Google's Project Suncatcher reached orbit on Thursday: a SpaceX Falcon 9 carried a Planet-built prototype satellite with four Trillium TPUs — Google's first in-orbit AI compute test. Contact was confirmed and the satellite is operating as expected; the chips run in roughly 15-minute bursts on about 1 kW of solar power, with a purpose-built two-satellite compute demo expected next year.
OpenAI's security posture was front and center this week. The company disrupted a coordinated model-distillation campaign that extracted protected model reasoning — 16,000+ requests from 4,000+ accounts before it was shut down. Separately, a safety leader quit warning the company's culture is "broken," following last week's reported split with three safety researchers over mishandling of sensitive information.
Beyond that: Sam Altman is publicly championing OpenAI's new Dot agent, NVIDIA cut Arabic dialect word error rates nearly in half with Nemotron 3.5 fine-tuning, Perplexity open-released a context-aware embedding model, Anthropic keeps pushing AI-for-science, DeepSeek shipped desktop apps, and StepFun's Step 5 Preview landed on the Vals leaderboard. OpenClaw also tightened ClawHub security this week by integrating Tencent's AI-Infra-Guard into every skill review.
SpaceX's Transporter-18 mission launched Alphabet's first in-orbit test for Project Suncatcher — a solar-powered prototype satellite built by Planet Labs carrying Google TPUs. Google's Travis Beals confirmed contact and that the satellite is operating as expected, calling it the first step toward scalable machine-learning infrastructure in space. Experts caution that capacity-constrained launches, radiation, cooling, and orbital debris remain unsolved hurdles.
OpenAI says it shut down a campaign to extract protected reasoning from its models, with a core cluster of activity linked to a Chinese startup. The activity surged to 16,000 requests from more than 4,000 users before being blocked, and OpenAI says it is strengthening defenses against adversarial distillation as model reasoning becomes a prime target.
A senior OpenAI safety figure resigned publicly, warning the company's internal culture is broken. The report also notes OpenAI scrapped the release of a next-generation model after researchers raised safety concerns during internal testing — a decision the company has framed as deliberate caution rather than retreat.
The Tarbell Center for AI Journalism is launching an institutional grantmaking fund to back newsrooms and ambitious AI reporting projects. Executive director Cillian Crosson says the grants aim to strengthen independent coverage of AI at a moment when the industry's pace is outstripping newsroom capacity.
SiMa.ai raised a $150M Series C at a $1.45B valuation, bringing total funding to $500M. The company builds chips and software so robots, drones, cameras, and other devices run AI directly on-device — a bet that physical AI demand for edge inference silicon keeps compounding.
03 / In their own feeds
Ten voices on X.
Selected posts from the edition, with links to the originals.
Demis Hassabis · @demishassabis01 / X
"Up, up, and away. 🚀 Today, in partnership with @planet, we launched a prototype satellite carrying four TPUs into orbit on @SpaceX"
"We've integrated @TencentHunyuan's AI-Infra-Guard (AIG) into ClawScan, the open-source command-line tool that powers security review on ClawHub. Every skill and plugin uploaded to ClawHub now runs through AIG as part of the review pipeline."
"A speech model can support Arabic and still struggle with local dialects. Fine-tuning NVIDIA Nemotron 3.5 ASR cut word error rate on Najdi and Hijazi Saudi Arabic from 55% to 30%."
"dot is my favorite openai product so far! it is amazing to me that each day it feels noticably better as it learns more of my workflow and style. having it do the stuff i don't like doing…"
"Please check out @DeepSeekHarness. We just released packaged desktop versions for macOS and Windows; Linux users can get it from the @deepseek-ai/dsh package on npm."
"Retrieval systems often split long documents into chunks, but this strips away the surrounding context. Contextual embedding models fix this by encoding the whole document once and pooling chunk vectors afterward." — announcing pplx-embed-v2-context-9b-preview, with the highest average nDCG@10 on ConTEB and 8x less storage per vector.
"When AI-generated content has become so realistic, how do we verify what's made by human, machine, or nature? We explore the vital role of provenance — and how imperceptible watermarking can safeguard both digital media and biology."
"In physics, an 'impedance mismatch' occurs when two systems each work well but are poorly matched. In this Science Blog guest post, Harvard physicist Matthew Schwartz argues that something similar is happening with AI and physics."