Today in AI — August 9, 2026
A day when AI's economics got loud: a hedge fund bet $400 million on chip hardware days after nearly collapsing, China's model race pivoted from cheap tokens to premium pricing, and Washington proposed taxing usage itself — while AI-assisted research toppled another decades-old open problem.
Models & Research
Zhipu confirmed GLM-5.3 is about to launch, with Chinese media reporting the next flagship jumps past the trillion-parameter mark — and Goldman Sachs raised the company's full-year revenue forecast by 35 percent, lifting its estimate for China's AI model market to $13 billion. The upgrade lands as DeepSeek announced API price hikes, the first crack in the low-price war that defined China's model market.
Tencent Hunyuan's WorldClaw turns a single text prompt into a freely explorable 3D world, with planning agents deciding what goes where and every object left as an editable mesh — explicit 3D you can fly through and edit, not camera-locked video.
Google DeepMind open-sourced WeatherNext 2 and WeatherNext Cyclones, the models behind a Nature paper showing AI cyclone forecasts can buy an extra day of warning — open weights give local forecasters the same tool Google operationalized in October.
A ChatGPT-assisted counterexample toppled the 30-year-old HRT conjecture in harmonic analysis: four mathematicians exhibited a smooth function and 12 time-frequency shifts whose weighted sum vanishes exactly, with AI supplying the initial strategy and numerics while the final proof was written by hand. Terence Tao called the disclosure "responsible."
An independent run reproduced DeepSeek V4 Flash's 82.7% Terminal-Bench 2.1 score to the decimal with fully public tooling and a complete job ledger — the first outside audit of a flagship agentic benchmark claim.
A community researcher packaged DeepSeek V4 Flash with a latent-reasoning head as a runnable model: a small head compresses reasoning into a 1024-dimensional latent instead of emitting tokens, with one latent step standing in for roughly six reasoning tokens.
A Chinese team says it ran the full 671B DeepSeek-V3/R1 on the Lingsheng supercomputer using only CPUs — 16 nodes of the LX2 CPU platform matching an 80-GPU cluster's output throughput at high batch sizes, a demonstration that CPU-only hardware can serve large-model inference.
OpenAI says two API settings tripled GPT-5.6 Sol's ARC-AGI-3 score, from 7.8% under the official harness to 38% with retained reasoning and compaction enabled — the latest evidence that benchmarks measure the harness as much as the model.
Industry
Situational Awareness poured $400 million into Source Foundry, a stealth chip-manufacturing startup targeting ASML's lithography monopoly — days after the fund nearly collapsed in a margin-call spiral, bringing its total commitment to $500 million. Sequoia-backed Source Foundry is one of several startups trying to break the "chip wall" with new lithography physics.
SAP froze most travel and hiring because of AI's soaring cost, with exceptions made only for AI-related roles and travel — and a current employee says the bans remain in force a month later, even as the company rolls out a new internal AI tool to its entire workforce.
Amazon started building a $2 billion AI data center in Gilroy, California that residents never got to vote on, under zoning rules set in 1981 that required no public meetings and no council vote — the public comment window lapsed back in 2024.
London's King's Cross is now one of the world's top AI hubs, rivaled only by San Francisco and Beijing — DeepMind, OpenAI, Anthropic, and Meta among the tenants, with AI companies leasing more than a million square feet of office space in the city since June.
DeepSeek is taking a strategic stake in robot maker Unitree's Shanghai IPO, paying about 141 million yuan for a placement locked up for 36 months — the brain-meets-body playbook for China's robot-AGI ambitions.
CICC is the quiet winner of China's AI listing boom, the Financial Times reports: the state-owned bank sponsored the multibillion-dollar IPOs of memory chipmaker CXMT and optical-transceiver maker Zhongji Innolight, with deal volume reaching $11.5 billion this year.
LinkedIn added a "Seems like AI slop" report button as part of a crackdown on machine-written posts — new classifiers, private notices to overly AI-dependent members, and a retreat from its own AI rewriting tool, after research found up to 41 percent of long-form LinkedIn posts appeared fully AI-generated.
PwC's survey of 4,454 CEOs found 56 percent still see no measurable revenue or cost benefit from AI — the returns gap that is fueling both worker anxiety and legislative intervention.
The Verge's Emma Roth argues AI-writing detectors are creating a new era of distrust: the tools remain unreliable, yet educators and publishers are adopting them anyway, poisoning classrooms and newsrooms with blanket suspicion.
Policy
House Democrats introduced the AI Tax and Work Protection Act, a levy on tokens and AI product revenue that would fund a WPA-style federal jobs program, with the rate rising and falling with unemployment. It has no path through a Republican-controlled House, but it is the first serious federal attempt to tax AI usage itself — and today's deep dive argues the arc from joke to legislation is the real news.
The Institute for Progress made the case for building "pacing" tools for automated AI research now, in a Noahpinion essay that previews 23 policy recommendations — arguing for targeted thresholds on the riskiest automated R&D rather than a broad halt.
Jill Lepore's new book argues Big Tech is quietly building an "artificial state" — government without consent, rule by algorithms and corporations — and her frame sharpens today's Gilroy and AI-tax stories: the machinery of the artificial state arriving without a vote, and a state buying back its own capacity.
Tools
A Kansas City researcher put anti-surveillance patterns on Kickstarter after demonstrating at Def Con that printed patterns can stop AI cameras from detecting people, faces, and vehicles — his noRecognition system defeated all 11 open-source detection systems he tested, including software that powers Flock license plate readers.
OpenChamber, an open-source agentic development environment built on the OpenCode SDK, hit the Hacker News front page — the same agent workspace on desktop, browser, phone, and inside VS Code, with on-device sessions and roughly 7,800 GitHub stars.
Who should decide what AI costs society — the labs, the market, or Congress? Write to us via our contact page.
Sources: OpenChamber (GitHub)
