Morning Digest, July 8, 2026

12 newsletters, 12 overlapping stories

Top Stories

Anthropic finds a hidden “workspace” inside Claude

(4 newsletters)

Anthropic published research describing what it calls “J-space,” a small set of internal neural patterns that emerged during training and act like a silent mental workspace where Claude can reason through intermediate steps without ever writing them into its chain of thought. Researchers can read what appears there and even edit it to change the model’s answers, which opens a path to catching deception, hidden goals, or test-awareness before a model speaks. Anthropic is careful to say this does not prove Claude is conscious, but it shifts AI safety from watching what models say toward inspecting what they may be thinking.

Meta enters the AI image race with Muse Image

(2 newsletters)

Meta released Muse Image, the debut model from Alexandr Wang’s Superintelligence Labs, which opened at No. 2 on Arena’s text-to-image and editing leaderboards behind only OpenAI’s GPT Image 2. It is free inside Meta AI and rolling out across Instagram and WhatsApp, with the ad platform to follow, and Meta teased a Muse Video model that previewed at No. 3. Having previously outsourced creative AI to Midjourney and Black Forest Labs, Meta now has a strong in-house option for its social, advertising, and chat needs.

Microsoft cuts 4,800 jobs, with Xbox taking most of the hit

(2 newsletters)

Microsoft laid off 4,800 people, about 2.1% of its workforce, with two-thirds of the cuts landing on Xbox, which is shedding 3,200 roles and five studios by mid-2027. New gaming CEO Asha Sharma told staff the business “is not healthy,” noting Xbox runs at margins three to ten times lower than comparable units and lost 64 cents per dollar invested in studios. It is a striking reversal less than three years after the $69B Activision Blizzard deal, as Microsoft’s buy-everything gaming era gives way to letting studios go.

A “Stargate for data” as the bottleneck shifts from compute to data

(3 newsletters)

OpenAI’s Will Depue, a core Sora creator, argued that compute is no longer what holds AI back; data is. Labs are burning through the roughly 300 trillion tokens of quality public text on the open web, and he projects they will spend north of $100B a year on private data by 2030. His proposed fix is a moonshot effort to go collect the high-quality private datasets models still cannot learn from, positioning data as a strategic asset on par with major compute investments.

Tencent’s Hy3 headlines a surge of cheap Chinese open models

(3 newsletters)

Tencent released Hy3, a 295-billion-parameter mixture-of-experts model with 21 billion active parameters, Apache 2.0 licensed and claimed to match flagship models two to five times its size. It arrived alongside Meituan’s 1.6-trillion-parameter LongCat-2.0, notable for being trained end-to-end on 50,000 Chinese chips rather than Nvidia GPUs. The releases underscore how fast open, low-cost Chinese models are becoming serious agent infrastructure, and US firms are increasingly evaluating them as OpenAI and Anthropic costs climb.

Beijing weighs restricting foreign access to top Chinese AI models

(2 newsletters)

Chinese commerce officials reportedly met with ByteDance, Alibaba, and Z AI to discuss limiting overseas use of their strongest models, including Qwen, Doubao, and GLM-5.2. Options under consideration span bars on public release, domestic-only limits, startup-funding restrictions, and penalties for leaking proprietary AI. The talks mirror Washington’s June export controls on Anthropic’s Fable and Mythos, and would make model access a two-sided political risk: Western users could lose the increasingly popular Chinese models as quickly as they gained them.

TeraWulf signs $19B AI data center lease with Anthropic

(2 newsletters)

TeraWulf signed a 20-year lease with Anthropic expected to generate about $19 billion in contracted revenue, covering a Kentucky campus with roughly 401 MW of critical IT load. Initial capacity is targeted for the second half of 2027, with full buildout by 2028. The deal sent TeraWulf shares higher and is another marker of the enormous infrastructure commitments underpinning frontier model training.

Also Worth Knowing

Quick Hits

Shower Thoughts

Birds probably don’t enjoy roller coasters. Their head stabilizers cancel out half the experience.