Week in Review — Kimi K3 at 2.8 Trillion, WAICO Gets 29 Signatures, and Gemini 3.5 Pro Slips (July 12–18, 2026)

A dense week, and an unusually well-documented one — four of the five major stories came with primary sources attached, which is not always how this goes. The exception is the one that moved the most market value.

Below is the week as it actually happened, with a note at the end about the parts that turned out not to have happened at all.

Moonshot ships Kimi K3

On July 16 Moonshot AI released Kimi K3, and the specifications are genuinely striking: a sparse mixture-of-experts architecture with roughly 2.8 trillion total parameters, a one-million-token context window, native vision input, and reasoning that cannot be turned off — the model ships with a single effort setting, and that setting is maximum. Pricing is three dollars per million input tokens and fifteen per million output.

The strategic detail that matters most is the one most easily garbled: the weights were not released at launch. API and app access shipped on July 16; open weights are promised by July 27. If that commitment holds, K3 becomes the largest open-track model ever released by a wide margin. If it slips, it was an API launch with good marketing. Those are very different outcomes and the week's coverage frequently collapsed them.

Worth watching: whether the open-weight release lands on schedule, and what the first independent evaluations say once anyone outside Moonshot can run the model against their own harness rather than the vendor-reported launch table.

Twenty-nine countries found WAICO

On July 16 in Shanghai, twenty-nine countries signed the agreement founding the World Artificial Intelligence Cooperation Organization, with UN Secretary-General António Guterres in attendance. The founding membership leans heavily toward the Global South — Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia and Pakistan among them. Headquarters will be in Shanghai.

The following day, July 17, Xi Jinping delivered the opening keynote at the World AI Conference — his first in-person appearance at the event since it began in 2018, which is the clearest available signal of how much weight Beijing is putting behind this. Xi described WAICO as "a major move by China to answer the call of the Global South."

Note the sequence, because it was widely reported backwards: the organization was founded on the 16th, the day before the keynote. Xi did not announce it in the speech; he ratified something already signed.

The substantive question is whether WAICO becomes a standards body with actual convening power or remains a diplomatic signal. Twenty-nine signatures is a real number. What matters next is whether any of them adopt binding technical standards through it.

Gemini 3.5 Pro is months behind

On July 16, Bloomberg reported that Google is months behind schedule on Gemini 3.5 Pro, citing ten current and former employees speaking anonymously. The attributed causes: coding capability that failed to improve as hoped after a late-June update to the training data, and engineers hitting capacity constraints from internal competition for compute. Sundar Pichai had signalled at the May developer conference that the model would ship in June.

Google's only on-record comment was that it is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners." Alphabet fell roughly four and a half percent on the report.

That is the entirety of what is sourced. It is also, notably, less than what most coverage of this story contained by the end of the week — see the closing section.

Fireworks AI raises at 17.5 billion

Fireworks AI announced a 1.505 billion dollar Series D on July 15 at a 17.5 billion dollar valuation, led by Atreides Management, Index Ventures and TCV, with Nvidia, Lightspeed, Bessemer and Menlo participating.

The number to hold onto is 17.5 billion, from the company's own announcement. Bloomberg had reported in late May that the round was in talks at 15 billion, and at least one trade publication published 17 billion this week. All three figures are now in circulation; only one of them is the actual valuation.

The broader read: inference serving continues to be priced as an independent infrastructure layer rather than a feature of the model labs, which is the bet this round is underwriting.

Stanford economists say act now

On July 13 the Stanford Digital Economy Lab published "We Must Act Now: A Statement on AI's Transformation of the Economy," signed by more than 200 economists and AI researchers including sixteen Nobel laureates. The statement argues AI "may become radically more powerful over the next 10 years" and could drive "an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame."

The signatory list is the story. Daron Acemoglu and Simon Johnson — 2024 Nobel laureates and, until recently, among the more prominent skeptics of aggressive AI-displacement forecasts — both signed. When the skeptics sign the urgency letter, the distribution of expert opinion has moved.

Also this week

OpenAI's GPT-Live, the full-duplex voice architecture that processes input while generating output, shipped on July 8 — a week before this reporting window. It appeared in several roundups this week as a fresh release. It was not.

The part that did not survive verification

A closing note, because it turned out to be this week's most instructive story.

The Gemini reporting above is thin on technical detail by design — Bloomberg's sources gave one capability area, coding, and one structural cause, internal compute contention. By July 18, a dozen publications were carrying substantially more: that the model failed specifically on long-horizon reasoning, that Google had scrapped a base model, that this was a third missed deadline, that there had been a July 17 launch date, that the failures involved recursive tool-calling collapse and SVG structural consistency, and that the model sits in a limited enterprise preview.

None of that appears in the original reporting. None of it is attributed to any source. Several claims contradict each other across outlets. As far as I can determine, it was generated downstream to fill word count.

I nearly published three of those claims myself before an adversarial verification pass killed them, which is the subject of this week's article on how the AI news supply chain manufactures the facts it reports. The short version for digest readers: multiplicity is not corroboration anymore. Ten outlets asserting the same unsourced detail is one claim with nine copies, and the copies are better optimized for search than the original is.

Every claim in this digest was checked against a primary source or a named institution with accountability for it. Where I could only get to anonymous sourcing — the Gemini story — I have said so explicitly and stopped there.