Anthropic and the Pentagon: Tensions Over AI Sonnet 4.6

An Episode Rich in AI News

The 235th episode of the LWiAI podcast was recorded on February 27, 2026, providing an overview of recent developments in the field of artificial intelligence. Hosted by Andrey Kurenkov and Jeremie Harris, this episode focuses on the major news from the past week. Andrey Kurenkov took the opportunity to announce that his startup, Astrocade, is expanding and actively recruiting. Astrocade, which aims to become the "YouTube of gaming," is looking for talent in various fields, including engineering, marketing, and product management, and invites interested individuals to contact them directly.

Listeners are encouraged to share their questions and comments via the provided email addresses, fostering ongoing interaction with the podcast community.

Highlighted Technological Advances

In this episode, several updates on models and tools are discussed. Anthropic's Sonnet 4.6 model is highlighted for its ability to handle a context of one million tokens, showing strong results on the ARC-AGI-2 test. Google also presented its Gemini 3.1 Pro model, which has made significant advancements on the same test while offering multimodal demonstrations. xAI, Elon Musk's company, launched the beta version of Grok 4.2, a multi-agent debate tool. Anthropic unveiled Claude Code "Remote Control," and Perplexity introduced "Computer," a multi-agent coordinator.

Strategic Moves in the Business Sector

On the business front, a report indicates that Meta is in discussions for a chip deal with AMD that could reach $100 billion. This agreement would include stock incentives and could transform Meta's quest for a "personal superintelligence." Additionally, MatX, a rival startup to Nvidia, has raised $500 million to develop specialized chips for transformers, expected by 2027. World Labs has also raised $1 billion, including $200 million from Autodesk, to integrate global models into 3D environments. Another startup secured $100 million to develop AI capable of simulating and predicting human behavior.

Infrastructure and Geopolitical Issues

In the realm of infrastructure, the Stargate data centers intended for OpenAI are experiencing delays due to disagreements between OpenAI, Oracle, and SoftBank over control of these facilities. Meanwhile, China has announced its intention to increase its production of advanced chips, aiming to produce 100,000 wafers per month of 7nm and 5nm chips, with a goal of 500,000 per month by 2030, despite challenges in yield and tooling.

Research and Security: Crucial Advances

Discussions on research highlighted the effectiveness of masked updates in adaptive optimizers and the use of "deep reflection tokens" to measure the reasoning effort of language models. The attractor behaviors of models, mechanistic interpretability of counting and line breaks, as well as methods for evaluating task difficulty relative to human time horizons were also addressed. Tensions surrounding the contract between Anthropic and the Pentagon were discussed, along with the distillation attacks identified by Anthropic and OpenAI's efforts to counter the malicious use of AI.

Thanks to Sponsors

The episode was supported by several sponsors, including Box, ODSC AI, and Factor, each offering special promotions to podcast listeners.

Detailed Timestamps

The episode is structured with precise timestamps, allowing listeners to easily navigate between different sections, ranging from light discussions to in-depth analyses of technological and geopolitical news.

Politics and Security: Critical Issues

Anthropic's CEO, Amodei, stated that the Pentagon's threats do not change the company's position on AI. Additionally, Musk's xAI has reached an agreement with the Pentagon to use Grok in classified systems. Discussions also covered the detection and prevention of distillation attacks, as well as OpenAI's efforts to disrupt the malicious use of AI, as detailed in a recent report.