Politico: Anthropic has been quietly shopping its own AI bill to Congress

Politico obtained a confidential draft bill that Anthropic is pitching to lawmakers. It would let the federal government block a model’s release only on a “foreseeable, imminent risk” of $1B+ in damages or 50+ people killed or seriously injured, and only in three scenarios (bioweapons, a model escaping developer control, or a cyberattack); enforcement runs solely through the Commerce Department, which must first get a federal court order. In exchange for transparency duties (published safety frameworks, transparency reports, whistleblower protections, incident reporting), the draft preempts state laws on disclosure, third-party evaluation, and incident reporting — meaning most of California’s SB 53 and similar state rules would stop applying once a bill like this passes. Anthropic confirmed to Politico that the draft reflects the Advanced AI Framework it has promoted since June; critics, Zvi Mowshowitz’s “unacceptable and troubling” read among them, note the enforcement terms are weaker in key ways than the bipartisan FRONTIER Act already sitting in the House (see also Zvi’s analysis and the leaked PDF).

Anthropic’s own workforce data: junior technical hiring fell from 27% to 9%

The same day the bill leaked, Anthropic published “How employment is changing at Anthropic,” built on internal hiring records, calendar data, and Claude usage telemetry. Junior staff in technical departments fell from 21% to 14% since early 2025, and the junior share of new technical hires dropped from 27% to 9%; by August, agents inside the company were logging six work-hours per human work-hour, sessions spent managing Claude rather than doing the work rose from 10% of interactive use in fall 2025 to 22% in spring 2026, and median meeting time was up 20% year over year. The report’s own reading: human review and sign-off are the new bottleneck, with non-technical staff needing over half a work week just to read Claude’s output, and it concedes an AI-native lab doesn’t generalize to the wider economy. Still, in the debate over whether AI eats junior jobs first, 27%-to-9% is a more direct data point than any public statement (see also TechCrunch).

Meta deletes the word that alarmed publishers: “train”

Nieman Lab reported on October 15 that Meta’s updated Commercial Terms, effective October 30, would have businesses using its ads products “authorise” Meta’s crawlers to collect content from their own websites and use it to “provide, develop, improve, research, train and maintain” Meta products; Cornell law professor James Grimmelmann read that as permission to train AI on publishers’ sites. Within two days Meta removed “train” and narrowed the clause to “advertising and business products,” with spokesperson Andy Stone saying on X that actual crawler behavior is not changing (see also TechCrunch, the terms preview). The training-data fight is moving into boilerplate: one word in terms businesses must accept is something Meta could later point to as contractual permission to train (whether that would hold up as valid consent is an open legal question), and it got caught only because someone read the document line by line.

One year of Tinker, with a cost ledger for fine-tuning

Thinking Machines’ October 15 anniversary post for its fine-tuning platform Tinker says it grew from 14 text-only models at launch to more than 30 (now including image and audio), and gives customer numbers: Bridgewater’s text-to-SQL model beats Fable 5 and GPT-5.6 Sol on accuracy while running at 15% of their cost, and Glean’s search agent got twice as fast while using a quarter fewer tokens. These are vendor-reported, but they are rare concrete reference points for the tune-a-small-model-versus-call-a-frontier-API decision; worth checking against your own workload before drawing conclusions.

Research radar

Behavioral Compositionality of Sparse Autoencoder Representations

Interpretability work often assumes SAE feature effects on behavior add linearly; this paper’s SAR metric (SAE Additivity Residuals) measures how much behavioral variation that assumption leaves unexplained beyond a chosen interaction order. Learned SAE coordinates turn out far more additive than random rotations of the same subspaces, but not fully: adding pairwise interactions lifted rank fidelity from 0.59 to 0.78 in steering experiments. If you intervene on SAE features, single-feature effects are not enough; joint explanations need interaction terms.

Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL

In multi-reward RL training (say, separate rewards for tool-call correctness and output format), each reward’s advantage is usually spread evenly across the whole response, diluting the signal onto tokens that reward’s verifier never examines. MRTPO replays the response with and without each reward’s verifier information, uses per-token Jensen-Shannon distance to measure which tokens that reward actually concerns, and reassigns advantage accordingly, with no auxiliary model or labels. It converges faster and scores better on text-to-SQL, multi-constraint instruction following, and tool calling; directly applicable if you train with RLVR or multiple verifiers.

Today in one line: the most reliable statement of an AI company’s governance position is the bill text it hands to legislators, not its safety blog: the bill says who can enforce, and at what threshold.