Skip to content
Source · Daily Brief

AI Daily Brief — 28 February 2025

Friday closed both Open Source Week and February with a thoroughgoing infra finale. DeepSeek capped its five-day batched release with 3FS — the Fire-Flyer File System hitting 6.6 TiB/s aggregate read on a 180-node cluster — plus Smallpond, a DuckDB-on-3FS data framework, both MIT-licensed. Humane’s AI Pin servers went dark at 3 PM ET, deleting all user data permanently. Microsoft confirmed via the M365 blog it would retire Skype on May 5, 2025 — 14 years after the $8.5-billion acquisition — pushing free users to Teams. Sam Altman went on CNBC and conceded GPT-4.5 is “giant and expensive,” “won’t crush benchmarks,” and that OpenAI staggered rollout because it is “out of GPUs.” Tencent’s Hunyuan Turbo S — a hybrid Mamba-Transformer “fast-thinking” model with 44% lower first-word latency and pricing at 0.8 / 2 yuan per million tokens — went live on Tencent Cloud API, pitched as a DeepSeek R1 rival. Tongcheng Travel became the first Chinese online travel agency to fully integrate DeepSeek across its products.

Top stories

  • DeepSeek closes Open Source Week with 3FS and Smallpond. Fire-Flyer File System hits 6.6 TiB/s aggregate read on a 180-node cluster; Smallpond is a DuckDB-on-3FS data framework. MIT-licensed on GitHub. Caps a five-day cadence — FlashMLA, DeepEP, DeepGEMM, DualPipe/EPLB, 3FS/Smallpond — covering kernels, comm, parallelism, file system, and data stack. via DeepSeek GitHub · via Xinhua recap
  • Humane AI Pin servers go dark at 3 PM ET. Calling, messaging, AI queries, and cloud access all cut. User data on Humane servers scheduled for permanent deletion. Follows HP’s $116M asset acquisition (Feb 18). via TechCrunch
  • Microsoft retires Skype on May 5, 2025. M365 blog post pushes free users to Teams; ends Skype Credit and subscription sign-ups for new customers. Fourteen years after the $8.5B acquisition. via Microsoft 365
  • Altman concedes GPT-4.5 is “giant and expensive.” On CNBC: “won’t crush benchmarks”; OpenAI staggered rollout because it is “out of GPUs.” API pricing $75 / $150 per million tokens — the most expensive flagship to date. Pricing/cost debate became the dominant frame entering March. via CNBC
  • Tencent Hunyuan Turbo S goes live on Tencent Cloud API. Hybrid Mamba-Transformer “fast-thinking” model. Sub-second replies, 44% lower first-word latency, priced 0.8 / 2 yuan per Mtok — pitched as a DeepSeek R1 rival. One-week free trial. via Business Standard
  • Tongcheng Travel becomes first Chinese OTA to fully integrate DeepSeek. The Chengxin Large Model embeds R1 across products, marking the first online travel agency in China to embed the model platform-wide. via Xinhua

Who shipped

DeepSeek closed OSW. Tencent shipped Hunyuan Turbo S commercial availability. OpenAI‘s Altman shipped a press positioning. Microsoft shipped a sunset notice. Anthropic, Google DeepMind, Meta, xAI, and Mistral made no dated launches; Anthropic’s Claude 3.7 Sonnet aftermath continued.

Open-source pulse

DeepSeek’s five-day cadence — kernels (FlashMLA, DeepGEMM), communication (DeepEP), parallelism (DualPipe + EPLB), storage and data (3FS + Smallpond) — covered the full infra stack required to serve frontier MoE reasoning models at scale. Tencent’s Turbo S launch added the second Chinese reasoning competitor to R1’s pricing pressure.

Money, infra & hardware

GPT-4.5 reception coverage solidified the cost/value framing line versus Claude 3.7 Sonnet’s $3/$15. MIT Tech Review noted Altman walking back the “win benchmarks” expectation; testers preferred GPT-4.5’s “vibe” for chat over GPT-4o while it trailed o3-mini on math and science. via MIT Tech Review

By the numbers

  • 6.6 TiB/s — 3FS aggregate read on 180-node cluster
  • 5 consecutive DeepSeek OSW releases
  • 14 years — Skype era ending May 5
  • 44% — Hunyuan Turbo S first-word latency reduction
  • 0.8 / 2 yuan per Mtok — Turbo S pricing
  • Most-mentioned company: DeepSeek

Compiled by AI Feed’s editor from verified web sources for 28 February 2025.