Salesforce Executives Signal Declining Trust in Large Language Models
Salesforce executives acknowledge a decline in confidence in large language models (LLMs) over the past year, citing randomness and instruction‑following failures. The company is pivoting its Agentforce platform toward simple, rule‑based automation while restricting generative AI in certain scenarios. Salesforce highlights persistent “drift” issues where AI agents lose focus during off‑track queries. Despite the shift, Agentforce remains on track for more than $500 million in annual revenue.
Report: OpenAI May Embed Sponsored Content Directly Into ChatGPT Responses
OpenAI is actively prototyping ways to embed sponsored content directly into ChatGPT answers, including woven‑into‑response ads and sidebar placements. Internal mockups show both immediate product recommendations and post‑click suggestions for travel or retail queries. The company is also exploring the...
A Zelda Puzzle Proves AI Models Can Crack Gaming Riddles that Require Thinking Six Moves Ahead
Modern language models are now capable of solving multi‑step visual puzzles, as demonstrated by a color‑changing Zelda shrine challenge. Google Gemini 3 Pro, OpenAI GPT‑5.2‑Thinking, and Anthropic Claude Opus 4.5 were tested on the same screenshot, with GPT‑5.2 solving it consistently and quickly, Gemini 3 Pro...
Zhipu AI Challenges Western Rivals with Low-Cost GLM-4.7
Zhipu AI unveiled GLM-4.7, a large language model tuned for autonomous programming and "vibe coding" website creation. The model introduces Preserved Thinking to maintain reasoning across extended dialogs and builds on Interleaved Thinking from GLM-4.5. It achieved a 73.8% score...
Alibaba's New Qwen Models Can Clone Voices From Three Seconds of Audio
Alibaba Cloud’s Qwen team unveiled two AI voice models: Qwen3‑TTS‑VD‑Flash, which crafts custom voices from detailed textual prompts, and Qwen3‑TTS‑VC‑Flash, which can clone a speaker’s voice from just three seconds of audio and render it in ten languages. Both models...
Ex-Tesla AI Chief Andrej Karpathy Shares Four Tips for AI Startups Competing with OpenAI
Former Tesla AI chief Andrej Karpathy argues that AI startups should view themselves as vertical specialists rather than direct rivals to large language‑model labs. He cites Cursor, an AI‑powered code editor, as proof of a new "LLM app" layer that...
GPT-5 Allegedly Solves Open Math Problem without Human Help
Swiss mathematician Johannes Schmitt announced that GPT‑5 independently solved an open problem in algebraic geometry, delivering a novel proof that draws on techniques from a different subfield. The solution appears in a newly posted arXiv paper that mixes contributions from...
Google Locks in New Energy Reserves for Its AI Expansion
Alphabet’s Google unit is buying clean‑energy developer Intersect for $4.75 billion in cash, assuming its debt, to secure roughly $15 billion of energy and data‑center projects. The acquisition targets projects that will deliver about 10.8 GW of renewable capacity by 2028—more than twenty...
Yann LeCun Calls General Intelligence "Complete BS" And Deepmind CEO Hassabis Fires Back Publicly
Yann LeCun, Meta’s departing chief AI scientist, dismissed the notion of "general intelligence" as meaningless, calling it "complete BS" on a recent podcast. DeepMind CEO Demis Hassabis publicly rebuked LeCun on X, arguing that LeCun confuses general intelligence with universal...
Kling 2.6 Adds Voice Control and Motion Upgrades as AI Video Tools Race Toward Realism
Kuaishou's Kling 2.6 video generator now offers voice control and upgraded motion control, allowing creators to add custom‑trained or uploaded human voices and achieve more precise full‑body, hand, and facial movements. The voice feature supports speaking, narration, singing, rapping and...
Nvidia Wants to Create Universal AI Agents for All Worlds with NitroGen
Nvidia unveiled NitroGen, an open‑vision action model designed to serve as a universal gaming agent. The model was trained on 40,000 hours of gameplay footage from over 1,000 titles, using YouTube and Twitch videos with visible controller overlays to extract...
Alibaba's Qwen Releases AI Model that Splits Images Into Editable Layers Like Photoshop
Alibaba’s Qwen unit unveiled Qwen-Image-Layered, an AI model that decomposes photos into editable RGBA layers. The system can split an image into three or eight transparent layers, and each layer can be further broken down recursively. Users can resize, recolor,...
China Wins the Open Model Race and the Price to Pay Goes Beyond Economics
In 2025 Chinese developers surpassed U.S. providers in open‑source AI model downloads, capturing 44 percent of the market according to the Economies of Open Intelligence report. Alibaba's Qwen family and Deepseek together generated over 1.2 billion downloads, while Meta and Google fell...
Anthropic's Claude Opus 4.5 Can Tackle some Tasks Lasting Nearly Five Hours
Anthropic's Claude Opus 4.5 set a new benchmark on METR’s evaluation, achieving a 50 percent time horizon of roughly 4 hours 49 minutes. This metric indicates the longest task length the model can solve half the time, surpassing all previous records. At a stricter...
Google's Open Standard Lets AI Agents Build User Interfaces on the Fly
Google unveiled the open‑source A2UI (Agent‑to‑User Interface) standard, letting AI agents generate graphical UI elements on demand via JSON streams instead of HTML or JavaScript. The protocol enables on‑the‑fly forms, buttons, and widgets that render natively within any host app,...
Google Releases FunctionGemma to Bring AI Commands to Smartphones
Google unveiled FunctionGemma, a function‑calling‑optimized variant of the compact Gemma 3 270M model, designed to run directly on Android smartphones. The on‑device AI can translate natural‑language prompts into executable commands, such as creating calendar events or manipulating game elements, demonstrated in the...
OpenAI Brings Cheaper Subscription Tier "Go" To More Markets
OpenAI is rolling out its low‑cost ChatGPT Go subscription to more than 70 additional countries, extending the tier that debuted in India earlier this year. In Germany the plan is priced at €8 per month and now includes image generation, file...
OpenAI Updates Codex Model, Adds Trusted Access Program for Cyber Defense
OpenAI unveiled GPT-5.2-Codex, an autonomous software-agent model that adds context compression and improved image processing. Benchmarks show modest accuracy gains—56.4 % on SWE-Bench Pro and 64 % on Terminal-Bench 2.0—over the standard version. The company also launched a trusted access program, allowing vetted...
Meta Preps "Mango" And "Avocado" AI Models for 2026
Meta is preparing two new AI models, codenamed “Mango” and “Avocado,” slated for release in the first half of 2026. Mango will specialize in image and video generation, while Avocado is a language model optimized for programming tasks. The projects...
GPT-5.2 Tops OpenAI's New FrontierScience Test but Struggles with Real Research Problems
OpenAI introduced the FrontierScience benchmark, a two‑part test featuring Olympiad‑level problems and open‑ended PhD‑level research challenges. GPT‑5.2 topped the leaderboard, reaching 77% accuracy on the Olympiad set and 25% on the research set, while Gemini 3 Pro trailed closely on Olympiad...
New OpenAI Platform Teaches Newsrooms How to Use AI Tools
OpenAI has unveiled the "Academy for News Organizations," a new learning platform designed to help journalists and publishers adopt artificial‑intelligence tools. Developed with the American Journalism Project and the Lenfest Institute, the on‑demand program offers practical training, translation examples, and...
OpenAI Launches App Submissions and Rolls Out Store in the New Year
OpenAI announced that developers can now submit ChatGPT applications for inclusion in a new directory accessed via the Tools menu. Submissions will undergo a review process, and approved apps can be launched directly in conversations using the “@” command. A...
OpenAI Reportedly Seeking up to $100 Billion in New Funding Round
OpenAI is in early discussions to raise as much as $100 billion, which could lift its valuation to roughly $750 billion—a 50% jump from its October share sale. The company reports an annualized revenue run rate of $19 billion and aims for $30 billion...
Terence Tao Proposes "Artificial General Cleverness" As a More Honest Label for What AI Actually Does
Renowned mathematician Terence Tao has argued that the term “artificial general intelligence” overstates what current systems can achieve. He proposes calling the capability “artificial general cleverness” (AGC), describing it as the ability to solve complex problems using improvised, sometimes random,...
The $10 Billion Loop: Amazon Could Pay OpenAI so OpenAI Can Pay Amazon
Amazon is in advanced talks to invest at least $10 billion in OpenAI, a move that could lift the startup’s valuation past $500 billion. The cash infusion is aimed at covering OpenAI’s soaring server expenses, including a $38 billion cloud agreement with Amazon...