Samwise Tech/AI/Robotics Newsletter
Saturday, April 18, 2026
Anthropic Launches Claude Opus 4.7, Outpacing GPT-5.4 and Gemini 3.1 Pro on Agentic Coding
Anthropic released Claude Opus 4.7 on April 16, narrowly reclaiming the top position among publicly available large language models. The new model outperforms OpenAI's GPT-5.4 and Google's Gemini 3.1 Pro on agentic coding, scaled tool use, agentic computer use, and financial analysis benchmarks. A standout improvement is image support: Opus 4.7 now processes photos up to 2,576 pixels on the long edge — more than three times the resolution cap of previous Claude models. The model also handles complex, hours-long autonomous tasks with greater consistency and includes cybersecurity safeguards developed from lessons learned during the restricted Claude Mythos Preview program. Pricing remains unchanged from Opus 4.6.
Sources: VentureBeat
Anthropic Builds Its Most Powerful AI Model, Then Withholds It Over Zero-Day Exploit Risk
Anthropic announced it has built its most capable AI model to date — Claude Mythos Preview — but will not release it publicly because it can autonomously identify and exploit software vulnerabilities at unprecedented scale. In controlled testing, Mythos Preview discovered thousands of previously unknown zero-day flaws across major operating systems and web browsers, including a 27-year-old OpenBSD bug and a 17-year-old FreeBSD remote root exploit. The UK's AI Safety Institute found the model completed a 32-step simulated enterprise network attack — the first AI system to do so. Anthropic launched Project Glasswing, a $100 million initiative giving access to 40 critical-infrastructure partners including Microsoft, Apple, Google, and CrowdStrike.
Sources: TechCrunch, VentureBeat
CVPR 2026 to Showcase How AI Is Driving the Next Wave of Embodied Robotics
The 2026 IEEE Conference on Computer Vision and Pattern Recognition, scheduled for June 3–7 in Denver, is set to spotlight the fastest-moving convergence in technology: AI driving physical robots into the real world. Preview sessions published this week outline a programme heavy on embodied AI, multi-modal reasoning, spatial computing, and robotic manipulation. The conference will host the ManipArena Competition, a real-robot challenge testing models across 20 real-world manipulation tasks and evaluating physical reasoning, generalization, and decision-making under uncertainty. Workshop speakers include researchers from leading institutions working on next-generation robot learning, with virtual access available for remote participants.
Sources: Robotics and Automation News
Snap Cuts 16 Percent of Workforce as AI Generates 65 Percent of New Code
Snap, the parent company of Snapchat, is cutting approximately 1,000 employees — 16 percent of its global workforce — as artificial intelligence takes over software development. CEO Evan Spiegel cited rapid advances in AI as making it possible to operate with smaller, focused teams. The company disclosed that AI now generates more than 65 percent of new code written at Snap. The restructuring will reduce annualized costs by more than $500 million by the second half of 2026, establishing a clearer path to net-income profitability. Affected US employees receive four months of severance, healthcare coverage, and equity vesting. Snap's stock rose roughly 7 percent on the announcement.
Sources: TechCrunch
Stanford AI Index 2026: China Has Nearly Closed the US Performance Gap as AI Agents Lag Human Experts
Stanford University's 2026 AI Index Report documents a striking shift in global AI competition: the performance gap between the top US and Chinese AI models has shrunk to just 39 Arena points, with Anthropic's Claude Opus 4.6 leading China's Dola-Seed 2.0 by a margin researchers describe as historically narrow. The 400-plus-page report also finds that the best AI agents score roughly half as well as PhD scientists on complex, multi-step scientific workflows — a finding that tempers optimism about near-term autonomous research capabilities. Total global AI compute has increased 30-fold since 2021, while training the largest frontier models now generates upward of 72,000 tonnes of carbon-equivalent emissions.
Sources: MIT Technology Review, Nature
Google's Gemini AI Blocked a Record 8.3 Billion Bad Ads in 2025 While Reducing Advertiser Bans
Google reported that its Gemini AI models blocked a record 8.3 billion policy-violating ads globally in 2025, even as the number of advertiser bans declined. The company attributes the divergence to AI detecting and blocking harmful campaigns earlier — before they generate significant reach — rather than reacting after the fact. Generative AI tools used by scammers to produce deceptive content at scale have driven bad-ad volumes higher, but Google says Gemini now identifies coordinated campaign patterns across large networks and shuts them down faster than human review teams. The result reflects a growing dynamic: AI being deployed to police outputs produced by other AI.
Sources: TechCrunch
China's Canton Fair 2026 Reveals Humanoid Robots Hitting the Mass Market at Under $75,000
The 139th China Canton Fair, which opened in Guangzhou on April 15, showcased a milestone in humanoid robotics: Chinese manufacturers are now selling production-ready robots at prices ranging from $29,000 to $73,000, compared to $100,000 to $117,000 at last spring's edition — a drop of more than 35 percent in twelve months. The fair's dedicated service-robot pavilion gathered 46 enterprises demonstrating systems already deployed in warehouses, factories, and logistics environments. Highlights include Ti5 Robot's T230 line capable of lifting up to 88 lbs., ChangingTek's left-right dexterous X2 hand for aerospace applications, and PHYBOT's C1 model capable of playing badminton against human opponents.
Sources: Robotics and Automation News
Tech Pulse
Top Models (SWE-Bench Verified): Claude Mythos Preview (93.9%) | GPT-5.3 Codex (85.0%) | Claude Opus 4.5 (80.9%)
Market: NVIDIA $4.94T | Alphabet $3.89T | Apple $3.81T
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.