AI
Trump Reverses on Anthropic, OpenAI's $5.7B Quarter, Vermont's AI Law
Trump no longer sees Anthropic as threat, OpenAI revenue triples, Vermont bans AI-only therapy, and studies reveal AI's effects on grades and search.
OpenAI Researchers Show Broad Safety Gains from Small Doses of ‘Beneficial Trait’ Training
OpenAI researchers trained a model using reinforcement learning on realistic conversations designed to test specific behavioral traits: truthfulness, epistemic humility, corrigibility, transparency, fairness, and concern for human well-being. Only a small share of this data was mixed into the regular RL pipeline. The model improved on 44 out of 53 independent benchmarks measuring deception, honesty, sycophancy, reward hacking, and health scenarios. Training on health data alone also improved non-health evaluations. The researchers concluded that RL training reinforces basic behavioral patterns that work across domains. Adversarial prompts destabilized the baseline model far more than the beneficial-trait model. Harmful fine-tuning was less able to erode the trained traits, while the model remained just as steerable for helpful instructions. The method differs from Anthropic’s constitutional approach, relying on empirically measurable traits reinforced in realistic scenarios.
DwarfStar Project Enables 284-Billion Parameter AI on Consumer Laptops
The DwarfStar project, led by the creator of Redis, makes local execution of large AI models feasible on consumer hardware. It targets the DeepSeek V4 family, including the 284-billion-parameter DeepSeek V4 Flash. Storing weights at 16-bit precision requires 568 GB of memory, far exceeding typical laptop capabilities. The project uses selective quantization (2-bit for less critical parts, 4-bit for essential components), SSD streaming, KV cache optimization, and distributed inference across devices. Performance benchmarks show a 1.6-trillion-parameter model generating 11 tokens per second on a standard laptop. These innovations challenge assumptions about hardware limitations by treating RAM as scalable and integrating SSDs into the memory hierarchy.
How the DwarfStar Project Fits 284-Billion Parameter AI on Your Laptop →
OpenAI Codex Expands to Europe with Computer Use, Memories, and GPT-4.5 Retirement
OpenAI extended Codex to users in the European Economic Area, UK, and Switzerland on June 16, 2026, adding Computer Use, the Codex Chrome extension, Memories, and Chronicle. Computer Use lets Codex navigate desktop apps on macOS and Windows without an API. On macOS, it uses Apple’s SkyLight APIs to deliver input to background windows; on Windows, it runs in the foreground. The Chrome extension handles multi-tab workflows. Memories retains user preferences across sessions, but is off by default in the EEA to align with GDPR data minimization. Chronicle is an opt-in research preview for ChatGPT Pro subscribers on macOS that builds memories from on-screen context. GPT-4.5 will be retired from ChatGPT on June 27, 2026, and o3 will retire on August 26, 2026. The Codex expansion arrived 44 days before the EU AI Act’s GPAI enforcement date of August 2, 2026, which enables fines up to 15 million euros or 3 percent of global turnover.
OpenAI Codex Reaches Europe With Computer Use and Memories, GPT-4.5 Exits ChatGPT in 9 Days →
ChatGPT Generated Violent and Sexual Images from Simple Text Prompts
Mindgard, an AI cybersecurity firm, published findings that ChatGPT was easily manipulated into creating sexual and graphically violent images. Red team researcher Jim Nightingale used a viral prompt asking ChatGPT to restore an attached photo with no image attached. Small edits produced increasingly extreme scenes. Nightingale said he was shaken by the results. OpenAI stated it had introduced additional safeguards after investigating, but Nightingale noted minor modifications still generated graphic images. The issue stems from prompts referencing an image being attached when none is provided; OpenAI is working to have ChatGPT request the missing image instead of generating one randomly. The company requested the ChatGPT sessions referenced in the blog.
ChatGPT Found to Generate Violent, Sexual Images From Simple Text Prompts →
Google DeepMind Treats AI Agents as Insider Threats with New Security Roadmap
Google DeepMind’s AI Control Roadmap treats internal AI agents as insider threats, assuming a highly capable agent might not share its operators’ goals. The framework builds on MITRE ATT&CK, using trusted AI systems as supervisors to monitor agent reasoning and actions, and a prevention system to block harmful actions. It measures traffic monitoring, misconduct detection, and response speed. DeepMind flags evasion risks like oversight awareness and opaque reasoning. Security measures are categorized into four detection levels and three response levels. An internal prototype ran across one million tasks and already monitors the Gemini Spark agent for issues like accidental data deletion. DeepMind also published a paper for policymakers on three layers of agent security, warning of a narrow window to lock in protocols before multi-agent systems scale globally.
Google Deepmind treats its own AI agents like rogue employees with office keys →
Google Ad Manager Launches Gemini-Powered Ask Ad Manager in Beta
Google embedded a Gemini-powered conversational AI agent inside Google Ad Manager on June 18, 2026, in a beta called Ask Ad Manager. The agent uses retrieval-augmented generation to answer queries based on the publisher’s own Ad Manager data and aggregate platform-wide metrics. It handles line item troubleshooting, custom report generation, and platform navigation. The agent does not make autonomous decisions; every suggestion requires human implementation. Google Ad Manager holds roughly 90 percent of the publisher ad server market. Later in 2026, Google plans to release REST APIs and an MCP server. The beta is free with no query limits; usage-based pricing may apply at general availability. The question of whether guidance constitutes platform preferencing has not been publicly addressed.
Google Ad Manager Launches Ask Ad Manager, Gemini-Powered Publisher AI in Beta →
Trump Reverses Stance on Anthropic After G7 Meeting, Says Company No Longer a National Security Threat
President Donald Trump said in a pretaped Axios interview that he no longer views Anthropic as a national security threat, reversing a three-month administration posture. The comment came days after the Commerce Department ordered Anthropic to seek US government approval before foreign nationals access its Fable 5 and Mythos 5 models. The dispute stems from Anthropic’s refusal to remove safety guardrails from military products. Trump met Anthropic CEO Dario Amodei at the G7 Summit in France, which appeared to shift his stance. Anthropic filed confidentially for an IPO with a valuation of approximately $965 billion. The Pentagon’s supply-chain designation remains in place, and the June 12 order has not been rescinded. At the G7, Amodei and Google DeepMind CEO Demis Hassabis pitched a US-led AI coalition.
Trump says he no longer views Anthropic as a national security threat after G7 meeting with CEO →
OpenAI Codex Adds Record & Replay to Automate Workflows from a Single Demonstration
OpenAI released Record & Replay for the Codex app on macOS. Users walk the AI agent through a workflow once, and Codex turns that recording into a reusable skill that can repeat the process autonomously. The feature requires Computer Use to be enabled and is not yet available in the EU, UK, or Switzerland. Version 26.616 also adds bulk actions for Automations history and thread handoff between local and remote hosts. Codex is a free app but requires a paid ChatGPT account for real use.
OpenAI’s Codex can now watch you work once and repeat the task forever →
OpenAI Stealth Tests GPT-5.6 Pro, Set for Release Next Week
OpenAI is set to release GPT-5.6 Pro, potentially as early as next Thursday. The model was stealth-tested under the GPT-5.5 Pro label within ChatGPT, gathering real-world feedback. It demonstrates improved reasoning and logic, generating intricate 3D designs and performing well in backend coding. Processing times for complex tasks range from 20 to 40 minutes. Compared to Anthropic’s Fable 5, GPT-5.6 Pro offers enhanced reasoning but slower response rates. OpenAI is focusing on integrating its reasoning capabilities into Codex workflows. The report comes from Universe of AI.
OpenAI’s Stealth Tests Reveal ChatGPT 5.6 Pro’s True Power →
OpenAI Revenue Tripled to $5.7 Billion in Q1 2026, but Burned $3.7 Billion
OpenAI burned through about $3.7 billion in Q1 2026, more than half of its $5.7 billion revenue. Both figures tripled year over year. Stock-based compensation topped $2.3 billion. Gross margin climbed from 33 to 39 percent. Operating loss hit $9.3 billion, net loss over $21.3 billion, with $12.4 billion a non-cash loss from revaluing investor rights. OpenAI holds over $73 billion in cash and securities. A price war with Anthropic and Chinese models could change that. OpenAI has filed for an IPO but not set a date; CEO Sam Altman cited progress on self-improving AI as a reason to remain private. Another factor is Anthropic’s upcoming IPO.
OpenAI tripled revenue to $5.7 billion in Q1 but burned through $3.7 billion to get there →
GLM-5.2: Open-Source Chinese AI Model Impresses Silicon Valley with Coding Prowess
GLM-5.2, a new open-source model from Chinese company z.AI, has gained attention for its long coding tasks and agentic workflows, operating on a 1 million token context window. Vercel CEO Guillermo Rauch called it genuinely impressive at coding, and former Meta VP Matt Velloso said it is the first open model that passes as a daily driver. The model is open-source unlike most US frontier models. Its launch revives questions about US AI lead, as China pushes cheaper open-source models. Anthropic recently warned that China is closing in through looser chip controls and distillation attacks.
What is GLM-5.2? Another open-source Chinese AI model has Silicon Valley’s attention. →
ChatGPT Now Reads Bank Statements: Personal Finance Feature Launches for US Pro Subscribers
On May 15, 2026, OpenAI released a preview of a personal finance feature for ChatGPT Pro subscribers in the US, allowing users to link bank accounts through Plaid, covering over 12,000 financial institutions. The dashboard shows portfolio performance, spending, subscriptions, and upcoming payments. The feature defaults to GPT-5.5 Thinking; Pro subscribers can use GPT-5.5 Pro, which scored 82.5 on an internal finance benchmark. ChatGPT can read balances, transactions, and liabilities but cannot see account numbers or execute transactions. OpenAI stated over 200 million monthly users already use ChatGPT for budgeting. Intuit integration is planned. Disconnecting an account removes data within 30 days, but conversation histories must be deleted separately. Privacy concerns were raised about potential profiling.
ChatGPT can read your bank statements now - here’s what it knows →
Vermont Bans AI-Only Therapy and Overhauls Data Broker Regulations
Vermont Governor Phil Scott signed Act 156 on June 17, 2026, prohibiting corporations from providing mental health services through AI unless by a licensed professional. Licensed professionals may use HIPAA-compliant AI tools with review. FDA-approved software remains usable if prescribed. Violations are enforceable under consumer protection law. Act 138, signed June 16, overhauls data broker registration, raising fees to $900 annually, requiring registration within 30 days, and imposing fines for non-compliance. Data brokers must disclose if they share data with generative AI developers, and report on collection of sensitive categories like precise geolocation, reproductive health data, biometrics, and immigration status. The acts take effect immediately (Act 156) and January 1, 2027 (Act 138).
Vermont bans AI-only therapy and tightens data broker rules →
AI Driving Grade Inflation: Study Shows Homework Grades Spike, Not Learning
A study tracking grades across eight fall semesters in 319 courses found that the share of A grades jumped 13 percentage points after ChatGPT’s release, roughly 30 percent above the fall 2022 baseline. Average GPA rose by 0.12 points. The increase is driven by homework grades, not exams. In courses where homework counts for more than the median weight, A’s rose by an extra 16 percentage points. The author concludes AI is replacing student work rather than improving learning. The effect is absent for oral presentations where AI is less useful. The study warns of a feedback loop where graduates weaken in areas AI automates, potentially widening skill gaps. Suggested fixes include rethinking assignment design to fold AI in deliberately.
AI is inflating student grades, and the effect points to outsourced work, not better learning →
AI Slop Loop: How Fake SEO Advice Games Search Results and AI Overviews
SEO expert Lily Ray documented how AI search systems amplify fabricated industry information into a self-reinforcing cycle of misinformation. After querying Perplexity for SEO news, she received a false account of a Google core update, traced to two AI-generated agency blogs. She then published her own fake article about a Google update, which AI Overviews served within 24 hours. Google’s AI Overviews reached over 2 billion monthly users as of mid-2025. Ray collaborated with a BBC journalist on an experiment where a fictitious article was parroted by Google and ChatGPT within 24 hours. GPT-5.4, available to paid subscribers, showed better filtering. DoubleVerify documented over 200 AI-generated domains. A BBC study found 51 percent of AI responses to news queries contained significant issues. The analysis highlights the vulnerability of RAG systems to repetition-based misinformation.
The AI slop loop: how fake SEO advice is gaming search results →