Tokens & Signals for 3/6/2026. We scanned ~605 Twitter accounts, 13 subreddits (0 posts), Hacker News (10 stories), 10 newsletters, 10 podcasts, and leaderboard data for you. Estimated reading time saved: ~19 hours.
gpt-5.4gpt-5.4-xhighclaude-opus-4-6-thinking-autogemini-3.1-pro-previewphi-4-reasoning-vision-15bflashattention-3openaianthropiccursortogether-aiproton-mailfbimicrosoftvllm-projecttencentvelacomputer-usecoding-agentsmultimodalityon-device-aivibe-codingsupply-chain-riskprivacymodel-infrastructuresamalioronaiomarsar0aakashgupta
Best to Build With Today
* Coding — gpt-5.4-xhigh (per Artificial Analysis coding leaderboards).
* Reasoning — claude-opus-4-6-thinking-auto (leads LiveBench in reasoning performance).
* General Chat — gemini-3.1-pro-preview (tops the Chatbot Arena and Artificial Analysis rankings).
* Efficiency — phi-4-reasoning-vision-15b (best for local/edge multimodal tasks).
Deeper Dives
🚀 Products & Launches
OpenAI Releases GPT-5.4
OpenAI's latest flagship is rolling out now, and the headline feature is native computer-use — meaning it can actually interact with desktop apps directly. Pair that with a 1M token context window and you've got something built for serious production workloads, not just demos.
Why it matters: We're moving from API-based text generation to direct desktop automation.
� Hacker News
news.ycombinator.com/item?id=47265045
Cursor Launches Cloud Agents
Cursor is moving AI agent execution off your machine and into the cloud. The big win here is breaking past the local context bottleneck — now the tool can manage large software projects autonomously without choking on your laptop.
Why it matters: It turns your editor into an autonomous "factory" for building software.
� Twitter� Hacker News
x.com/LiorOnAI/status/2029648291563196489
Anthropic Makes Claude Memory Free
Claude's persistent memory — the feature that lets it remember your preferences and project details across sessions — is now free for everyone.
Why it matters: It's the ultimate quality-of-life update, making the AI feel like a partner that actually knows how you work.
x.com/aakashgupta/status/2029783984247648514
💼 Industry & Business
Anthropic Designated Supply Chain Risk
The U.S. Department of War has labeled Anthropic a "supply chain risk" after a dispute over safety guardrails for military applications. Anthropic is planning to fight it in court.
Why it matters: It's an unprecedented use of procurement power to force alignment from a top AI lab.
� Twitter� Hacker News
x.com/AnthropicAI/status/2029719864533721481
Together AI Hits $1B ARR
Together AI has crossed the $1B Annual Recurring Revenue milestone at a $7.5B valuation — solid proof that demand for managed, performant model infrastructure is very real.
Why it matters: The "picks and shovels" companies are becoming the biggest winners in the AI gold rush.
Proton Mail and the FBI
Proton Mail complied with a Swiss court order to hand over recovery metadata, which the FBI then used to identify a protester. They're clear that email contents stayed encrypted and out of reach — but this is a good reminder that even privacy-first tools have limits when international courts get involved.
Why it matters: Privacy claims hit a brick wall when they meet international legal jurisdictions.
� Hacker News
news.ycombinator.com/item?id=47267628
🧠 Models & Research
Microsoft Releases Phi-4-Reasoning-Vision 15B
Microsoft's new 15B model uses a "mid-fusion" architecture to bridge vision and language, and it's smart about when to trigger chain-of-thought — only when it's actually needed. That makes it surprisingly efficient for edge devices.
Why it matters: Smaller, highly capable models are the key to bringing real reasoning power to local hardware.
x.com/omarsar0/status/2029926242640912429
vLLM Triton Attention Backend
The vLLM project shipped a Triton-based attention backend that matches FlashAttention-3 performance in just 800 lines of code — with massive speedups on AMD hardware to boot.
Why it matters: Easy multi-hardware support is critical for keeping production costs down.
x.com/vllm_project/status/2029919035924828234
🔥 Takes & Drama
Filtering "Vibe-Coding" Pull Requests
A new protocol is gaining ground to help maintainers automatically reject low-quality, AI-generated "vibe-coding" contributions before they clog up the queue.
Why it matters: As AI-generated spam scales up, open-source maintainers need automated gatekeeping just to stay sane.
� Hacker News
news.ycombinator.com/item?id=47267947
Launches
AI Twitter Recap
Closing thought: The industry is hitting a real inflection point. We're not just chatting with models anymore — we're handing off desktop tasks to them in the cloud, and the tools are starting to feel genuinely factory-level. Stay tuned.