Tokens & Signals for 4/28/2026. We scanned ~1,200 Twitter accounts (1269 tweets), 13 subreddits (65 posts), Hacker News (12 stories), 6 newsletter posts, 6 podcast episodes, 190 Discord messages, and leaderboard data for you. Estimated reading time saved: ~12 hours.
* OpenAI is pushing hard on "GPT-5.5" and Codex to make agentic workflows the industry standard — and they're clearly gunning for the enterprise dev market. x.com/OpenAINewsroom/status/2049163410424647808
* Security researcher Sagitz disclosed CVE-2026-3854, a critical RCE vulnerability in GitHub that lets attackers run unauthorized commands across millions of repositories with a single git push. x.com/sagitz_/status/2049153195243372569
* DeepSeek V4 Pro just dropped — 75% price cut, 8x faster inference. Cost-sensitive agentic workflows are already migrating. x.com/thdxr/status/2048991814183473302
* Google officially reversed its 2018 "no AI for the military" policy and is now partnering with the Pentagon to deploy air-gapped AI infrastructure for classified defense analysis. theverge.com/ai-artificial-intelligence/919494/...
* Anthropic's Claude Code now sends mobile push notifications, so developers can finally kick off a long-running task and actually step away from the terminal. x.com/ClaudeDevs/status/2049154855143649315
* @karpathy on the DeepSeek price war: "75% cheaper + 8x faster. Incumbents have to decide: compete on economics or die on quality premiums."
* Mitchell Hashimoto is leaving GitHub after 18 years — and people are treating it as a symbolic gut-punch to GitHub's "developer-first" identity. x.com/mitchellh/status/2049213597419774026
* @nathanbenaich on the ProfluentBio/Eli Lilly deal: "AI is moving from generating ideas to generating functional biological machines for massive biopharma outcomes." x.com/nathanbenaich/status/2049162488475009140
* NVIDIA's Nemotron-3 Nano Omni (30B) uses A3B quantization to squeeze multimodal reasoning — video, audio, text — directly onto edge devices. huggingface.co/unsloth/NVIDIA-Nemotron-3-Nano-O...
* YC's Summer 2026 "Request for Startups" is out, and they're betting big on physical-world agents and autonomous infrastructure. x.com/benln/status/2049117209750900865
Best to Build With Today
* Coding — gpt-5.2-codex for general integration, or claude-opus-4-7 for complex logic.
* Reasoning — gpt-5.5-xhigh based on the latest LiveBench math and data analysis scores.
* Chat — gemini-3.1-pro currently leads the overall Chatbot Arena leaderboard.
* Open-source — NVIDIA-Nemotron-3-Nano-Omni-30B-A3B for edge-native, multimodal agent workflows.
* Value pick — DeepSeek-V4-Pro for coding inference given the massive price-to-performance advantage.
Deeper Dives
💼 Industry & Business
Google Partners with Pentagon
Google has formalized a multi-year deal to provide air-gapped AI infrastructure for classified Pentagon analysis — a full 180 from the company's 2018 policy banning AI military work. Make of that what you will.
� Twitter� Hacker News
Elon Musk in OpenAI Trial
Musk took the stand over 2017 internal emails about OpenAI's transition from nonprofit to capped-profit structure. The trial is dragging early governance fights into the open — specifically how the balance between safety research and building commercial products was negotiated behind closed doors.
� Twitter
ProfluentBio and Eli Lilly
ProfluentBio is using its generative "OpenCRISPR" platform to help Eli Lilly discover and design custom gene-editing proteins. It's a clear example of how proprietary generative AI can carve out a real competitive edge in long-cycle drug development.
� Twitter
YC Publishes Summer 2026 Startup Requests
Y Combinator's latest RFS makes one thing obvious: they're done with LLM wrappers. The focus is now on AI agents for physical-world tasks, biology, and high-efficiency compute. This is basically a map of where the next wave of venture money is heading.
� Twitter
🧠 Models & Research
DeepSeek V4 Pro Released
DeepSeek launched V4 Pro with a "Chat Prefix Completion" mechanism that caches system prompts to cut latency. Add a 75% price drop and 8x speed boost, and you've got a model that's actively pulling agentic workflows away from pricier alternatives.
� Twitter
NVIDIA Nemotron-3 Nano Omni
NVIDIA dropped a 30B multimodal model that uses "Adaptive 3-Bit" (A3B) quantization to stay sharp while dramatically shrinking its memory footprint for edge deployment. Text, vision, audio, video — all in a single reasoning loop.
� Twitter� Reddit
GitHub Security Vulnerability
Researcher Sagitz disclosed CVE-2026-3854, an RCE vulnerability that lets attackers execute unauthorized commands through specific repository workflow manipulations — bypassing traditional code checks entirely. GitHub is now doing a major internal review.
� Twitter
Talkie 13B Model
Researchers released Talkie, a 13B model trained exclusively on data from before 1931. The idea is to study how historical worldview and deliberate training data curation actually shape how an AI reasons. Niche, but genuinely interesting.
� Twitter� Reddit
🚀 Products & Launches
OpenAI Codex and GPT-5.5
OpenAI is aggressively scaling Codex and folding in GPT-5.5 to reduce developer friction. The play is enterprise-grade security and positioning OpenAI as the default infrastructure layer for agentic development.
� Twitter
Claude Code Push Notifications
Anthropic's Claude Code now pings your phone in real time — kick off a task in the terminal, walk away, and get notified when it's done. It makes the whole thing feel less like a chatbot and more like an actual teammate handling work in the background.
� Twitter
Funding & Deals
* ProfluentBio entered a multi-billion dollar partnership with Eli Lilly to use AI to design custom proteins for gene editing.
* Anthropic joined the Blender Foundation as a corporate patron, funding 3D creative infrastructure to secure long-term model integration.
Launches
* DeepSeek V4 Pro — High-performance, low-cost model featuring optimized multi-turn inference.
* NVIDIA Nemotron-3 Nano Omni — Compact multimodal model optimized for real-time edge agent workflows.
Closing thought: The shift from "chatbots" to "autonomous agents" is playing out at every level — from YC's new priorities to how Claude and OpenAI are retooling for speed. Between Google's Pentagon pivot and the mass migration toward DeepSeek, the industry isn't really about chatting with AI anymore. It's about commanding it to handle real-world, infrastructure-level work.