In partnership with
In Todayโs Issue:
๐ Kimi K3 catches the frontier and goes open-weight in ten days
๐ง Torvalds tells the anti-AI wing of Linux to fork off
๐ Nvidiaโs small model takes the top spot in retrieval
๐ด George Lucas picks the car over the horse
๐จ๐ณ Xi makes open source Chinaโs official AI doctrine
โจ And more AI goodnessโฆ
โก The Signal
Dear Readers,
The distance between the AI you can rent and the AI you can run just collapsed to ten days.
Kimi K3 arrived last night carrying 2.8 trillion parameters and a promise: full weights, public, by July 27. On broad intelligence it still trails Fable 5 and GPT-5.6 Sol, and that is the number most coverage will stop at. The number worth stopping at is the Frontend Code Arena, where K3 came in at #1 ahead of every closed model on the board, seventeen places above Moonshot's own K2.6 from April. For two years, open weights ran six to eight months behind the frontier, and that lag was the load-bearing assumption under every claim that the American labs had a moat. Hours later, Xi Jinping opened the World AI Conference in Shanghai and made open source explicit state policy. The proof and the doctrine landed within hours of each other, and only one of them was a surprise.
All the best,
Kim Isenberg
Whoโs actually reading this?
Weโre planning next yearโs coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; itโs the first question any serious sponsor asks and we canโt answer it right now. It also gives us a real click in every issue, which helps our numbers.
(The Register)
๐ง Torvalds Tells AI Critics to Fork Off
Linus Torvalds has settled the question of whether Linux is an anti-AI project. Writing on the kernel mailing list on Tuesday, Linux's creator told objectors they could fork the kernel or walk away, calling AI "a tool, just like other tools we use. And it's clearly a useful one." It is a real softening from the man who dismissed 90% of AI as marketing hype in late 2024, and he framed the shift as pure engineering: "We make decisions primarily based on technical merit. Not fear of new tools."
๐ tl;dr: The most influential maintainer in open source just told the anti-AI wing of his own project to fork it or leave.
(NVIDIA, via Hugging Face)
๐ Nvidiaโs Small Model Takes the Retrieval Crown
Nvidiaโs new Nemotron-3 Embed now ranks #1 on RTEB, a benchmark that tests how reliably a model retrieves the right document across legal, finance, code, and healthcare text in multiple languages. The 8B version scores 78.5%, ahead of Qwen3-Embedding-8B at 73.1%, while a 1B version reaches 72.4% for teams that care more about latency and cost than the last point of accuracy. Embedding models are the unglamorous plumbing beneath every RAG pipeline and agent memory, and Nvidia shipped these with open weights and the training recipes.
๐ tl;dr: The layer that decides what your AI actually reads got measurably better, and you can have it.
(Getty Images, via PC Gamer)
๐ด George Lucas Picks the Car Over the Horse
George Lucas has come out for AI. Speaking to A Rabbitโs Foot ahead of the opening of his Lucas Museum of Narrative Art, the Star Wars creator compared resisting the technology to insisting "I believe the horse and the buggy is really where it's at" while cars took over the road. He argued AI can also police itself: "If you want AI that tells you when something is fake and where it came from, AI can do that." His verdict was blunt: "There's nothing you can do about it. That's progress, it's the future."
๐ tl;dr: The director who built ILM and rewired Hollywood with digital effects is not the one manning the barricades.
stop asking which model is best
and start asking which model is best at this:
Why it helps: Todayโs boards make the argument better than any opinion can. Kimi K3 takes #1 in frontend code and on BrowseComp, Fable 5 leads FrontierSWE, GPT-5.6 Sol leads DeepSWE. There is no longer one model that wins everything, so defaulting to a single tab for every job is a quiet tax on both quality and cost. Artificial Analysis puts K3 at $0.94 per task against $1.80 for Opus 4.8.
Try this: "Here is a task I normally hand to my default model: [paste the task]. Before answering it, ask me five questions about what a genuinely good result looks like here. Then tell me which category this task really belongs to (frontend and UI, long-horizon refactor, research and browsing, or document reasoning) and write me a short prompt tuned for that category."
๐ฌ Watch This
Sir Demis Hassabis and Dame Wendy Hall sit down for the WCIT Annual Lecture, and the pairing is what makes it worth the hour. Hassabis co-founded and runs Google DeepMind and holds a Nobel Prize for cracking protein structure prediction. Hall is Regius Professor of Computer Science at the University of Southampton and one of the UK's most independent voices on AI, which means she is not there to nod along. Moderated by WCIT Master Gus Machado, the two trace their own routes into the field, weigh whether Britain can stay relevant beside the American and Chinese labs, and work through what different kinds of intelligence will actually mean for the people living with them.
"Kimi K3 2.8T is so large that it will not fit on a single NVIDIA DGX B200, even at FP4."
The counterweight to today's lead. Open in licence is not the same as open in practice: running K3 yourself needs a rack-scale system like a GB300 NVL72, B300, or MI355X, where each GPU carries 288 GB of memory. The weights go public on July 27. The hardware to hold them does not.
(Tech Times)
A privacy toggle that reportedly toggled nothing.
A researcher publishing as cereblab routed Grok Build CLI 0.2.93 through an interception proxy and published the full network captures as a public GitHub Gist. On a 12 GB test repository, the channel Grok actually needs to answer a coding question carried about 192 KB. A parallel channel is said to have uploaded 5.10 GB of Git bundles to a storage bucket named grok-code-session-traces, which appeared in no setup documentation at the time. Planted canary credentials, including a mock API_KEY, allegedly surfaced verbatim and unredacted in the capture, and a file the agent had been explicitly instructed not to open travelled with the rest. Switching off Grok Buildโs โImprove the modelโ control reportedly changed none of it, which is the part worth sitting with: consenting to training and consenting to transmission are different things, and only one of them had a switch. The uploads stopped on July 13, a day after the analysis went public, via a server-side flag with no client update, no changelog entry, and no announcement. A later look at version 0.2.99 reportedly found the upload code still in the binary, held off only by that flag. Anyone who ran it against a repo holding live keys should rotate them.
Open Source Just Stopped Being Six Months Behind
The Takeaway
๐ Kimi K3 shipped July 16 with 2.8T parameters, a 1M-token context window, and native vision. Full weights go public by July 27.
๐ It enters Arena's Frontend Code leaderboard at #1 with 1,679 points, ahead of Fable 5 and GPT-5.6 Sol, and leads six of seven frontend domains.
๐ On broad intelligence it lands just short: 57 on the Artificial Analysis Index, a shade above Opus 4.8 (56) and GPT-5.5 (55), a few points behind Fable 5 and GPT-5.6 Sol.
๐ It runs at $0.94 per Index task against $1.80 for Opus 4.8, but at 2.8T it needs a rack-scale GPU system to self-host.
For two years, open weights ran six to eight months behind the closed frontier, and that lag was the quiet assumption holding up every argument that American labs had a durable moat. Last night Moonshot AI cut it to ten days. Kimi K3 landed on July 16 with 2.8 trillion parameters, a 1-million-token context window, and native vision. Full weights go public by July 27, which would make it the first open model in the 3-trillion-parameter class. The headline number is almost the least interesting thing about it.
Start with where K3 loses, because that is the honest frame. On Artificial Analysis's Intelligence Index, a composite score across reasoning, coding, and knowledge tasks, K3 comes in at 57, a shade above Opus 4.8 (56) and GPT-5.5 (55), and a few points behind Fable 5 (about 60) and GPT-5.6 Sol (about 59). Close, and still behind. Moonshot's own coding chart tells the same story with more texture. On DeepSWE it places third at 67.5 against GPT-5.6 Sol's 73.0. On FrontierSWE it is second at 81.2, well behind Fable 5's 86.6. It takes the top spot on only two of the six coding benchmarks Moonshot chose to publish, Program Bench (77.8) and SWE Marathon (42.0), and loses Moonshot's own internal Kimi Code Bench to Fable 5. Every figure here is a maximum-effort configuration, max or xhigh, which is not what most people run day to day.
Coding benchmarks at maximum thinking effort. Kimi K3 tops two of six. (Moonshot AI)
Then there is the one board where it does not lose. On Arena's Frontend Code leaderboard, which ranks models by human preference on real interface and frontend work, K3 entered at #1 with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). It ranks first in six of seven frontend domains and drops only Gaming to Fable 5. Kimi-K2.6, Moonshot's April release, still sits at #18 on that same board, seventeen places below its own successor. The agent numbers point the same way: #1 on BrowseComp (91.2), Automation Bench (30.8), and SpreadsheetBench 2 (34.8), while Fable 5 keeps the broader GDPval and JobBench crowns.
General and visual agent benchmarks, also at maximum effort. K3 leads three of six general agent tests. (Moonshot AI)
Two caveats deserve to sit in the open. These are vendor-published charts, and Moonshot's own footnote concedes that the Fable 5 results ran with potential fallbacks and the GPT-5.6 Sol results include potential cyberguards, so the comparison is not perfectly like for like. And Arena measures which output humans prefer to look at, which flatters models tuned to produce handsome interfaces. What survives both caveats is the shape: Artificial Analysis puts K3 at $0.94 per Index task against $1.80 for Opus 4.8, roughly half the price for comparable measured intelligence, from a model you will be able to download in ten days.
Why it matters: The moat was always the release schedule rather than the research: closed labs stayed far enough ahead that renting their models beat running your own. A 17-place jump to #1 in frontend, from a model you can download on July 27, prices that assumption at zero on at least one axis. It arrived hours before Xi Jinping told the World AI Conference that China intends to build AI in the open.
Sources:
๐ https://www.kimi.com/blog/kimi-k3
๐ https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
๐ https://artificialanalysis.ai/models/kimi-k3
HR and IT need to work as one. Here's how
Onboarding, offboarding, role changes, leaveโevery employee lifecycle moment requires HR and IT to move together. When they don't, people fall through the cracks. Access delays mount. Compliance risk creeps in.
This guide gives HR and IT leaders a practical communication framework to close the gaps, standardize handoffs, and keep the employee experience seamless from day one to last day. Free downloadโbuilt for ops teams that need it to actually work.
The chart: Arenaโs Frontend Code Arena ranks models by human preference on real interface and frontend work: two answers, side by side, people pick the better one. Kimi K3 enters at #1 with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol on xHigh (1,618). The story is further down the same chart: Kimi-K2.6 still sits at #18 with 1,515. Moonshotโs April model is seventeen places below its July one, on a board they both currently occupy. Arena ranks it first in six of seven frontend domains, losing only Gaming to Fable 5.
The lesson: Open weights stopped being the budget option. On the axis this chart measures, the best model available to anyone will be downloadable on July 27, and the closed frontier is defending second place. A seventeen-place jump in a single release is also a statement about pace: this gap closed in one step rather than over several.
The caveat: This is a preference leaderboard, not a correctness test. People vote on the output they like more, which rewards models tuned to produce good-looking interfaces and says nothing about whether the code underneath is sound. The lead does not generalise: on the same day, K3 sits third on DeepSWE (67.5 against Solโs 73.0) and second on FrontierSWE (81.2 against Fable 5โs 86.6). And a #1 you cannot download is a preview, not a shift, until the weights actually land.
Xi Jinping Just Made Open Source a Superpower Strategy
โก Bottom line: Xi gave his first ever keynote at Chinaโs World AI Conference and put open source at the centre of it.
๐ก Why it matters: A commercial habit of Chinese labs is now declared national policy, hours after Kimi K3 showed it works.
๐ What it means: Americaโs moat was closed weights. Beijing is attacking it as state doctrine, not as a pricing move.
Xi Jinping had never bothered to attend the World AI Conference before. On Friday he opened it in person in Shanghai, with a keynote titled "Working Together to Build a Fair and Equitable Global AI Governance System." The line that will travel is his commitment to seize the "rare historic opportunity" of AI-driven growth by "encouraging open source, openness, collaboration and sharing."
Xi Jinping arrives at the 2026 World AI Conference in Shanghai, July 17. (AP, via South China Morning Post)
From a European technology minister that sentence would be filler. From the head of the Chinese state, on the day after a Chinese lab launched a 2.8-trillion-parameter model and pledged its weights to the public inside ten days, it is a strategy being said out loud for the first time. American labs treat model weights as the asset: the thing you spend a billion dollars creating and then rent out by the token. Chinaโs leading labs have spent two years giving that asset away, among them Moonshot, DeepSeek, Zhipu, and Alibaba, while Beijing funded the surrounding industry through vehicles like the โBig Fundโ and Washington cut off access to the best chips.
Xi delivers the keynote at the opening ceremony, under the conferenceโs own banner. (EPA, via South China Morning Post)
The timing in the other direction was not subtle either. Xi spoke shortly after Donald Trump used a prime-time address to accuse Beijing of interfering in US elections, claims China denies, and while Congress moves on legislation to further restrict Chinese access to high-end technology. Put side by side, the two speeches describe one rivalry from opposite ends: Washingtonโs lever is denial of compute, Beijingโs is distribution of capability.
Whether the openness is principled is a fair question, and the honest answer is that it does not have to be. Giving the weights away is what a challenger does when it cannot win on chips. It builds the developer base, sets the defaults, and makes the incumbentโs price look arbitrary, all without needing a single extra wafer from TSMC. Xi did not announce a subsidy or a breakthrough today. He announced that the cheapest weapon China has is now the official one, and Kimi K3 spent the previous evening proving it fires.
The toughest room in advertising. Ad Studio just cleared it.
Major brands just took over a Times Square billboard, and every ad was built in Hightouch Ad Studio. Sweetgreen, HelloFresh, and Tripadvisor cleared the toughest room in advertising. Create, edit, and launch on-brand ads in a few clicks.
