Government documents caught hallucinating citations & China backs national AI champions - AI News (May 8, 2026)

First up: a very real-world AI governance mess. South Africa’s Department of Home Affairs suspended two officials after discovering what it described as AI-style “hallucinations” in a reference list attached to a major white paper on citizenship, immigration, and refugee protection. The department pulled the standalone reference list, apologized, and said it will add AI declarations and automated checks to its approval process—plus a wider review of past policy documents. The takeaway is simple: when credibility is the product, even a sloppy references section can undermine an entire institution’s work, and it’s pushing governments toward formal “AI usage” controls rather than informal guidance.

Now to China’s AI race, where the money is getting bigger and more politically meaningful. DeepSeek is reportedly in talks to raise funding from government-backed investors, with some discussions valuing the company around fifty billion dollars—far above earlier ranges that were reportedly much lower. In parallel, Moonshot AI—the company behind the Kimi chatbot—raised a massive new round led by Meituan’s venture arm, valuing it above twenty billion, with reports pointing to rapidly growing recurring revenue. Together, these moves show capital concentrating into a small set of perceived national champions. And in a world of export controls and tighter access to advanced chips, that kind of backing isn’t just about valuation—it’s about securing compute, infrastructure, and staying power.

Let’s talk infrastructure—because the next limiter on AI progress is often not the model, it’s the plumbing. OpenAI and NVIDIA both highlighted Multipath Reliable Connection, or MRC, a new networking approach meant to keep giant GPU clusters running at high utilization even when networks get congested or links flap. The notable part isn’t just performance claims—it’s that the spec is being published through the Open Compute Project, aiming for broader adoption across vendors. Why this matters: frontier training is increasingly constrained by networking reliability and tail latency. If the industry can standardize a sturdier Ethernet-based fabric for AI factories, it reduces the odds that “one bad link” slows down tens of thousands of GPUs waiting on each other.

On inference—where most AI products actually spend their time—there’s a new open-source entrant optimized for agent-style workloads. The LightSeek Foundation announced TokenSpeed, positioning it as an inference engine tuned for long contexts and heavy, sustained token generation, like coding assistants and autonomous agents. They’re claiming meaningful throughput and latency improvements in early testing, while also being clear it’s still being hardened for production. The bigger point is the trend: as agents become normal, inference efficiency stops being a nice-to-have and becomes a line item you feel in power, GPU budgets, and user experience.

A related warning came from ServiceNow researchers working on online reinforcement learning pipelines. They reported that moving from an older vLLM backend to the newer vLLM V1 led to major training divergence—because small differences in inference-side log probabilities can poison the learning signal. Their conclusion is blunt: before you “fix RL,” you may have to fix inference correctness and parity, because caching, scheduling, and numerical details can quietly turn into model-behavior changes. It’s a reminder that in modern AI systems, training and serving aren’t separate worlds anymore—especially when the model learns from what it just served.

Speaking of strain: the business model for AI is being stress-tested by agents that don’t behave like humans clicking around. One analysis of recent plan changes argues that old subscription designs are breaking under long-running, parallel agent sessions. We’ve seen rapid shifts: tighter limits, sudden policy enforcement on agent harnesses, and a general move toward usage-based billing. The meta-lesson is that capability has outpaced metering. Providers are now rebuilding “monetization layers”—entitlements, rate limits, and pricing logic—as core infrastructure, because without it, every surge becomes a public pricing crisis.

On the enterprise distribution front, Alphabet is reportedly in talks with private equity firms—Blackstone, KKR, EQT—about broad Gemini access deals spanning their portfolio companies. It’s a platform-style bet: make procurement easy and let consultancies or internal teams handle deployment, rather than embedding large engineering squads into each client like some rivals do. If this lands, it could become a powerful channel—thousands of companies at once. The tradeoff is also clear: lighter-touch distribution can scale fast, but you may learn less about real workflows than you would by being deep inside deployments.

A quick look at evaluation: two new benchmarks are trying to measure what people actually want from agents—end-to-end work, not just clever answers. Meta’s ProgramBench asks agents to recreate complete software projects from a compiled executable and documentation, without access to the original code. Early results are brutally low, which is kind of the point: it’s meant to expose the gap between coding snippets and real system-building. In legal AI, Harvey open-sourced its Legal Agent Benchmark, built around realistic “client matters” with strict pass/fail rubrics. The shift here is important: as agents move into high-stakes domains, the industry needs evals that punish almost-right outputs, because in law, security, and finance, “almost” can be the failure mode.

Now, the cultural side effects. One story noted that writers are deliberately changing their style—adding typos, slang, or an exaggerated voice—just to avoid being accused of using AI. At the same time, another commentary argues online communities are being flooded with low-effort AI-generated posts and projects, raising the moderation burden and driving out experienced contributors. Together, these signals point to the same problem: trust is being taxed from both directions. People are pressured to “prove they’re human,” while communities struggle to keep signal-to-noise high when content generation is cheap and verification is expensive.

Two more quick items to close. First, a sober take on robotics: an essay argues that “world models” could be as transformative for robots as LLMs were for text—but the bottleneck is data friction. Real-world interaction data is hard to gather, expensive, and messy, so progress may be determined as much by operations and data pipelines as by model architecture. And finally, AI’s ripple effects are hitting consumer hardware. A report cited by PC industry watchers says motherboard sales are dropping sharply as chip and component supply is squeezed by AI demand, pushing prices up and making DIY upgrades less attractive. It’s another reminder that the AI boom isn’t contained to data centers—it’s reshaping the entire tech supply chain.