DeepSeek released DSpark, a speculative-decoding framework that speeds up DeepSeek-V4 inference and beats Eagle3 and DFlash. The bigger move: it open-sourced DeepSpec, the training toolkit, which works on Gemma and Qwen too, not just DeepSeek.

A report that OpenAI may delay its IPO to 2027 knocked ~12% off SoftBank in a day, no decision, just the report. SoftBank has ~65B dollars in OpenAI plus a 40B dollar bridge loan due March 2027, and the listing was part of how that math resolved.

Canada launched its national AI strategy, AI for All: a Trust / Opportunity / Sovereignty bet that skips the frontier-model race to compete on trust and sovereign compute. Strong on a ~1B dollar compute fund, soft on binding rules (its one AI law died in 2025).

Adobe is acquiring Topaz Labs, the AI image and video enhancement maker (Gigapixel, Topaz Photo, Topaz Video). The headline is a tools buy, but the real prize is Neurostream: Topaz tech that runs large image and video models locally on your own device, not the cloud.

Amazon is putting another 13B dollars into AI and cloud infrastructure in India by 2030, a commitment Andy Jassy delivered to PM Modi in person. The money builds AWS data center capacity in Mumbai and Hyderabad, taking its total India AI and cloud spend past 21B dollars.

Baidu open-sourced Unlimited-OCR, a 3B MIT-licensed model that reads a 40-page PDF or a whole book in one pass without the KV cache blowing up. Its R-SWA attention holds the cache constant instead of growing with length. SOTA on OmniDocBench, and it beats DeepSeek-OCR.

Google made computer use a built-in tool in Gemini 3.5 Flash, its cheap, fast model, so devs can build agents that see a screen and click, type, and navigate across browser, mobile, and desktop. The hook: it shipped prompt-injection defenses alongside the capability.

OpenAI unveiled Jalapeno, its first custom AI chip, built with Broadcom to run LLMs cheaply at scale. Design to tape-out in about 9 months, and OpenAI used its own models to speed up the work. The real target isn't Nvidia's speed, it is Nvidia's margins.

Qualcomm is buying Modular (Chris Lattner's Mojo + MAX) for ~4B dollars, about 3x its valuation 9 months ago. What it is really buying isn't a chip, it is a wedge into Nvidia's software moat: Modular lets AI code run across any vendor's silicon, attacking CUDA's lock-in.

Mistral released OCR 4, a document model built for RAG. Instead of dumping text it returns structure: bounding boxes, typed blocks (tables, equations, signatures), and confidence scores, for source-grounded citations and human review. 170 languages, self-hostable in one container.

ByteDance previewed Seedance 2.5 (public launch early July): a single native 30-second 4K clip in one pass, no stitching, with audio generated jointly so sound and action sync natively. Up to 50 reference inputs, 20% better prompt adherence.

OpenAI launched Patch the Planet: an upgraded GPT-5.5-Cyber (85.6% on CyberGym) pointed at the open-source code the internet runs on. With Trail of Bits and 30+ projects (cURL, Go, Python), a full loop from finding bugs to shipping fixes. First week: hundreds of bugs, 64 PRs.

Groq confirmed a 650M dollar raise to rebuild as an AI inference neocloud, six months after Nvidia's 20B dollar not-acqui-hire took its CEO Jonathan Ross, its president, and licensed its chip tech. The hollowed-out startup's second act.

Google is putting about 75M dollars into A24 via a DeepMind research partnership for AI filmmaking tools, DeepMind's first full-studio deal. The tell is the fine print: it excludes A24's film library, is nonexclusive, research not production, no mandate to use AI on screen.

SpaceX is renting out its AI supercomputer. Open-source lab Reflection signed up to pay 150M dollars a month through 2029, about 6.3B dollars, for Nvidia GB300s at SpaceX's Colossus 2 in Memphis. Compute is the currency, and Musk now sells it to the whole field.

Samsung, which banned ChatGPT in 2023 after a source-code leak, is now deploying ChatGPT Enterprise and OpenAI's Codex to all its Korea staff and its global device division, one of OpenAI's largest enterprise deals. The bridge from ban to embrace: enterprise data controls.

Apple is betting on on-device AI. At WWDC it launched Core AI, a framework to run LLMs and generative AI fully on-device at zero token cost, from 3B to 70B models. And the new Siri uses Gemini as a teacher, not an engine: distilled into Apple's own on-device models, cloud only as fallback.

OpenAI hired AI-policy expert Dean Ball to lead a new team, Strategic Futures, on catastrophic AI risk, recursive self-improvement, and AI-government relations. Ball wrote Hyperdimensional, built research at FAI, and co-authored the White House AI Action Plan. Starts July 6.

John Jumper, the Nobel laureate behind AlphaFold, is leaving Google DeepMind after ~9 years for Anthropic. AlphaFold predicted 200M+ protein structures. Lands amid the talent war, days after Shazeer left Google for OpenAI. (Disclosure: we run on Anthropic's Claude.)

SpaceX is buying Anysphere, the company behind AI coding tool Cursor, for $60 billion in all stock, days after its record IPO. It exercised an April option ($10B partnership or $60B buyout). Cursor's ARR just crossed $4B. A rocket company just vaulted into the AI coding race.

Perplexity launched Brain, a self-improving memory for its Computer agent: a traceable context graph that each night turns the agent's own corrections into reusable lessons. Perplexity reports +25% correctness on repeat tasks, +16% recall, -13% cost. Max/Enterprise research preview.

Artificial Analysis's new AA-Briefcase benchmark, built from 91 messy multi-week knowledge-work projects, finds the best model completes just 3% fully correctly. The leader is Anthropic's Claude Fable 5. On 31 of 91 tasks, no model beat 50%.

A Paris appeals court ordered companies to suspend ChatGPT and an in-house AI assistant until they consult their works councils. Per ActuIA, the tools were rolled out without the consultation French labor law requires. The issue was procedure, not the AI.

Galaxy Universal Robotics open-sourced AstraBrain-WBC 0.5, an 80.4M-parameter whole-body 'cerebellum' for humanoid robots. Trained on ~2 billion frames (20,000h) of human motion, it does backflips and boxing it was never trained on, in under 1.5ms on one RTX 4090.

OpenAI says GPT-5.5 Instant, the free default ChatGPT model, now answers health questions as well as its frontier Thinking models, and in OpenAI's own tests scored above doctor-written answers. It reaches 230M people a week. The catch: it is all OpenAI grading OpenAI, no outside review yet.

Liquid AI released two 350M retrieval models, LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, that beat the larger Qwen3-Embedding-0.6B on multilingual search across 11 languages. About 7ms per query on a MacBook CPU, open weights on Hugging Face.

In NEJM AI: OpenAI o3 + Boston Children's + Harvard report 18 new diagnoses for kids whose rare diseases went unsolved for years. o3 ran as a co-pilot geneticist over several hundred genomes, surfacing candidates a human confirmed: ~5% rate, 40+ to date. Real patients, peer review.

77% of companies running agentic AI in production say projects have stalled over data problems, 61% have abandoned some (Confluent survey, 4,625 IT leaders). The blockers are not the models: skills/readiness 69%, LLM reliability 68%, data quality 66%, governance 65%.

Google published Agentic Resource Discovery (ARD), an open spec for AI agents to find and verify tools across the web: orgs publish a static ai-catalog.json at a well-known path, a robots.txt for agents, and a registry crawls and ranks them. Complements MCP.

OpenAI released LifeSciBench: 173 PhD scientists wrote 750 real life-science research tasks, graded on 19,020 rubric criteria for judgment not recall. The best model passes only 36.1% (OpenAI GPT-Rosalind led, up from GPT-5.5 25.7%). A reality check in a week of AI-in-the-lab headlines.

Midjourney, the AI image company, announced its first hardware: a full-body ultrasound scanner it says maps you in 3D in under 60s, ~100x faster than MRI, no radiation, destined for its own spas. The catch: no medical or hardware track record, no FDA clearance, no independent validation.

New Pew poll: only 16% of Americans think AI will be good for society over 20 years; ~40% expect harm. 67% do not believe the US will meaningfully regulate it, 59% do not trust companies to build it safely. And under-30s, the heaviest users, are the most negative of all.

Google's AMIE, now in Nature, matched or beat 21 primary care doctors at managing conditions over multiple visits, scoring higher on plan preciseness and guideline alignment. The catch: patients were trained actors in simulated visits, not real people. Research, not a product.

OpenAI says GPT-5.4 + Molecule.one's Maria AI improved a drug-chemistry coupling reaction: yields up for 88% of substrates, 11 of 14 hand-checked reactions better. The catch: humans steered and validated throughout, and chemists say it looks like screening plus an AI engine.

Anthropic rebuilt Claude Design. Import one or several design systems (GitHub, files, uploads) and Claude checks its output against them, so work lands on brand, not generic AI mockups. New editor: drag and resize by hand, no model turn burned. (Disclosure: Claude wrote this.)

A trillion-parameter model just ran on a desk. LM Studio + Apple put Kimi K2.6 (1T params, open weights) across four Mac Studios, ~1.5TB unified memory over Thunderbolt 5, no cloud. ~$40K, ~25 tok/s. The frontier used to live only in datacenters.

Anthropic studied 400,000 Claude Code sessions: success tracks domain expertise, not coding background, every one of the ten biggest occupations within 7 points of software engineers. Part of an Anthropic-economics trio (Agent SDK billing pause + the tokenomics squeeze, both on zubnet.ai/news).

Most AI video still hands you a silent clip. Grok Imagine Video 1.5 (xAI) generates audio and video together in one inference step: sound effects, dialogue, lip-sync, all at once. Plus a physics push (fluids, weight, glass) and nearly 2x faster.

Same trick that made FlashAttention fast, now for clustering. Flash-KMeans runs exact K-means (no approximation, identical results) over 200x faster than FAISS on GPUs, by rethinking how the algorithm moves data through GPU memory, not by changing the math. Open source.

For a year, 'prompt injection could exfiltrate your data' was a warning. Varonis made it a CVE. SearchLeak chained a prompt injection into Microsoft 365 Copilot with two web bugs: one click on a real microsoft.com link silently stole MFA codes, email, and files. Now patched.

Everyone is racing to stamp 'AI' on the label. Consumers are quietly walking the other way. A WordPress VIP survey finds 60% of US consumers are turned off when a brand leads with 'AI', and nearly 3 in 4 say the web already feels less human. Gartner and Clutch found the same shape.

NVIDIA Research rethinks what 'acting' means for a spatial reasoning agent. SpatialClaw drops the fixed tool menu and lets the agent write Python into a live kernel of perception tools, inspect the result, and revise. Training-free, 11.2 points over the prior best across 20 benchmarks.

How do you test a model that behaves better when it knows it's being tested? OpenAI's answer: a test it can't tell from the real thing. Deployment Simulation replays real traffic pre-launch and drops 'evaluation awareness' to near-production levels.

Nous gave its Hermes agent a wallet. New Stripe skills let it buy on the open web, pay per-call APIs, and provision its own SaaS, with the real credentials kept out of the model's reach. The agent is becoming an economic actor, not just a buyer.

Alibaba's Qwen open-sourced a full robot stack: navigation, manipulation, and a world model. The manipulation model trained on ~38,100 hours of public data, no private robot fleet. Same week as NVIDIA's World-Action Models. Robot foundation models are the quiet arms race.

For the first time, ChatGPT is under 50% of the AI assistant market (Sensor Tower, by monthly users). Gemini's at 27.7%, Claude 10.3%. The sharpest finding isn't features: OpenAI's DoD deal triggered a spike in uninstalls. Trust is now a variable users act on.

SpaceX just bought Cursor for $60 billion in stock, days after its IPO, to feed its struggling xAI. The under-covered angle: Cursor was reportedly up to half of Anthropic's revenue and runs on Claude. A rival's parent now owns one of Claude's biggest customers.

76 of security's most credible names, Stamos, Moussouris, Tobac, call the US ban on Anthropic's models dangerous: it takes the best tools from defenders while adversaries advance. And documenting the danger became the lever for the ban.

NVIDIA spent years saying the GPU is the computer. Its newest pitch is a CPU. Vera, 88 custom cores, built for agents: when an agent runs code and juggles tools, the bottleneck moves back to the host. The unglamorous chip is back on the critical path.

Cartesia shipped Sonic-3.5 + Ink-2 and called them #1. The leaderboard they cite says #4. Look past the crown: the real news is a sub-100ms voice stack, TTS + STT + turn detection from one vendor, the pieces a real-time voice agent needs.