OpenAI prices GPT-6 Sol at $2/$10 per million tokens and Luna at $0.10/$0.50, about 50% cheaper than GPT-5.6. - The Deep View
Claude Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index, priced at $4 per million input tokens and $20 per million output tokens. - Artificial Analysis
Anthropic thunder stole OpenAI’s big day, with Opus 5.5 exceeding Fable 5.1. Luna-tic pricing…Luna keeps owning the low end, beating everything on price-performance, including Qwen-27b, which is small on parameters, but seems to use a lot of compute. MiMo 2.6 survives the frontier onslaught, but GPT-6 matches most of the mid-tier Chinese contenders like GLM, Kimi.
A reporter gave Meta's new Muse AI agent access to bank accounts and email to handle calls, bookings and errands for two weeks. - The New York Times
what could go wrong
unstoppable force of AI meets immovable object of consumer laziness. Everybody is Lazlo Hollyfeld from Real Genius now who used 80s AI to win sweepstakes, and Capital One is not amused.
Wargames: “we ought to take the men out of the loop”
Google says PROMPTFLUX malware asks Gemini to rewrite its own code hourly to evade detection - Fox News
Continuous Mythos red-teaming? Palo Alto’s Unit 42 says its new service has already found more than 10,000 vulnerabilities, with 37% rated critical - TNW | Security
Trump told the UN General Assembly the US rejects a 'globalist scheme' to control super intelligence (new name for AI) unchecked. - Fox Business
Sam Altman pitched Trump a US-China AI safety deal, but distrust and internal divisions make agreement unlikely, analysts say. - The Washington Post
don’t get hung up on stuff like technical degrees
Anthropic cofounder Jared Kaplan warned that letting AI train itself could trigger an uncontrollable 'intelligence explosion' between 2027 and 2030. - Business Insider
Predicting AI catastrophe probabilities, or 'P(doom)', has become mainstream, with predictable consequences for public discourse. - Financial Times
Pluralities across the US, Canada, UK, France, Spain and Germany favor pausing AI development over safety and job-loss fears. - Business Insider
72% of people across 37 countries feel curious or excited about AI, versus 41% worried, primarily in US and West, Gallup finds - The New York Times
AI disasters may be inevitable as adoption spreads to critical systems, but nuclear power's safety culture offers a governance model - Transformer News
Chernobyl FTW
Baselayer raised a $35 million Series A led by M13 to verify AI agents for financial institutions and commerce - SiliconANGLE
Apple is reportedly developing a circular AI Pin wearable with on-device visual intelligence and hands-free Siri, expected as early as 2027. - Geeky Gadgets
Most employees who save time using AI for routine tasks don't redirect it toward higher-value strategic work - Fast Company
Dallas Fed analysis finds AI-exposed Texas graduates faced lower post-2022 employment odds and wages. - The Wall Street Journal
Johns Hopkins researchers found GPT-4, Llama, Gemini, and Mistral write less formal, less professional responses when prompts give away that women wrote them - Business Insider
Stanford dining used AI to race swap white male student for black female in ad: dude felt ‘Silenced and erased’ - New York Post
A University of Chicago professor published roughly 200 papers and 14 books in a year using AI; a platform pulled 257 of his papers over quality concerns - The Washington Post
Jensen Huang: I’m so cuddly, do I look like Dr. Frankenstein? Jensen talks to Ezra Klein, does not want you to be afraid of A.I. - The New York Times
Oncologists say AI aids cancer detection and treatment, but a cure within 5-10 years is unrealistic - the Guardian
Building investing Superintelligence - Colossus
Andreessen Horowitz is investing $35 million in Horowitz Andreessen Academy, a two-year unaccredited college alternative for ages 16 to 22. - The Wall Street Journal
Don’t tell Shakespeare to sort your inbox. Asking an LLM, is this an AI headline, answer one character only, Y or N, is the dumbest, brutest force way to solve a classification problem.
A Jev approach eliminates a lot of the LLM overhead that is useless for this problem, and gives you confidence levels and probabilities. For e.g. spam detection if you decide obvious cases with Jev and route the harder ones to a full-blown LLM model with deep understanding, you will definitely get big cost and performance benefits.
ofc if you are dealing with large volume, you may get even bigger improvements by bootstrapping some labeled data with Jev and LLMs, and then training a proper classifier with e.g. XGBoost and other traditional ML classifiers.
Follow the latest AI headlines via druce.ai on Bluesky
