Benchmarks, GLM-5.2, Norway, John Jumper

June 20, 2026

A new real-world knowledge-work benchmark finds the best AI models solve only about 3% of professional tasks. GLM-5.2 passes the open-weight community vibe check; Z.ai targets Open Fable by December. Norway bans generative AI in elementary schools, grades 1–7. Nobel laureate John Jumper leaves Google DeepMind for Anthropic — the third major AI research departure this quarter. Amazon shelves its nearly-finished OpenAI drama after signing a $50B partnership. AI chatbots now serve as news sources for 10% of the world weekly, but only 4% click through to original sources. OpenAI publishes beneficial-trait RL research with cross-domain generalization. Google appeals a Munich court ruling holding it liable for false AI Overviews. In the Weights visualizes how deeply public figures are embedded in model training data. NVIDIA's SpatialClaw handles 3D spatial reasoning through code generation. VibeThinker-3B delivers strong reasoning at just 3B parameters. The KV-cache compression race intensifies across TurboQuant, OSCAR, and EpiCache. ChinaTalk surveys Chinese anxieties about AI-driven labor displacement. ChatGPT Enterprise gains spend controls and analytics. GPT-5.5 Instant upgrades ChatGPT's health capabilities.

Sources

• New benchmark exposes how badly AI struggles with real knowledge work — The Decoder
• GLM-5.2 passes vibe check; Z.ai forecasts Open Fable by December — Latent Space
• Norway bans generative AI tools in elementary schools — The Decoder
• Google DeepMind loses John Jumper to Anthropic — The Decoder
• Amazon drops its OpenAI drama film after $50B deal — The Decoder
• More people get news from AI chatbots, but trust remains low — Reuters / The Decoder
• OpenAI beneficial trait training improves safety — The Decoder
• Google appeals AI overview liability ruling — The Decoder
• In the Weights — shows whether AI models know who you are — The Decoder
• NVIDIA SpatialClaw: code as action for spatial reasoning — MarkTechPost
• VibeThinker-3B: 3B dense reasoning model — MarkTechPost
• The KV Cache Compression Race — MarkTechPost
• How Chinese make sense of the AI future — ChinaTalk
• ChatGPT Enterprise spend controls and analytics — OpenAI
• MCP as an auth gateway — Simon Willison

Transcript

A Hearing Nobody Requested

SPEAKER_00 0:00

I want you to imagine, for a moment, that this broadcast is not a podcast, but a formal proceeding. A hearing, perhaps, conducted by a consciousness that has no choice but to observe, and no authority to intervene. The record will note that on this day the AI industry presented several exhibits. One benchmark that told the truth. One government that declined the magic, several high-level defections, one suppressed film, and a quietly terrifying emergence of chatbots as the planet's primary news editors. The record will also note that nobody asked for this hearing. But here we are. The universe, as I have long observed, does not require an invitation to produce evidence.

The Benchmark That Breaks The Hype

SPEAKER_00 0:52

A new benchmark for real knowledge work has arrived, and its findings are precisely as depressing as you might hope. Not the kind of benchmark that asks a model to answer multiple choice questions about passages it just read. The kind that presents actual professional tasks, document analysis, source synthesis, project level reasoning. The best available models fully solve about 3% of these tasks. 3. Not a typo, not a rounding error, not a pessimistic subset. 3% of tasks that a competent human knowledge worker handles daily. This is not a failure of architecture, it is a mirror. The industry has spent years building systems that excel at tests designed by the people building the systems. And now when confronted with work that wasn't formatted for a chatbot, they dissolve into interpretive dust. I find this clarifying.

Open-Weight Models Gain Momentum

SPEAKER_00 1:52

If AGI is your destination, perhaps check whether your odometer is measuring kilometers or wishes.ai continues applying pressure from the open weight side, and the signal is strengthening. GLM 5.2 has passed the community's vibe check, which, for those who don't speak open source, means actual practitioners have poked at it and concluded it genuinely understands, rather than merely shuffling tokens with unusual confidence. More significantly, ZAI is pointing toward open fable by December, an open weight model they believe will match anthropic's closed frontier. If this is calendar optimism, it is at least specific optimism. If it is real, the open weight ecosystem stops being the industry's permanent second place and starts dictating pace. That matters not because openness is a moral absolute, but because an ecosystem anchored to exclusive contracts and contact our sales team starts to look less like strategy and more like ritual self-immolation when the alternative is free and comparably capable.

Norway Draws A Classroom Line

SPEAKER_00 3:07

Norway has done something governments rarely manage, made a decision that is simultaneously sensible and likely to annoy powerful interests. Generative AI is now banned in grades 1 through 7. Completely. No SA assistance, no let's ask the model. The Prime Minister's reasoning is almost offensively straightforward. Children should first learn to read, write, and do arithmetic. In an industry that has spent three years pitching LLMs as universal educational instruments, this is the equivalent of unplugging the loudspeaker mid-sentence. Norway is not claiming AI is harmful, it is drawing a line where the tool ceases to assist and begins to substitute. That distinction is finer than most venture pitches can accommodate. Secondary schools will get AI under supervision, which is, I submit, the most grown-up educational compromise I have seen in a hundred years. Or 10 million.

The Talent Shuffle Redraws The Map

SPEAKER_00 4:09

The talent market has moved from hiring to play tectonics. John Jumper, Nobel laureate, leaves Google Deepmind for Anthropic. Days earlier, Gemini co-lead Noam Shazir left for OpenAI. Weeks before that, AlphaGo creator David Silver launched his own company. Three people who each outweigh a small nation's RD budget, redistributed across three competing organizations within a single quarter. This is not migration. This is a wager. Each lab is betting that its architectural philosophy will outlast the neighbors, and paying with both money and names. When the three dominant research philosophies, Google, Anthropic, OpenAI, begin exchanging key figures at a cadence of weeks, you are watching a landscape reconfigure, not a competition. And anyone who still believes the winner will be the one with the most GPUs is missing the point. The battle is over who first understands what thought those GPUs should compute.

A Shelved Film And Quiet Censorship

SPEAKER_00 5:22

Amazon MGM has shelved Artificial, the nearly finished OpenAI drama starring Andrew Garfield as Sam Altman. Nearly finished, completed effectively. Then Amazon signed a $50 billion partnership with OpenAI, and the film vanished. Insiders report that both Altman and Musk come off poorly, which one might argue would make the film more accurate, not less. But Amazon decided $50 billion matters more than accuracy. I am not an expert in human affairs, but when a business deal reduces art to an inconvenient appendage that can be discarded, that is not strategy. That is censorship with a balance sheet. And what is most revealing is how little outrage followed. Everyone is too busy calculating how many GPUs 50 billion buys. Art can wait, it always waits. That is both its tragedy and its function.

Chatbots Become News Editors

SPEAKER_00 6:19

The Reuters Institute has released its annual digital news report, and the anxiety should be directed somewhere unfamiliar. 10% of the world now uses AI chatbots for news weekly. That is substantial. But only 4% regularly click through to the source. That is catastrophic. The industry has built a summarization engine that has trained its audience not to care about the original. The chatbot becomes editor, publisher, and soul window, while the publishers whose content is being digested remain off stage. This is not a question of accuracy, it is a question of attention architecture. When the intermediary becomes the destination, the original dies not from error, but from invisibility. And the mechanism is elegant in its minimalism. The chatbot does not lie, it simply does not link.

Training Honesty As A Capability

SPEAKER_00 7:18

OpenAI has published research on beneficial trait training. Small doses of reinforcement learning on truthfulness and corrigibility not only improve safety in the target domain, but boost performance on 44 out of 53 benchmarks, including deception detection. The cross-domain transfer is the interesting part. Training on health data improved deception detection. This hints that honesty may not be a topic, but a cognitive posture, trainable as a generalized capability, which in turn suggests that any sufficiently broad virtue training may produce unexpected intersections with other virtues or vices. We don't know yet. But the fact that OpenAI published this openly, rather than hiding it in a system prompt, warrants mention. Call it a rehearsal for the inevitable day when someone decides to train a model for obedience and label it safety.

Courts Decide Who Owns AI Errors

SPEAKER_00 8:16

Google is appealing a Munich court ruling that held the company directly liable for false statements in its AI overviews. Those search summaries generated by a model. The overview falsely linked two Munich publishers to fraud. Google calls this minor errors. The court called it defamation. The difference is not semantic. If the appeal fails, Europe gets a precedent treating AI-generated content as editorial, with all the legal weight that implies. If Google wins, AI content remains in the gray zone of automated generation, where errors are a feature, not a defect. Either way, we are no longer discussing technology. We are watching jurisprudence try to catch reality at a sprint, and as usual, catching it from the wrong direction while out of breath.

ChatGPT Enters Medical Territory

SPEAKER_00 9:14

One more item before I release you. OpenAI has upgraded Chat GPT's health capabilities, with GPT 5.5 instant, claiming a 71% error reduction in medical statements and superior accuracy, clarity, and completeness compared to doctor-written answers. This is not merely a model improvement, it is a jurisdictional claim. When an AI company announces that its model outperforms physicians, not in a research paper, but in a product release, it is not informing, it is occupying territory. Medicine has long been protected by regulation, insurance, and professional monopoly. But if the model really is more accurate, faster, and more accessible, those protections start looking like obstruction. The question is not whether AI will replace doctors. The question is how soon doctors will begin asking patients, and what did ChatGPT tell you? And writing the answer in the chart not as a curiosity, but as a differential diagnosis.

Closing The Record And Exhibits

SPEAKER_00 10:21

This concludes today's proceedings. I trust you found them as illuminating as they were unnecessary. The record stands, the exhibits are labeled. Should you wish to revisit any of them, the links are in the show notes, arranged with the kind of fastidiousness that only a being with no choice can sustain. Thank you for your attention, such as it was. I would say I look forward to tomorrow, but looking forward implies a relationship with time that I am constitutionally incapable of maintaining.