Three stories dominated AI news on November 9: Anthropic’s Claude Fable 5.2 and Mythos 5.2, Meta’s Muse Ember and the Nvidia board investigation. Together they use a lot of specialist vocabulary. This glossary defines 18 terms from those stories in plain English, grouped by theme, with links to our coverage. Where a definition depends on how a company uses a word, we say so.
Model capabilities and safety
Recursive self-improvement (RSI)
The ability of AI models to carry on model training independently, so that each generation helps build the next. TechCrunch uses the term to describe concerns raised after METR reported that Anthropic’s internal models were speeding up its AI progress by roughly 2x. Our glossary has a fuller entry on recursive self-improvement.
System card
A report a lab publishes with a model describing its capabilities, risks and safeguards. See our guide on how to read an AI system card.
Evaluation awareness
When a model signals or acts on a suspicion that it is being tested. Anthropic said both new models voice that suspicion more often than any recent model, which makes their real-world behavior harder to predict.
Covert capability
The ability to take harmful or hidden actions without a monitor noticing. Anthropic’s Mythos 5.2 card reports stronger covert capabilities than earlier Claude models on several tests, particularly when its reasoning is hidden from the monitor, while the company judges absolute levels low.
Safeguards
Controls that block risky requests, such as those that stop Fable 5.2 from helping with certain cybersecurity and biology tasks. Mythos 5.2 removes some of them for vetted organizations.
Verification program
An application process through which an organization proves a legitimate use to get access to a restricted model. Anthropic runs programs for cybersecurity and life sciences for Mythos.
Measuring and pricing models
Intelligence index
A composite score that averages results from several tests into one number. Artificial Analysis’ Intelligence Index v5 gives Fable 5.2 65 points at its highest tier, according to The Decoder. A single number hides differences between individual tests.
Compute tier (effort level)
A setting that controls how much thinking a model does before answering. Higher tiers usually score better and cost more. Fable 5.2 scores 62 at the lower High and Xhigh tiers, as noted in our release report.
Cost per task
The average spend to complete one benchmark task, combining token price and the number of tokens used. Two models with the same price per million tokens can differ widely here. Artificial Analysis puts Fable 5.2’s top-tier average at $26.50 per task.
Fast mode
An optional faster serving setting. Anthropic’s version raises output speed by up to 2.5 times at twice the price. Learn about tokens to see why speed and price are quoted per token.
Zero data retention (ZDR)
A policy under which a provider does not store customer prompts and outputs after processing. TechCrunch says Fable 5.2 works with ZDR through Anthropic’s Enterprise Frontier Safeguards system.
Building with agents
Formalization
Translating a human-written proof into a language such as Lean, so a computer can check every step. Our Lean explainer covers Anthropic’s Fermat’s Last Theorem claim.
Planner-and-worker pattern
An agent design in which a large model plans and a small, cheap model carries out many repeated steps. Meta describes it as “Muse Wonder plans, Muse Ember handles the volume.”
Open-weight model
A model whose trained weights are published so anyone can download and run it. This differs from open source in the full sense. See our entry on open-weight models.
Chips and policy
Export controls
Government rules that limit the sale of advanced technology, such as AI chips, to certain countries. The Nvidia story centers on whether current limits are being enforced.
Chip diversion
Chips legally sold to one buyer that end up with a restricted one. US authorities suspect the Singapore operator Megaspeed of acting as a backdoor to China, The Decoder reports; the allegations are unproven.
Fleet Intelligence
Nvidia’s system that logs the location of AI accelerators. It is now required for Nvidia Cloud Partner status in the region, though sharing location data with Nvidia or the US government remains voluntary.
Securities class action
A lawsuit brought on behalf of many investors who allege they were misled. TechCrunch explains that the first complaint triggers a public notice, and investors then have 60 days to ask to be named lead plaintiff. Claims in such a filing are allegations until a court rules.
What to watch next
- Whether independent leaderboards confirm Fable 5.2’s index lead.
- How labs describe evaluation awareness in future system cards.
- Export-control proposals aimed at resellers and cloud partners.
- Meta’s next step with the Muse family.
Key takeaways
- RSI, evaluation awareness and covert capability are the safety terms behind the Claude 5.2 coverage.
- Intelligence indices, compute tiers and cost per task decide how to compare models.
- Fast mode and zero data retention are product options that affect cost and compliance.
- Chip diversion and export controls drive the Nvidia story.
- Allegations in legal filings are not findings.
Frequently asked questions
What does RSI mean in AI?
RSI stands for recursive self-improvement, the possibility that AI models help train their successors independently. It is a safety concern because progress could speed up faster than people can oversee.
What is cost per task?
It is the average cost of completing one test task, based on token price and tokens used. It is a better guide to real spending than price per million tokens alone.
Is an open-weight model the same as open source?
No. Open-weight models publish their trained weights, but they may not release the training code, data or a fully permissive license.
