This Week in AI: Google Just Got Beaten by a Free AI Model.... Kimi K3

[00:00:00]This is your AI news rundown for the week of July 18th, 2026. 15 stories straight through. A flagship model that still hasn't shipped. Open AI's first piece of hardware. A 27 billion parameter model squeezed onto a phone.

[00:00:13]Meta pulling a tool after backlash. And 16 Nobel laureates sounding the alarm.

[00:00:18]Let's get into it. First up, Google's Gemini 3.5 Pro. Despite all the buzz, it has not launched. Google blew past its original June target. And according to reporting, missed another deadline this month with zero official specs, no model card, no pricing.

[00:00:33]Bloomberg says the hold up is disappointing coding performance and checkpoints that actually regressed.

[00:00:38]Reporting adds the rebuilt model still hallucinates and trails Open AI's GPT 5.6.

[00:00:44]But that part is from leaks, not Google.

[00:00:46]The confirmed core, it hasn't shipped.

[00:00:49]And Google is now the only major lab without a 2026 flagship out the door.

[00:00:52]But while Google stalls, someone did ship. Moonshot AI launched Kimmi K3, a 2.8 trillion parameter open model available now through its API with a million token context window. Moonshot says it beats GPT 5.5 and Claude Opus on coding and independent leaderboards back part of that. It reportedly tops Arena's web dev coding board ahead of GPT 5.6 and Claude Fable 5. The open weights land July 27th. Next, Open AI just shipped its first hardware and it's a $230 keypad. The Codex Micro, a limited run programmable macro pad built with Work Louder, is designed to drive Open AI's Codex coding agents. It's a novelty, not their marquee device. That screenless AI companion is still years out. But the timing is a story. It dropped just 5 days after Apple sued Open AI alleging ex-Apple staff took trade secrets to build Open AI's hardware. So, Open AI's very first gadget is landing in the middle of a lawsuit over how it builds gadgets. Then there's a 27 billion parameter model that fits in about 4 GB. Prism ML's Bonsai 27B is a natively low-bit build of Qwen. Its 1-bit version is just 3.9 GB. And it's not theoretical, it runs up to 163 tokens per second on an RTX 5090 and up to A7 on an Apple M5 Max. That's a frontier class model running on consumer hardware released under Apache 2.0. The compression tax that made big models desktop only is shrinking fast.

[00:02:13]Everybody suddenly wants their own AI chips. Google's is real and shipping its eighth generation TPUs designed in-house.

[00:02:22]Apple is different. It's only reportedly hunting AI chip acquisitions to cut its Nvidia dependence per the information.

[00:02:28]No target, no deal, no Apple statement.

[00:02:31]And in the background tool giant ASML has signaled price hikes officially just pricing power with a reported 10% figure ASML hasn't confirmed. Google shipping, Apple shopping, and the whole supply chain feels it. Meanwhile, Hyundai workers walked out but not purely over robots. On July 13th, unionized workers in Ulsan began a three-day partial strike as Bloomberg reported. The primary fight is bonuses and stalled wage stocks. Job guarantees against Hyundai's AI-driven humanoid robots, Boston Dynamics' Atlas, are a real demand, but a secondary one. It's an important nuance. This isn't the first AI strike, it's a bonus dispute with an AI clause. And as talks deadlocked it escalated toward longer daily stoppages.

[00:03:10]Also this week, TikTok Shop is banning AI from live streams. New quality rules bar AI-generated voices, pre-recorded audio, and static or slideshow visuals from shopping live streams and now require real-time human narration, but note what it's not. It's not a blanket ban on AI influencers. Avatars just can't film more than half the screen and AI is still fine for scripting and editing. Why do it? This lands against a projected $23.4 billion US TikTok Shop and survey data showing shoppers are about four times more likely to trust a brand less when they spot AI. Overseas, the global AI order is visibly splitting in two. On July 15th, China's regulator finally approved Apple Intelligence after a roughly 22-month wait, but only with Alibaba's Quinn powering the on-device AI and Baidu handling visual search. American AI running on Chinese models to clear China. And that very same week, Xi Jinping launched a 29-nation World AI Cooperation Organization in Shanghai, a rival block to the US-led order. Two stories, one message, the AI world is picking sides.

[00:04:12]ChatGPT's voice now talks like a person.

[00:04:15]On July 8th, OpenAI launched GPT Live, a full-duplex voice model that listens and speaks at the same time, so it can be interrupted and answer naturally, replacing the old advanced voice mode.

[00:04:25]The mini version is now the free default. Harder questions get handed to GPT 5.5 in the background. In testing, users preferred it over the old voice mode about 76% of the time. Meta, meanwhile, built a tool to put your face in AI images then pulled it. Its Muse image model let people generate images using the likeness of anyone with a public Instagram account. Opt out, not opt in, just tag them, no notification.

[00:04:47]Private accounts and minors were excluded, but public citizen Sag Aftra and the talent agency CAA pushed back hard. And within days, Meta disabled the feature saying it missed the mark. So, this one's already over, but it's is a preview of the consent fights ahead. And there's a second Meta story, 8,000 jobs cut for AI that its own CEO admits isn't working yet. That's about 10% of the workforce gone in May while Meta guides to between 125 and 145 billion dollars in 2026 AI spending. Then at a July 2nd town hall, Zuckerberg conceded that over the prior 4 months, AI agent development hasn't really accelerated in the way that we expected. Record spending, deep cuts, and an honest admission the payoff isn't here yet. China's biggest fintech is pouring money into humanoid robots.

[00:05:31]Alibaba affiliate Ant Group led a 500 million yuan round, about 73.6 million dollars, in robotics startup Zeroth, which says it already has orders for more than 30,000 units. And this isn't a one-off, it's Ant's 12th robotics deal since the start of last year. Pair it with that Hyundai strike and the theme is clear. Humanoid robots are scaling fast and the money knows it. Here's a weird one, an AI failure mode and an open source fix.

[00:05:55]Reasoning models sometimes fall into doom loops, repeating themselves until the context window runs out. Liquid AI just open source anti-doom, a training technique aimed at exactly that. On a third-party 4 billion model, it cut the doom loop rate from 22.9% to 1%. On one of Liquid's own checkpoints from 10 10 point code on 10 10 point down to 1.4.

[00:06:15]Codes on GitHub Apache licensed. A small concrete look at how these models actually break and get patched.

[00:06:20]>> Then there's a startup that raised $52 million to read your brain. Israeli company Hemisphere came out of 6 years of stealth to launch Descartes, a 6 billion parameter foundation model that decodes non-invasive brain activity EEG into quantitative diagnostics for psychiatric and neurological conditions.

[00:06:37]The hook, it was co-founded by a co-inventor of Apple's Face ID. The pitch is reading the brain like a blood test from a 15-minute headset scan, a foundation model pointed at the mind.

[00:06:47]And the big one, 200 economists and 16 Nobel laureates just warned the world about AI. Organized through Stanford's Digital Economy Lab and released July 13th, the statement "We must act now" argues AI could drive a transformation larger than the Industrial Revolution, but compressed into a far shorter time frame with real risk of mass job displacement and widening inequality.

[00:07:08]When that many economists and that many Nobel laureates sign the same letter, it's worth reading. That's 15 stories the pattern this week. Huge spending, honest admissions, the payoff isn't here yet. Robots and chips scaling fast and a lot of very smart people asking everyone to slow down and think. That's the rundown. See you next week.