What's New

AI developments from the last 24 hours

GLM-5.2: A Chinese Lab Releases What Reviewers Call the Best Open-Source AI Model Ever

China's Z.ai (formerly Zhipu AI) released GLM-5.2, and independent reviewers are calling it the most capable open-weights model yet built. On Artificial Analysis's closely watched Intelligence Index it's now the top-ranked open model, and on several long-horizon coding benchmarks it beats OpenAI's GPT-5.5—at roughly one-sixth the cost—while trailing Anthropic's Claude Opus 4.8 by about a single point. Developer Simon Willison called it "probably the most powerful text-only open weights LLM," and practitioners describe it as the first open model that feels "frontier-adjacent" in everyday use.

Anyone can download it. Z.ai posted the full weights of the 753-billion-parameter model on June 16 to Hugging Face (and ModelScope) under an unrestricted MIT license—no registration, no usage limits, free even for commercial use. The catch is hardware: the full model is a roughly 1.5-terabyte download that needs about eight high-end GPUs to run at full precision, though compressed ("quantized") versions run on a beefy single workstation, with day-one support in popular tools like Ollama and vLLM.

That it's Chinese matters, because Washington has spent the past year treating Chinese models as a security problem. Lawmakers have pushed bills like the "No Adversarial AI Act" to bar federal agencies from using models built in China, and DeepSeek—the last Chinese model to stun the industry—has been banned from government devices in more than a dozen U.S. states over fears its API routed user data to mainland China. GLM-5.2 partly sidesteps that specific worry: because the weights are open, a U.S. company or agency can run it entirely on its own servers, sending nothing to China—even as the political climate around "Chinese AI" stays toxic.

Why it matters: The gap between the best closed models and the best free, downloadable ones just shrank to roughly a rounding error—and a Chinese lab closed it. For governments and regulated industries weighing "best model" against "model we control," a near-frontier system you can run in-house, license-free, makes the sovereign-AI option far more viable. It also tightens the bind for U.S. labs heading toward IPOs: their pricing power erodes every time an open challenger matches them at a fraction of the cost.

Source: Simon Willison · VentureBeat · TechTimes

---

Hyundai Takes Full Ownership of Boston Dynamics, Plans Factory Robots by 2028

Hyundai is paying $325 million to acquire SoftBank's remaining 9.65% stake in Boston Dynamics, making the robotics company a wholly owned subsidiary. The deal, expected to close June 22, comes as Hyundai prepares to deploy Boston Dynamics' Atlas humanoid robot in its own factories. The automaker plans to start with parts-sequencing tasks at its Georgia EV plant by 2028, expanding to heavier operations by 2030. Boston Dynamics' CEO said Atlas would need to learn new factory tasks in one to two days and achieve 99.9% reliability to be production-ready.

Why it matters: This signals that the long-promised era of humanoid robots in manufacturing is moving from demos to deployment timelines—and that major industrial players are betting their own capital on it working.

Discuss on Hacker News · Source: startupfortune.com

---

What's Controversial

Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community

Norway Restricts AI Tools in Elementary Schools

Norway has imposed significant restrictions on AI in elementary schools, though specific policy details remain unclear. The move positions Norway among the first countries to limit AI tools for younger students. Community reaction is divided: some compare this to 1990s school internet bans that now look shortsighted, while others note Norway simultaneously offers 'Sikt AI for schools'—suggesting the policy may be more nuanced than a blanket prohibition. Critics question whether the rules distinguish between AI for tutoring versus AI as a homework shortcut.

Why it matters: As schools worldwide struggle to set AI policies, Norway's approach—and whether it proves protective or counterproductive—will be watched closely by education systems debating similar restrictions.

Discuss on Hacker News · Source: reuters.com

---

What's in Academe

New papers on AI and its effects from researchers

Ugandan Farmers Test Cooperative Platform for Better Market Access

Researchers developed Farmer Connect, a mobile-first platform designed to help smallholder maize farmers in Uganda coordinate sales through cooperatives, track earnings transparently, and connect with buyers. The system addresses a persistent problem: small farmers often lack bargaining power and market access that larger operations take for granted. The pilot implementation covered about 85% of identified user requirements and demonstrated the platform could support core workflows for group-based crop marketing. The cooperative model lets farmers pool resources and negotiate collectively rather than selling individually at whatever price middlemen offer.

Why it matters: This is a template for how AI-adjacent digital tools might reshape agricultural supply chains in developing economies—relevant for anyone tracking how technology platforms are being designed for populations that enterprise software typically ignores.

Source: arxiv.org

---

Two New Studies Tackle the Bottlenecks Holding Back Phone-Operating AI Agents

Two research teams took aim at different obstacles facing AI agents that can operate your smartphone—tapping, typing, and navigating apps on your behalf. The first, MobileForge, attacks the data bottleneck: training such agents normally demands expensive human-written instructions and demonstrations, but MobileForge generates its own training tasks and grades itself, reaching 67% task completion on the AndroidWorld benchmark (78% in an enhanced version)—nearly matching models trained on proprietary labeled data. The second, MemGUI-Agent, attacks the memory bottleneck: rather than passively piling up screen history until key details get buried, it treats remembering as deliberate actions the agent takes, keeping only what matters across long, multi-step tasks like booking travel. Its team released a dataset of nearly 3,000 annotated task sequences and an 8-billion-parameter model that beats comparable open models.

Why it matters: As companies race to ship AI assistants that actually do things in apps—not just answer questions—two of the biggest blockers are the cost of training them and their tendency to lose the thread over long tasks. These papers chip at both, a sign the "agent that runs your phone" is inching from demo toward something dependable.

Source: arXiv (MobileForge) · arXiv (MemGUI-Agent)

---

Adaptive AI Tutor Improves Student Completion Rates in Early Testing

Researchers built an AI tutoring system that dynamically switches its teaching approach based on how a conversation is going—and tested it with real high school students. The system analyzes 14 pedagogical features from the conversation transcript to select the most effective prompting strategy. In A/B testing with 656 tutoring conversations across 359 students, the adaptive system reduced session length by about 3 turns while maintaining or improving exercise completion. The version using randomized prompt selection achieved a 28% exercise completion rate versus 20% for the static baseline.

Why it matters: This is early evidence that AI tutors can learn to adjust their teaching style on the fly—a capability that could make educational AI substantially more effective than today's one-size-fits-all chatbot tutoring.

Source: arxiv.org

---

AI Personality Tests Mostly Measure Survey Quirks, Not Real Traits

A study testing 56 large language models with standard personality assessments found that their apparent psychological profiles are largely illusions. The culprit: directional response bias—LLMs tend to drift toward one end of rating scales regardless of what's being asked. This bias accounts for 81-90% of variation between models, compared to just 9-16% in humans. The researchers showed they could manufacture different personality profiles simply by choosing different survey items. More capable models showed less bias, but none eliminated it entirely.

Why it matters: Companies using personality-style assessments to evaluate or select AI models—or researchers claiming to have measured AI 'values'—may be measuring artifacts of how models respond to scales, not genuine model properties.

Source: arxiv.org

---

What's On The Pod

Some new podcast episodes

AI in Business — How Commerce Leaders Avoid Renewal Traps and Vendor Drag - with David Cost of Rainbow Apparel