Social Media Briefing — March 27, 2026

Karpathy envisions a future where an agent can handle the entire DevOps lifecycle - from browsing services, reading docs, getting API keys, to deploying to production. He argues the hardest part of building software isn't code but assembling services (payments, auth, DB, security, domains).

Anthropic announces peak-hour rate limit adjustments for Claude: during weekdays 5am-11am PT / 1pm-7pm GMT, 5-hour session limits will be consumed faster for free/Pro/Max users. Weekly limits remain unchanged.

Following yesterday's Social announcement, Chollet details ARC-AGI-3 human testing methodology: ~500 testers, no qualifications required, $115-140 incentive, 10 testers per environment, 2+ must independently clear all levels. Most cleared by 5+ testers.

Andriy Burkov explains a paper proposing 'learned experiences' as prompt-based lessons instead of finetuning. Using DeepSeek-V3.1-Terminus with just 100 examples and $18 in API calls, it outperforms models finetuned with thousands of examples at >$10k cost. Method is modeled after GRPO but operates on prompt text rather than weights.

Delangue highlights trend of major companies (Pinterest, Airbnb, Notion, Cursor, Intercom) moving from API-based AI to training/using open models in-house, predicting majority of AI workflows will be in-house open-source.

Guillaume Lample (Mistral co-founder) announces Voxtral TTS - Mistral's first speech model with SOTA performance, lower cost, very low latency, using a novel architecture combining auto-regressive semantic token generation with flow-matching for acoustic tokens.

MistralAI introduces Voxtral TTS: an open-weight text-to-speech model with realistic emotional expression, 9 languages, ultra-fast latency, and easy voice adaptation.