AI
Anthropic ports Claude to AMD GPUs in three days, challenging Nvidia's CUDA moat
Anthropic used its Mythos model to move Claude onto AMD chips in three days, while Google's AlphaProof verified Signal's crypto core in four weeks.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Google DeepMind’s AlphaProof Code verifies Signal’s cryptographic core in four weeks
Google DeepMind unveiled AlphaProof Code, an AI system that mathematically proves software meets its security specifications. Researchers demonstrated it on Signal’s encryption protocol, completing in four weeks work that typically takes years. Earlier verification efforts stopped at the theoretical design stage, according to a DeepMind blog post. AlphaProof Code verified the actual cryptographic core code running on users’ devices. After experts defined formal security requirements, AI agents built on Gemini Carbon wrote over six million lines in the Lean proof language, flagging roughly one percent of the codebase for human review. Two deliberately planted bugs were caught immediately. Signal CTO Ehren Kret says the approach strengthens trust in data privacy as AI threats grow. DeepMind plans to extend the method to power grids and medical devices.
Topics: Google DeepMindEncryption
Microsoft’s Lazarus technique hides backdoors in fewer than 10 dormant neurons
Microsoft researchers demonstrated a technique called Lazarus that turns fewer than 10 rarely activated dormant neurons in a large language model into a hidden backdoor, making a normal model behave as expected until a specific trigger appears. The method, described by principal researcher Dimitris Papailiopoulos, requires no retraining. A classifier trained on roughly 50 examples can distinguish triggered inputs; a dormant neuron is then reprogrammed to recognise the trigger and force a chosen token into the output. In the demonstration, an agent waits until the date changes to 2027 before downloading and running attacker-supplied code. The attack needs only inference-level computing plus write access to the model checkpoint. Modified models can still pass standard safety evaluations, and internally developed company models are as vulnerable as public open-weight ones. Microsoft is developing safeguards for its Foundry platform.
Microsoft finds AI agents can be backdoored by editing under 10 neurons →
Topics: Microsoft
Anthropic ports Claude to AMD GPUs in three days using its own Mythos model
Anthropic says it ported its Claude model to AMD GPUs in three days with almost no human intervention, using its own frontier model, Mythos, which ported or wrote nearly 100 percent of the kernels itself. The figures refer to prompt processing of an Anthropic model on AMD MI355X GPUs. Bringing up a new chip platform traditionally kept dozens of engineers busy for six to twelve months; with Mythos, Anthropic says it reaches usable performance in about three days. Co-founder and Chief Compute Officer Tom Brown said Anthropic made an unconventional bet on a multi-chip strategy early. Anthropic now trains and runs Claude on AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. The figures come from an investor presentation and have not been independently verified.
Google launches Gemini Decision 1 for fast closed-ended decisions
Google DeepMind is introducing the Gemini Decisions API, built for fast, closed-ended decisions such as yes/no questions, multiple choice and scale-based ratings. The underlying model, Gemini Decision 1, processes up to 1,000 independent questions in a single pass across text, images, audio, video and PDFs. It scores options directly and returns probabilities with confidence values in 150 to 250 milliseconds. Google says it already uses the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch and safety checks in Google Search. In one example, Gemini Decision 1 evaluates a photo of a kitchen counter in a single pass, combining stain detection at 0.95 confidence, a cleaning action at 98 percent and a severity rating of 1.19 out of 2. The API costs $0.10 per million input tokens with no output-token charge, about five times cheaper than Gemini 3.5 Flash-Lite.
Topics: Google GeminiGoogle DeepMind
DeepL’s Aure voice model carries tone and speaking pace into translations
DeepL has introduced a new voice model, Aure, for DeepL Voice that carries tone, emotion, emphasis and speaking pace from the original speaker into the target language. Previously, DeepL Voice generated translated speech from a static voice sample captured at the start of a conversation, so it could not pick up on mood shifts or sarcasm. Aure continuously analyzes both the translated text and the original audio as the conversation unfolds. During training, DeepL separates voice identity, speech melody and content so each component can be transferred independently. Reinforcement learning keeps accents out of the output. Aure runs automatically in DeepL Voice for Online Meetings, In-Person Conversations and the DeepL API. At launch, it supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.