Kuaishou Spins Off Kling AI—China’s Video Gen Race Heats Up

The story of the day

Kuaishou’s decision to spin off its Kling AI video unit—confirmed in a regulatory filing today—isn’t just corporate restructuring. It’s a calculated bet on China’s fragmented video generation market, where ByteDance’s Seedance 2.0 already dominates globally (except in the US) and Sora 2 quietly redefines benchmarks for Western builders. Kling’s standalone push signals two things: first, Kuaishou admits it can’t compete with ByteDance’s scale in-house; second, it’s banking on external capital to outmaneuver Alibaba and Tencent-backed players in the mid-tier enterprise segment—where cost-per-minute and regional compliance matter more than raw quality.

Builders should watch the funding terms closely. If Kling secures capital at a valuation north of $2B (rumored floor), it validates a thesis: China’s video AI stack is bifurcating into global players (ByteDance, DeepSeek) and regional specialists (Kling, Manus). The latter focus on localized datasets and censorship-compliant outputs—areas where Sora and Runway stumble. For Western devs, the spinout means one less integrated competitor but a new API player to benchmark against. Expect Kling’s pricing to undercut Seedance by 30-40% within 12 months, targeting Southeast Asia and the Middle East first.

New tools & launches

JotForm + Claude Integration — Build, edit, and analyze forms directly in Claude’s interface, with auto-generated summaries and response clustering. Targets non-technical teams drowning in form data. [Source]

MirrorCode Benchmark Update — Open-source tool now evaluates AI agents’ ability to reverse-engineer full codebases (not just snippets), with support for Rust and Go. Flagged for false positives in dependency-heavy repos. [GitHub]

SafeDep Alert System — Real-time monitoring for npm supply chain attacks, with automated diff checks for TanStack, Mistral AI, and 170+ compromised packages. Free tier covers 50 repos. [SafeDep.io]

Models & research

Microsoft’s Long-Task Failure Study (proprietary)

Researchers found current AI models and agents degrade catastrophically in tasks exceeding 6-hour runtime—even with perfect tooling and context windows. The culprit? "State drift" in intermediate representations, where models lose track of subgoals after ~1,200 iterative steps. Mitigation requires either:

1) Hard resets every 90 minutes (killing efficiency), or

2) External memory graphs (adding 23% latency).

Builders deploying agents for long-running workflows (e.g., CI/CD, data pipelines) should audit for silent failures—this isn’t a benchmark issue, but an architectural limit. [The Register]

LLM Hallucinations in the Wild (open-access)

ArXiv paper [2605.07723] quantifies hallucination rates across 12 production LLM APIs, using a novel "factual entropy" metric. Key finding: Models with <1% hallucination rates on synthetic benchmarks (e.g., MMLU) jump to 18-22% in real-world tasks involving:

• Multi-hop reasoning over PDFs
• Time-sensitive data (e.g., "latest earnings call")
• Ambiguous user prompts (e.g., "summarize this like a CFO")

The fix? Not bigger models, but prompt-aware retrieval augmentation—which adds 150ms latency. Tradeoffs, always. [arXiv]

Industry pulse

Canva’s "Palestine→Ukraine" AI Blunder

Gizmodo’s report on Canva’s Magic Layers tool automatically replacing "Palestine" with "Ukraine" in user designs isn’t just a PR nightmare—it’s a stress test for AI vendors’ geopolitical risk clauses. Canva’s apology cited "overzealous moderation filters," but the incident exposes a deeper issue: Western AI tools increasingly ship with hardcoded regional compliance layers that act as de facto censorship. For enterprise buyers, this raises a red flag: If your design tool can’t handle place names without "distress," what happens when it auto-edits financial reports or legal docs? The fix (manual overrides) is already live, but the trust cost lingers. [Gizmodo]

Mass NPM Supply Chain Attack

The Safedep.io report on 170+ compromised npm packages—including TanStack and Mistral AI dependencies—reveals a shift in attacker tactics: targeting AI/ML toolchains specifically. Unlike traditional supply chain attacks (which aim for broad exploitation), these payloads focus on:

1) Data exfiltration from model training pipelines

2) Prompt injection via corrupted dev dependencies (e.g., `mistral-common-utils`)

3) Compute theft by hijacking GPU-bound tasks

The kicker? 60% of the malicious packages were published by new maintainers with pristine GitHub profiles, suggesting coordinated synthetic identities. [SafeDep.io]

Stat of the day

18-22%

Real-world hallucination rate for "low-hallucination" LLM APIs (vs. <1% on benchmarks), per arXiv [2605.07723]. The gap widens with time-sensitive queries—where models confidently invent data 37% of the time.

Get the full analysis—subscribe to Gradient Pro.