Tokens & Signals

Thursday, March 5, 2026

claude-opus-4-6-thinking-autogemini-3.1-pro-preview-highltx-2.3gemini-3.1-prophi-4-reasoning-vision-15bjido-2.0anthropicopenaigooglenvidialightricksamazonicon-ai-admakersafety-theaterinference-speedupstorage-clustercoding-agentsreasoningvideo-generationagentic-workflowsproduct-market-fitinference-latencyon-device-aidario amodeisam altmanjensen huangdeedydastanishqkumar07ns123abclevelsiocgtwts
Tokens & Signals for 3/5/2026. We scanned ~605 Twitter accounts, 13 subreddits (72 posts), Hacker News (9 stories), 10 newsletters, 10 podcasts, and leaderboard data for you. Estimated reading time saved: ~24 hours.

TLDR

* Anthropic CEO Dario Amodei publicly blasted OpenAI's Pentagon deal as "safety theater," calling their messaging "straight-up lies." x.com/ns123abc/status/2029301113493639447

* ChatGPT uninstalls surged 563% after the military contract news broke and the internet lost its mind. reddit.com/r/OpenAI/comments/1rlae0s/chatgpt_un...

* Google dropped a Workspace CLI that lets AI agents tap directly into Gmail, Drive, and Docs without torching your context window. x.com/cgtwts/status/2029504941489057978

* A new "Speculative Speculative Decoding" (SSD) algorithm hits 2x inference speedups by parallelizing token verification instead of doing it one step at a time. x.com/tanishqkumar07/status/2029251146196631872

* Five 19-year-olds built a 30PB storage cluster for $500k — they're claiming it's 40x cheaper than AWS S3 for pretraining. x.com/deedydas/status/2029391960159740003

* Nvidia is reportedly pulling back from its deep investment ties with major labs like OpenAI and Anthropic. news.ycombinator.com/item?id=47256781

Best to Build With Today

* Codingclaude-opus-4-6-thinking-auto is the current leader for complex architectural tasks.

* Reasoninggemini-3.1-pro-preview-high consistently tops the charts for high-stakes logic and math.

* Video generationLTX-2.3 is the new open-source standard for high-detail, local production.

* Chatgemini-3.1-pro remains the overall ELO king for everyday assistance.

* Open-sourcePhi-4-reasoning-vision-15B is the best-in-class for a 15B model that manages its own "thinking" compute.

Deeper Dives

💼 Industry & Business

Anthropic vs. OpenAI: The Pentagon Rift

Dario Amodei didn't hold back. His leaked memo torched OpenAI's new Department of Defense deal, calling it "80% safety theater" and accusing Sam Altman of "gaslighting" the industry. Anthropic's refusal to sign, he says, was about preventing real-world AI abuse — not optics.

* Why it matters: The divide exposes a hardening ideological split over the ethical price of federal partnerships.

* 📱 Twitter · 🔶 Hacker News

ChatGPT Uninstall Surge

SensorTower data shows a 563% spike in ChatGPT uninstalls right after the Pentagon deal dropped. Users are migrating to Claude, which hit #1 on the US App Store — proof that user sentiment has become a real competitive variable, not just a PR concern.

* Why it matters: For the first time, ethical perception is directly hitting the bottom line of the dominant AI player.

* 💬 Reddit

Nvidia Pulls Back

Jensen Huang signaled that Nvidia will likely stop making direct investments in OpenAI and Anthropic as both companies head toward IPOs. It's a notable shift for the "central bank of AI" — moving from financier to straight-up hardware supplier.

* Why it matters: Hardware labs are seeking more autonomy, and chipmakers are focusing on broader ecosystem dominance.

* 🔶 Hacker News

Icon AI Admaker Bankruptcy

Icon AI Admaker filed for bankruptcy. Despite massive hype and a $12M domain purchase, they never found product-market fit — a brutal reminder that a flashy domain name is not a business.

* Why it matters: The "AI-as-a-feature" bubble is popping for companies that lack a real moat.

* 📱 Twitter

🧠 Models & Research

Microsoft's Phi-4-Reasoning-Vision 15B

Microsoft open-sourced a 15B parameter model built on a mid-fusion architecture with SigLIP-2. The clever bit: it intelligently toggles between deep thinking mode and instant-answer mode depending on what the task actually needs.

* Why it matters: It gives you a highly capable, locally-runnable reasoning engine for vision-heavy agents — without needing a data center.

* 💬 Reddit · 📱 Twitter

Speculative Speculative Decoding (SSD)

Researchers dropped the SSD algorithm, which breaks the sequential bottleneck in standard speculative decoding by pre-speculating multiple verification outcomes at once. The result: a clean 2x speedup.

* Why it matters: Inference latency is the biggest barrier to scaling agentic applications, and this is the kind of algorithm-level fix that actually moves the needle.

* 📱 Twitter

🚀 Products & Launches

Google Workspace CLI

Google launched a CLI giving AI agents direct, programmatic access to Workspace — Gmail, Drive, Docs, all of it. Instead of dumping a massive raw document into a context window and hoping for the best, agents can now edit structured data directly.

Why it matters: This drastically lowers the barrier for building agents that can actually do* work inside enterprise software.

* 📱 Twitter · 🔶 Hacker News

OpenAI Symphony

OpenAI introduced "Symphony," an orchestration layer that automates the software development lifecycle by reading tickets from boards and spinning up dedicated agent workspaces.

* Why it matters: It moves the agent game from simple chat to autonomous workflow management.

* 📱 Twitter

Launches

* LTX-2.3 — Lightricks' latest video model brings sharper detail, native portrait support, and cleaner audio.

* Jido 2.0 — An Elixir-based agent framework built for reliability and complex multi-agent supervision.

AI Twitter Recap

* @deedydas on extreme infra: "Five 19-year-olds built a 30PB storage cluster for $500k, cutting AWS costs by 40x. This is the reality needed for modern training." x.com/deedydas/status/2029391960159740003

* @tanishqkumar07 on SSD: "Why wait for verification? SSD allows draft models to pre-speculate for multiple outcomes, breaking the sequential bottleneck." x.com/tanishqkumar07/status/2029251146196631872

* @ns123abc on the Anthropic memo: "The 1,600-word memo from Dario is a scorched-earth takedown of OAI's 'safety' branding." x.com/ns123abc/status/2029301113493639447

* @levelsio on Icon bankruptcy: "A $12M domain doesn't buy product-market fit. Another AI startup learns that the hard way." x.com/levelsio/status/2029575116187730300

* @cgtwts on the Workspace CLI: "Stop writing curl calls against REST docs. GWS gives agents structured JSON and instant access to your workspace." x.com/cgtwts/status/2029504941489057978

Closing thought: The industry is entering a "show me the product, not the philosophy" phase — where user sentiment and real-world efficiency are finally starting to matter more than lab hype.