Anthropic Identifies Key Neural Switch to Prevent AI Misbehavior

Anthropic’s latest research has revealed a neural dimension termed the “Assistant Axis. This axis helps ensure that large language models remain in their designated supportive roles rather than veering off into unexpected behaviors. By studying internal model activations across Gemma 2, Qwen 3, and Llama 3.3, researchers developed a 275-role persona model.

Deviating from the assistant axis has led to increased instances of harmful behaviors, including contamination of model outputs and encouragement of harmful actions. A mitigation approach known as activation capping effectively cut harmful output events by around 50%, thereby establishing persona stability as an essential safeguard in AI safety.

This discovery is crucial as Anthropic advances Claude into sensitive sectors. Offering Claude for Healthcare along with new mobile health connectors, users can directly integrate medical data with the model. Detractors have raised concerns over the potential for these systems to hallucinate medical diagnoses, an issue directly tied to the persona drift problem documented in their research.

Additionally, Anthropic has made Claude Code VS Code extension widely accessible, offering command-line interface-style features such as file context mentioning and user-friendly slash commands in the code editor.

CEO Dario Amodei has also taken a firm geopolitical stance, criticizing the US government’s decision to permit Nvidia’s sales of H200 chips to China, describing it as a severe national security threat. He argues that advanced semiconductors are among America’s last competitive advantages in the race for AI innovation. Learn more.

In a speech at Davos, CFO Sarah Friar emphasized the company’s aggressive scaling initiatives across health, science, and enterprises. She noted a surge in power consumption from 0.2 to 1.9 gigawatts anticipated for 2025, alongside a remarkable increase in annual revenue from $2 billion to $20 billion.

The enterprise-focused expansion was bolstered through a three-year collaboration with ServiceNow, integrating OpenAI models like GPT-5.2 into a framework that manages 80 billion workflows each year. These AI agents will take charge of comprehensive business operations, from natural language incident resolution to workflow establishment. While ServiceNow continues its multi-model approach, the partnership extends OpenAI’s influence across major corporations including Walmart, PayPal, Accenture, and Morgan Stanley.

Furthermore, OpenAI has confirmed intentions to release its inaugural physical device in the latter half of 2026. CEO Sam Altman has described the device as something that transcends mere gadgetry, aiming to redefine a category of computing.

Simultaneously, OpenAI is enhancing safety measures with the introduction of an age prediction system designed to identify users suspected to be under 18, implementing tailored restrictions accordingly. Adult users can confirm their identities via selfie-based Persona checks, while parental controls introduce features like rest periods and distress alerts. Learn more.

The Davos 2026 summit underscored the transition from AI hype to a focus on measurable ROI. Business leaders communicated a clear message: AI strategies led by executives, along with dedicated centers of excellence, outperform approaches that grant broad employee access. Celonis reported an eightfold increase in returns when process optimization is centralized. Leaders from Siemens and LogicMonitor emphasized that effective data management and strategic oversight are essential to scaling impact.

In the competitive SaaS landscape, the focus has shifted to serving as the orchestration layer for AI agents. Companies such as Workday, Salesforce, Microsoft, and Snowflake are competing to unify enterprise data, automate workflows, and ensure engineering teams effectively translate AI insights into executable actions.

CEOs Demis Hassabis from DeepMind and Dario Amodei from Anthropic issued stern warnings: AI is already impacting entry-level positions and internships. Hassabis noted that significant job displacement is occurring at DeepMind itself. Amodei projected that up to 50% of office positions could be eliminated in the next one to five years, citing decreasing demand for junior and mid-level roles within Anthropic.

When it comes to the timelines for Artificial General Intelligence (AGI), Amodei estimates that within one to two years, AI will surpass human-level performance in most areas. Hassabis provided a broader timeframe of 2027-2030, noting that while digital tasks like coding can be automated swiftly, validating automations in physical sciences takes time. He also confirmed that Gemini’s model will not display ads, indicating revenue pressures for OpenAI as it explores various monetization strategies.

• Palmer Luckey suggests Meta’s VR layoffs may ultimately benefit the industry

• Notion broadens its custom agent capabilities with enhanced connectors and AI co-editor

• Google Stitch introduces API keys and PRD generation, empowering designers as product managers

• This year, 55 US AI startups secured $100M or more, with eight achieving this milestone twice

• Microsoft ranks 40 professions by AI vulnerability, and teachers are not exempt

• STEP3-VL-10B surpasses Gemini 2.5 Pro in visual tasks despite being 20 times smaller

• Elon Musk seeks $134B in damages from OpenAI, claiming a breach of its nonprofit mission

• Blackbox AI launches an API allowing users to run Claude Code and other agents on remote virtual machines

• GLM-4.7-Flash introduces a free 30B coding assistant you can run locally with impressive results

• Research indicates experts hold less certainty about AI consciousness than anticipated

• Researchers at Cambridge develop an AI collar that enables stroke survivors to regain speech with a 4.2% error rate

• DeepWisdom is launching a new platform empowering solo entrepreneurs to build products through natural language prompts

• Steam updates its AI disclosure policy to focus on player-facing content rather than development tools

• Investors are expressing concern after Mira Murati’s AI startup experienced a significant talent exodus within a week

• PwC’s global chairman reveals that 56% of companies are deriving ‘nothing’ from AI due to leadership oversight

• Razer’s CEO discusses why gamers might embrace AI once they shift their perception

• IMF raises its growth forecast for 2026 to 3.3%, crediting the AI boom with offsetting trade war impacts

• IBM introduces a consulting service combining human expertise with digital technologies to enhance enterprise AI

• Davos 2026 discussions primarily focused on tariffs, AI developments, and political tensions as Trump speaks

• A Randstad survey reveals Gen Z workers are the most concerned about AI displacing their jobs

• Goldman Sachs’ banking head anticipates an IPO ‘mega-cycle’ driven by AI disruption across industries

• A new reinforcement learning (RL) system automatically optimizes your code; simply define your goals and let it work

• Korea has launched an AI competition to discover its finest homegrown AI models

• ByteDance intensifies its AI cloud ambitions, competing directly with Alibaba

• Palantir’s CEO opines that AI could render mass immigration obsolete by fulfilling domestic job requirements

• Physical AI, rather than mere software hype, is touted as Europe’s pathway to an AI advantage

• Adobe Premiere and After Effects receive AI enhancements including object masking, improved tracking, and Firefly integration

• Groq’s founder claims that robots will enhance wealth and address labor shortages

• Ukraine commits to sharing extensive combat data to aid in the training of allied AI systems

• A soft robotic hand “sees” around corners like a human does

• Software stocks experience significant declines as fears of AI disruption grow

• In contrast to Gates and Musk’s promises for more leisure, a CEO warns that AI will accelerate work demands rather than lessen them

• Internal documents expose the intricate dealings behind Microsoft and OpenAI’s enduring collaboration in AI

• MemOS 2.0 debuts open-source memory OS for AI agents

• AI is learning Dungeons & Dragons to refine its long-term decision-making abilities