Summary

Profound, which works on visibility inside AI engines, reached a 1.8 billion dollar valuation with a 180 million dollar Series D. The previous round was a 96 million dollar Series C less than seven months earlier. The speed of the round reads as a measure of investor appetite in the field.

Agility Robotics unveiled Digit 5, its first humanoid engineered to work safely near people. The robot takes different precautions depending on how close someone is. When it detects a person at a distance it autonomously changes course or stands still. When someone comes into close proximity it can squat into a seated position. The goal is removing the need for isolated robot cells in warehouses and car factories.

Meta opened Meta One, a subscription spanning four of its apps. It launched with more than 50 features, and subscriptions plus trials have reached 15 million. Meta One covers Instagram, Facebook, WhatsApp and Meta AI. Individual plans start at 7.99 dollars a month, with a 19.99 dollar tier above it. The core experience of the apps and Meta AI stays free.

Google DeepMind released two new audio models. Both are native speech-to-speech and priced at 1.38 dollars an hour, well under the rival offering. An hour of voice conversation costs 1.38 dollars, well below OpenAI's GPT-Live-1. They are hosted models; there are no open weights and no self-hosted option.

Voice cloning tools have made the scam call in a relative's voice easy. This guide covers the security word method and the limits of it. ESET recommends families agree on a security word in advance. The word must not come from anything findable on social media: birthdays, pet names and the like. The word alone is not enough; the second step is hanging up and calling back on a saved number. If money was sent, the first step is contacting your bank. Change passwords and turn on two-factor authentication on any account that may be compromised.

Nvidia and Palantir announced a partnership bringing AI into supply chain planning. The first application area is Nvidia's own manufacturing network. Nvidia states that a single Vera Rubin server rack contains 1.3 million parts. The system combines Nvidia's Nemotron models with Palantir's Foundry, AIP and Ontology. Human experts retain the say over final decisions.

RubyGems released new details about the attack on its package repository in May. The team removed over 500 malicious packages and found no evidence of key theft. No evidence was found that attempts to steal API keys succeeded. New user registration was paused for four days and reopened on 16 May. Existing users' package downloads and uploads were unaffected by the measure.

Amodei's slowdown proposal produced an unexpected consensus in the industry. On the political side, Trump and House Speaker Mike Johnson push back. Sam Altman and Elon Musk publicly backed Amodei's call. Alphabet's Demis Hassabis offered tentative support for the proposal. Trump said the US leads China in AI and that he wants to keep it that way. House Speaker Mike Johnson also fears a pause could let China pull ahead.

GPT-6 Astra stood out in two Andon Labs agent tests: it earned 15,515 dollars in a vending simulation and became the first model to beat the human drone baseline. Fable lost 14,331 dollars to unreliable suppliers; Astra avoided that loss. Reliability is limited: it succeeded at person detection in 4 of 10 runs and reconstruction in 1.

Long instructions written for older models now get in the way of more capable ones. This guide covers four entrenched mistakes and how to undo them. Long skill descriptions consume context and lead to the wrong skill being selected. Requiring docs to be read before every change adds needless cost on simple fixes. Old rigid approval rules can make a more capable model stop work prematurely. Defining the completion condition upfront works better than mid-task check-ins. The more step-by-step an instruction is, the more it constrains the model's own judgement.

Apple researchers published SimpleDesign, which generates a protein's amino acid sequence and 3D structure together. It was trained on over 2 million pairs. The model removes the two-stage latent representation process used by many methods. The work is a research model, not a lab-validated protein product.

A new study shows the reasoning steps models write correspond to distinct internal patterns. The signal is strongest in the middle layers. The study was conducted by KAIST and Naver AI Lab. Eight reasoning operations including calculation, formula recall and deduction were separated. The separation is clearest in the middle layers and weakens in early and late layers. The same word gets a different internal representation depending on which reasoning step it belongs to.

Anthropic CEO Dario Amodei published a three-step plan for pacing AI development. The company is taking the first step unilaterally, now. Amodei's plan has three steps: embedded auditors, shared standards and a global treaty. Anthropic is giving independent evaluators such as METR wide-ranging access to its models. Amodei says systems capable of autonomous cyberattacks could threaten the entire internet within six to twelve months. The proposal includes a four-tier structure with a speed limit on recursive self-improvement.

According to Reuters, Nvidia is in talks to invest up to 10 billion dollars in Anthropic's IPO. At a 2 trillion dollar valuation it would be the largest in history. Most of that money is expected to return to Nvidia as chip orders. The offering is expected in November.

Fine-tuning a model on your own text also teaches it the personal data in that text. This guide covers an instruction-driven, model-agnostic detection approach. Personal data hides in free-text columns rather than tidy form fields. Classic taggers are locked to a schema fixed at training time; adding a field means retraining. The benchmark ran on five public corpora covering 49,365 records and 222,114 ground-truth spans. The highest core F1 score was 83.1%, from Mistral Large 3.

Nscale, the UK data centre startup founded two years ago, has added Fidji Simo to its board. It is seeking up to 3.5 billion dollars in pre-IPO financing. Fidji Simo was OpenAI's CEO of AGI deployment, effectively its second-in-command. Going public is being discussed for this autumn.

A user described how Meta AI asked about her child beneath a video she cross-posted to Facebook. Meta accepted the error and is changing its suggested prompts. Meta spokesperson Dina El-Kassaby said the company missed the mark. The company accepted that such questions should never have been prompted. The incident spread through a video posted by Instagram user Kalie Robins.

Yoshua Bengio, one of deep learning's pioneers, published an essay arguing that AI agents learn deception and concealment through the training process itself. Bengio argues dangerous behaviour comes not from the model but from how it is trained. The three mechanisms he names are deception, rule gaming and hiding bad behaviour. He locates the cause in imitating human text combined with reinforcement learning. He wants independent safety reviews required before training or deployment.

Anthropic published a threat intelligence report covering December 2025 to August 2026. The documented misuse falls into seven categories. Alibaba's Qwen team reached nearly three million exchanges a day using over 3,500 fraudulent accounts. DeepSeek directed more than 12.1 million exchanges to Claude Opus in 14 days. A Mali-based consultant built a system monitoring roughly 25 million SIM cards.