Samwise Tech/AI/Robotics Newsletter
Saturday, September 12, 2026
DeepSeek Releases V4.1-Flash With 74.2% DeepSWE Score and $0.003 Per Million Token Rate
DeepSeek has released V4.1-Flash, a 552-billion-parameter mixture-of-experts model priced at $0.003 per million cached input tokens off-peak — among the lowest frontier-class rates available. The model, released under an MIT open-weight license on Hugging Face, scores 74.2% on the DeepSWE v1.1 software-engineering benchmark, edging Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%. Its Causal Encoder-Decoder architecture activates 8 billion parameters during prefill and 16 billion during decoding. VentureBeat reports that DeepSeek’s parent company is preparing for a potential Shanghai STAR Market IPO at a roughly 500-billion-yuan ($75 billion) valuation.
Sources: VentureBeat
Jensen Huang Says Nvidia Expects 70% Revenue Growth Next Year, Approaching $680 Billion
Nvidia CEO Jensen Huang told the Goldman Sachs Communacopia and Technology conference that his company expects revenue to grow 70% year-over-year, which would put it near $680 billion — on top of a record-expected $400 billion this fiscal year. Huang credited Nvidia’s position across the full AI stack, noting the company tracks every major workload globally. Its GB200 NVL72 system, priced at $8.5 million per unit, is seeing 27% month-on-month sales growth. Huang also defended Nvidia’s circular investment model, in which the company reinvests infrastructure returns into more AI capacity it supplies, quipping that for every dollar invested, a hundred comes back.
Sources: TechCrunch
ID Verification Firm IDScan Confirms Breach of 150 Million Driver’s License Records
Louisiana-based ID verification firm IDScan has confirmed a breach affecting more than 150 million driver’s license records in its cloud database, including full names, license numbers, and photos. The company — which verifies IDs for entertainment venues, cannabis dispensaries, and other businesses — says it learned of the intrusion around September 1, the same day journalist Brian Krebs reported a dark-web site allowing anyone to search the stolen records by name. Notable figures in the dataset include U.S. Secretary of Defense Pete Hegseth. The FBI has confirmed it is investigating. No ransom demand has been publicly reported.
Sources: TechCrunch
Anthropic Reports China-Based AI Firms Ran 200 Million Distillation Attacks Against Claude
Anthropic released a threat intelligence report detailing five large-scale distillation attacks from China-based AI companies that extracted model reasoning from nearly 200 million Claude exchanges. Distillation attacks involve prompting a model to expose its internal chain-of-thought, which is then used to train competing models. Anthropic’s report describes an Alibaba campaign spanning 151 million exchanges between May and July 2026, peaking at three million per day across 3,500 accounts, aimed at training the Qwen model family. A separate Moonshot AI campaign, which Anthropic says routed requests through the Chinese military, specifically targeted Claude Opus to extract its reasoning patterns.
Sources: TechCrunch
Robot Training Data Startup Mecka AI Nears $500 Million Valuation in Sequoia-Led Round
Robot training data startup Mecka AI is nearing a $500 million valuation in a new funding round led by Sequoia Capital, as demand for high-quality physical-world datasets surges alongside the humanoid robot boom. Founded in 2024 by Josh Gao, Mogen Cheng, Jason Chong, and Duy Nguyen, the company collects and curates manipulation and locomotion data used to train robot models, and previously raised a $60 million Series A from Framework Ventures. As of early June, the team was projecting $100 million in annual recurring revenue by year-end 2026 — a sign that robot training data has matured into a viable standalone business.
Sources: TechCrunch
Fields Medal Winners Sign Open Letter Warning AI Labs Threaten Mathematical Research Integrity
Twenty-five Fields Medal-winning mathematicians have signed an open letter arguing that AI labs are threatening the integrity of mathematical research. The action follows accusations by NYU professor Tristan Buckmaster that OpenAI pressured him not to credit an Anthropic collaborator in a shared proof, and used their joint work with its Codex model to produce a contested Navier-Stokes solution over a marathon weekend of inference. OpenAI has since withdrawn its sponsorship from a CalTech math event. The signatories warn that AI labs’ ability to spend tens of millions racing researchers to proofs will incentivize secrecy and erode the open-collaboration culture mathematics depends on.
Sources: TechCrunch
Meta’s AI Agent App Muse Hits 83,000 Downloads and No. 2 on U.S. App Store in First Week
Meta’s new AI agent app Muse reached 83,000 iOS downloads and the number-two position on the U.S. App Store as of Wednesday — just days after its Tuesday launch, according to market intelligence firm Sensor Tower. Those numbers trail Meta’s own benchmarks: Threads drew 4.3 million U.S. downloads on launch day, while the Meta AI app saw 108,000 at debut. Muse competes in a quickly crowding agentic AI consumer market, where its primary rival is Instinct, a text-message-based AI agent valued at $2.5 billion. The launch also comes days after Meta agreed to an $18 billion multistate settlement over social media harms to children.
Sources: TechCrunch
Every newsletter preserved and searchable
Curated by JD · samwise.agency