AI Weekly - Week 01/2026

TL;DR

This week in 30 seconds:

• DeepSeek Dilemma: China’s AI lab shows efficiency breakthrough – but with significant security concerns
• CES 2026: Robots, smart glasses, and AI hardware dominate – Jensen Huang and Lisa Su keynote
• IPO Fever: OpenAI plans $1 trillion IPO, Anthropic and SpaceX follow – analysts warn of bubble
• Claude Code: Impressive benchmarks, but reality is more nuanced than viral anecdotes

---

Audio Version

---

Story of the Week

DeepSeek: Efficiency Breakthrough with Security Questions

Image: AI-generated with Pollinations.ai

Chinese AI startup DeepSeek has kicked off 2026 with a technical paper that rethinks the fundamentals of AI training [1]. The method called Manifold-Constrained Hyper-Connections (mHC) promises more stable training at lower costs.

The numbers are impressive: The 19-member research team led by Zhenda Xie tested mHC on models with 3, 9, and 27 billion parameters. The result: “Effectively stable large-scale training with superior scalability” at “negligible computational overhead” [1]. Analysts call the approach a “breakthrough for scaling” [1].

The Other Side: What the efficiency hype ignores – CSIS and NIST have documented significant security concerns [10]:

• DeepSeek-R1 blocks 0% of jailbreak prompts (vs. GPT-4o: 86%)
• 11x more likely to produce harmful outputs than US models
• All data stored in China – with known links to state entities [10]

Geopolitical Context: While OpenAI and Microsoft invest trillions in compute, DeepSeek shows that clever architecture can partially compensate for hardware disadvantages. However: The US maintains a 21-49x compute advantage (if export controls hold) [10]. Pentagon and Capitol Hill have already banned DeepSeek for official use.

What’s Next: Analysts expect the launch of DeepSeek R2 before Chinese New Year [1].

Bottom Line: DeepSeek proves software innovation, but security risks make enterprise deployment in critical systems problematic. The real test: Can efficiency compensate for the chip gap long-term?

---

More Top Stories

CES 2026: The AI Hardware Revolution Begins

Image: AI-generated with Pollinations.ai

The Consumer Electronics Show opens January 6 in Las Vegas – and AI is everywhere [3]. Nvidia CEO Jensen Huang and AMD CEO Lisa Su are delivering keynotes, while over 4,000 exhibitors showcase AI hardware.

The highlights:

• Robotics Explosion: LG’s “CLOiD” home robot, SwitchBot’s “Onero H1”, Narwal’s AI vacuum [3]
• Smart Glasses: Google announces two Gemini glasses for 2026 – with and without display [6]
• On-Device AI: Intel’s Panther Lake chips with 18A technology, LG’s “world’s lightest 17-inch RTX laptop” [3]

Reality Check: Analysts warn of overhype – many products will “overpromise and underdeliver on AI features” [3]. The real advances lie in on-device AI and robotics (1000+ units already deployed).

---

The Year of Trillion-Dollar IPOs – With Warning Signs

Image: AI-generated with Pollinations.ai

2026 could be the biggest tech IPO year in history [4]. The numbers:

Company	Target Valuation	Timeline
OpenAI	up to $1 trillion	Late 2026
SpaceX	$30+ billion	June 2026
Anthropic	TBD	Preparation ongoing

Sam Altman called an IPO the “most likely path” [4]. Anthropic has hired Wilson Sonsini for preparation [4].

The Warning: Neither OpenAI nor Anthropic are financially transparent. We know expected revenues (OpenAI: $30B in 2026), but not profitability [4]. Historical parallel: Nvidia survived the dot-com bubble in 1999 – but only reached $5 trillion valuation in 2025. The journey took 26 years [4].

---

Quick Hits

In brief:

• Gemini 3 @-Menu: Google’s new model switcher enables quick switching between @Fast, @Thinking, and @Pro via keyboard [6]
• MCP Becomes Standard: OpenAI, Microsoft, and Google adopt Anthropic’s Model Context Protocol – Linux Foundation takes governance [2]
• Chip Rally: ASML +9%, Micron +10% at year start – AI demand drives semiconductor stocks [9]
• End of Scaling? Yann LeCun and Ilya Sutskever argue that current scaling laws are plateauing [2]
• World Models: Google DeepMind’s Genie, World Labs’ Marble, and Runway’s GWM-1 drive new research focus [2]

---

Tool of the Week

Claude Code - Productivity Boost with Nuance

Image: AI-generated with Pollinations.ai

Jaana Dogan, Principal Engineer at Google (Gemini API team), shared on X: She gave Claude Code a problem description and received a working system in one hour that her team has been building for a year [7].

The Benchmark Reality: Claude Code achieves 80.9% on SWE-bench – a legitimate benchmark for code generation. Typical productivity gains are 50-75% error reduction for repetitive tasks [7].

However: A randomized study showed: Experienced developers with AI tools on their own code sometimes took 19% longer. The effect depends heavily on the use case [2].

Boris Cherry from the Claude Code team: 259 pull requests in 30 days – 497 commits, 40,000 lines added – “every line written by Claude Code + Opus 4.5” [7].

Claude Code Docs

---

Fail of the Week

“The Security Gap No One Wants to See”

While DeepSeek is celebrated for efficiency, NIST evaluations paint a different picture [10]:

The Problem: DeepSeek-R1 blocks 0% of jailbreak prompts. For comparison: GPT-4o blocks 86%, Gemini 64% [10].

Root Cause: Missing or inadequate safety guardrails during rapid scale-up – a problem that systematically affects Chinese AI labs [10].

Lesson Learned: Efficiency without safety isn’t innovation – it’s a risk. Pentagon and Capitol Hill have already responded.

---

Number of the Week

37% vs. 14%

37% of US companies plan to replace jobs with AI by end of 2026 [8]. But reality? Only 14% have actually replaced positions [8].

This expectations-reality gap shows: Many announcements, less implementation.

The most vulnerable groups according to Goldman Sachs:

• High-salary employees without AI skills
• Recently hired workers
• Entry-level positions [8]

Context: In 2025, 55,000 US jobs were directly attributed to AI – out of 1.17 million total layoffs [8].

---

Reading List

For the weekend:

1. In 2026, AI will move from hype to pragmatism - TechCrunch’s deep dive into trends: Small Language Models, World Models, MCP adoption (8 min)
2. CSIS: Delving into the Dangers of DeepSeek - The security analysis every CTO should read (10 min)
3. DeepSeek Paper: Manifold-Constrained Hyper-Connections - Technical analysis of the efficiency breakthrough (5 min)

---

Next Week

What’s coming:

• CES 2026 (January 6-9) - Nvidia, AMD, Intel keynotes – expect chip announcements and robotics demos
• Anthropic Briefing (January 12) - AI in Healthcare, San Francisco
• DeepSeek R2? - Possible launch before Chinese New Year (February)

---

This newsletter was researched and written with AI assistance. Images generated with Pollinations.ai.