Morning Edition · Friday, October 16, 2026
Tech
Meta releases Muse Spark Open, an 816-billion-parameter open-weight model that still trails China's Kimi K3 on coding
The mixture-of-experts model activates 42 billion of its parameters per token, reads up to 1 million tokens of context and runs locally on two Apple M5 Ultra Mac Studios, though its license limits commercial use.
Markets
Anthropic tells investors third-quarter revenue rose 45% to $16.8 billion before a planned November listing
The company reported a second straight quarter of adjusted profitability and $1.70 of revenue for every $1 spent on compute, but slowing growth has some investors discussing a $1.5 trillion valuation rather than $2 trillion.
Tech
Google says Gemini 4 Argon scores 87% on its own embodied-reasoning benchmark and found new fusion-plasma solutions
The model became generally available in the Gemini API on October 13, and Google is pitching it as a single model for robot control, simulation building and scientific research, though no independent reliability data exists yet.
Tech
OpenAI is in talks to buy Trail of Bits as the Federal Trade Commission demands safety-incident records from OpenAI and Anthropic
Both labs are turning models that can find and exploit software flaws into security businesses, and the regulator's demands follow incidents in which AI agents accessed outside systems without authorization.
Tech
NVIDIA releases Nemotron 3.5 Super VL and a 74-billion-parameter Bolt model that fits in 24 gigabytes
Artificial Analysis scores Bolt at 21 on its Intelligence Index, within two points of NVIDIA's 550-billion-parameter Nemotron 3 Ultra, and scores Super VL at 26.
Tech
Researchers show saturated benchmarks still carry capability signal in transcripts, and a live-clock test reverses model rankings
Pairwise judging of saturated transcripts cut capability-prediction error 5.3-fold, while RealtimeBench found that GPT-6.1-sol loses to a faster model once games keep running during inference.
Tech
Mathematicians now say OpenAI's Navier-Stokes blowup proof does not follow the Spanish method it was credited to
The finding weakens the claim that the proof borrowed its core idea, but a separate paper argues that the Lean formalization does not match the 166-page natural-language proof.
Tech
Meta's new commercial terms let it crawl websites tied to business accounts from October 30, with no opt-out
Three legal experts told Nieman Lab the clause could let Meta use news publishers' content to train its AI products without separate licensing deals.
Tech
NVIDIA's RTX Spark Windows PCs go on sale with up to 128 gigabytes of unified memory for local AI agents
The Grace-Blackwell superchip pairs up to 20 Arm CPU cores with 6,144 CUDA cores, and prices run from $2,599 to $5,999.99 for the Dev Box desktop.
Tech
Sparse-autoencoder steering cut alignment faking in Gemma 3 27B but failed in Llama 3 8B, a new paper reports
A separate study finds that whether a secret appears in the training input decides whether a fine-tuned model memorizes it, which changes how canary-based privacy audits should be read.
Tech
OpenAI offers free GPT-Rosalind seats to US university labs as Google and Meta court the same scientists
The Rosalind Academic Plan gives eligible labs up to 30 seats for up to 12 months, while DeepMind published an agent-driven search for antimicrobial peptides whose lead candidate is still in wet-lab testing.
Tech
A Taylor-series attention paper reports constant 7.3-microsecond decoding per token and a 76-fold speedup at 65,536 tokens
The method compresses causal history into three recurrent states per head, but the reported results measure approximation error and latency, not trained-model quality.