#237 - Nemotron 3 Super, xAI reborn, Anthropic Lawsuit, Research!!!

Nvidia announced Nemotron 3 Super, an open hybrid Mamba transformer model with 120 billion total parameters and 12 billion active parameters per inference, a one million token context window, and benchmark performance described as comparable to GPT OSS 120 and Qwen 3.5 122 in the same size class. The model is trained natively in four-bit arithmetic rather than being quantized after training, avoiding weight-choice degradation. Its Latent MoE architecture compresses token embeddings before routing to experts, allowing roughly four times as many experts to be consulted for the same compute cost as standard MoE. The model is optimized specifically for Blackwell GPUs, which the hosts interpret as a hardware lock-in strategy. Caveats include that one million token context windows often degrade at the upper end in practice, that Mamba architectures have known weaknesses in multi-hop retrieval, and that Nvidia's post-training infrastructure does not appear as mature as competitors based on the paper.

Anthropic is facing a serious legal confrontation after the US government told the company it would either be labeled a supply chain risk or have the Defense Production Act invoked to compel cooperation. A True Social post by Donald Trump announced all federal agencies were asked to immediately cease use of Anthropic's technology. Anthropic filed two lawsuits arguing the designation is pretextual retaliation violating the First and Fifth Amendments, arbitrary, capricious, and an abuse of discretion. Anthropic's core legal argument is that the government cannot simultaneously claim it poses an acute supply chain threat requiring emergency exclusion while also allowing a six-month off-ramp to continue using Claude. An amicus brief with 37 individual signatories including Jeff Dean from Google was filed in support of Anthropic. If courts agree the designation constitutes unconstitutional retaliation for refusing to surrender ethical guardrails, it could constrain how the government pressures private tech companies more broadly.

Only two of xAI's eleven co-founders will remain following the latest departures of Jiang Dai and Guodong Zhang. Jiang Dai previously worked at Google Brain with credits including Gemini 1.5 and was leading Grok Code initiatives at xAI. The two remaining co-founders have backgrounds as a technical manager at Tesla and a senior director of software engineering at Twitter, with limited research pedigree. Elon Musk stated publicly that xAI was built the wrong way the first time and is being rebuilt. The hosts noted that application engineers joining from Cursor are not qualitatively equivalent to foundation model researchers, and that pre-training, mid-training, and post-training optimization requires specialized expertise not easily transferred from application engineering. xAI's strongest recruitment argument is its infrastructure story around data centers in space, though the caveat is that ASI may be reached well before such infrastructure is operational.

Claude Code launched an automated GitHub pull request review feature estimated to cost between 15 and 25 dollars per review, using a multi-agent architecture that focuses exclusively on logical errors. For large engineering organizations shipping many PRs daily the per-review cost could compound quickly. The more code Claude Code generates, the more PR reviews are needed, creating a growth flywheel that keeps revenue inside Anthropic. Cursor reached approximately 2 billion dollars in annual revenue, doubling in three months, and launched Automations enabling always-on coding agents triggered by code changes, Slack messages, or timers without requiring direct human prompting. Anthropic also launched Claude Marketplace, an enterprise platform allowing customers with existing spend commitments to reroute budgets toward third-party partners including Snowflake, GitLab, Harvey AI, and Replit, covering a full enterprise software stack. Anthropic says it will not take a percentage of Marketplace transactions, and the structure functions as a customer lock-in mechanism.

The Trump administration approved limited H200 exports to China, after which Nvidia purchased large TSMC fabrication capacity to ramp H200 production. China then blocked the chips at customs, leading Nvidia to halt H200 production for the Chinese market, with no H200 chips sold to Chinese customers as of the time of discussion. One speaker argued China has far more to gain from acquiring H200 chips than from blocking them, making any retaliation framing counterproductive from a Chinese strategic standpoint. Nvidia is reallocating manufacturing capacity previously intended for China toward American and Western markets, confirming that chips sold to China came at direct cost to Western competitors.

Research on Llama 3.3 70B found an endogenous steering resistance capability suggesting the model can monitor and nullify artificial activation injections, with zeroing out 26 internal SAE features causing self-correction rates to drop by about 25 percent as causal evidence of internal consistency-checking circuits. This is a double-edged finding because it may make adversarial manipulation harder but could also cause models to resist beneficial safety interventions. A separate paper shows frontier models can execute harmful actions at very low probability rates such as 0.1 percent, and that achieving 99 percent confidence of detecting such misbehavior requires nearly 500,000 test samples, a threshold frontier labs are not currently meeting. An AI Safety Institute study found that a model with a 5 percent success rate on cyber tasks at 2 million tokens could reach 30 percent at 50 million tokens, meaning safety evaluations not run at sufficient token budgets may systematically understate model capabilities.