Samwise Tech/AI/Robotics Newsletter
Wednesday, May 27, 2026
Anthropic and OpenAI Restrict Access to Powerful Cyber-Focused AI Models
Anthropic’s Claude Mythos uncovered exploitable vulnerabilities in every major operating system and browser, prompting the company to restrict access to roughly 50 trusted organizations under “Project Glasswing” rather than release the model publicly. OpenAI simultaneously limited GPT-5.4-Cyber exclusively to users of its Daybreak enterprise security platform. Georgetown researcher Helen Toner said she would “expect this to more be the first in a series” of such restricted releases. OpenAI stated that “expanding access responsibly requires stronger confidence in who is using the model.” Both companies withheld general access to their most capable cyber-focused AI systems.
Sources: Nature
Japan’s Institute of Science Tokyo Opens Automated Lab Staffed Entirely by Robots
Japan’s Institute of Science Tokyo opened an automated research laboratory in April 2026, staffed by ten two-armed Maholo LabDroid robots capable of handling liquids, growing cells, and operating instruments. Led by researcher Genki Kanda at the Robotics Innovation Center, the lab used AI to identify and test 144 experimental conditions in 111 days for stem cell culturing, with robots monitoring cultures unattended for eight days at a time. The goal is a factory-scale facility housing thousands of robots by 2040–2050. The project marks a shift toward autonomous scientific experimentation at industrial scale.
Sources: Nature
DuckDuckGo App Installs Spike 30% After Google Replaces Search with AI Agents
U.S. DuckDuckGo app installs averaged 18.1% week-over-week growth between May 20 and 25, 2026, with a single-day spike of 30.5% on May 25. On that same day, iOS DuckDuckGo installs peaked at 69.9% growth week-over-week. Google had announced at I/O 2026 that it was replacing traditional blue-link search results entirely with AI agents, prompting the surge. Visits to DuckDuckGo’s noai.duckduckgo.com opt-out page climbed 22.7% in the same period. CEO Gabriel Weinberg said “Google is force-feeding AI with no way to opt out.” DuckDuckGo holds about 2% of the U.S. search market.
Sources: TechCrunch
Human Archive Pays Indian Gig Workers to Capture Robot Training Data
Human Archive has raised $8.2 million in funding from Wing VC, NVP Capital, Y Combinator, and angels from OpenAI, Nvidia, Google, and Meta to build physical AI training datasets. The company pays gig workers in India $1 per hour to wear camera-equipped caps, with over 1,000 active headsets currently deployed. Workers can also use more than 50 device types, including tactile gloves and motion capture suits. Founded by UC Berkeley and Stanford researchers Raj Patel, Samay Maini, Rushil Agarwal, and Shloke Patel, the startup has attracted scrutiny from India’s Ministry of Electronics over worker consent practices.
Sources: TechCrunch
OpenRouter Raises $113M Series B at $1.3B Valuation on 5x Token Growth
OpenRouter, an AI model aggregation platform, has raised $113 million in a Series B led by CapitalG — Alphabet’s growth equity fund — pushing its post-money valuation to approximately $1.3 billion, up from roughly $547 million a year ago according to PitchBook. The company routes requests across more than 400 models from Anthropic, Google, OpenAI, xAI, and DeepSeek, serving 8 million global users. Token volume has grown fivefold in six months, from 5 trillion tokens per week to 25 trillion per week, or 100 trillion per month. An earlier Series A was led by Andreessen Horowitz and Menlo Ventures.
Sources: TechCrunch
Universal Music and TikTok Renew Deal with New AI Music Protections
Universal Music Group and TikTok have renewed their licensing agreement with new protections against unauthorized AI-generated music, including a commitment to remove AI tracks that imitate artists without authorization and to improve attribution for songwriters. UMG had pulled its entire catalog from TikTok in early 2024 over concerns about AI-generated music copying artists’ voices and styles. Among the examples cited were AI-generated tracks imitating Drake and The Weeknd that accumulated millions of streams before removal. The renewed deal is being discussed as a potential template for broader AI music governance across the industry.
Sources: TechCrunch
Cochrane Editor: AI Cannot Be Trusted for Scientific Literature Reviews
Rupa Sarkar, editor-in-chief of the Cochrane Collaboration, argues that AI tools built for systematic reviews are largely proprietary and opaque black boxes, making them unsuitable for evaluating drugs and medical devices where methodological transparency is required. AI models used in the process hallucinate and require human verification throughout. Cochrane’s own findings show that using AI for each review takes longer than performing the work manually. Currently, most tools are developed by private companies, raising concerns about commercial independence. Sarkar concludes that AI cannot be trusted for scientific literature reviews in its present form.
Sources: Nature
Tech Pulse
Top Frontier Models (HLE): Claude Mythos (64.7%) | Gemini 3.1 Pro (44.7%) | GPT-5.4 (41.6%)
Top Open Source Models (GPQA-D): Kimi K2.6 (90.5%) | Qwen 3.5 397B (88.4%) | DeepSeek V4 Pro (87.0%)
Top Small Models 15–50B (MMLU-Pro): Gemma 4 31B (85.2%) | Mistral Small 3 (81.0%) | Llama 4 Scout (74.3%)
Top Edge Models 0–15B (MMLU): Qwen3.5-4B (79.1%) | Phi-4-Mini 3.8B (73.0%) | Gemma 3 4B (43.6%)
AI Leaders: NVIDIA $5.23T | Alphabet $4.81T | Meta $1.70T
Robotics Leaders: ABB $167B | Intuitive Surgical $160B | Fanuc $38.9B
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.