#5 Decentralized promises, unstable judges, and smarter trade-offs

#newsletter #ai

Zuckerberg says AI will set us free. However, his track record says otherwise

A manifesto written by Mark Zuckerberg about his perspective on the upcoming years with AI. He offers an optimistic perception of how humans will benefit from AI, proposing an individual empowerment philosophy as the source of prosperity.

Meta is leading several projects across different areas, and Zuckerberg makes a reasonable case for the investment the company has been making.

By providing free or affordable access to these AI tools, he argues, anyone can get the most out of them. Decentralizing power from a small group of people and giving it to everyone will enhance their lives, their economics, and their education.

It's a good manifesto, but I still have a lot of doubts about this. It's no secret his company is one of the "Big Seven." They rule over everyone in the world with products that are, in a sense, silent weapons of control. Facebook, Instagram, and WhatsApp are everywhere. These tools were built with the help of the best neuroscientists to capture and hold people's attention. America is doing what it does best, trying to monopolize AI power while distributing controlled and monitored versions of its models to the rest of the world. I do believe AI will benefit us a lot, but I can't take my eyes off the moves the "Big 7" are making.

The Future is for EveryoneRead Mark Zuckerberg’s piece on our values for building superintelligence for everyone.meta.com

Your LLM judge isn't as confident as it looks

This paper stress-tests top LLM models on their accuracy when used as "LLM-as-Judge". The Meta researchers behind it, built what they call the Wiggle Framework, which tests 9 frontier models across 14 different judging tasks to see how stable their verdicts really are.

The framework checks three things: how consistent a model is when re-prompted with the same question, how it holds up against a single pushback, and how it holds up under sustained-pressure challenges.

Models showed low variation when simply asked the same question again, but that stability collapsed under pressure. Verdicts flipped 25% to 71% of the time when faced with static pushback, and 62% to 91% of the time when facing an adversarial persuader actively arguing against the original verdict.

When a model's verdict did change under pressure, it almost always moved away from the ground truth, not toward it. A concerning finding…

Jagged Judges: Epistemic Stability Under Perturbation, Pressure, and PersistenceLLM judges have become central infrastructure for model evaluations, online grading, and reward modeling. Judges are typically validated by accuracy on golden data, but accuracy says little about whether they are stable under re-prompting, challenge, or sustained pushback. We introduce the…arXiv.org

Smarter at math, clueless about facts

I came across this blog post that argues that AI labs are deliberately trading factual knowledge for reasoning power in smaller models.

Based on the post, small reasoning models are getting dramatically smarter per parameter. GLM-5.2 scores 99.2% on AIME 2026 with just 40 billion active parameters, far less than GPT-4 needed in 2023. But ask these same models a basic factual question and the picture flips. Qwen3.5's smaller models hallucinate 80 to 82% of the time on pure knowledge questions.

The reasoning is, facts take up a lot of space in a model's weights, while reasoning skills compress much better and transfer well through training. So labs are stripping out deep factual knowledge and leaning on external tools like search and retrieval to supply facts at runtime instead.

The author's key point is that facts go stale but procedures don't, and this setup also fixes a lot of hallucination. A wrong fact in a document can be found and corrected while a wrong fact baked into a model's weights can't.

Models Are Getting Dumber on Purpose - Walter van der GiessenReasoning scores climb while per-token compute drops. Labs are stripping world knowledge out of models, and I think that's the right trade.Walter van der Giessen - dev blog

SpaceXAI launches the autonomous agent Grok Bot

SpaceXAI has launched Grok Bot, a system of persistent AI agents that run on dedicated cloud computers and can interact with websites, apps, and inboxes. Unlike coding-focused agents like Claude Code or Codex, Grok Bot targets broader business workflows.

Each Bot remembers context and preferences and can even learn a task just by watching it get done once. Multiple Bots can also run in parallel, communicate with each other, and split work between specialized agents, looping in a human only when a decision is needed.

The launch follows SpaceX's recent acquisition of Cursor, bringing their AI efforts under one roof. Grok Bot is currently in beta for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers.

SpaceXAI Launches Grok Bot for Autonomous AI AgentsSpaceXAI has introduced Grok Bot, a system of persistent AI agents that operate on dedicated cloud computers and can interact with websites, applications, inboxes, and other tools.InfoQ