7:57 PM
Today in brief.
Pika gave AI video its voice back
Pika shipped four audio models at once - soundtrack, speech, SFX and music - and demoed them by fixing the Will Smith spaghetti clip three years after it became the benchmark for bad AI video.
3 sources
🏛️ Policy
Anthropic published its risk report and everyone read the redactions
Anthropic's second Responsible Scaling risk report landed the same day as a widely-shared claim that they run an internal model well ahead of anything they ship - so the document about what they know is being read next to the rumor about what they're holding.
🔧 Tech
Opus 4.6-class coding now runs on a machine you can buy
Z.ai shipped GLM-5.3 with open weights and frontier coding scores, and people are already running it locally at Opus-4.6-max quality - the gap between the frontier and your own hardware just collapsed again.
🤖 AI
A 27B model is claiming state of the art on agentic coding
Qwen3.8-27B is a compact open-weight multimodal model claiming state-of-the-art results in agentic coding, computer use and browser tasks - at a size that fits on one consumer card.
🔧 Tech
X's anti-slop system is a behavioral model, not a content filter
Someone read X's open-source algorithm and found TweetSpamBot judges accounts on up to 512 recent behaviors - posting bursts, quote-post patterns, timing, dwell - rather than trying to detect whether the text was AI-written.
Perplexity put its search engine inside anyone's agent
Perplexity's Search SDK is now callable from inside any agentic harness, which turns their whole product into a component you can drop into your own agent loop.
3 sources 1:22 AM
🔧 Tech
Two tools that cut your agent's token bill
A Claude Code hook set that cuts grep tokens by 42%, a batched-questions trick that speeds up design work, and Anthropic's own guide on getting more out of a session - all landed the same day.
🤖 AI
A 150M parameter model hit 29.5% on ARC-AGI for $0.0007 a task
A recurrent model with 150 million parameters - a rounding error next to a frontier model - scored 29.5% on ARC-AGI-1 at seven hundredths of a cent per task, which says the scaling story is not the only story.
💼 Business
OpenAI and Anthropic are in a price war because of China
The two biggest US labs are cutting prices as Chinese open-weight models close the capability gap - the open-weights wave is now showing up on American invoices.
🔬 Science
Google made encrypted AI fast enough to use
Google says it made homomorphic encryption practical for AI - running a model on your data without ever decrypting it, which has been theoretically possible and practically far too slow for a decade.
A neurosurgery resident used GPT-5.6 Sol to prove a 20-year-old conjecture
A neurosurgery resident at Peking College Hospital used GPT-5.6 Sol to prove a two-decade-old conjecture in numerical linear algebra - not as a math project, but as a side quest inside his transcranial ultrasound research.
No stories in this category for this edition.