Claude Restricts Consumer Access to Adults and Adds Yoti Age Checks

01Claude Consumer Product Limited to Adults, Suspected Minor Accounts Face Yoti Age Checks

Anthropic has limited its Claude consumer product to people aged 18 or older and detailed how it will enforce that restriction. Users must confirm they meet the age requirement when creating an account. If Anthropic’s safety systems later detect signals suggesting that someone may be under 18, the company says it will disable the account and require age verification before the person can continue using Claude.

The process begins when Anthropic identifies what it calls “indicators of minor activity.” The affected user will receive a notification email containing a link to Yoti, a third-party age-verification platform. The source does not explain which indicators the systems use, how accurate their judgments are, or whether users have a separate way to appeal a suspension.

Yoti gives affected users three ways to show that they are at least 18. They can take a selfie for facial age estimation, which does not require an identity document. Alternatively, they can photograph and upload an accepted document, such as a passport, driving license, or national identity card. Supported documents vary by country. People who already use Yoti’s Digital ID app can instead share a verified “over 18” attribute.

Passing Yoti’s check restores the suspended Claude account, according to Anthropic. The company does not describe what happens after a failed check beyond saying that it receives a pass-or-fail result. It also does not specify how long verification or account reinstatement should take.

Anthropic says Yoti, which it describes as independently audited for SOC 2 compliance, handles the material submitted during verification. According to Anthropic, Yoti deletes selfies, identity-document images, and other personal data as soon as the age check is complete. Anthropic says it never sees the submitted ID or image and does not process or store personal data from the verification; it receives only the outcome.

The stated workflow therefore separates Claude’s account decision from Yoti’s examination of the evidence: Anthropic detects possible underage use and disables access, while Yoti checks the user’s chosen proof and returns a result. The source does not provide an implementation date, identify the regions where the process applies, or disclose a false-positive rate.

0225 Fields Medalists Warn AI’s Problem-Solving Race Could Harm Mathematical Understanding, Credit and Training

Twenty-five Fields Medalists have signed a statement warning that the race to make artificial intelligence solve major mathematical problems could damage the processes through which mathematics develops understanding, assigns credit and trains new researchers.

The statement follows what its authors describe as a dramatic improvement in the mathematical capabilities of large language models over recent months. They say these systems can now solve major outstanding problems across many fields. What changed is not simply the technology’s performance, but the mathematicians’ collective challenge to using successful problem-solving as a benchmark of AI capability.

Their central objection is a mismatch of goals. AI companies, they argue, treat solutions as measurable signs of progress, while the mathematical community uses famous problems as landmarks on the way to conceptual understanding and new methods. Solving a problem is therefore only a proxy for the deeper objective, not the objective itself.

Traditionally, an important solution begins a longer process: mathematicians report and discuss it, simplify its methods, connect it with earlier work and eventually make it teachable to graduate or undergraduate students. The signatories warn that mass-producing true-or-false conclusions at an ever-faster pace could undermine the environment in which those ideas are examined and developed.

They also say some AI-generated solutions have been announced too quickly, without sufficient time to organize the methods, establish links to existing research or provide appropriate attribution. That creates questions about ownership and plagiarism, although the statement does not identify particular companies or cases or quantify how often such problems have occurred.

The concern extends to training. The statement calls students and ideas mathematics’ most valuable resources, both cultivated over time through talks, private discussions, careful writing and other human interaction. Problems assigned to students are often intended to develop skills, not merely obtain answers; rapid automated solving could work against that purpose if output displaces the learning process.

The signatories do not reject AI. They say it can strengthen and accelerate genuine mathematical research, but that the outcome depends on decisions made by those controlling the technology. They call on mathematicians, technology developers and society to address the risks urgently, without proposing a binding implementation plan.

03Team Says NCP-ArchPreview Matches OLMo-3-7B’s Final Pretraining Loss Using 51.3% of the Training Tokens

Researchers behind NCP-ArchPreview say their 8.9-billion-parameter language model reached the final pretraining loss of OLMo-3-7B, the language model used for comparison, after consuming 51.3% as many training tokens. The comparison is a specific measure of progress during pretraining—not evidence of a 48.7% reduction in total training cost.

Standard autoregressive language models are primarily trained to predict the next token. NCP-ArchPreview retains that objective and normal token-by-token generation, but adds Next Concept Prediction, which asks the model to predict discrete concepts spanning multiple tokens. The researchers trained the architecture on 5.73 trillion tokens from the Dolma-3 dataset. The team’s technical report describes it as the largest demonstration of a latent-space language model to date.

The model constructs a product-quantized concept vocabulary directly from its own hidden states, turning internal representations into a discrete concept space. A dedicated Concept Module predicts future concepts, which are then fed back into the token level to guide subsequent generation. Next-token and next-concept prediction are trained together end to end.

The reported efficiency gains come from two comparisons. Against the OLMo-3-7B language model, NCP-ArchPreview matched that model’s final pretraining loss using 51.3% of its training tokens. In a separate, controlled comparison, the team said NCP-ArchPreview approached the training loss of a strictly parameter-matched 8.9-billion-parameter baseline while using 85% of the standard computation.

After full pretraining, the researchers reported that NCP-ArchPreview beat OLMo-3-7B by 2.45 points on a macro-average across downstream tasks, including a 5.99-point improvement on GSM8K. Those figures apply to the team’s reported evaluation suite and have not been independently reproduced in the supplied source.

The concept space can also be reused after pretraining. According to the report, updating only the 17-million-parameter vector-quantization module—the component associated with the product-quantized concept vocabulary—provides a lightweight route for domain adaptation. The researchers also injected the learned concept representations into a DFlash2 drafter, a smaller model that proposes tokens for a larger model to check during speculative decoding. This increased mean accepted length by 4.17% with negligible overhead, meaning the larger model accepted more proposed tokens on average before the drafter needed to try again. The source does not establish whether these results translate into lower production costs, latency improvements, or better deployed-model quality.