AI Models

Meta Just Fired MORE Employees...

Published
Mar 27, 2026
Duration
13:24
Module
AI Models
Click to load the YouTube player

Video summary

Companion notes

Meta cut 4,000 staff — roughly 20% — and AI is the main reason, not just post-pandemic overhiring.

## Why the layoffs are AI-shaped The hosts frame Meta as planning 5–10 years out, comparing it to Tencent paying for 6-month MBA sabbaticals. Routine work is pattern-matching, and the models are learning those patterns. Meta staff engineer John Kim still reviews code personally but uses Claude Code as an assistant — proof the right workflow already exists inside the company. Brownfield code (legacy code with heavy interdependencies) is where AI still struggles, but that gap is closing fast.

## Claude Code is outshipping Meta Anthropic's Claude Code shipped 20+ features in March with a core engineering team of roughly 30 people. New capabilities — Telegram access, memory storage, and "daydream" persistence — now overlap with OpenClaw. The hosts are already migrating skills to Claude Code because it surfaces every tool call and skill invocation, giving far more visibility than OpenClaw.

## Gemini 3.1 Flash is a price play, not a quality play Google's new light model is $0.50 per million context tokens. The hosts call out that this matches what Chinese labs charge for similar performance, and that Gemini 3.1 Pro is rarely used for coding — Claude and GPT still lead. Voice and Nano image integration are the only real differentiators.

## Money, valuations, and the bubble question Reflection AI raised $2 billion from Nvidia at a $25 billion valuation. The hosts compare today's AI capex to the dot-com era: inflated individual company valuations, but the underlying infrastructure (the internet) was real. Same logic applies here — Claude and similar tools let you ship product "so fast, so quickly."

## Security and the open-source supply chain A compromised LiteLLM package exposed roughly 97 million API keys. The team caught it within one hour, but the transitive-dependency structure (one large project pulling hundreds of small ones) means a single malicious package cascades. Sora's decline is also attributed to ByteDance's SeedDance 2.0 — Chinese video models are currently ahead of the West, priced at about $1 per 5-second clip.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=74xdgNfsPlk