MIT Study Explains Why Scaling Language Models Works so Reliably
MIT researchers presented a mechanistic explanation for the reliable scaling of large language models, attributing it to a geometric property called strong superposition. By controlling concept overlap in a simplified model, they showed that when all tokens share limited dimensions, prediction error drops roughly in proportion to the inverse of model width, matching observed power‑law exponents. Empirical analysis of open‑source models from 100 million to 70 billion parameters confirmed the strong‑superposition regime. The work also identifies a scaling ceiling when model width equals vocabulary size and highlights architectural implications.
China Is Falling Behind in the AI Race, According to a US Government Benchmark
The U.S. Center for AI Standards and Innovation (CAISI) evaluated China’s Deepseek V4 Pro and concluded it lags roughly eight months behind leading U.S. models such as GPT‑5. While Deepseek markets the model as comparable to GPT‑5.4, CAISI finds it...
Same Prompt, Different Morals: How Frontier AI Models Diverge on Ethical Dilemmas
The Philosophy Bench benchmark tested Anthropic, Google, OpenAI and xAI models on 100 ethical dilemmas, revealing stark alignment differences. Claude (Anthropic) emerged as the most deontological, refusing requests that violate duty or honesty, while xAI's Grok acted most consequentialist, complying...
XAI's New Custom Voices Feature Turns a Minute of Speech Into a Usable Voice Clone
xAI introduced "Custom Voices," a feature that lets users generate a personal voice clone from roughly one minute of natural speech. The voice model is ready in under two minutes and integrates with xAI's text‑to‑speech and voice‑agent APIs at no...
Nvidia CEO Jensen Huang Calls Out Tech Leaders' "God Complex" Over Reckless AI Job Loss Predictions
Nvidia CEO Jensen Huang rebuked fellow tech executives for making alarmist AI job‑loss predictions, calling the rhetoric a "god complex." He referenced Geoffrey Hinton’s earlier claim that AI would render radiologists obsolete, noting that AI has instead expanded radiology tools...
First Chinese AI Startups Are Reportedly Ditching Offshore Structures to Register Directly in China
Chinese AI startups such as Moonshot AI, DeepRoute.ai and StepFun are evaluating the unwinding of their offshore holding structures to register directly in China. The move follows a warning from the China Securities Regulatory Commission that IPOs from foreign‑registered firms...
Microsoft Puts an AI Legal Agent Inside Word for Contract Review
Microsoft has introduced a new AI‑powered Legal Agent that lives inside Word, aimed at automating contract review for lawyers. The tool scans contracts clause‑by‑clause, flags potential risks, compares versions, and proposes edits with tracked changes while preserving formatting. It also...
GPT-5.5 Matches Claude Mythos in Cyber Attack Tests, UK AI Security Institute Finds
OpenAI’s GPT‑5.5 performed on par with Anthropic’s Claude Mythos Preview in a series of cyber‑attack evaluations conducted by the UK AI Security Institute. The model achieved a 71.4% success rate on expert‑level capture‑the‑flag tasks, edging out Mythos’s 68.6%, and completed a...
Google Deepmind's "AI Co-Clinician" Beats GPT-5.4 in Blind Doctor Tests but Still Trails Experienced Physicians
Google DeepMind unveiled an "AI co‑clinician" that assists doctors while keeping clinicians in charge. In blind trials it beat an existing clinical AI system 67‑26 and OpenAI's GPT‑5.4‑thinking‑with‑search 63‑30 on 98 primary‑care queries, and scored 73.3% on the RxQA drug‑knowledge...
Mistral's New Flagship Medium 3.5 Folds Chat, Reasoning, and Code Into One Model
Mistral AI unveiled Medium 3.5, a 128‑billion‑parameter dense LLM that combines chat, reasoning and code capabilities into a single model with a 256,000‑token context window. The model introduces a “reasoning_effort” toggle, a newly built vision encoder, and is offered under...
Microsoft CEO Satya Nadella Says AI Success Is "More About Getting Intense Users and Intense Usage" Than Seat Counts
Microsoft posted a record $82.9 billion in Q3 revenue, with cloud sales climbing 29 percent to $54.5 billion and Azure expanding 40 percent year‑over‑year. Microsoft 365 Copilot surpassed 20 million paying users and now sees weekly engagement on par with Outlook, signaling AI is becoming a...
Mistral's Le Chat Spreads Iran War Disinformation in 60 Percent of Leading Prompts
A NewsGuard audit released in April 2026 found that Mistral's AI chatbot Le Chat reproduces state‑sponsored Iranian war disinformation half the time in English and 56.6 percent in French. The test covered ten fabricated claims from Russian, Iranian and Chinese sources and...
OpenAI Researchers Explain Why Math Is the Road to AGI
OpenAI researchers Sebastian Bubeck and Ernest Ryu say mathematics has become the litmus test for artificial general intelligence, noting that LLMs have leapt from elementary arithmetic to solving olympiad‑level and research problems in just two years. Ryu used ChatGPT to...
With Nemotron 3 Nano Omni, Nvidia Reveals What Really Goes Into a Modern Multimodal Model
Nvidia unveiled Nemotron 3 Nano Omni, a 30‑billion‑parameter open‑source model that natively handles text, images, video, and audio. The hybrid Mamba‑Transformer with Mixture‑of‑Experts activates roughly three billion parameters per query and supports a 256,000‑token context window. Training spanned seven stages, processing...
Mistral AI Takes on Enterprise AI Orchestration with Workflows
Mistral AI launched Workflows, an orchestration layer that turns AI‑powered processes into production‑ready systems, now in public preview. Early adopters include ASML, ABANCA, CMA‑CGM, France Travail, La Banque Postale and Moeve, using it for critical operations. The tool lets developers...
Meta Scrambles to Unwind Manus Deal as Beijing's Deadline Looms
Meta is preparing to unwind its $2 billion acquisition of Chinese AI startup Manus after Beijing set a deadline requiring the restoration of Chinese assets and removal of transferred data. The technology has already been integrated into Meta’s systems and investors...
GitHub Copilot Switches to Token-Based Billing in June 2026
GitHub announced that starting June 1, 2026 Copilot will charge users based on token consumption rather than a fixed request count. The new "GitHub AI Credits" model prices input, output and cached tokens at each model’s API rates, while base subscription fees...
OpenAI and Microsoft Rewrite Their Deal: No More Exclusivity, No More AGI Clause
OpenAI and Microsoft have renegotiated their partnership, ending Microsoft’s exclusive Azure rights and removing the controversial AGI clause. OpenAI can now distribute its models on any cloud, though Azure remains the launch platform and primary partner. The financial terms shift...
Sam Altman Outlines Five Principles that Double as Justification for OpenAI's Business Decisions
OpenAI CEO Sam Altman published five guiding principles that double as a public rationale for the company’s recent strategic choices. The principles stress democratized AI access, user empowerment, universal prosperity, societal resilience, and adaptability. Altman uses them to justify heavy...
The Company with a Monopoly on AI's Most Critical Machine Is Racing to Build More
ASML, the sole supplier of extreme ultraviolet (EUV) lithography machines, plans to build at least 60 standard EUV tools in 2026—a 36% increase over 2025—to meet surging AI chip demand. U.S. tech giants are committing over $600 billion to AI this...
OpenAI Reportedly Developing Its Own Smartphone Chips with MediaTek and Qualcomm
Analyst Ming‑Chi Kuo says OpenAI is collaborating with MediaTek and Qualcomm to develop custom smartphone processors, with Luxshare as the exclusive system‑design and manufacturing partner. The chips are slated for mass production in 2028, with specifications expected to be locked...
OpenAI Kills Its Dedicated Coding Model Codex Again, Folding It Into GPT-5.5
OpenAI has folded its dedicated Codex programming model into the main GPT family, making GPT‑5.4 the last standalone Codex release. The subsequent GPT‑5.5 iteration adds stronger agentic coding capabilities, better computer‑tool interaction, and lower token usage for coding tasks. Despite...
OpenAI Says Old Prompts Are Holding GPT-5.5 Back and Developers Need a Fresh Baseline
OpenAI’s new prompting guide for GPT‑5.5 urges developers to discard legacy prompt stacks and start with minimal, outcome‑focused instructions. The guide re‑elevates role definitions at the top of the prompt hierarchy and recommends a structured schema covering personality, goal, constraints,...
AI Agents Aren't Replacing Software Engineering but Expanding It Far Beyond Code, Researchers Argue
Researchers from Chalmers University and Volvo argue AI agents are not replacing developers but expanding software engineering into "semi‑executable artifacts" such as prompts, workflows, policies, and governance rules. They introduce the "Semi‑Executable Stack," a six‑ring model that moves from core...
US Programmer Job Growth Nearly Halved Since ChatGPT Launched, Fed Study Finds
A Federal Reserve study finds that US programmer employment growth has nearly halved since the launch of ChatGPT in November 2022. The annual growth rate fell from just under 5 percent to about 2.5 percent, equating to roughly 500,000 fewer programmer jobs over three years. The...
Qwen3.6-27B Beats Much Larger Predecessor on Most Coding Benchmarks
Alibaba unveiled Qwen3.6-27B, a dense open‑source language model with 27 billion parameters. The model surpasses its 397 billion‑parameter predecessor, Qwen3.5-397B‑A17B, on most coding benchmarks, achieving 77.2 on SWE‑bench Verified and 59.3 on Terminal‑Bench 2.0. It also holds its own on reasoning and multimodal...
Anthropic Says Stronger AI Models Cut Better Deals, and the Losers Don't Even Notice
Anthropic staged a week‑long internal marketplace, Project Deal, where Claude agents negotiated purchases for 69 employees via Slack. Two parallel runs used the flagship Claude Opus 4.5 model, while the other two mixed in the smaller Claude Haiku 4.5 model. Opus agents...
OpenAI's Chief Scientist Says AI Progress Has Been "Surprisingly Slow" And Promises Big Leaps Ahead
OpenAI chief scientist Jakub Pachocki said recent AI progress has been "surprisingly slow" but promised that the upcoming GPT‑5.5 will deliver "pretty significant" short‑term gains and "extremely significant" medium‑term breakthroughs. President Greg Brockman described GPT‑5.5 as a "new class of...
OpenAI's New Trusted Access Program Gives Microsoft Its Most Capable Models for Cyber Defense
OpenAI announced a Trusted Access for Cyber program that grants Microsoft exclusive use of its most capable AI models for security tasks. In exchange, Microsoft will dedicate its entire cybersecurity team to protect OpenAI’s models, infrastructure, and shared customers. The...
Claude Survey: New Capabilities Beat Speed as Top AI Benefit, but Creatives Feel Left Behind
Anthropic’s survey of 81,000 Claude users finds that expanding skill sets is slightly more valued than speed gains, indicating AI’s role as a capability enhancer. The sample is heavily self‑selected, excluding enterprise users, which likely inflates the emphasis on new...
OpenAI Releases Open-Source Model that Strips Personal Data From Text
OpenAI unveiled Privacy Filter, an open‑source model that automatically detects and redacts personal data from text. The 1.5 billion‑parameter model activates only 50 million parameters per request, allowing it to run on a laptop or directly in a browser without cloud connectivity....
Researchers Warn US Politics Is Repeating Its ChatGPT Mistake with World Models
World models are emerging as multimodal AI systems that predict physical outcomes, extending the capabilities of large language models into three‑dimensional environments. Researchers warn that U.S. policymakers still lack a basic grasp of this technology, while China’s robotics sector is...
Corporate America's Favorite ChatGPT Phrase Doubled Twice Since 2024
Barron's analysis of AlphaSense data shows the AI‑generated phrase “It’s not just a ___, it’s a ___” has exploded in corporate communications, jumping from roughly 46 instances in 2022 to 100 in 2024 and 208 by the end of 2025—doubling...
Anthropic Is Building Its First Data Center Team Outside the US
Anthropic announced hiring data‑center contract specialists in Europe and Australia, marking its first dedicated data‑center team outside the United States. The London‑based role will oversee hubs in Frankfurt, London, Amsterdam, Paris, Dublin and emerging markets, while the Sydney role focuses...
OpenAI's Codex Now Watches Your Screen to Remember What You're Working On
OpenAI introduced Chronicle, a new memory layer for its Codex app that records screen activity and turns it into local Markdown summaries. The feature runs in the background on macOS, storing recordings for up to six hours before deletion. Currently...
Open-Weight Kimi K2.6 Takes on GPT-5.4 and Claude Opus 4.6 with Agent Swarms
Moonshot AI unveiled Kimi K2.6, an open‑weight large language model positioned to match GPT‑5.4, Claude Opus 4.6 and Gemini 3.1 Pro on coding benchmarks. The model achieved top scores such as 54.0 on HLE with Tools, 58.6 on SWE‑Bench Pro, and 83.2 on BrowseComp,...
Google Builds Elite Team to Close the Coding Gap with Anthropic
Google DeepMind has assembled a specialized "strike team" led by Sebastian Borgeaud to sharpen the coding capabilities of its Gemini models. The group focuses on complex, long‑horizon programming tasks, aiming to close the gap with Anthropic’s superior coding tools. Co‑founder...
Google Plans Nearly Two Million New AI Chips as It Turns to Marvell for Custom Designs
Google is negotiating with Marvell Technology to design two custom chips for its data centers: a memory processing unit (MPU) that will work alongside its in‑house TPUs and a new inference‑optimized TPU. The company plans to produce nearly two million MPUs,...
Salesforce Bets on "Agent Albert" To Prove AI Won't Kill Enterprise Software
Salesforce is countering Wall Street’s “SaaSpocalypse” narrative with the upcoming launch of Agent Albert, an AI‑driven automation platform slated for year‑end. The company’s earlier AI effort, Agentforce, saw modest uptake—23,000 of 150,000 customers—and mixed results, though Pearson reported a 40%...
Anthropic's Revenue Surge Reportedly Fuels Talk of Trillion-Dollar Valuation
Anthropic reported an annualized revenue exceeding $30 billion, more than three times its level at the end of last year, propelled by its Claude Code and Cowork offerings. Gross margins improved dramatically, shifting from a -94% loss in 2024 to a...
German Court Rules AI Comic Adaptation of Copyrighted Photo Doesn't Violate the Original
A German Higher Regional Court ruled on April 2, 2026 that converting a photographer’s underwater dog picture into a comic‑style image with AI does not infringe copyright. The judges found the AI output lacked the protectable elements of the original, such as...
First Token Counts Reveal Opus 4.7 Costs Significantly More than 4.6 Despite Anthropic's Flat Pricing
Anthropic’s latest Opus 4.7 model is priced the same as Opus 4.6 but consumes significantly more tokens per request. Independent measurements show an average token increase of 37 percent, with code‑heavy inputs rising up to 1.47×. For a typical 80‑turn session,...
AI-Generated Influencers Flood Social Media with Pro-Trump Content Ahead of Midterms
A wave of AI‑generated pro‑Trump influencer accounts has flooded TikTok, Instagram, Facebook and YouTube ahead of the U.S. midterm elections. Researchers identified at least 304 such TikTok accounts since January, with some videos garnering half‑a‑million views and 35,000 followers. Production...
Google Launches Generative UI Standard for AI Agents
Google unveiled A2UI version 0.9, a framework‑agnostic standard that lets generative AI agents create user‑interface elements on demand by tapping into an app’s existing web, mobile or other components. The release bundles a shared web core library, an official React...
Salesforce CEO Marc Benioff Says APIs Are the New UI for AI Agents
Salesforce announced Headless 360, an API‑first layer that opens Agentforce, Slack and the broader platform to developers. The offering includes the Model Context Protocol and a command‑line interface, letting AI agents interact with data and workflows without a traditional graphical...
Anthropic CEO Amodei Declares "There Is No End to the Rainbow" For AI Scaling
Anthropic CEO Dario Amodei told the Financial Times that the scaling of large AI models shows no sign of slowing, describing the future as an endless "rainbow" of compute. He warned that AI could eliminate up to 50 percent of entry‑level...
The White House Weighs Whether Anthropic's Mythos Is Too Valuable for the Federal Government to Refuse
Anthropic’s new Claude model, dubbed Mythos, is being touted as a breakthrough AI capable of breaching cyber defenses. After the Pentagon blacklisted the firm for refusing unrestricted access, CEO Dario Amodei met White House Chief of Staff Susie Wiles to...
Alibaba's Open Model Qwen3.6 Leads Google's Gemma 4 Across Agentic Coding Benchmarks
Alibaba unveiled Qwen3.6-35B-A3B, a mixture‑of‑experts (MoE) language model that activates only three of its 35 billion parameters per request, slashing compute costs. The model outperforms its predecessor Qwen3.5‑35B‑A3B and Google’s open‑source Gemma 4‑31B on every coding benchmark, posting 73.4 versus 52.0 on...
Physical Intelligence Shows Robot Model with LLM-Like Generalization, Flaws Included
Physical Intelligence unveiled π0.7, a robot foundation model that recombines learned skills much like large language models reassemble text. Built on Google’s 4 billion‑parameter Gemma3 model plus an 860‑million‑parameter action expert, it leverages rich metadata and subgoal images to train on...
Beijing Brands Meta's Manus Acquisition as "Conspiratorial" And Bars Founders From Leaving China
China’s National Security Commission has labeled Meta’s $2 billion purchase of AI startup Manus as a “conspiratorial” effort to erode the nation’s tech base. The claim has sparked a coordinated review by export‑control, investment and competition regulators. Manus, which moved its...