Artificial Curiosity Labs
Go back

AI News Digest

AI News Digest — 2026-10-20

Top Stories

Nscale and Anthropic are reportedly in talks for yet another $45B compute deal (techcrunch.com)

The prospect of a $45 billion compute deal between Nscale and Anthropic signals that the infrastructure race among frontier AI labs shows no signs of slowing, with single deals now approaching the scale of entire cloud computing markets from just a few years ago.

AI Solves a Major Unsolved Math Problem. Not Everyone Is Happy (IEEE Spectrum)

An AI system’s solution to a Millennium Prize problem marks a landmark achievement for AI in science, yet its reception in the mathematics community highlights real tensions over the role and recognition of machine-generated proofs.

Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool (techcrunch.com)

AWS launched TorchNeuron, a native PyTorch library that lets developers compile models directly for its Trainium and Inferentia chips without touching Nvidia’s CUDA, potentially undermining the platform lock-in that has made Nvidia dominant in AI training and inference.

Tenstorrent hits $6.2B valuation to scale open-weight AI chips (TechCrunch)

Tenstorrent’s $1.5 billion raise at a $6.2 billion valuation signals strong institutional appetite for alternatives to Nvidia’s proprietary ecosystem, as the open-weight chip movement—which Tenstorrent advances with its RISC-V–based architecture—positions itself to capture demand from enterprises seeking cost-effective, programmable AI compute without vendor lock-in.

AI assistant startup Town nabs unicorn valuation 4 months after launch (TechCrunch)

Town has achieved a $1B+ valuation with a $108M Series B merely four months after launching, marking one of the fastest climbs to unicorn status in the AI assistant sector and signaling intense investor appetite for early-stage AI tools.

Funding & Capital

What moved

Partnering with Blacksmith (Sequoia Capital)

Sequoia’s backing of Blacksmith signals that the bottleneck in AI-assisted development is shifting from code generation to code validation, as the rise of agent-created software demands infrastructure that can verify and ship trustworthy code without slowing down the pace of creation.

China’s AI microdrama boom joins a global surge in mobile viewing (TechNode)

Tencent adding AI-generated series to its WeChat short-drama service signals that generative video is moving into mainstream mobile entertainment distribution, not just experimental demos, and it arrives amid a broader global surge in mobile viewing.

Dream Labs raises $90M to help build more intelligent robots with simple video feeds (siliconangle.com)

Dream Labs’ $90 million raise to develop a universal robot foundation model that converts simple video feeds into actions underscores a rapid escalation in the race to create generalist, API-accessible robot intelligence rather than task-specific automation, with the disclosed customer pipeline of 3,500 companies suggesting that commercial demand for adaptable physical AI is both immediate and substantial.

Also tracking

Foundation Models

What moved

China Telecom Open-Sources Xing4.0-29B-Intent to Turn Commands Into Structured Parameters (Pandaily)

China Telecom’s release of the 29-billion-parameter Xing4.0-29B-Intent model under an open license, together with a full SDK and 2.75 million training samples, gives developers a ready-made pipeline for translating natural-language commands into structured API parameters, significantly reducing the data-engineering and compute costs that have kept robust intent extraction limited to large platforms.

StartLux-Decision v1.1 Tops Decision Index 0.3.2 With Open Models That Think Only When Unsure (pandaily.com)

StartLux-Decision v1.1 outperforms Decision Index 0.3.2 by employing an adaptive reasoning mechanism where the model allocates additional compute only when it encounters uncertainty, rather than applying heavy computation to every input. This design, implemented in open-weight models, directly tackles the high inference cost of reasoning-heavy systems, making advanced problem-solving more accessible and practical for developers seeking efficient, self-regulating AI.

Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second (marktechpost.com)

Inception’s Mercury 3 diffusion LLM claims a generation speed of 1,068 tokens per second and a context window of 1 million tokens, directly challenging the dominance of autoregressive models by demonstrating that diffusion-based architectures can achieve competitive benchmark performance at much higher throughput.

Also tracking

Infra & Compute

What moved

Stars Micro System Unveils C82-4G PCIe 6.0 Switch, an 800G AI NIC and a Multi-Mode RAID Chip (Pandaily)

Stars Micro System’s trio of chips signals a methodical push to reduce China’s reliance on foreign interconnects and storage controllers inside AI infrastructure; the 144-lane C82-4G PCIe 6.0 switch at 64 GT/s directly competes with Broadcom’s equivalents for GPU-to-GPU and CPU-to-accelerator bandwidth, the 800G AI NIC addresses east-west traffic bottlenecks in GPU clusters, and the multi-mode RAID chip provides an in-house alternative to Broadcom and Microchip for NVMe/SAS/SATA attach, which together could lower system cost and export-control exposure for domestic server builders.

SpacemiT Pairs V100 RISC-V Server CPU With Three Servers, Kunlunxin M100 Validation and Linux (Pandaily)

SpacemiT’s V100 RISC-V server CPU is now paired with three servers and validated with Kunlunxin M100 accelerators under Linux, a combination that signals growing RISC-V maturity for AI workloads and suggests the architecture is moving closer to production server deployments.

Spain-Based Startups Advance Regional Innovation With Barcelona Supercomputing Center (NVIDIA Blog)

The accelerator programme’s provision of sovereign compute resources through the Barcelona Supercomputing Center directly addresses European startups’ need for high-performance infrastructure without dependency on US-based hyperscalers, marking a tangible investment in regionally controlled AI development capabilities that could shift the balance of innovation gravity within the EU.

Also tracking

Robotics & Physical AI

What moved

DAMO Academy’s XuanTie E908A Is an Automotive RISC-V CPU IP Aimed at ASIL-B and ASIL-D Designs (Pandaily)

DAMO Academy’s XuanTie E908A marks the first RISC-V CPU IP announced with a roadmap to ASIL-D functional safety certification, the highest automotive integrity level required for fully autonomous systems. The chiplet-based design integrates AI acceleration and consistent ISA extensions, signaling that RISC-V can now meet the redundancy and fault-tolerance demands previously locked to proprietary architectures, potentially lowering silicon costs and supply-chain risks for next-generation vehicle intelligence.

AgiBot’s Lingxi X2 Handles 46,686 Patient Queries at Peking University People’s Hospital Front Desk (Pandaily)

AgiBot’s Lingxi X2 sustained ~707 daily patient interactions across 66 days in a live hospital setting, signaling that general-purpose humanoid robots are approaching operational reliability for high-volume public-facing roles rather than remaining in controlled demonstrations.

X Square Robot Launches X2Real, an Evolving Simulation Benchmark It Says Tracks Real-Robot Results (Pandaily)

X2Real promises a simulation benchmark with a 0.96 correlation to real-world robot performance across 44 tasks, directly tackling the costly sim-to-real gap that slows deployment of learned policies on physical hardware.

Also tracking

Applications

What moved

Tencent’s Marvis AI Butler Lands in the Kylin Software Store With Support for Domestic CPUs (Pandaily)

Marvis is explicitly tuned for Kylin’s V10 SP1 desktop environment — the operating system widely mandated in China’s government and state-owned enterprises — and supports locally made chips including LoongArch, ARM, and x86 architectures, meaning Tencent is now embedding its AI assistant directly into the workflow layer of the country’s official IT stack.

Tencent Launches Free WorkBuddy Education Edition With 2,000 Monthly Credits for Every Teacher (Pandaily)

Tencent’s free WorkBuddy Education Edition removes a key cost barrier by giving every teacher 2,000 monthly credits, and with backing from top Chinese universities it could meaningfully scale AI adoption in classrooms without per-teacher fees slowing uptake.

Replika’s founder built the AI assistant she pitched at Disrupt 12 years ago (TechCrunch)

Eugenia Kuyda, founder of the AI companion app Replika, has launched Wabi, a free personal AI assistant — a direct realization of the concept she first pitched at TechCrunch Disrupt 12 years ago, marking her strategic expansion from emotional AI into general-purpose assistance.

Also tracking

Voice & Speech

What moved

Ant Group Open-Sources Ming-3.0-tts, Small Speech Models With Voice Design and Fine Control (pandaily.com)

Ant Group’s open-sourcing of Ming-3.0-tts puts compact, multilingual text-to-speech models with voice cloning and fine-grained paralinguistic control into the public domain, lowering the barrier to customizable speech synthesis from a major Chinese fintech player and broadening the options for developers seeking small, steerable voice models.

Introducing Gemini 3.8 Live with Live Avatar (Google DeepMind)

The rollout of live avatar and voice capabilities in Gemini 3.8 Live, integrated directly within the Gemini app, marks a shift toward AI that communicates with visual and emotional immediacy, not just text. By synchronizing expressive, low-latency digital faces with real-time spoken dialogue, this upgrade makes interactions feel more human and can improve engagement for tasks like learning, assistance, and companionship, while signaling a broader industry move from functional chatbots to presence-driven interfaces.

Gemini 3.8 text-to-speech says hello (Google DeepMind)

Google DeepMind has introduced Gemini 3.8 text-to-speech, broadening its voice AI portfolio with new synthesis capabilities.

Agents & Tooling

What moved

Meet Drex Max: Nace AI’s Multimodal Decision Model for Computer-Use Agents (marktechpost.com)

Nace AI released Drex Max, a 6.7-billion-parameter multimodal decision model specifically designed for computer-use agents, achieving an 86.4% accuracy rate on the screenspot-pro grounding benchmark for GUI element identification while operating at sub-70-millisecond response times in cloud inference. The model’s architecture cleanly separates decision-making from execution, accepting interleaved screen-history inputs to reduce downstream agent failures, and was trained from a Llama 4 Maverick base on a 100,000-sample expert-action dataset covering 35 diverse applications and websites—a design that skips intermediate planning steps to directly produce executable actions.

Lawaken Opens Pre-Sale of X-AI Memory Glasses From ¥699 With MCP Agent Hooks (pandaily.com)

The pre-sale positions the X-AI Memory Glasses as a hardware entry point for MCP-compatible AI agents, with native hooks to multiple LLM platforms and a Workbuddy agent for long-context memory and task automation starting at ¥699. This open-architecture approach—enabling direct agent interaction beyond a single assistant—could accelerate consumer-grade agent adoption by decoupling the wearable from any one AI provider and making persistent, contextual assistance accessible at a mass-market price.

Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’ (techcrunch.com)

Microsoft is publicly reclassifying AI agents as akin to trusted but ultimately fallible insiders, warning that frontier models remain too unreliable to grant fully autonomous access to sensitive systems.

Also tracking

Policy & Safety / Other

What moved

Mamdani says feds have abdicated on AI, urges New Yorkers to push back (techcrunch.com)

New York City Mayor Mamdani’s direct accusation that federal authorities have abdicated responsibility on artificial intelligence oversight reflects a tangible fracture in U.S. tech governance, as the vacuum forces cities to contemplate their own regulatory frameworks for a technology that fundamentally transcends municipal borders.

Trump administration orders ban on DeepSeek in federal agencies (siliconangle.com)

The Trump administration’s ban on DeepSeek in federal agencies is the first formal US federal prohibition targeting a specific foreign AI model, citing multiple threat vectors and setting a precedent for how AI-related national-security orders may be structured going forward.

Embedded AI evaluators won’t be an easy solution, Hugging Face’s policy chief says (techcrunch.com)

Hugging Face’s policy chief Irene Solaiman cautioned that embedding AI systems as evaluators of other AI—a model increasingly floated for safety oversight—introduces profound technical and governance challenges, arguing it shifts rather than resolves risks and lacks the reliability needed for real-world deployment; the stance from a platform central to open-source AI development intensifies scrutiny on the viability of self-regulatory safety architectures.

Also tracking