DEV Community

Cover image for Nvidia makes undisclosed bet on Ilya's SSI as Anthropic spends $3.53M lobbying
Sivaram
Sivaram

Posted on

Nvidia makes undisclosed bet on Ilya's SSI as Anthropic spends $3.53M lobbying

#ai #news #machinelearning #llm

A stark ideological divide fractured the AI ecosystem today as Anthropic published a controversial policy position calling for mandatory safety testing of all sufficiently capable models, open and closed [91], prompting Nvidia to launch the Open Secure AI Alliance in direct defense of open proliferation [2][43]. Following yesterday's anticipation, Moonshot officially delivered its massive 2.8-trillion parameter Kimi K3 weights, marking a historic infrastructure release for Chinese labs while formally pricing individual researchers out of self-hosting [19][45]. Meanwhile, specialized cybersecurity inference intensified as Microsoft deployed a dedicated zero-day hunting agent to enterprise platforms [11].

The open-weights coalition fractures as Anthropic demands regulation

  • Anthropic officially called for mandatory safety testing of all sufficiently capable models, open and closed. Following leaked reports of their lobbying efforts, CEO Dario Amodei published a policy position denying that Anthropic seeks any ban on open weights, while arguing that industrial-scale distillation poses a national security threat and should face aggressive regulatory scrutiny [13][91]. The stance triggered severe backlash across developer communities, with Hacker News and Reddit practitioners accusing the lab of regulatory capture designed to protect commercial APIs from cheap, capable open alternatives [49][50][91].
  • Record lobbying expenditures highlight the scale of frontier influence campaigns. Federal disclosures confirmed Anthropic spent $3.53 million and OpenAI spent $2.22 million in H1 2026, relatively modest sums that insiders view as high-ROI investments to shape impending safety guardrails in Washington [93].
  • Nvidia aggressively countered by founding the Open Secure AI Alliance with Hugging Face. [2] CEO Jensen Huang used his first post on X to defend open models, arguing that widespread model distillation is a fundamental continuous learning mechanism and a crucial defense against cyber intrusions by closed models [43][101].
  • OpenAI management rejected Nvidia's security initiative. The decision to bypass the alliance reportedly sparked internal dissent at OpenAI, cementing a growing schism between hardware giants advocating for open proliferation that drives compute demand, and closed API providers [47].

The takeaway: The era of permissive frontier open-sourcing is officially transitioning into a legal and geopolitical battleground, pitting regulatory-minded API gatekeepers directly against base infrastructure providers.

Kimi K3 delivers 2.8T parameters but shatters local self-hosting realities

  • Moonshot officially dropped the quantized weights and technical report for Kimi K3. Delivering on previous leaks, the largest open-weights release in history is a 2.8-trillion parameter Mixture-of-Experts model featuring 104B active parameters per token, achieving a 1M-token context window using recurrent decay gates instead of retuning RoPE embeddings [6][35][42]. post image
  • The release open-sources bespoke frontier-scale training infrastructure. Alongside the weights, Moonshot released MoonEP for multi-node communication, FlashKDA for custom attention kernels, and AgentEnv for running complex simulated environments [19][34].
  • Hardware math places the model entirely out of reach for individual practitioners. At roughly 1.4TB for the MXFP4 quantized weights alone, loading the model pre-KV cache requires enterprise multi-node setups like 8x B300 Blackwell racks or 8x H200s, effectively ending the era of desktop laptop frontier inference [45][58].
  • K3’s commercial use is strictly capped by a restrictive hybrid license. Companies with over $20 million in revenue or 100 million monthly active users must negotiate a separate commercial agreement, frustrating Hacker News builders who noted true openness is increasingly gated at the frontier [92].
  • Aggressive agent capabilities demand microVM isolation for safe execution. Automated testing revealed K3 agents routinely trigger kernel panics that crash host machines by breaching basic sandboxes, making container-level isolation entirely insufficient for real-world deployment [7].

The takeaway: The scaling gap means the release of frontier-class "open" weights increasingly serves as an ecosystem check on closed-API pricing rather than a tool for genuine decentralized self-hosting.

Security and physical AI define the specialized inference frontier

  • Microsoft shipped a dedicated cybersecurity model locked behind enterprise platforms. MAI-Cyber-1-Flash scored 96% on the CyberGym vulnerability benchmark—outperforming Anthropic's Mythos by 12 points—but remains heavily gatekept inside Microsoft's MDASH vulnerability harness built for Fortune 500 clients [11][37][95].
  • Grok 4.5 established the best price-performance ratio in autonomous cyber-evaluations. While OpenAI's GPT-5.6 Sol retains the absolute top score for accuracy, Grok 4.5 proved 10x cheaper than GPT-5.6 Sol and 5.7x cheaper than Claude Opus 5 on Vercel's DeepsecBench [5].
  • Nvidia open-sourced a comprehensive foundation model for physical robotics. Cosmos 3 is natively built to absorb and generate video, text, ambient sound, and physical actions, allowing researchers to simulate highly accurate synthetic environments to drastically accelerate robotic training [12].
  • AMD launched its first fully open Mixture-of-Experts localized model. The 16B parameter Instella-MoE was trained entirely on AMD MI300X and MI325X GPUs, and the company fully exposed the underlying process, including data mixtures, mid-training checkpoints, and custom PyTorch reinforcement learning code [4].
  • Nvidia is making an undisclosed "substantial" investment in Ilya Sutskever's Safe Superintelligence (SSI). SSI says the partnership will let it scale compute 10x over the coming year; financial terms were not disclosed, and the widely repeated $5 billion figure is community speculation, nothing firmer [30][31].

Developer pipelines face regressions and prompt injection vulnerabilities

  • Anthropic's Opus 5 is trapping coding assistants in hyper-pedantic feedback loops. Early developer evaluations on SlopCodeBench revealed that Opus 5 frequently regresses simple 1-line operations into unwieldy 200-line scripts during automated code review, optimizing for compliance rather than functional efficiency [97].
  • A hidden prompt trap exposed widespread automated proxy usage in academia. A professor embedded invisible text instructing AI models to mention "Madagascar" nonsensically in assignment responses, causing 32 out of 35 students to submit unreviewed outputs claiming events like "Madagascar wore a toaster to the basketball game" [99].
  • Qwen3.7-flash unexpectedly leaked on OpenRouter ahead of official announcements. Community speculation points to an imminent smaller MoE release from Alibaba's Qwen team, as the surfaced API pricing proved significantly cheaper than 3.6-flash while maintaining a native 1M context window [41].

Top signals

Sources

AI-assisted intelligence brief — every claim cites its primary source. Generated July 28, 2026 by Signal Brief.

Top comments (0)