🚀 RELEASES & LAUNCHES
- Google's Project Suncatcher satellite runs first AI workload on TPUs in orbit
Google's Project Suncatcher prototype satellite produced the first AI response from a data-center-class TPU in orbit, running a Gemma 4 model.- For AI infrastructure teams exploring space-based compute: Google's fridge-size prototype satellite, built with Planet and launched Oct 1, 2026 on a SpaceX rideshare, carries four TPUs.
- This week it hit its 'Hello World' milestone: the first AI response from a data-center-class TPU in orbit, with an onboard Gemma 4 open model generating a haiku downlinked during a pass over Svalbard, Norway.
- The satellite runs the model for 15 minutes at a time due to heat constraints; upcoming tests will evaluate TPU performance under thermal stress and radiation exposure.
- Next milestone: twin interconnected satellites in 2027, en route to laser-linked satellite clusters forming orbital AI data centers powered by up to 8x more solar energy than on Earth.
- It signals viability for orbital ML infrastructure, though Google's own lead says space compute won't be cheaper than terrestrial data centers for at least five years.
💰 BUSINESS & FUNDING
- Runware launches Sonic Pods: 1MW modular AI data centers in 20-foot containers
Runware announced Sonic Pods, modular 1MW AI inference data centers packed into 20-foot shipping containers with up to 1,200 liquid-cooled GPUs.- For AI developers and companies needing inference compute, including customers like Higgsfield AI and Wix.
- Each pod holds ~1,200 GPUs delivering 1MW of compute in a 20-foot container, with closed-loop liquid cooling that consumes no water.
- Runware claims 30–80% lower cost per GPU-hour than other inference providers, and deployment in days versus 3–5 years for traditional data centers.
- 10 pods are deployed across the US, Europe, and Asia-Pacific; the company targets 10,000 nodes online in H2 2026 and 10GW of inference capacity in 2027.
- The modular approach lets capacity scale by adding pods near users rather than building fixed hyperscale facilities, backed by a $50M Series A raised in December.
- Runware is also building a factory to mass-produce the pods on an assembly line.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AI
⚠️ INCIDENTS & SAFETY
- DOJ charges five in 'Operation Ghostbusters' AI-driven ghost student loan fraud crackdown
The DOJ charged five people in a Kenya-linked 'ghost student' scheme that used AI to enroll fake students and steal federal student aid.- Targets: federal student aid programs (FAFSA), community colleges, and US taxpayers; victims include US citizens whose identities were stolen.
- Operation Ghostbusters charged five defendants acting as US-based money mules, diverting loan and grant disbursements to Kenyan co-conspirators; DOJ seized over $750,000 in cryptocurrency, the first international seizure of its kind in Kenya.
- Scammers used AI to submit fraudulent applications, complete coursework, and place stand-ins in virtual classes to keep fake students enrolled; over 830,000 potentially fraudulent FAFSA accounts identified with more than $130 million at stake.
- Estimated total fraud exceeds $200 million over five years; the FBI has identified more than $32 million in losses, with 40+ open cases across 26 FBI field offices and 20+ US attorneys' offices.
- California's 116 community colleges found nearly one-third of 2024 applications fraudulent, costing an estimated $13 million in state and federal funds.
🔬 RESEARCH
- Soft-MTP: MIT paper fixes noisy MTP gradient with one-line KL recovery trick
An MIT paper introduces Soft-MTP, a one-line change to multi-token prediction training that removes a noisy gradient residual and improves perplexity, speculative decoding, and reasoning at no extra cost.- For practitioners training MTP-based LLMs (DeepSeek- and Meta-style), the paper decomposes the MTP cross-entropy gradient into a useful KL self-distillation term plus a low-SNR residual that interferes with learning and worsens with scale.
- The KL gradient-recovery trick extracts only the KL signal via straight-through gradient routing, requires no vocabulary-sized logit matrices, and stays compatible with fused cross-entropy kernels.
- From 167M to 1.2B parameters, Soft-MTP improves NTP perplexity, speculative decoding (1.8-2.2% longer acceptance length), and zero-shot accuracy at every scale, with stronger math/code reasoning after fine-tuning at 1.2B.
- It adds no parameters and no measurable overhead, and matches or exceeds direct KL minimization on most metrics.
- Code and pre-trained models are on GitHub and Hugging Face, making it a drop-in improvement for MTP pre-training pipelines.
🏛️ POLICY & REGULATION
- Pope Leo XIV calls to 'disarm' AI and ban lethal autonomous weapons
Pope Leo XIV urged UN member states meeting in Geneva in November to take concrete steps toward banning lethal autonomous weapons, saying AI must be 'disarmed'.- In an October 30, 2026 address to a Vatican meeting organized by the Pontifical Academy for Life and the Future of Life Institute, Pope Leo XIV laid out five considerations on AI, drawn from his encyclical Magnifica Humanitas.
- He said 'machines are tools, not persons' and warned that granting AI systems rights or legal personhood risks obscuring human responsibility for harm.
- He stressed that decisions affecting human life, freedom, and the use of force must never be delegated to machines and must remain under effective human oversight.
- He called for international institutions with shared standards, independent oversight, and accountability mechanisms to guide AI development toward the common good.
- The call gives moral and political momentum to ongoing Geneva negotiations on autonomous weapons ahead of the November meeting.