Issue #42 2 min read

AI Engineering Signal #42

Andrej Karpathy joins Anthropic

Share

Signals

Andrej Karpathy joins Anthropic

shifts frontier research weight toward Anthropic; update assumptions about which lab leads training research through 2027.

Web

Gemini 3.5 Flash costs roughly 30x more than Gemini 1.5 Flash

reprice any pipeline budgeted against older Gemini tiers before June billing hits.

Simon Willison

Google Antigravity 2.0 builds a working OS with 96 agents under $1K

sets a concrete cost-per-complexity benchmark for multi-agent orchestration procurement.

Web

Guardrails lift an 8B model from 53% to 99% on agentic tasks

invest in constraint layers before scaling model size when task accuracy gates deployment.

GitHub

Cerebras runs a trillion-parameter model at 1,000 tokens per second

latency is no longer the blocker for large MoE at scale; revise inference assumptions.

Web

Meta cuts roughly 8,000 roles, forcibly reassigns thousands more to AI

headcount reduction via AI is now active org design, not a future threat.

Web

Get signals like this in your inbox

Daily AI engineering intelligence. No noise.

[ Subscribe ]

The Take

Inference throughput is collapsing in price at the hardware layer while frontier API pricing climbs. Teams that locked budgets against last year's Gemini tiers or assumed model scale was the only path to accuracy need to reopen both cost models and architecture assumptions now.

Subscribe

Unsubscribe any time.

Related Signals