Anthropic says it moved Claude onto AMD GPUs in three days—down from the months and dozens of engineers typically required to bring up a new chip platform. The company says its Mythos frontier model ported or wrote nearly 100% of the required kernels with almost no human intervention.
Anthropic says it now trains and runs Claude across all four major chip families, challenging Nvidia’s CUDA software advantage. The company treats chips from different vendors and locations as one compute pool, allowing it to target price-performance and match hardware to training or inference. The figures come from Anthropic’s investor presentation and have not been independently verified. Nvidia currently earns margins above 70% on its AI hardware.
Why it matters: If Anthropic’s figures hold, Nvidia loses part of the switching-cost advantage behind its more-than-70% hardware margins, while AMD gains a demonstrated path into Claude workloads that previously took months to port.
Australia's government, led by Prime Minister Anthony Albanese, has announced legislation that would impose mandatory safety rules on frontier AI companies operating in the country. The rules target the most advanced models with the highest risk profiles and would require independent audits and ongoing safety checks—the first such rules Australia has applied to those models.
A new oversight body would work alongside the Australian AI Safety Institute to enforce compliance and hand out steep penalties. Frontier-model builders would also have to publish safety frameworks and immediately report security incidents, a requirement partly aimed at catching early warning signs from AI agents that go off the rails.
Why it matters: For frontier-model companies, Australia is no longer waiting for international negotiations to set the rules: operations there would face a domestic oversight body that can penalize violations.
Google launched the Gemini Decisions API, letting developers evaluate up to 1,000 yes/no, multiple-choice, or rubric questions in a single pass. The Gemini Decision 1 model handles text, images, audio, video, and PDFs, returning probabilities and confidence scores in 150–250 milliseconds. It costs $0.10 per million input tokens, with no output-token charge.
Unlike a general model that generates answers token by token, the endpoint scores candidate options directly and bills shared input once per request. Google says it is nearly five times cheaper across a shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. The API is available now in Google AI Studio, with integrations rolling out across OpenRouter, LangChain, Vercel, and others.
Why it matters: Google is turning bounded decisions into a batch-processing primitive rather than another chat endpoint. At $0.10 per million input tokens, the economics could pull routing and policy workloads away from Gemini 3.5 Flash-Lite—if Google's benchmark claims hold outside its shared test suite.
Nvidia’s board could decide as soon as Wednesday whether Jensen Huang remains CEO, as shares fell about 8% in after-hours trading. The Wall Street Journal reports that directors are preparing to determine Huang’s future; he has led the chipmaker for 33 years.
The reports involve Huang’s admitted affair with a woman purported by U.S. intelligence to be a Chinese spy and a second affair with an executive who reports to him. Nvidia has no obvious successor or succession plan. Huang says he delegates very little and has about 60 direct reports. Bloomberg reports the company is working with an outside firm to identify potential replacements.
Why it matters: For Nvidia investors, this is a succession test as much as a scandal: the company may need to replace a 33-year CEO without an obvious successor, while its unusually centralized structure includes about 60 direct reports to Huang.
AWS introduced a standalone Web Search API for Amazon Bedrock AgentCore, letting developers retrieve ranked, current web results with a single API call. The operation returns text extracts and, when available, titles, URLs, and publication times from an Amazon-operated web index.
The API targets custom agent loops, LangChain and LlamaIndex applications, dashboards, fact-checking pipelines, and monitoring systems that do not need AgentCore Gateway. Previously, developers using AgentCore Web Search had to create a gateway and search target, then call the tool through an MCP client.
Requests use standard AWS IAM credentials and SigV4 authentication, with no API key or resource provisioning. The Gateway remains AWS’s recommended route for multi-tool agent workloads.
Why it matters: AWS is turning web grounding into a lower-level building block for existing applications, reducing the infrastructure commitment required before a team can add current information. That gives developers a direct alternative to AgentCore Gateway while keeping both paths on the same search infrastructure used by Alexa+ and Kiro.
AWS added three checks to PyTorch DDP training loops on SageMaker HyperPod, targeting silent data corruption (SDC): a GPU returns a plausible but wrong number without raising an error. Non-finite gradient checks catch NaN or Inf; ring replay audits catch finite wrong values; weight fingerprints catch divergence after optimizer updates.
SDC can poison every replica while loss looks normal. During Meta’s 54-day Llama 3 405B pre-training run on up to 16,000 H100 GPUs, six of 419 unexpected interruptions were attributed to SDC.
AWS tested the approach with software-injected faults on four p5.48xlarge instances—32 H100 GPUs—and published the code. When a check fires, operators map the rank to a node, cordon it, replace the hardware, and resume from a pre-suspect checkpoint.
Why it matters: Meta’s 16,000-GPU training run shows why clean hardware telemetry is not enough: HyperPod operators now need application-level evidence that the computation itself is correct before trusting hours of accelerator time.
MIT researchers developed a training method that lets climate emulators learn from one full Earth system model run instead of the usual six to eight. These AI surrogates reproduce climate projections in seconds, allowing scientists to test thousands of emissions scenarios rather than wait months for each supercomputer-based simulation.
The method generates synthetic emissions paths, including abrupt changes such as sharply higher CO2 in 2050 or rapidly reduced methane in 2037, and iteratively tunes them against model error. The resulting emulator accurately predicts unfamiliar scenarios, including local pollution controls that traditional emulators failed to capture. The synthetic data need not look realistic; their purpose is to expose the climate system’s behavior more efficiently.
Why it matters: Noelle Selin’s MIT team is addressing the weak point in fast climate analysis: models that run quickly but fail outside their training scenarios. A one-run training pipeline could give policymakers more dependable projections for pollutant controls, power-grid planning, and public-health decisions.