Tokens & Signals for 5/19/2026. We scanned ~1,200 Twitter accounts (1271 tweets), 13 subreddits (66 posts), Hacker News (10 stories), 7 newsletter posts, 4 podcast episodes, 116 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* Andrej Karpathy is joining Anthropic's pre-training unit to lead a new team dedicated to recursive self-improvement—essentially teaching Claude to research and train itself. x.com/ClaudeDevs/status/2056753265346564259
* Google dropped the Gemini 3.5 series; the Flash model is out now, already beating 3.1 Pro on coding benchmarks and running 4x faster. x.com/Google/status/2056788000546386273
* @karpathy on the Antigravity app: "Finally a real desktop agent platform that isn't just another chat wrapper." x.com/Google/status/2056789045548896516
* Cerebras hit a staggering 1,000 tokens/s on a 1-trillion parameter model (Kimi K2.6), proving wafer-scale hardware can absolutely demolish standard GPU clouds. x.com/cerebras/status/2056778123329274279
* Google is taking AI off the screen with new intelligent eyewear, partnering with Samsung, Warby Parker, and Gentle Monster. x.com/Google/status/2056806066055455187
* OpenAI is tackling misinformation by baking C2PA metadata and SynthID watermarking into every image they generate. x.com/OpenAI/status/2056793648571011232
* @sama on where things are headed: "The talent moves are interesting, but the real race is who solves recursive self-improvement first."
Bytedance open-sourced "Lance," a 3B parameter multimodal model that handles video and image understanding and* generation in one tiny package. huggingface.co/bytedance-research/Lance
* Google is standardizing agentic e-commerce with new protocols (UCP and AP2), letting AI agents securely check out and track orders across retailers like Walmart and Shopify. x.com/Google/status/2056798660663365944
Best to Build With Today
* Coding — gpt-5.4-xhigh for autonomous agentic tasks; gpt-5.2-codex for raw coding logic.
* Reasoning — claude-opus-4-6-thinking-auto is currently the king of deep logic.
* Chat — gemini-3.1-pro holds the top spot for general conversation and world knowledge.
* Open-source — Lance (3B) is the new go-to for lightweight, on-device multimodal tasks.
* Value pick — Gemini 3.5 Flash is free in AI Studio and already outperforms older Pro models.
Deeper Dives
💼 Industry & Business
Andrej Karpathy Joins Anthropic
Karpathy is heading back into frontier R&D to lead a new pre-training team at Anthropic. His focus: recursive self-improvement — teaching models how to accelerate their own training cycles.
Why it matters: Karpathy has a reputation for actually shipping things, and his bet on recursive improvement signals we're getting serious about self-building models.
� Twitter� Reddit� Hacker News
Google Partners on Agentic Commerce Protocols
Google introduced the Universal Commerce Protocol (UCP) and Agent Payments Protocol (AP2). With Shopify, PayPal, and Walmart already on board, Google is quietly building the infrastructure for agents to function as real economic actors.
Why it matters: This is the gap between an AI that recommends a product and one that actually buys it for you.
� Twitter
OpenAI Adds SynthID and C2PA Verification
OpenAI is applying dual-layered provenance to all generated media — combining C2PA metadata with Google's invisible SynthID watermarks. They also launched a public tool so anyone can check whether a file came from their models.
Why it matters: In a world drowning in deepfakes, standardized, hard-to-strip provenance might be the only thing that keeps trust intact.
� Twitter� Hacker News
🧠 Models & Research
Cerebras Claims 1,000 Tokens/s for Trillion-Parameter Model
Cerebras is running Moonshot AI's massive Kimi K2.6 model in enterprise trials at nearly 1,000 tokens/s. By moving weights onto their wafer-scale hardware, they're outperforming GPU clouds by nearly 7x.
Why it matters: This is proof that custom silicon can completely rewrite the economics of inference at the trillion-parameter scale.
� Twitter� Reddit
New Paper: Code as Agent Harness
This research makes the case that we need to stop relying on loose natural language prompting for agents and move to structured code as the control layer.
Why it matters: If you want agents that actually hold up in production, you need the strict logic of code — not the slippery ambiguity of English.
� Twitter
🚀 Products & Launches
Google Announces Gemini 3.5 Model Series
Google's 3.5 series is live. The Flash model is free right now, 4x faster than other frontier models, and already outperforming 3.1 Pro on coding and agentic benchmarks.
Why it matters: Google is making a real push to ensure their most capable models are also the most efficient ones developers can actually use.
� Twitter� Reddit� Hacker News
Antigravity 2.0 Launches as Desktop App
Antigravity 2.0 has grown up — it's moved from the browser to a native desktop app, with multi-agent orchestration, native CLI/SDK access, and support for long-running background tasks.
Why it matters: Serious agents need local OS access to do anything useful; this makes your desktop the natural home for your agent workflows.
� Twitter� Reddit
Funding & Deals
* ElevenLabs — Closed a $500M Series D at an $11B valuation in February. They're currently leading the charge on interactive voice agents, including the newly launched Einstein agent.
Launches
* Gemini 3.5 Flash — High-efficiency, free model available now in AI Studio.
* Lance (3B) — Bytedance's open-source model for video/image understanding and generation.
* Einstein Voice Agent — ElevenLabs' interactive historical voice synthesis for education.
Closing thought: The industry is moving past the chatbot phase. With Karpathy building recursive systems and Google laying down actual commerce and OS-level agent infrastructure, the next year isn't about better talkers — it's about better doers.