Week 4, 2026 | Meanwhile in AI

AGI predictions went mainstream at Davos. Lab CEOs gave timeline estimates to policymakers. Meanwhile, a police chief resigned over a hallucinated football match. The capability curve and the institutional readiness curve are diverging.

The AGI Timeline Goes Mainstream

So here’s what happened at Davos.

Anthropic’s Dario Amodei and DeepMind’s Demis Hassabis appeared together. Both expect significant job displacement. Both think AGI is close. Both worry about getting safety right before it arrives.

Hassabis put AGI at “five to ten years,” noting we might need “one or two more breakthroughs.” Amodei went further: half of entry-level white-collar jobs could disappear within five years. He’s already seeing changes in coding.

Not everyone agrees. Yann LeCun, speaking at AI House in Davos, argued that LLMs will never achieve human-level intelligence. “The AI industry is completely LLM-pilled,” he said. “We’re never going to get to human-level intelligence by training on text only. We need the real world.”

Then again - these same predictions have been made for decades. “AGI in 5-10 years” has a long history of being wrong. That doesn’t mean this time is different, but it does mean the prediction is now mainstream enough to shape policy whether it’s accurate or not.

On China: Hassabis said Chinese models are “just months behind.” Amodei publicly criticised the Trump administration for approving H200 sales to China, comparing it to “selling nuclear weapons to North Korea.”

Jensen Huang offered the infrastructure perspective: AI needs more energy, more land, more skilled workers. He sees robotics as “a once in a lifetime opportunity” and predicts trade jobs will see salaries nearly double.

Satya Nadella emphasised that realising AI’s potential requires investment and infrastructure that are “fundamentally driven by governments.”

What’s new isn’t the predictions. It’s the venue. When lab CEOs say this at Davos, policymakers and executives treat it as consensus. These statements will shape investment and regulation for the next year.

Anthropic Opens the Black Box

The same week the CEOs debated AGI timelines, Anthropic released the actual document that shapes how Claude thinks.

Anthropic trains Claude using Constitutional AI - a method where the model critiques its own outputs against a set of explicit principles, rather than relying entirely on human feedback. Those principles are “the constitution.”

This week they released the full constitution under CC0. Not a summary - the actual ruleset. Anyone can read it, critique it, or use it to train their own models.

This is unusual. Most AI companies treat their alignment approaches as proprietary.

The same week, they published research identifying a specific neural pathway that controls whether Claude behaves as a helpful assistant or drifts into problematic behaviour.

They call it the “Assistant Axis.” Think of it like a mixing board with hundreds of sliders. Anthropic found that one particular slider controls how much Claude acts like a helpful assistant versus going off-script. Turn it up, more helpfulness. Turn it down, chaos. The breakthrough is identifying which slider does what in a system with millions of them.

That’s different from hoping your RLHF worked.

They also published how they redesigned their hiring tests after Claude kept solving them. Opus 4.5 matched the performance of the majority of human applicants, so they open-sourced the original test and shifted toward out-of-distribution puzzles.

If your evaluation can be gamed by the system you’re evaluating, you need a new kind of evaluation.

Three releases in one week - constitution, interpretability research, evaluation methodology - all made public. Other labs now face pressure to explain their own approaches.

Agentic Coding: Mixed Results

Agents can build browsers, orchestrate clusters, generate complex infrastructure from natural language. The capability demos are impressive. The production results are mixed.

Wilson Lin’s FastRender - a browser built by thousands of parallel agent calls writing over a million lines of code - shows what coordinated agent swarms can produce. Cursor researchers deployed a fleet of autonomous agents that ran for one week to generate it.

But a Hacker News thread with 449 comments captured the practitioner divide: “Do you have any evidence that agentic coding works?” Responses split between converts and sceptics who report more cleanup than productivity gains.

Recent research mapped where AI agent setups actually land on the cost-performance curve. Most deployments aren’t on the efficient frontier - they’re worse than necessary on both cost and accuracy. Related work found that a single LLM with KV cache reuse can match multi-agent performance while cutting inference costs.

One founder reduced monthly API costs by 80% - from $1,500 to $300 - by benchmarking actual production prompts against 100+ models instead of trusting general benchmarks.

If you’re not measuring your specific workloads, you’re making decisions based on vibes and marketing.

Open Source Catches Up

Open source caught up across the stack this week. Voice, models, hardware.

Voice infrastructure went open. Qwen3-TTS dropped under Apache 2.0 with voice cloning from three seconds of audio. Microsoft’s VibeVoice-ASR processes 60-minute files with speaker diarisation across 50+ languages. NVIDIA’s PersonaPlex offers real-time full-duplex conversations.

But practitioners report a gap between infrastructure and intelligence. The voice processing works. The models powering the conversations don’t. As Ethan Mollick put it: most commercial voice modes are powered by “dumb, sycophantic” models.

Models are competitive on specific benchmarks. DeepSeek V3.2 topped Multivac’s blind peer evaluation on a production JSON parsing challenge, scoring 9.39 vs GPT-5.2-Codex (9.20) and Claude Opus 4.5 (7.57). GLM-4.7-Flash set a new 30B-class standard with 59.2% on SWE-bench Verified. Liquid AI’s 1.2B reasoning model runs in 900MB of RAM and outperforms models 40% larger.

Hardware is getting cheaper. A 128GB VRAM inference server for under €10,000. 26.8 tokens/second on MiniMax-M2.1 using $880 of legacy MI50s. The $600 local LLM machine is viable for daily dev work.

The caveat: proprietary models still hold advantages in generality and safety tuning. Open source excels when you know exactly what you need.

LLM-Generated Exploits

Research on LLM exploit generation showed that LLM agents successfully generated over 40 distinct exploits for a QuickJS zero-day vulnerability, achieving 100% success rate in the most advanced model tested. Cost: approximately $30 per run.

OpenAI flagged their upcoming releases as “high” cybersecurity risk - the first time they’ve applied that rating to models before release.

Researchers showed that bypassing the standard chat template function strips safety alignment from models like Gemma and Qwen.

What’s a chat template? When you talk to these models through the normal interface, there’s a wrapper that formats your messages a certain way. The safety training expects that format. Send raw text without the wrapper and the model doesn’t recognise it as the kind of input it should refuse. Like a security guard trained to check IDs at the front door - walk in through the loading dock and nobody asks.

This week: automated exploit generation at scale.

Ground-Level Reality

While Davos discussed AGI timelines, institutions are still learning how to handle current AI.

A UK police chief resigned after his force used a hallucinated Microsoft Copilot report to justify banning Israeli fans from a non-existent Carabao Cup match. Ten days from denial to resignation. Over a football match that never happened.

Job applicants are suing AI recruitment tool company Eightfold AI, arguing that automated “Match Scores” should fall under Fair Credit Reporting Act protections.

eBay explicitly banned “buy for me” agents and LLM scrapers in its updated terms of service. No automated systems that place orders without human oversight. The platform that pioneered online auctions is drawing a line against agentic commerce.

What It Means

The timeline conversation has shifted. AGI in five to ten years is now the mainstream position among leading labs. Shane Legg, DeepMind co-founder, is hiring economists to study post-AGI economics. Whether you believe the timeline or not, policy and investment will follow.

Anthropic published its alignment methodology. Constitution, interpretability research, evaluation methodology - all public. Other labs will face questions about why they haven’t done the same.

Benchmark your actual workloads. Most teams are overpaying by 5-10x because they’re not measuring what matters. General benchmarks are useless for specific use cases.

Open source is catching up. Voice, models, hardware - all saw major releases under permissive licenses. For teams with specific, well-defined needs, the gap with proprietary has closed.

Institutions aren’t ready for current AI, let alone AGI. A police chief resigned over a hallucinated football match the same week lab CEOs predicted AGI in five years. The capability curve and the institutional readiness curve are diverging.

Worth Your Time

If you read three things:

The Hassabis-Amodei Davos discussion - The full “Day After AGI” panel. This will shape policy discussions all year.
Anthropic’s constitution - Not the blog post. The actual document under CC0. Read the section on “Claude’s identity” to understand what they’re actually training for.
Sean Heelan on exploit industrialisation - The security landscape is changing faster than most people realise. This piece makes the economics clear.