Can LLMs run on ultra-low-bit memory without tanking accuracy? Researchers from Together AI, University of Sydney, and UIUC present OSCAR — a method that uses offline, attention-aware covariance analysis to design fixed rotations and clipping thresholds for 2-bit KV cache
AI
-
Mythos break expected progress line in System Card
By
–
Returning to the System Card, here is one of my favorite graphs, and it shows how the Mythos category models break the expected line of progress. As always with AI, we reach higher levels ahead of time.
-
Google Unveils Gemini 3.5 Live Translate Model
By
–
Google has released a new Gemini 3.5 Live Translate model that supports low-latency translation across 70+ languages. The model is now available in Preview on AI Studio and APIs, and Google Meet will soon integrate this model for live translation as well.
-
My Hermes Agent using Claude Fable 5 to write git commit
By
–
My Hermes Agent using Claude Fable 5 to write git commit -m "fixed" pic.twitter.com/Xq5Vk7Oyvq
— Shubham Saboo (@Saboo_Shubham_) 9 juin 2026My Hermes Agent using Claude Fable 5 to write git commit -m "fixed"
-
Agent verifies brand logos and details in ads for $5–$8
By
–
The agent > locks verified brand details, logos, and claims as constraints > then checks every frame against them with its own vision intelligence to catch logo drift and product errors before delivery. A finished ad runs about $5 to $8 and returns dozens of
-
Creatify Agent: Research, Direct, and Launch Ads in One Conversation
By
–
Creatify Agent can now research a brand, direct the ad, and connect to Meta, TikTok and Google to launch it, all from a single conversation.
— 🚨 AI News | TestingCatalog (@testingcatalog) 9 juin 2026
The agent leads the work and brings the marketer in at the checkpoints that matter: strategy, scripts, casting. https://t.co/fS7tsGoRAs pic.twitter.com/60rNuxJJdXCreatify Agent can now research a brand, direct the ad, and connect to Meta, TikTok and Google to launch it, all from a single conversation. The agent leads the work and brings the marketer in at the checkpoints that matter: strategy, scripts, casting.
-
Cost frontiers: more expensive and superior model, Fable surpasses Opus 4.8
By
–
The cost frontiers, as expected, show an upward and rightward shift, meaning we have a more expensive and notably superior model (the low of Fable remains well above the x-high of Opus 4.8)
-
Claude Fable 5 released with safeguards
By
–
This is a super exciting release – Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that *qualitatively* also, this is a major-version-bump-deserving step change forward
-
Benchmark tables: the model dominates in everything evaluated
By
–
Here are the complete benchmark tables. The model basically dominates in everything evaluated in these tables (Humanity Last Exam, CriPT, ArxivMath, HealthBench, etc.)
-
Language Swap by Pika MCP lets you speak any language in videos
By
–
Take your content global, already!
— Pika (@pika_labs) 9 juin 2026
The Language Swap skill via Pika MCP swaps the language you’re speaking in any video, making it look and sound like you speak fluent…anything. pic.twitter.com/kApd9yEbvXTake your content global, already! The Language Swap skill via Pika MCP swaps the language you’re speaking in any video, making it look and sound like you speak fluent…anything.