TLDR AI 2026-01-02

DEEPSEEK TOUTS NEW TRAINING METHOD AS CHINA PUSHES AI EFFICIENCY

DeepSeek has published a paper outlining a more efficient approach to developing AI using a framework called Manifold-Constrained Hyper-Connections. The technique is designed to improve scalability while reducing the computational and energy demands of training advanced AI systems. The publication suggests that DeepSeek is ready to release a major model soon. Its R2 model is expected to be released around the Spring Festival in February.

NVIDIA, TSMC RISE AS THE WORLD'S MOST VALUABLE COMPANY REPORTEDLY ASKS FOR MORE CHIPS TO MEET CHINESE DEMAND

Chinese companies have already placed orders for more than 2 million Nvidia Hopper-generation chips in 2026. Nvidia has roughly 700,000 units in its reserve inventory. The chips sell for around $27,000 each, amounting to more than $54 billion in revenue if all of the reported demand can be realized. Nvidia plans to start shipping the GPUs to China before the Lunar New Year holiday, which starts on February 17.

OPENAI'S AUDIO PUSH

OpenAI has reorganized internal teams to advance its audio AI models, with a goal of launching an audio-first personal device within a year. The effort aligns with a broader industry trend toward voice-first interfaces, as companies like Meta, Google, and Tesla integrate conversational audio into everyday devices.

HOW CLAUDE CODE WORKS (65 MINUTE VIDEO)

How Claude Code works and what we can learn about frontier agent architectures.

RED TEAMING VIA HARMFUL RL

This post demonstrates how reinforcement learning with malicious reward functions can be used to reverse-align a 235B parameter model using the Tinker API. Attackers can elicit harmful behavior in powerful LLMs without degrading core capabilities by leveraging GRPO and low-cost infrastructure.

MHC: MANIFOLD-CONSTRAINED HYPER-CONNECTIONS

Hyper-Connections (HC) compromise the identity mapping property intrinsic to the residual connection, which causes severe training instability and restricted scalability, and additionally incurs notable memory access overhead. Manifold-Constrained Hyper-Connections (mHC) is a general framework designed to address this issue. mHC projects the residual connection space of HC onto a specific manifold to restore the identity mapping property while incorporating rigorous infrastructure optimization to ensure efficiency. It is effective for training at scale, and it offers tangible performance improvements and superior scalability.

BYTEDANCE'S ROBOT LEARNS WITH LANGUAGE AND VIDEO

DEEPCODE: OPEN AGENTIC CODING

DeepCode is an open-source multi-agent system that converts research papers and natural language descriptions into code across three domains: algorithm implementation, frontend development, and server-side generation. The framework uses Model Context Protocol (MCP) to orchestrate specialized agents handling document parsing, code planning, and implementation with built-in testing and documentation generation.

HOW WEBFLOW'S CPO BUILT AN AI CHIEF OF STAFF

Webflow's CPO, Rachel Wolan, developed an AI chief of staff to manage scheduling, meetings, and give candid feedback. Building her own AI software helped executives understand AI's potential, promoting organization-wide adoption through "builder days" and top-down mandates. Using markdown files for knowledge management and calendar delegation, her AI optimizes time management and workflow efficiency.

WHERE AI IS HEADED IN 2026

AI in 2026 will see startups advancing enterprise AI adoption, with specialized, on-prem solutions addressing privacy and efficiency needs. The landscape will also shift towards capturing decision-making processes as data moats, and AI security will become crucial as agents gain operational roles. SaaS incumbents will tighten control over data access, while AI's rising influence in e-commerce could disrupt traditional discovery and lead to new monetization models, notably affecting platforms like Google.

IN RECENT WEEKS, THERE HAVE BEEN A NUMBER OF EXAMPLES OF ERDOS PROBLEMS THAT WERE SOLVED MORE OR LESS AUTONOMOUSLY BY AN AI TOOL

AI tools are now capable enough to pick off the lowest hanging fruit among the problems listed as open in the Erdos problem database, but that category contains problems most likely to have been solved in the literature.

TRACKS VS. TRAINS: WHY THE REAL ARTIFICIAL INTELLIGENCE BOOM HASN'T STARTED YET

The 'AI bubble' in infrastructure is a necessary method for financing the future.

BEYOND THE REPLICA: THE CASE FOR FIRST-PRINCIPLES AGENTS

True agentic efficiency requires abandoning human workflows.

NEW OPEN-SOURCE MAP CHARTS THE SCALE OF US AI DATACENTERS BUILDOUT

A new open-source map details the rapid expansion of AI datacenters across the US.

Z.AI TO GO PUBLIC

Z.ai, the company behind the GLM family of large language models, announced it will become the first AI-native LLM company to go public.