Epoch AI tested Pangram, GPTZero, and Originality.ai against 495 human passages from 99 authors, all written before ChatGPT's release. AI text from plain prompts was caught almost every time (miss rate topping out at 0.7%), but feeding the model five sample passages and asking it to imitate the author pushed the average miss rate to 13%—and to 24%–29% for scientific writing.
Bloomberg reporter Mark Gurman revealed in the July 19 Power On newsletter that Apple is piloting a system called Live Notes at some retail stores: with the consent of both the employee and the customer, it records the Genius Bar conversation, and AI transcribes it and generates a summary, which the employee can edit before saving into an internal system. Recordings aren't retained long-term, and store managers can't access the transcripts, but some employees worry it may be used for performance evaluation in the future—Apple says it won't, and whether the pilot will expand is currently unclear.
Anthropic's free-use period of Fable 5 for paid subscribers ends at 23:59:59 Pacific on July 19, already the third extension in five weeks. After it expires, Fable 5 switches to prepaid-credit billing ($10/M input, $50/M output). The community is guessing at three outcomes: another extension, the start of billing, or seizing the moment to release Opus 5—the "Claude Honeycomb EAP" that briefly appeared in Cursor earlier was widely read as a signal of Opus 5, none of it officially confirmed.
OpenAI updated the ChatGPT desktop app (macOS/Windows) intensively this week: on July 16 it added a global switcher between ChatGPT and Codex, merged "recent conversations," brought Projects to the desktop, and enabled cross-device cloud sync for Work conversations; on July 15 it expanded the custom-instructions limit from 1,500 to 5,000 characters (Plus and enterprise/education editions); and on July 18, Codex 0.144.6 corrected the context-window labeling of three GPT-5.6 models to 272,000 tokens, and strengthened dangerous-command blocking in 0.144.5.
The rumored July 17 launch of Gemini 3.5 Pro didn't happen. Bloomberg, citing people familiar, says the model was pushed back several months because its coding ability fell short of internal expectations; the news sent Alphabet's stock down about 4.4% Thursday, wiping out nearly $200 billion in market value in a single day, and down over 2% again Friday. Google responded that the model is still in testing with partners, and gave no new date.
TSMC's latest Q2 2026 results show: $40.2 billion in revenue, a 67.7% gross margin, net profit up 77.4% year over year, and AI/HPC already 66% of quarterly revenue; Q3 guidance of $44.6–45.8 billion, full-year revenue growth guidance raised to over 40%, and an additional $100 billion investment announced in Arizona. Management also confirmed N3-process capacity is sold out through year-end—the AI compute supply bottleneck won't ease in the short term.
Moonshot released Kimi K3 on July 16: a 2.8-trillion-parameter MoE, 1-million-token context, natively multimodal, billed as the largest open-weight model to date. After the anonymous checkpoint was unmasked, it topped Arena.ai's frontend coding leaderboard with 1679 Elo, surpassing Claude Fable 5 and GPT-5.6 Sol; the full weights and technical report are slated for July 27, and the open source license terms haven't been announced.
TSMC reported Q2 2026 results on July 16: net profit of NT$706.6 billion (about $22 billion), up 77% year over year, far exceeding analyst expectations and a fifth straight quarterly record; HPC (including AI chips) is already 66% of revenue. The company also raised its full-year USD revenue growth guidance from 30% to over 40%, and again increased capital expenditure.
Lithography giant ASML released its Q2 2026 report on July 15: revenue of €9.3 billion and net profit of €2.92 billion, both beating analyst expectations, with full-year revenue guidance raised to €43–45 billion. The CEO explicitly attributed the momentum to AI-related investment driving advanced logic and memory chips, and announced a 30% expansion each for EUV and DUV capacity.
Several media outlets and leak accounts say Google is targeting July 17 to officially launch Gemini 3.5 Pro, with rumored specs including a 2-million-token context window and a Deep Think reasoning mode. To stress: the launch date, specs, and benchmarks all come from third-party leaks; as of July 15 Google has given no official confirmation, and no corresponding model appears in the API docs.
Bloomberg reported on July 14 that OpenAI's first hardware device is a movable, screenless smart speaker: with a built-in mechanical structure that can move on its own, headlining a "human-like AI companion that lives in your home," able to proactively learn user habits and connect to digital-life data like email, built with help from several former Apple iPhone/Mac engineers. Release date and price are both unannounced; the news comes amid Apple suing OpenAI for stealing trade secrets, which OpenAI denies.
On one side, OpenAI is gradually pushing GPT-5.6's three tiers toward availability; on the other, the public availability of Anthropic's strongest model, Mythos 5, keeps sliding back under government frontier-model review and its own safety evaluation. As of July 14, the difference in the two companies' cadence is caused more by "review and compliance" than by pure technology. Per official disclosures.
Showing 12 of 287 stories
