Good morning. It’s Wednesday, February 18th.

On this day in tech history: In 2020, Microsoft released Turing-NLG, a 17-billion-parameter behemoth in natural language generation. Preceding GPT-3's hype, this encoder-decoder juggernaut fine-tuned on vast corpora for zero-shot tasks, pushing autoregressive limits. Its Mixture-of-Experts scaling foreshadowed sparse activation efficiencies, enabling today's trillion-param models to tackle long-context coherence without quadratic attention blowups.

In today’s email:

You read. We listen. Let us know what you think by replying to this email.

Anthropic has launched Claude Sonnet 4.6, a major upgrade designed for autonomous enterprise tasks. The model features a massive 1-million-token context window and native "computer use" capabilities, allowing it to navigate software interfaces and manage complex workflows just like a human. However, there is a significant limitation: the expanded 1-million-token window is currently a beta feature available only via the API, while standard web and app users remain on the previous limits.

A key technical advancement is the new dynamic filtering system for web research. Instead of processing raw HTML, Claude now writes Python code to parse and pre-filter data before it enters the context window. This boosted accuracy by 11% and reduced token waste by 24% in benchmarks. However, the high intensity of these "agentic" tasks carries a significant compute cost; during testing, Sonnet 4.6 consumed 280 million tokens, nearly five times more than its predecessor.

This push for autonomy coincides with a massive bet on India. Anthropic officially opened its Bengaluru office this week, its second Asian base, to support a market where technical tasks drive nearly half of all Claude workloads. The company’s India run-rate revenue has doubled since October 2025, with enterprise clients like Air India and Cognizant using Claude.

However, this transition toward "hands-off" AI has met resistance from the developer community. The latest Claude Code update (v2.1.20) now hides granular file activity by default, replacing detailed logs with vague summaries like "Read 3 files." While Anthropic argues this reduces terminal clutter as the model becomes more autonomous, developers have expressed concern that the lack of transparency makes it harder to audit security, verify context, and manage the high token costs associated with the new architecture. Read more.

The company announced a multiyear partnership with NVIDIA to build AI-optimized data centers across cloud and on-premises environments. Millions of Blackwell and Rubin GPUs will be paired with Arm-based Grace CPUs in the first large-scale Grace-only rollout, tuned for performance per watt with deep hardware–software codesign.

Meta is standardizing on GB300 systems and using Spectrum-X Ethernet for low-latency, high-throughput AI networking, while NVIDIA Confidential Computing powers privacy-conscious AI features in WhatsApp. The goal is to scale next-generation recommendation and personalization models to billions of users.

On the agent side, Meta-owned Manus launched its “Agents” mode on Telegram. The autonomous system handles multi-step tasks, research, data processing, and document creation, with Manus 1.6 Max for complex workflows and Lite for simpler queries. Telegram-first testing lets Meta explore agent reliability while the acquisition remains under regulatory review in China.

Finally, Meta may bring facial recognition to its Ray-Ban and Oakley smart glasses with a “Name Tag” feature, identifying contacts and public profiles. The move revives privacy debates after Facebook retired similar tech in 2021. Read more.

The era of the "all-knowing" chatbot is ending. xAI just launched Grok 4.20 Beta, and it’s moving from simple chat to multi-agent collaboration.

Grok 4.20 kills the single-model bottleneck. Instead of one AI guessing an answer, a specialized council of four agents processes every query in parallel. Grok coordinates, Harper fact-checks, Benjamin handles the logic and code, and Lucas provides the creative edge. They debate internally before you see a single word. This peer-review architecture isn't just a gimmick; it’s a high-speed filter designed to kill hallucinations and logical errors before they reach the UI.

With a 2 million-token context window and the Agent Tools API, Grok 4.20 isn't playing with consumer-grade toys. It executes code, crawls the live web, and runs financial simulations with a level of synthesis single models can't touch.

This release marks xAI’s fastest iteration cycle yet, following Grok 4.1 by just three months. Access is restricted to SuperGrok subscribers ($30/month) and X Premium+ members, signaling that the target isn't the casual user; it's the professional who needs a digital staff, not a search engine. Read more.

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide.

📄 OneVision-Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence

📄 Categorical Flow Maps ​​

📄 Less is Enough: Synthesizing Diverse Data in Feature Space of LLMs

📄 PISCO: Precise Video Instance Insertion with Sparse Control

📄 DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!