TechCapsules

Anthropic launched Mythos 5.2 and Fable 5.2, claiming gains in coding, reasoning, vision, and clearer user communication. The new Fable model now costs $8 per million input tokens, undercutting GPT‑6 Astra, and delivers 30 % faster responses, with a premium option for even quicker inference. Fable 5.2 supports zero‑data‑retention policies via Enterprise Frontier Safeguards, addressing enterprise privacy concerns. The release follows a METR report suggesting Anthropic’s models may be approaching recursive self‑improvement, raising evaluation challenges.

Ron Howard and Akiva Goldsman have announced a television adaptation of Daniel H. Wilson’s 2011 New York Times bestseller Robopocalypse. The project, previously stalled after a 2009 DreamWorks acquisition and a 2018 Michael Bay involvement, will see Howard direct the pilot and Goldsman serve as showrunner. Imagine Entertainment and Weed Road Pictures will produce, with Spielberg’s Amblin Entertainment as executive producers. The series will explore a rogue AI, Archos, threatening humanity, echoing current AI anxieties.

Samsung is preparing to upgrade its Bixby assistant with AI‑agent capabilities in the upcoming One UI 9.5 release. Reports from SamMobile indicate Bixby can now create “actions,” allowing users to schedule automated tasks such as daily currency conversions. Users retain control with options to delete, pause, or pin actions, and a log may track activity. Early testing suggests future functions could include booking hotels and other services, potentially rolling out by 2027.

Apple has introduced a paid “Intelligence usage pack” for Creator Studio subscribers who exceed the built‑in limits of its AI‑driven features. The $49.99 one‑time purchase, first visible in Logic Pro’s beta usage window, adds extra capacity to generate images, shapes, slide decks and presenter notes beyond the monthly allotment of 50 images, 250 shapes, 50 presentations or 700 notes. The pack rolls over unused credits, expires after 12 months, and is tied to the purchaser’s iCloud account [...]

Boston Dynamics released a video showing its Atlas humanoid equipped with wheels on its feet, used as a research testbed to explore the blend of deterministic physics‑based control and whole‑body reinforcement‑learning (RL) policies. Controls engineer Andy Wu attached off‑the‑shelf scooter rims to Atlas’ ankle actuators, allowing the robot to roll on flat terrain and switch to articulated legged motion for obstacles. The experiment, trained entirely in simulation, demonstrated real‑time proprioceptive feedback processing, dynamic pose and force adjustments, [...]

NVIDIA unveiled GR00T 2, an upgraded world‑action model, and Isaac Lab 3.0, the release of its robot‑learning framework, at the Conference on Robot Learning in Austin. GR00T 2 was trained on 100,000 hours of egocentric robot and human data and can complete 40 of 50 benchmark tasks without task‑specific demonstrations, achieving 71 % success with a single demonstration and 95 % with five. The model ranks in the top three across vision‑language‑action benchmarks and supports an API for sequencing skills and error [...]

Google is preparing to roll out two re‑engineered TPU accelerators—Sunfish TPU 8t and Zebrafish TPU 8i—later in 2026, with broader deployment slated for Q1 2027 and a multi‑region ramp through 2028. The 8t, fabricated by Broadcom and TSMC, targets AI training, featuring a 3‑D torus interconnect linking 9,600 chips and delivering 121 exaflops at FP4 precision, plus 216 GB HBM3E per device. A derivative, the 8tx, swaps the torus for a lower‑latency “boardfly” fabric and replaces Google’s Axion ARM CPU with an AMD EPYC [...]

HTEC’s latest research shows that semiconductor firms are ahead of most industries in AI adoption, with 48% labeling themselves as AI‑native or scaling AI versus 44% overall. Investment growth mirrors the broader market, as 78% report increased AI spending. The sector distinguishes itself by prioritizing strategic alignment and demanding formal ROI models before project approval—45% and 36% respectively. However, once AI initiatives launch, measurement fragments: 38% rely on case‑by‑case reporting and only 28% use enterprise dashboards, compared [...]