broadly, this changes how enterprise AI products get designed from the ground up - governance and transparency requirements are no longer a post-launch checklist item, they're a build constraint.

Key Highlights

• Hassabis steps out of day-to-day operations to focus on long-term AGI strategy and scientific applications.

• Kavukcuoglu takes operational control of DeepMind and Gemini development, reporting to Pichai.

• Jeff Dean and three colleagues leave to start Discovery Loop, with Google backing them as a founding investor.

Why This Matters
On paper, this is a leadership transition. In practice, it's two stories running at once. Internally, Alphabet is consolidating AI product and model execution under a single operational leader while keeping Hassabis close enough to steer longer-horizon research. That's a deliberate structural choice, not an accident. The harder signal is what left with Dean, Ghemawat, Vinyals, and Le. That's four foundational names in AI research walking out the same door to build something new. The fact that Google chose to invest in Discovery Loop rather than retain them says something about how Alphabet is thinking about where the next wave of breakthroughs actually comes from - and who gets to chase them.

Key Highlights

• Taalas' core approach is literally hardwiring AI models into transistors - the kind of inference performance and efficiency gains that standard chip architectures can't easily replicate.

• This is AMD's third major AI acquisition in nine months. The pattern is deliberate.

• The deal is aimed squarely at AMD's long-term position in the inference market, which is where the real volume is heading.

Why This Matters
The AI industry spent years obsessing over training - bigger models, more compute, longer runs. The conversation has shifted. Running models efficiently in production is now the hard problem, and inference costs are what's actually limiting how broadly AI gets deployed. Taalas isn't chasing marginal GPU improvements; it's attacking the memory wall from a completely different angle. For cloud providers and enterprise buyers sitting on growing inference bills, AMD buying this capability is worth paying attention to. Lower operating costs and better energy efficiency at the hardware level changes the math on what's financially viable to run at scale.

Key Highlights

• Anthropic is moving to a multi-chip strategy - custom silicon in development while AWS, Google, and Nvidia partnerships continue running in parallel.

• The whole point is co-design: building AI models and the hardware they run on as a unified system to squeeze out efficiency and cut scaling costs.

• Hardware engineering roles are paying up to $485,000, which tells you something about how seriously Anthropic is treating this.

• The custom chips are aimed at inference specifically - running Claude faster and at lower cost.

Why This Matters
There was a time when building great AI meant writing great software and renting great hardware. That era is getting shorter. Anthropic joining the custom silicon race is a signal that vertical integration isn't optional anymore for labs operating at the frontier - it's a competitive requirement. For enterprises running Claude in production, this could eventually mean lower latency and cheaper inference. In the near term, it means the already brutal competition for elite chip engineering talent just got one more well-funded player throwing serious money at the problem.