Top Stories
Anthropic’s Mythos Spooked DeepSeek Into a $7.4B Raise
DeepSeek is raising $7.4 billion in outside capital for the first time, and the proximate cause is Anthropic’s Mythos frontier preview, according to reporting from The Information. The fund-raise is a sharp break with DeepSeek’s prior posture as a low-overhead, largely self-funded Chinese lab and a signal that a single US frontier preview is now enough to rearrange the open-weight race.
Mythos’s release to trusted organizations has been the through-line of the past week: MIT Technology Review’s Download notes a Chinese model reportedly matching Mythos’s security capabilities, the kind of parity claim that typically lands right before a fundraising push. If the DeepSeek round closes at $7.4B, it will be the largest single Chinese AI capital event of the cycle and a marker that the open-vs-closed contest is now moving on the investor side, not just the weights side.
California Cuts Anthropic a Half-Price Deal While the Feds Move Against It
Governor Gavin Newsom has signed a first-of-its-kind partnership giving every California state agency and local government access to Claude at half price, plus training and support. The discount follows Newsom’s March executive order on responsible AI use and is targeted at drafting documents and analyzing information across state services, TechCrunch reports.
The timing is the story. California’s CIO told TechCrunch the federal “supply-chain risk” designation of Anthropic “just didn’t come up” in negotiations, even as the Pentagon has signed with OpenAI instead. A state-vs-federal split on which frontier model governments are allowed to buy is now an explicit procurement question, not a hypothetical.
Managers Catch 18% Fewer Errors When AI Is Framed as a “Coworker”
A Boston University field study of 1,261 managers found they caught 18% fewer errors when AI tools were presented as agentic “AI employees” rather than chatbots, and were 44% more likely to escalate questionable work. Emma Wiles, the BU business professor who led the study, frames it as direct evidence that the “AI coworker” branding is a measurable safety regression, not a vibes problem.
The study lands inside a week when enterprises are racing to roll out agent frameworks under exactly that branding. The counter-data point will be useful for safety, legal, and audit teams trying to slow agent rollouts until monitoring catches up. A companion Stanford survey of 1,500 workers across 104 jobs adds more weight to the same direction of travel.
Adopters of AI Saw Headcount Rise 10.2%, Not Fall
New data from Ramp and Revelio Labs covering nearly 22,000 companies shows “high-intensity adopters” of AI tools saw headcount climb 10.2% and entry-level roles rise 12%, undercutting the simple “AI replaces workers” framing that has been hardening into US policy debates over the past quarter.
The story is not “AI is job-creating” - the data is correlational and the companies adopting AI are also the ones hiring fastest for other reasons. But it does mean the policy debate over AI displacement is running ahead of the underlying labor data. Expect this number to be cited heavily in the next round of AI-jobs hearings.
Meta Limits Internal Use of Claude Code and Codex Over Distillation Fears
Internal Meta documents reviewed by The Information reportedly restrict engineer access to rival coding agents Claude Code and Codex so that their outputs cannot contaminate Meta’s own training data. The policy treats Meta’s own staff as a distillation risk - a rare concrete datapoint on how a frontier lab is operationally handling the data-leakage problem internally.
It also sets up an awkward optics problem: Meta is publicly arguing against open-weight distillation risks while, per the documents, restricting competitor tools to stop the reverse flow. Both moves are defensible in isolation; together they will be used by both sides of the open-vs-closed debate.
Supreme Court Extends Carpenter to Short-Term Location Data
In Chatrie v. United States, the US Supreme Court ruled that people retain a Fourth Amendment-protected expectation of privacy in app-generated location records even when those records are short-term and shared with third parties. The EFF frames the decision as extending the 2018 Carpenter v. United States ruling and rejecting geofence warrants as dragnet searches of innocent bystanders.
The court did not decide whether the specific warrant at issue was reasonable or whether the good-faith exception applied; that question is remanded to the Fourth Circuit. But the underlying holding - that location data revealing physical movements is constitutionally protected regardless of duration - will now constrain every AI-driven location-analytics product that quietly depends on bulk short-term surveillance.
EFF Asks Pritzker to Veto Illinois’ Age-Gating Bill
The EFF has formally asked Governor JB Pritzker to veto Illinois HB 5511, a device-level age-gating framework that would require platforms to collect and share users’ ages and strip features like personalized feeds and overnight notifications from young users absent “verifiable parental consent.” The bill is modeled on California’s AB 1043 and New York’s SAFE for Kids Act, neither of which has gone into effect or been tested in court.
EFF’s argument is that the bill would dismantle online anonymity, jeopardize data security, and threaten open-source ecosystems that lack the compliance budget to implement device-level age verification. The Illinois vote is the cleanest test of whether age-gating legislation can survive digital-rights pushback in a third state.
TIDAL Will Not Pay Royalties for Wholly AI-Generated Music
TIDAL has announced that, effective July 15, it will not pay royalties on tracks it identifies as wholly AI-generated, will not allow AI-generated music to be sold direct-to-fan, and will route detection through an external partner. The policy stops short of removing AI-tainted music entirely - the platform cited artists’ freedom to use AI tools and listeners’ autonomy.
The carve is narrower than Spotify’s de-monetization stance and harder for uploaders to game. For the long-running debate over unattributed AI music, TIDAL is now the third major DSP to draw a hard line, and the first to publish a clean technical definition of “wholly AI-generated.”
Ornith-1.0 Adds a Fully MIT-Licensed Coding Model Family
DeepReinforce has released Ornith-1.0 under MIT license, with four sizes (9B Dense, 31B Dense, 35B MoE, 397B MoE) built on Gemma 4 and Qwen 3.5. Simon Willison ran the 35B Q4_K_M GGUF via LM Studio hooked up to the pi.dev agent harness, hitting roughly 103 tokens per second, calling the model “state-of-the-art among open-source models of comparable size on coding benchmarks.”
For local-AI tinkerers, the Ornith release is notable for two reasons: full MIT licensing on a frontier-adjacent coding model, and a size ladder that includes a 397B MoE tier that still fits in the upper end of consumer hardware with the right quant. It adds to the small but growing list of openly licensed coding models that can be self-hosted end-to-end.
Quick Hits
-
Cursor ships a mobile app: A new Cursor Mobile app lets users spin up or steer coding agents from a phone. Boris Cherny of Anthropic, quoted in the piece, said “most of my coding now is on my phone.”
-
Base44 launches Base1: Wix-owned vibe-coding platform Base44 has launched its first in-house model, Base1, betting that owning the inference stack is the only path to defensibility against Lovable and the frontier labs.
-
Warner to unveil an AI agent bill: Senator Mark Warner is set to introduce a framework specifically targeting AI agents, the first move in what is likely to become a year of agent-specific federal legislation.
-
Amazon re-prices its Anthropic deal: Anthropic and Amazon have renegotiated their cloud and model arrangement with Amazon taking on more cost for using Anthropic’s models.
-
Tencent signs $3B CXMT memory deal: Chinese memory maker CXMT has locked in roughly $2.94B in long-term server DRAM supply to Tencent, the first major hyperscaler commitment to a Chinese alternative to Samsung and SK Hynix.
-
Open-source lift from the clampdown: Closed frontier models being restricted from certain US deployments is reportedly pushing developer demand toward open-weight alternatives.
-
Supermicro raided in Taiwan: Taiwanese prosecutors have raided Super Micro offices as the Nvidia-chip-smuggling-into-China probe widens.
Worth Watching
The Mythos ripple effect. Anthropic’s preview has now produced measurable downstream action: a DeepSeek fund-raise, a Chinese model reportedly matching its security capabilities, and procurement fights at the state level. The next data point to watch is whether the Mythos 5 general release, when it lands, settles the open-vs-closed contest or intensifies it.
State-versus-federal AI procurement. California’s half-price Anthropic deal is the cleanest instance so far of a state explicitly defying the federal posture on which frontier model its government buys. Other states with their own AI executive orders (New York, Massachusetts, Washington) are likely to face the same pressure; expect more announcements before the end of the quarter.
Agent-specific federal legislation. Warner’s bill is the opening move in what looks like the first serious attempt to write rules for autonomous AI agents rather than generative models generally. The EFF-versus-Trump-administration pattern from the past quarter will recur here, with privacy and open-source groups lined up against industry lobbyists favoring narrow disclosure rules.
The labor data is going to keep contradicting the policy debate. Ramp-Revelio’s 10.2% headcount rise at AI adopters is the third high-quality dataset this quarter pointing in the same direction. Until the policy debate catches up to the data, the AI-jobs story will stay politically untethered from what is actually happening in payroll data.