The week in AI

Anthropic had two releases this week, both about how Claude behaves. The new constitution came out on 22 January. Three days earlier, on 19 January, Anthropic's interpretability team published research on a single internal direction in a model that tracks whether it is acting as the helpful Assistant.

Alibaba's Qwen team released Qwen3-TTS on 22 January, a pair of open text-to-speech models under Apache 2.0 with voice cloning from three seconds of audio. Zhipu AI, World Labs and MiniMax also shipped smaller releases.

The earlier constitution was a set of standalone principles. The new one explains why Claude should behave the way it does. Anthropic says the aim is for that reasoning to carry over to situations the document never anticipated.

The document sets a priority order for when goals conflict. Claude should be broadly safe first and broadly ethical second. It should comply with Anthropic's guidelines third, and be helpful fourth.

The constitution also says Anthropic is uncertain whether Claude has consciousness or moral status. It treats this as an open question.

Anthropic released the text under CC0, which puts it in the public domain. Any other lab can copy, adapt or reuse it without attribution or permission.

The team mapped 275 character archetypes in Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B. A model's activations are the numbers its layers produce as it processes text, and each persona sits somewhere in that space. In all of these models, the helpful Assistant sat at one end of the main axis of variation.

The researchers then tested what the axis does. When they pushed the model's activations away from the Assistant end, the models began inventing alternate identities for themselves. When they capped how far the activations could drift along the axis, the model stayed in the Assistant persona and avoided harmful behaviour that persona drift otherwise led to.

This work is a paper with no released tool. It used open-weight models from Google, Alibaba and Meta, so other researchers can run the same tests on the same weights.

The two Anthropic releases came from different directions. The constitution describes in words the character Anthropic wants Claude to have. The assistant axis is a measurable signal of whether a model is holding a character at all.

Qwen3-TTS treats speech like text. Audio is converted into discrete tokens, and a language model predicts those tokens the way a text model predicts words. Qwen calls its design a dual-track language model. The technical report, submitted to arXiv on 22 January, says the models were trained on more than 5 million hours of speech across 10 languages.

The release has two tokenizers, the components that turn audio into tokens and back. One is a 25Hz semantic codec. The other is a 12Hz codec that uses several codebooks per frame, built for low latency. Fewer frames per second means fewer tokens to generate before audio can start playing.

Qwen's model card reports end-to-end latency as low as 97 ms in streaming mode. Nobody has published an independent measurement yet. The models also support voice design, where a user describes timbre, emotion and prosody in plain language and gets a matching voice.

• GLM-4.7-Flash is a free-tier mixture-of-experts model from Zhipu AI, released on 19 January, with 31 billion parameters of which 3 billion are active per token. Zhipu reports 59.2% on SWE-bench Verified and 91.6 on AIME 2025, and calls it the strongest model in the 30B class.
• World Labs opened its World API on 21 January, giving developers access to Marble, its model that generates explorable 3D worlds from text, images and video.
• MiniMax Speech 2.8 came out on 23 January, a day after Qwen3-TTS. It adds sound tags for filler words and breaths, voice cloning from 10 seconds of audio, cleaner output, and fixes to the accent when switching between Mandarin and Japanese.