In partnership with
In Todayโs Issue:
๐จ Meta's Muse crashes the image top two
๐ Beijing weighs walling off its best AI
๐ง MiniMax preps a 2.7-trillion-parameter model
๐ Grok 4.5's public launch, reportedly tomorrow
โจ And more AI goodnessโฆ
โก The Signal
Image generation just stopped being a moat and became a feature.
Meta shipped Muse Image, its first in-house model, straight into Instagram and WhatsApp, and it debuted at #2 on the public Text-to-Image Arena, behind only OpenAI's GPT Image 2. A year ago Meta was renting this capability from Midjourney and Black Forest Labs. Now every major lab has a competitive image model, so quality alone no longer wins; distribution, price, and control do. That is why the same week brings Beijing weighing limits on who can use its best models abroad and xAI rushing Grok 4.5 out the door. The interesting question is no longer image quality. It is who decides whose face ends up in the picture.
All the best,
Kim Isenberg
(South China Morning Post)
๐ง MiniMax Preps a 2.7-Trillion-Parameter Giant
China's MiniMax is preparing its biggest model yet, and it plans to give it away. According to The Information, the Shanghai startup is building a 2.7-trillion-parameter model, known internally as M3 Pro, that would be the largest from any Chinese lab and could arrive as soon as Q3 as an open-source release. It dwarfs MiniMax's current flagship M3 (428 billion parameters) and takes direct aim at rivals Zhipu, DeepSeek, and Moonshot.
๐ tl;dr: A free, giant Chinese model would pour more fuel on the global rush to cheap open-weight AI.
(Reuters)
๐ Beijing May Wall Off Its Best AI
China is weighing limits on the world's access to its most capable AI models, a mirror image of Washington's chip controls. According to a Reuters exclusive, the Ministry of Commerce has spent the past month studying curbs on frontier models from Alibaba (Qwen), ByteDance (Doubao), and Zhipu AI (GLM), plus making AI-model theft a national-security offense. Officials suggest any rules might apply only to future models, and it is unclear whether they take effect at all.
๐ tl;dr: China's cheap, capable open models won global share; now Beijing is debating whether to keep them home.
๐ญ Claude Fable 5 Is Free, But Not for Long
Anthropic's promotional free window for its Fable 5 model is closing. Through July 12 at 11:59 PT, subscribers on Pro, Max, Team, and eligible Enterprise seats can use Fable 5 at no extra cost, up to 50% of their weekly limits, with nothing to activate. After that, continued use needs separate usage credits, and free-plan users and API access stay excluded.
๐ tl;dr: Want to test-drive Fable 5 on your current plan? The free ride ends July 12.
With Meta, OpenAI, Google, and xAI all shipping strong image models this week, the quality of your results now comes down to the prompt you write.
Why it helps: The newer models plan before they draw, so they reward you for specifying shot, light, and mood, the details where amateur and professional outputs diverge.
Try this: Paste this into Meta AI, GPT Image, or Nano Banana: "A product photo of [your object] on matte concrete, soft window light from the left, shallow depth of field, 50mm lens, muted earth tones, one small hard shadow, photorealistic. Then give me the same shot in a bright, playful pop-art style."
๐ฌ Watch This
In my latest conversation, I sit down with Sam Stanwyck, who leads NVIDIA's quantum computing product team, for a refreshingly hype-free look at where the technology actually stands. We get into why NVIDIA is not building its own quantum computer, why the future is hybrid (CPUs, GPUs, and QPUs working together), and how today's AI and accelerated computing already handle the hard parts: calibration, simulation, control, and error correction. Stanwyck built quantum hardware at Rigetti and control systems at Keysight, so he treats this as a real engineering stack rather than a magic box. Worth 30 minutes for the grounded version of the quantum story.
"It's important that AI is distributed between all countries โ that there isn't one or two countries which are much stronger than the others."
โ Clรฉment Delangue, CEO of Hugging Face
(eMerge Americas)
SpaceXAI looks ready to take Grok 4.5 public as soon as tomorrow. According to a post from @SpaceXAI, strong feedback from its beta program has cleared the model for a public release, and Elon Musk has said the launch is set for July 9. The company positions Grok 4.5 as an Opus-class model that is faster, more token-efficient, and cheaper to run. Worth remembering: it has sat in private beta at SpaceX and Tesla since late June, and none of its performance claims have been independently benchmarked yet.
Meta Ships Its First Image Model, and It Lands at #2
The Takeaway
๐ Meta Superintelligence Labs shipped Muse Image, its first in-house text-to-image model, into Meta AI, Instagram Stories, and WhatsApp on July 7.
๐ It debuted at #2 on the public Text-to-Image Arena with a score of 1,280, behind only OpenAI's GPT Image 2 (1,385).
๐ Muse ends Meta's reliance on licensed models from Midjourney and Black Forest Labs, folding image generation into apps used by billions.
๐ The launch drew instant backlash: Muse can pull any public Instagram user's photos into AI edits by default, with opt-out buried in settings.
Meta just stopped renting its imagination. On July 7, Meta Superintelligence Labs, the unit run by Alexandr Wang, released Muse Image (codenamed "Mango"), its first image-generation model built in-house. It is free for everyday use inside the Meta AI app, Instagram Stories, and WhatsApp, with Facebook, Messenger, and the Advantage+ ad tools to follow. For a company that reaches more than three billion people a day, that distribution is the whole point.
(Meta)
The model is built to reason before it draws. It plans a layout, pulls in live web context, blends several photos into one scene, renders legible text and even QR codes, and lets you sketch edits directly onto a result instead of starting over. On the public Text-to-Image Arena, a leaderboard where people blind-vote on outputs, Muse Image entered at #2 with an Arena score of 1,280, behind OpenAI's GPT Image 2 (1,385) and just ahead of Reve 2.0 and Google's Nano Banana 2. Meta says Muse beats Nano Banana 2 on editing, though that claim is its own.
The context matters. A year ago Meta had no competitive image model and licensed the capability from Midjourney and Black Forest Labs. Muse ends that, and it arrives with Meta's data advantage baked in: it can tailor images using what your account already reveals about you, and it can @-mention public Instagram accounts to fold their photos into a generation.
Why it matters: That last feature is also the catch. Because photo-tagging is on by default, Muse can place real people's faces into synthetic images unless they opt out, and Meta says users will not be notified when it happens. One widely shared critique called it "a privacy landmine waiting to detonate." Meta has closed the quality gap; whether users accept the terms is the open question.
Sources:
๐ Meta Newsroom
๐ Axios
๐ CNBC
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: The Text-to-Image Arena ranks image models by Arena Score, an Elo-style rating built from head-to-head human votes on blind outputs. OpenAI's GPT Image 2 leads at 1,385. Meta's brand-new Muse Image debuts at #2 with 1,280, edging Reve 2.0 (1,271) and Google's Nano Banana 2 (1,270); xAI's Grok Imagine sits back at #10 (1,229).
The lesson: Below OpenAI, the field is a near-tie. Just 48 points separate #2 from #9 (1,280 down to 1,232), so a first-generation model from Meta landing second on day one shows how quickly image quality has commoditized. The differentiator now is distribution, where Meta's built-in audience dwarfs standalone rivals.
The caveat: Arena Score measures crowd preference on paired samples, not accuracy, safety, or prompt-faithfulness, and rankings still shift as models gather votes. GPT Image 2's 105-point lead is real, but one number hides where each model actually wins, whether that is text rendering, photorealism, or editing.
๐ง The Chip Built for AI That Works One Step at a Time
โก Bottom line: NVIDIA unveiled Vera, a data-center CPU built to make AI agents that work step by step run noticeably faster.
๐ก Why it matters: Agent loops stall when each step waits on the last, so raw single-core speed becomes the bottleneck, no matter how many cores you add.
๐ What it means: As software shifts from one-shot answers to multi-step agents, the CPU beside the GPU quietly sets how fast they feel.
An AI agent runs like a relay race. Each runner has to finish before the next can start, so adding more runners does not speed things up; you need each runner faster. That is the problem NVIDIA's new Vera CPU is built to solve. Modern agents work in loops: call a tool, read the result, decide the next step, and repeat. Because each step depends on the one before it, what matters is how fast a single core finishes one step. Piling on more cores does not help.
(NVIDIA)
For a decade, data-center chips optimized the opposite thing: pack in more cores to run many independent jobs cheaply. That is great for serving web pages and useless for an agent waiting on itself. NVIDIA says Vera's custom "Olympus" core delivers 50% more instructions per cycle than its previous Grace chip and about 1.8x the sustained per-core performance of a comparable x86 server on loaded agent workloads. Early tester Perplexity clocked a real coding workflow running 1.5x faster, with sandboxes starting up 1.9x quicker.
The benchmark is the smaller story. The bigger one is a shift in what a chip is even for. While GPUs get the spotlight, they sit idle whenever the CPU is busy running tool calls, executing code, or fetching data between model calls. As more of AI becomes agents that loop for minutes at a time, the unglamorous CPU next to the GPU increasingly decides how responsive the whole system feels. A skeptic will note these are NVIDIA's own numbers on NVIDIA-picked workloads, with independent tests still to come.
See what enterprise-ready AI support looks like
How are leading teams getting AI support to work? We're breaking down the playbook, live July 9.
