# Daily Summary for 2026-03-02 ## 2026-03-02 16:01:13 # AI Digest - March 2, 2026 ## Industry News - **Cal AI acquired by MyFitnessPal**: Two 17-year-old high school students built Cal AI to $50M+ ARR in 18 months, helping millions lose weight with AI-powered calorie tracking. MyFitnessPal acquired the company while one founder was still in school. [link](https://x.com/zach_yadegari/status/2028473704359874652) - **Qwen 3.5 Small Models (0.8B-9B) released**: Alibaba's Qwen3.5-9B beats their previous 235B-22AB model and GPT-OSS 120B on GPQA, showing strong reasoning about complex scientific topics—while running on a laptop. [link](https://x.com/Alibaba_Qwen/status/2028478905229639684) - **Pentagon designates Anthropic as supply chain risk**: DoW banned contractors from working with Anthropic after the company refused to remove red lines on autonomous weapons use, marking an unprecedented government retaliation against a US AI company. [link](https://x.com/deanwball/status/2028464782622195992) - **OpenAI raises $110B from Cerebras partnership**: OpenAI and Cerebras signed multi-year agreement to deploy 750 megawatts of wafer-scale systems, with Amazon, NVIDIA joining the cap table in what may be the largest AI infrastructure deal to date. [link](https://x.com/andrewdfeldman/status/2028481771679961266) ## Tips & Techniques - **Use prompt caching as core infrastructure**: Claude Code is built around prompt caching—static content first, dynamic last—to keep costs manageable. One engineer noted cache operations account for 99% of visible costs but enable the entire UX. [link](https://x.com/law_wu/status/2028493125283365028) - **Small models getting scary good**: Qwen's 9B model beating 120B models shows distillation and architecture improvements matter more than raw parameter count for many real-world tasks. [link](https://x.com/nathanhabib1011/status/2028498736230625360) - **Local AI infrastructure matters now**: With cloud AI services facing geopolitical and reliability risks, running models locally on consumer hardware (even 9B models at 2,629 tok/s) provides genuine independence. [link](https://x.com/danielhanchen/status/2028496924815581508) ## New Tools & Releases - **Telegram bots now support streaming responses**: All Telegram chatbots can stream responses in real-time, improving UX for AI assistants built on the platform. [link](https://x.com/durov/status/2028493506273018065) - **Claude Code gets Figma integration**: You can now generate Figma design files from Codex by saying "send my app to Figma"—arguably more useful than implementing frames from scratch. [link](https://x.com/derrickcchoi/status/2028498260084555783) - **skilld: Turn any NPM package into agent skill**: New tool crawls docs, release notes, and GitHub issues to generate versioned best practices your AI agent actually follows—no vendors, no maintenance cost. [link](https://x.com/harlan_zw/status/2028482370933068273) - **Ginkgo Cloud Lab service launches**: GPT-5 can now order experiments from Ginkgo's autonomous lab, bridging AI agents directly to wet lab automation for biological research. [link](https://x.com/jrkelly/status/2028495941184168429) ## 2026-03-02 16:01:14 ## Research & Papers - **WarpSpeed beats decade of GPU kernel optimization**: DoubleAI's AI system using deep search and verification comprehensively surpassed human experts on GPU kernel optimization—a task only a handful of experts globally can do well. [link](https://x.com/YoavLevine/status/2028500436848660973) - **DECODE: Universal deconvolution framework**: New deep learning framework unifies deconvolution across transcriptomics, proteomics, and metabolomics, achieving top performance across 15 datasets where existing tools failed—especially for metabolomics with only hundreds of features. [link](https://x.com/bravo_abad/status/2028499248774291551) - **BullshitBench v2 shows models aren't improving at sycophancy resistance**: One of few benchmarks where models generally aren't getting better (except Claude), revealing persistent challenges in AI truthfulness and calibration. [link](https://x.com/petergostev/status/2028495417785430496) --- *Curated from 500+ tweets across AI research, engineering, and industry communities* --- ## Emerging Trends ✨ **Cal AI MyFitnessPal Acquisition** (185 mentions) - NEW Cal AI, a calorie tracking app built by teenage founders using AI, has been acquired by MyFitnessPal after reaching $50M ARR in 18 months. The story is generating significant attention as an example of AI-enabled entrepreneurship and distribution-first growth strategy. ✨ **Qwen 3.5 Release** (95 mentions) - NEW Alibaba's Qwen released small multimodal models (0.8B, 2B, 4B, 9B) with 262K context, built-in thinking capabilities, and strong performance benchmarks. The release is generating excitement in the open source AI community for on-device and edge computing applications. 🔥 **Anthropic DoD Dispute** (520 mentions) - RISING The Department of War has designated Anthropic as a supply chain risk after the company refused contract terms around autonomous weapons and surveillance. This is generating massive debate about AI ethics, military use, and government overreach, with OpenAI offering alternative terms. 🔥 **OpenAI $110B Funding Round** (340 mentions) - RISING OpenAI announced a $110 billion funding round from Amazon, NVIDIA, and SoftBank, one of the largest financing rounds in tech history. Discussion focuses on burn rates, path to profitability, and competitive positioning against Anthropic. 🔥 **Iran Conflict Escalation** (210 mentions) - RISING Military strikes on Iran are underway with missiles reported over Dubai and other locations. The situation is being closely monitored across social media, with people building real-time OSINT dashboards and discussing the geopolitical implications for AI infrastructure.