The Week in AI: January 19-25, 2026 (Week 4)

This week’s news: January 19-25, 2026 (Week 4)

The week was dominated by the World Economic Forum in Davos, where AI company leaders clashed over chips for China and ads in chatbots. Anthropic published a new “constitution” for its Claude model, and OpenAI began estimating users’ ages in ChatGPT. Apple is reportedly preparing to turn Siri into a full-fledged chatbot built on Gemini. Investors kept pouring money mainly into companies that speed up running existing models. And the controversy over nonconsensual sexual images generated by the Grok chatbot continued.

Davos: AI Takes Center Stage

Amodei: Selling Chips to China Is Like Selling Nukes to North Korea

Anthropic CEO Dario Amodei sharply criticized the US government’s decision to allow sales of Nvidia H200 chips to Chinese customers in an interview with Bloomberg on January 20. He called it “crazy” and compared it to selling nuclear weapons to North Korea (Bloomberg).

The remark drew attention partly because Nvidia is a major investor in Anthropic and announced an investment of up to 10 billion dollars just a few months earlier (TechCrunch).

The Dispute over Ads in ChatGPT

OpenAI CFO Sarah Friar defended the plan to introduce ads into the free version of ChatGPT in Davos on January 21. According to her, the company needs a “strong business model” to be able to offer AI to people who do not pay for it (Bloomberg).

Google DeepMind CEO Demis Hassabis told Axios on January 22 that he was surprised how early OpenAI turned to ads. He said Google has no plans for ads in its chatbot for now (TechCrunch).

OpenAI’s First Device Coming This Year

OpenAI Chief Global Affairs Officer Chris Lehane said in Davos on January 21 that the company wants to unveil its first hardware device in the second half of 2026. According to leaks, it could be a pair of earbuds with a custom processor for on-device AI (TechCrunch).

Models and Research

Anthropic Rewrote Claude’s “Constitution”

On January 21, Anthropic published a new version of the document it calls a constitution. It is the text used to train the behavior of the Claude model: instead of a list of rules, it explains why the model should behave in a certain way (TechCrunch).

The document runs to roughly 80 pages and has four main parts: safety, ethics, compliance with the company’s guidelines, and genuine helpfulness. One sentence attracted attention: “Claude’s moral status is deeply uncertain,” an open admission that the company does not know whether the model can have some form of consciousness.

New Open Models from China

January 19: Z.ai (formerly Zhipu) released the lightweight open-weight model GLM-4.7-Flash, which can also run on your own computer (Gigazine).
January 22: Alibaba released the Qwen3-TTS family of voice models under the Apache 2.0 license. They convert text to speech in ten languages and can clone a voice from a short recording (Qwen GitHub).

Microsoft Unveiled the Rho-alpha Robotics Model

On January 21, Microsoft Research showcased Rho-alpha, a model that turns natural language instructions into movements of robotic hands. Besides images and text, it also works with tactile sensors. For now, it is available only through an early access program (TechRadar).

Research: Office Agents Still Falling Short

January 22: Mercor published the APEX-Agents benchmark, which measures how well AI agents handle real office tasks. The top performer, Gemini 3 Flash, succeeded on the first try in only 24% of cases, and GPT-5.2 in 23% (TechCrunch).
January 21: The GPTZero tool examined 4,841 papers accepted at the prestigious NeurIPS conference and found 100 fabricated citations in 51 papers. This is a typical example of so-called hallucinations, where AI invents nonexistent sources (TechCrunch).
January 25: According to The Guardian, the GPT-5.2 model cites Grokipedia, an AI-generated encyclopedia from xAI, in its answers (TechCrunch).

Products and Services

ChatGPT Began Estimating Users’ Ages

On January 20, OpenAI launched an age prediction feature in ChatGPT. Based on signals such as account age, typical usage times, or stated age, the system estimates whether the user is a minor (TechCrunch).

Accounts flagged as under 18 automatically get sensitive content filters. An adult who is misclassified can verify their age with a selfie through the partner Persona.

Apple Is Preparing a Siri Chatbot and Possibly an AI Pin

According to Mark Gurman of Bloomberg, Apple wants to turn Siri into a full-fledged chatbot, code-named “Campos,” in iOS 27 (report from January 21) (TechCrunch).

On January 25, Gurman added that Apple wants to unveil the new Siri built on Google Gemini models as early as the second half of February (TechCrunch).

The Information also reported on January 21 that Apple is developing a wearable AI pin the size of an AirTag with two cameras and three microphones, which could launch in 2027 (TechCrunch).

Google Connected AI Mode with Gmail and Photos

On January 22, Google expanded its Personal Intelligence feature to the AI Mode search experience. Google AI Pro and Ultra subscribers in the US can opt in to have answers draw on their Gmail and Google Photos as well (TechCrunch).

More Product News

January 21: YouTube CEO Neal Mohan announced that creators will be able to make Shorts this year using their own AI-generated digital likeness (TechCrunch).
January 22: Google now offers free SAT practice tests (the US college admissions exam) in the Gemini app (TechCrunch).
January 20: X published the code of its recommendation algorithm on GitHub, which the company says is built on a model derived from Grok (TechCrunch).

Business and Investment

The Inference Boom

Most of the money this week went to companies focused on inference, that is, running already trained models quickly and cheaply. Unlike training, it is an everyday cost that grows with every user query.

January 21: According to sources, the SGLang project spun out as the company RadixArk at a valuation of about 400 million dollars, in a round led by Accel (TechCrunch).
January 22: Inferact, the company founded by the creators of the popular open source tool vLLM, raised 150 million dollars at an 800 million valuation from Andreessen Horowitz and Lightspeed (TechCrunch).
January 23: Baseten raised 300 million dollars at a 5 billion valuation, in a round led by IVP and CapitalG with participation from Nvidia (Baseten).
January 22: Neurophos raised 110 million dollars to develop optical processors for inference, with the investment led by Gates Frontier (TechCrunch).

Other Major Rounds and Acquisitions

January 20: The startup Humans&, founded by alumni of Anthropic, xAI, and Google, raised 480 million dollars in a seed round at a 4.48 billion valuation. Investors include Nvidia and Jeff Bezos (TechCrunch).
January 21: OpenEvidence, an AI assistant for doctors, raised 250 million dollars at a 12 billion valuation, double its valuation in October (TechCrunch).
January 22: LiveKit, the technology behind ChatGPT‘s voice mode, raised 100 million dollars at a 1 billion valuation (TechCrunch).
January 22: Through a licensing deal, Google DeepMind hired Alan Cowen, CEO of the voice startup Hume AI, along with several engineers (TechCrunch).
January 23: Legal AI company Harvey acquired the startup Hexus (TechCrunch).
January 23: The leadership of AMI Labs, Yann LeCun‘s new company focused on “world models,” came to light. LeCun is chairman and Alex LeBrun is CEO (TechCrunch).

Data Centers and Energy

On January 20, OpenAI introduced its Stargate Community plan. It promises to pay for its own energy and new power sources at its data centers so that their operation does not raise electricity prices for local residents (The Register).

Elon Musk announced on January 20 that Tesla‘s restarted Dojo3 chip project will be used for space-based AI compute (TechCrunch).

Regulation and Law

Singapore Is the First to Set Rules for AI Agents

On January 22 in Davos, Singapore’s IMDA presented a governance framework for agentic AI, meaning systems that independently plan and carry out tasks on behalf of users. The voluntary framework addresses limits on agent autonomy, human accountability, and transparency toward users (MDDI Singapore).

Other Regulatory Steps

January 21: The European data protection authorities EDPB and EDPS issued a joint opinion on the proposal to simplify the rules of the AI Act as part of the Digital Omnibus package (EDPB).
January 21: The Canadian province of Ontario published joint principles from its privacy and human rights commissioners for the responsible use of AI (IPC Ontario).

Safety and Incidents

Grok under Pressure from US Attorneys General

On January 23, a coalition of 35 US state attorneys general called on xAI to stop the Grok chatbot from creating sexualized images of real people without their consent, including children. They also demand the removal of content already created and blocking of users who generated it (letter from the attorneys general).

At the same time, Grok returned to Asia: the Philippines announced on January 21 that it would lift its ban, and Malaysia lifted its ban on January 23 after X added safety measures (Bloomberg, Bloomberg).

Meta Paused AI Characters for Teens

On January 23, Meta paused teen access to AI characters in its apps worldwide until a new version with parental controls is ready. The move came a few days before a trial in New Mexico (TechCrunch).