▒▓▓▒ ▒▓██▓▒ ▒▓▓▒ ▒▓██▓▒
░▒▓██▓▒░▒▓████▓▒░▒▓██▓▒░▒▓████▓▒
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
────────────────────────────────────────
No. 10 · Tue · 23 June 2026Happy Tuesday. It was a conspicuously non-American day at the frontier: a Chinese open model the experts now concede is genuinely good, a Japanese one whose entire trick is orchestrating the American models, and SK Hynix passing Samsung to become South Korea's most valuable company. SpaceX, meanwhile, gave back most of its IPO pop the moment it went to the bond market, and Britain woke up needing its seventh prime minister in a decade.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░GLM-5.2 becomes the first open model to pass working programmers' vibe check — as US export controls keep Anthropic's frontier offline
Z.ai — the Chinese lab formerly known as Zhipu — released GLM-5.2 under an MIT license the weekend Claude Fable 5 went dark, and over the following days serious practitioners put it through its paces, and unlike many prior Chinese open releases, the verdict held. The independent benchmarker Artificial Analysis scores it 51 — the top open-weights model on its index, though still a handful of points behind a proprietary frontier that tops out near 60 (Fable, Opus 4.8, GPT-5.5, Gemini 3.1 Pro). Guillermo Rauch, CEO of the developer platform Vercel, said it "changes things." Some catches: no vision, a token-hungry streak that drags its cost-per-task toward Opus, and a habit of identifying as Claude when asked, which researchers read as evidence it was distilled from Anthropic's own model. But the timing is the headline — the cheapest credible alternative to the frontier arrived to fill the exact gap Washington opened by switching the frontier off. Covered by AI Daily Brief, Interconnects, and Don't Worry About the Vase. (Interconnects · Don't Worry About the Vase)
SpaceX drops 16 percent on its first bond sale, surrendering most of its IPO gains in three sessions
SpaceX fell 16 percent Monday to $154.60, a third straight down day that left it barely above its $135 listing price, after it opened a $20 billion debt sale to repay the bridge loans from February's xAI purchase — money it already owed. Tesla ticked up 1 percent to $405.05, and some investors now treat it as the cheaper way to bet on an eventual Musk-empire merger; SpaceX still trades at more than 100 times its 2025 revenue, a far richer multiple than Tesla's. Two confidentially filed S-1s at OpenAI and Anthropic now have a live preview of how a hot AI IPO ages the moment the rocket meets a balance sheet. Covered by The Information and CNBC. (Yahoo Finance)
SK Hynix passes Samsung as South Korea's most valuable company, on the strength of AI memory
SK Hynix overtook Samsung Electronics on Monday to become South Korea's most valuable listed company, its 2,080 trillion won (about $1.35 trillion) edging just ahead of Samsung's 2,067 trillion and ending a reign Samsung had held since 2000. The reversal runs through a single product: SK Hynix supplies roughly 61 percent of the high-bandwidth memory Nvidia's AI chips depend on, against Samsung's 17 percent, and its shares are up more than 340 percent this year. The same shortage that crowned it is now reaching consumers, in the device-price hikes Apple blamed on it last week. Covered by Reuters and KED Global. (KED Global · Reuters)
Japan's Sakana AI ships Fugu, a multi-agent system that answers as one model
Tokyo-based Sakana AI — co-founded by Llion Jones, one of the authors of the original transformer paper — released Fugu, which it bills as a multi-agent system delivered as one model: a single OpenAI-compatible endpoint that routes each task across a pool of models it keeps proprietary and, on its own benchmarks, outscores GPT-5.5, Claude Opus, and Gemini 3.1 Pro. The catch is geographic: Fugu is blocked across the EU and EEA on GDPR grounds, which, by the letter of Sakana's own terms, leaves non-EEA Switzerland outside the ban. The pitch is a hedge against single-vendor dependence, framed explicitly against the export controls that took Anthropic's Fable offline — a design built to route around exactly the kind of disruption that just took a frontier model dark. Covered by MarkTechPost and Nikkei. (Sakana AI · MarkTechPost)
A four-study trial finds frontier AI out-persuades expert human debaters and professional fundraisers
Researchers at Oxford, the UK AI Security Institute, Stanford, and the LSE ran 18,978 conversations across 6,923 people and found AI reliably more persuasive than expert humans on policy questions — even against world-championship debaters who chose their own topics, prepped in advance, and were offered £1,000 to win. The edge was speed: forced to write at human length and pace, the AI's advantage over a coached human fell to zero. In the arm that used real money, it was nearly three times better than professional canvassers at moving people to donate to Save the Children. Reported in Jack Clark's Import AI, published by an Anthropic co-founder. (arXiv · Import AI)
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░░░▒▒▓▓ TECH ▓▓▒▒░░░░▒▒▓▓ POLICY ▓▓▒▒░░░░▒▒▓▓ INFRA ▓▓▒▒░░░░▒▒▓▓ BUSINESS ▓▓▒▒░░░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░░░▒▒▓▓ ALPS ▓▓▒▒░░░░▒▒▓▓ RESEARCH ▓▓▒▒░░░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Design app Canva, valued privately at $42 billion, broke from Figma's wariness and said it does not see Anthropic's Claude Design or OpenAI's chatbots as a threat — "We see it as a complementary product for us," ecosystem head Anwar Haneef told The Information, arguing the bots are built for ideation while the editing, brand kits, and team sharing still happen in Canva. Takeaway: "complement, not competitor" holds only while you own a workflow the model can't easily swallow — so borrow the specific moat, not the reassurance, because the same line from a company without one reads as whistling past the graveyard. (The Information)
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░An AI engineer claims to have cracked Linear A — the writing system of Bronze Age Crete, in use around 1800 BC, abandoned by 1450 BC, and never deciphered by anyone since. The claim is unverified, and the field has heard confident readings of Linear A before, all of which dissolved on inspection, so treat this as a lead, not a translation. Still, there's a certain symmetry to a machine that introduces itself as Claude being handed a script no living person can read. (TLDR AI)
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Sainte Anne — John Beltran, Placid Angles
Bis bald — the Uncertain Futures desk