The pace slowed at WAIC by day two, but the week's more durable signals came from outside Shanghai. A former OpenAI CTO shipped the largest US open-weight model. Sixteen Nobel economics laureates put a deadline on workforce policy. And the head of Google DeepMind called for a regulatory body — not after harm occurs, but before the next frontier model ships.
Mira Murati Ships Inkling: 975 Billion Parameters, Fully Open
Thinking Machines Lab released Inkling on 15 July, making it available on Hugging Face under an Apache 2.0 licence. The model is a sparse mixture-of-experts system: 975 billion total parameters, 41 billion active per token, trained on 45 trillion tokens across text, images, and audio. It natively handles all four modalities without separate encoder modules.
Murati's framing is a direct challenge to the general-purpose closed-model thesis. Thinking Machines argues that enterprises ultimately care less about the smartest model available than one they can make their own. Inkling is wired directly into Tinker, the company's fine-tuning platform, and its architecture supports deep customisation without full re-training. The model's context window reaches one million tokens.
One design decision will draw attention: Thinking Machines trained explicitly for resistance to censorship, meaning the model is calibrated to answer politically or institutionally sensitive questions rather than reflexively refusing them. The company is careful to separate this from safety — Inkling scored 98.6 per cent on StrongREJECT, the benchmark for genuine harmful-request refusals. The bet is that over-refusal is itself a cost enterprises will no longer accept from closed APIs, and that a downloadable, modifiable model shifts the negotiation.
Sixteen Nobel Laureates: Act Now on AI's Economic Impact
A letter organised by economists Erik Brynjolfsson, Ajay Agrawal, Anton Korinek, and Tom Cunningham, published on 13 July by Stanford's Digital Economy Lab and now carrying over 2,000 signatures, warns that AI may bring "an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame." Among the sixteen Nobel laureates who signed: Daron Acemoglu, Simon Johnson, Paul Krugman, Ben Bernanke, and Joseph Stiglitz.
The signatories are not calling for a slowdown. They are calling for preparation: deeper research into AI's distributional economic impacts, and the policy frameworks and institutions that would allow productivity gains to be broadly shared rather than concentrated. The shift in tone matters. Acemoglu and Johnson won the 2024 Nobel in part for their work on how institutions determine whether technology benefits accrue broadly or narrowly — they are not reflexive pessimists. Their signature carries a specific credibility.
For an operator: the workforce-planning conversation has moved from conjecture to professional consensus. Waiting for further evidence before acting is now a position that requires defending.
PrismML Puts a 27-Billion-Parameter Reasoning Model on a Phone
PrismML released Bonsai 27B on 14 July under Apache 2.0, claiming it as the first 27-billion-parameter model capable of running on consumer mobile hardware. The 1-bit quantised variant compresses to 3.9 GB and runs at 11 tokens per second on an iPhone 17 Pro, and up to 87 tokens per second on an M5 Max. The model retains 90 to 95 per cent of full-precision benchmark performance across maths, coding, reasoning, and vision tasks.
Bonsai 27B is based on Qwen3.6 27B, supports the full 262,144-token context window of the base model, and handles text, vision, tool calling, and multi-step agentic workflows. CNBC reported that Apple is in discussions with PrismML about the compression technology, though no agreement has been confirmed by either party.
The operational implication extends beyond mobile. Cloud dependency for inference is not a permanent constraint — it is a consequence of weight size. A reasoning-capable model at 3.9 GB opens a category of offline, low-latency, and data-sovereign applications that cloud API economics cannot reach. For regulated industries with strict data residency requirements, this is the more significant development.
Demis Hassabis Proposes a FINRA for AI, Operational Before Year End
In a personal manifesto published 14 July, Google DeepMind CEO Demis Hassabis called for a US-led independent standards body to vet frontier AI models before release. His proposal, covered by Axios and TechCrunch, borrows from FINRA, the Financial Industry Regulatory Authority: frontier labs would initially share models voluntarily, up to 30 days before deployment, for safety testing covering cyber, biological, and deception capabilities. Once the regime proved reliable, mandatory participation would follow. Hassabis set an explicit target: operational before the end of 2026.
His stated motivation was the Anthropic Mythos and Fable export-control episode in June, in which the US government improvised a 2.5-week negotiation with no established rules or protocol. The absence of procedure imposed real cost on a US company and revealed a structural gap that no amount of voluntary commitment fills.
Two facts make this proposal consequential. First, it comes from the CEO of a frontier lab, not a regulator — a credibility signal that pure-advocacy proposals lack. Second, the White House voluntary frontier-model standards framework already exists. A FINRA-style body would formalise and extend that framework with independent testing authority rather than relying on self-assessment. The gap between those two positions is where the policy debate will be fought for the next six months.
SAP Closes €1 Billion Bet on Tabular Foundation Models
SAP confirmed the close of its acquisition of Prior Labs in July, completing a transaction announced in May at an investment of over €1 billion across four years. Prior Labs, based in Freiburg, Germany, will operate as an independent entity and is positioned to become a frontier AI lab for structured data.
Prior Labs pioneered Tabular Foundation Models: AI architectures purpose-built for rows and columns rather than text or images. Its TabPFN model series, published in Nature, set state-of-the-art results on tabular benchmarks across hundreds of independent academic studies. Practical applications include predictive maintenance, demand forecasting, customer churn analysis, and cash-flow prediction running directly on SAP data.
For most enterprises, structured data is the majority of AI-relevant data. The LLM-first framing of enterprise AI has consistently underweighted this reality: language models are not well-suited to structured prediction at scale. SAP's wager is that tabular foundation models will become the infrastructure layer for the share of enterprise data that never approaches a chat interface — which, by most estimates, is the larger share.
The through-line this week is legibility. Open weights make frontier-class AI legible and modifiable. Nobel economists make the labour-market stakes legible enough that policymakers can no longer plead uncertainty. Hassabis makes the governance gap legible enough to propose a specific institutional solution. And SAP makes the structured-data blind spot legible as a market. An operator who reads only the headline model releases is missing the deeper pattern: AI is being disaggregated, and the leverage lies in knowing which piece fits which problem.