TSMC holds its Q2 2026 earnings conference today (July 16) after posting record quarterly revenue of NT$1.27 trillion (~US$39.6B), up 36% YoY. June sales surged 67.9%—its biggest month ever—on 'extremely robust' AI chip demand as the main foundry for Nvidia's GPUs.
TSMC says it's sold out on its N3 node; analysts (LSEG) expect Q2 net profit to climb ~59% YoY.
Effective today (July 15), China's new "Anthropomorphic AI Interaction Services" rules force ByteDance's Doubao (300M+ users) and Alibaba's Qwen to shut down custom AI companion/agent features. Rules require anti-addiction prompts, instant-exit options, and real-time over-dependence detection—incompatible with persistent-memory companions.
Doubao keeps chat history viewable until Oct 15 before deletion; Qwen deletes configs immediately. Workplace agents stay exempt.
A German research consortium (Fraunhofer IAIS/IIS, DFKI, TU Darmstadt and others) released Soofi S, a fully open 31.6B-parameter MoE model with 3.2B active params and a hybrid Mamba-Transformer design. Trained on ~27T tokens using up to 512 Nvidia B200 GPUs on Deutsche Telekom's Munich AI cloud, it beats OLMo 3 32B and Apertus 70B on English (70.1) and German (79.1) benchmarks.
Weights, checkpoints, and full training code are public under OSI's Open Source AI Definition.
OpenAI CEO Sam Altman says he was 'pretty wrong' about AI's impact on jobs, reversing his mid-2025 warnings that entry-level roles faced mass layoffs. 'This is not what I expected,' he wrote, saying he's 'delighted to be wrong' as the predicted wave of entry-level cuts hasn't materialized.
He still expects future displacement alongside new roles, and flagged 'AI washing' — firms blaming layoffs on AI. Yale's Budget Lab found no AI-related job market shifts.
OpenAI is pushing ChatGPT from a productivity tool into household technology, hiring a dedicated product manager in San Francisco for family, caregiver, and older-adult experiences. Its user base is aging: 35+ users rose to 31% in Q2 (from 26% a year earlier), while 18-24 users fell to 29% (from 34%).
Nearly 1 in 4 US smartphone-using parents used ChatGPT in Q2, up from 16% a year ago.
Beijing Academy of AI (BAAI) unveiled Orca, a 'world foundation model' that matches specialized robotics systems on manipulation tasks despite training with zero action labels. It blends 'unconscious learning' from 125,000 hours of unlabeled video with 'conscious learning' from verbally described actions.
Orca-4B beat image generators like FLUX on benchmarks and showed stronger error recovery than the π0.5 baseline.
Apple has sued OpenAI in California federal court, alleging former Apple staff stole trade secrets to build OpenAI's consumer AI hardware. The suit names OpenAI Chief Hardware Officer Tang Tan (ex-Apple iPhone/Watch design VP) and engineer Chang Liu, who allegedly downloaded confidential files and kept an Apple laptop.
Apple says the material included engineering specs and a proprietary metal-finishing technique.
A Financial Times investigation found OpenAI and Google are supplying advanced AI models to Singapore-based subsidiaries of Alibaba, Baidu and Tencent, three firms on the Pentagon's '1260H' blacklist over alleged military ties. The sales don't breach US export controls, which target specific entities and mainland China but not Singapore.
OpenAI says it blocks direct China access but allows some Chinese-owned firms where safeguards apply.
_updates.jpg)
Mistral AI released Robostral Navigate, an 8B model that lets robots navigate from a single RGB camera plus plain-language instructions, no LiDAR or depth sensors. Trained entirely in simulation (~400,000 trajectories across 6,000 scenes), it hits 76.6% success on the R2R-CE unseen benchmark, beating the best single-camera method by 9.7 points and multi-sensor systems by 4.5.
It runs on wheeled, legged and flying robots and improves via online reinforcement learning.
SpaceXAI (xAI) launched Grok 4.5, its first model co-built with newly-acquired Cursor, targeting software engineering, legal, and financial analysis plus autonomous agents. Musk calls it 'roughly comparable to Opus 4.7, but much faster,' trained on tens of thousands of NVIDIA GB300 GPUs.
Pricing runs $2/$6 per M input/output tokens ($4/$18 premium). It rolls out across Cursor's desktop, web, iOS, CLI, and SDK after SpaceXAI's $60B stock deal for Cursor.
AI chip maker SambaNova has closed the first tranche of a $1B Series F at an $11B valuation, led by General Atlantic with Intel, BlackRock, T. Rowe Price, Vista Equity and Qatar Investment Authority joining.
The funds will scale operations and secure supply chain amid surging demand, as its next-gen SN50 chip ships in H2 2026 with SoftBank as first deployment partner and JPMorgan Chase as an inference customer. The round follows a $350M Series E in Feb 2026.
The US Commerce Department's Center for AI Standards and Innovation has cleared OpenAI to broadly release GPT-5.6, with a public launch expected this week, Axios reports. Last month officials forced a staggered rollout that limited early access to vetted partners under a June executive order giving reviewers up to 30 days to vet new models.
Rival Anthropic faced similar limits on its Mythos and Fable models, with Fable's restriction lifted last week.
Meta launched Muse Image, its first in-house AI image-generation model, built by Meta Superintelligence Labs and embedded in Meta AI and Instagram (Facebook/WhatsApp next). It supports text-to-image, image-to-image, and instruction-based region edits via text, sketches, or handwritten notes, and powers 30+ AI visual effects across Instagram and WhatsApp.
Meta claims it beats Google's Nano Banana 2 on several evals, trailing only OpenAI's latest GPT Image model.
Forterra has run the largest U.S. deployment of autonomous ground vehicles in active combat.
More than 100 of its Lancer robots (built on Polaris ATVs, Starlink-connected) have operated in Ukraine for nine months, logging 1,100+ missions, 2,500+ miles, 777,440 lbs of cargo hauled and 52 casualty evacuations. Ukrainian crews mostly teleoperate rather than run full autonomy, with losses when vehicles get stuck.
Forterra has raised $500M+ in VC.
OpenAI released two new Realtime API voice models: gpt-realtime-2.1 and gpt-realtime-2.1-mini. p95 latency is cut at least 25% via improved caching, and both now support configurable reasoning effort (minimal to xhigh), giving voice agents a 'thinking' mode plus tool use and better alphanumeric/noise handling.
The mini keeps prior pricing while adding reasoning: $10/1M fresh audio input ($0.30 cached), $20/1M output.
Anthropic published new interpretability research revealing a 'global workspace' inside Claude. Using the Jacobian lens (J-lens) technique, they found a 'J-space' of internal representations Claude broadcasts and reuses across reasoning steps.
In a swap test, replacing 'France' with 'China' in J-space made Claude correctly update the capital, language and continent at once. It also surfaced hidden model thoughts for safety monitoring.
China's new "Interim Measures for the Administration of Anthropomorphic AI Interaction Services" take effect July 15, forcing ByteDance's Doubao and Alibaba's Qwen to shut down user-created AI agent features that day. Existing custom agents will stop working; users can still view configs and chat histories during the transition and are urged to back up via export or screenshots.
Related data stays recoverable until Oct 15, 2026, then is permanently deleted.
Nvidia's next-gen "Kyber" AI rack system has been delayed more than 12 months to 2028, research firm SemiAnalysis reports. The Kyber NVL144 — built to house 2027 Rubin Ultra chips — is stalled by manufacturing issues with its 78-layer PCB midplane.
Nvidia also cancelled its NVL72x2 back-to-back rack design, limiting Rubin Ultra's scale-up, while the larger NVL576 is likely delayed too — just 3 months after Jensen Huang demoed Kyber at GTC.
The UN's first Global Dialogue on AI Governance opened today in Geneva, convening all 193 member states plus industry, academia and civil society through July 7. Co-chaired by El Salvador and Estonia, day one features the Independent Scientific Panel on AI — 40 experts led by Yoshua Bengio and Maria Ressa — warning AI capabilities are outpacing regulators, with the US and China holding ~90% of frontier compute.
No binding resolutions this session.
New benchmark DiscoBench, from Tencent Hunyuan and Tsinghua University, finds AI search agents fail not at retrieval but at spotting ambiguous queries and asking to clarify. When agents keep searching instead of asking, success drops to 51.9% vs 93.4% when they ask first.
Top base-test scorers: Doubao Seed 2.0 Pro (43.1%), Gemini 3.1 Pro (40.8%), Claude Opus 4.7 (39.8%). Ambiguity hints lifted the 10-model average from 28.6% to 33.7%.
