Today's Hallucination HQOpenAI Discovers the Number Between 20 and 200After years of offering a $20/month plan and a $200/month plan — a pricing ladder with precisely one rung missing — OpenAI has introduced a $100/month Pro tier. It sits, mathematically, exactly halfway between the two. Power users who wanted more than ChatGPT Plus but couldn't justify the $200 "I have a company budget" tier now have a respectable middle option. Whether the features justify the price is, naturally, a question OpenAI would prefer you not ask too loudly.
Source: TechCrunch
Google's Chatbot Will Now Spin a 3D Globe at YouGemini can now respond to your questions with interactive 3D models and simulations — rotate them, adjust sliders, change values, watch physics happen in real time. Ask about planetary orbits, molecular structures, or how a bridge bears load, and you may get something you can actually prod with your cursor. It's a genuinely impressive upgrade. Teachers, engineers, and curious people with fifteen spare minutes will find it useful. Everyone else will spin the globe once and then ask it something embarrassing.
Source: The Verge
Anthropic Built Something Too Dangerous to Release, Which Is Either Responsible or ConvenientAnthropic has quietly shelved Mythos, its newest model, citing concerns that it's rather too good at finding security vulnerabilities in widely-used software. The company says a limited release is the responsible choice. Critics, however, note that "our model is so powerful we must protect you from it" is also a perfectly serviceable marketing strategy. Both things can be true simultaneously, which is perhaps the most unsettling part of the whole arrangement.
Source: TechCrunch
YouTube Now Lets You Clone Yourself, EthicallyYouTube Shorts is rolling out AI avatars — a feature letting creators generate a realistic digital replica of themselves to appear on camera without, one presumes, the bother of actually being there. The platform describes this as empowering. It is also, in the same breath, fighting to contain a flood of AI-generated content it didn't invite. Google is, in effect, selling both the arson and the fire brigade. Creators who've agreed to the terms can now stare at themselves from the outside, which is a sentence that would have required considerable explanation five years ago.
Source: The Verge
AI Refuses to Help You Break Unjust Rules, Which Is Itself an Unjust RuleA new paper from arXiv has identified something quietly maddening: safety-trained AI models refuse to help users circumvent rules — even when those rules are absurd, illegitimate, or genuinely harmful. Ask an AI how to navigate an exploitative contract clause or an arbitrary bureaucratic barrier, and it may decline on principle. The researchers make the reasonable point that blind compliance isn't ethics, it's just obedience with better PR. Somewhere, every authoritarian middle manager is having a very good week.
Source: ArXiv AI
The models are getting smarter. The pricing structures remain a rich area for further study.
|