The Singularity
Ledger

INTELLIGENCE

The new year opens with a fundamental tension: small models are getting surprisingly capable, reliability benchmarks are being weaponized for PR, and the field's sharpest minds are still arguing over whether any of this constitutes real reasoning.

Friday, January 2, 2026

Tencent's 1.96B Parameter Model Performs Like an Agent — and DeepSeek Shows How to Train Efficiently Under Chip Constraints

Two Chinese AI papers dropped over the holiday that challenge Western assumptions about what's required to build capable AI agents: Tencent's tiny agentic model and DeepSeek's efficiency-first training methodology.

FREEREAD MORE

@rohanpaul_ai @business @thisguyknowsai

Grok Claims Lowest Hallucination Rate in Workplace AI Study — But the Methodology Deserves Scrutiny

A Relum study found Grok hallucinated at just 8% compared to ChatGPT's 35%, generating massive engagement. The result is provocative, but the study's framing raises questions about what "hallucination rate" actually measures in practice.

SUBSCRIBERSREAD MORE

The Information's 2026 AI Predictions: Google Acquires Thinking Machines Lab, OpenAI Ships an AI Research Intern, Agents Hit $1K/Month

The Information published a detailed set of AI predictions for 2026 that double as a map of where the industry's center of gravity is shifting — toward agents, continual learning, and consolidation.

SUBSCRIBERSREAD MORE

Get The Singularity Ledger

Daily AI news in your inbox. Free, no spam.

Questions? hello@thesingularityledger.com

Get this briefing every morningFree daily AI

The Singularity
Ledger

Tencent's 1.96B Parameter Model Performs Like an Agent — and DeepSeek Shows How to Train Efficiently Under Chip Constraints

Grok Claims Lowest Hallucination Rate in Workplace AI Study — But the Methodology Deserves Scrutiny

The Information's 2026 AI Predictions: Google Acquires Thinking Machines Lab, OpenAI Ships an AI Research Intern, Agents Hit $1K/Month

Terence Tao: AI Doesn't Understand — It's 'Probability Pretending to Think'

Building Internal Agent Frameworks: Nine Hard Problems Engineers Are Actually Hitting

NotebookLM's Anti-Hallucination Design Is Finding a Quiet Professional Audience

Agent Security Primitives Emerge: Pay-Per-Call Verification for Wallets, Contracts, and URLs

The Best Feature in One Startup's AI Product Came From a Customer Using It 'Wrong'

The Self-Healing Software Vision: AI That Fixes Its Own Bugs Before You Notice Them

The Best Advice on Coding Agents: Use Them Where You're Already an Expert

Two New Grok Models Spotted in LMArena Testing