Skip to content
Source · Daily Brief

AI Daily Brief — 21 August 2025

DeepSeek’s day. V3.1 formal launch with chip-sovereignty bombshell buried in release notes. Forbes verified Grok privacy disaster — 370K chats on Google. IBM+NASA shipped Surya. OpenAI still in GPT-5 recovery arc.

Top stories

  • DeepSeek officially launches V3.1. 685B-parameter MoE (37B active per token), 128K context, hybrid “thinking” + “non-thinking” modes toggled via chat template. Weights for Base and main on Hugging Face. Followed quiet WeChat tease on Aug 19. via DeepSeek API Docs
  • DeepSeek-V3.1 posts 71.6% on Aider, beating Claude Opus 4 on coding. 71.6% on Aider polyglot programming test (edges out Claude Opus 4). >40% jump over V3 and R1 on SWE-bench. Total testing cost ~$1 — roughly 68x cheaper than Claude Opus for the same eval. via TechTalks
  • DeepSeek signals upcoming Chinese AI chips with UE8M0 FP8 weight format. Buried in V3.1 release notes: model trained using “UE8M0 FP8” scale data format — 8-bit floating-point scheme with 0 mantissa bits, all 8 in exponent. DeepSeek said format is tailored for “the next-generation domestically built chips that will be launched soon”. Read by analysts as deliberate signal Huawei Ascend / Cambricon / Moore Threads accelerators about to arrive. via SCMP
  • DeepSeek V3.1 priced at $0.56 / $1.68 per M tokens. $0.56/M input (cache miss), $0.07 cache hit, $1.68/M output. Made model default behind deepseek.com chat. Pricing keeps it an order of magnitude below frontier US labs for both modes. via Hugging Face
  • Forbes investigation: ~370,000 Grok chats indexed on Google. Between Aug 20-22 verified that xAI’s Grok chatbot inadvertently exposed 370K+ user conversations to Google and other search engines. Grok’s “Share” button created public URLs without robots/noindex tags or any disclaimer. Indexed transcripts included medical questions, business data, at least one password, prompts violating xAI’s ToS (drug synthesis, an assassination plot against Musk). via Fortune

Who shipped

DeepSeek shipped V3.1 + chip-sovereignty signal. xAI shipped a privacy disaster. IBM + NASA shipped Surya.

Open-source pulse

IBM and NASA released Surya, open foundation model for solar weather, on Hugging Face — Sanskrit for “Sun”. Trained on high-resolution solar observation data to predict space-weather effects on Earth-orbiting and ground-based technology. Released on HF as community foundation, intended to be fine-tuned by domain researchers worldwide. via IBM Newsroom

Money, infra & hardware

Bloomberg/SCMP framed V3.1 as long-awaited successor — missing “R1” label conspicuously gone. SCMP: speculation V3.1’s hybrid reasoning mode is company’s replacement for standalone R2, folding reasoning into base series rather than shipping it as separate flagship. via SCMP

Quiet corners

GPT-5 fallout continues — OpenAI doubling rate limits, restoring GPT-4o after user revolt. Auto-router that broke for hours making GPT-5 “appear way dumber”. Reddit revolt over GPT-4o retirement (“losing a trusted friend”). NVIDIA Q2 FY26 print looms — earnings due Aug 27 with H20-to-China question front and centre. EliseAI $250M Series E from Aug 20 anchored application-layer funding.

By the numbers

  • 685B / 37B / 128K — V3.1 total / active / context
  • 71.6% / +40% / ~$1 / 68x — V3.1 Aider / SWE-bench gain over V3 / test cost / cheaper than Opus
  • $0.56 / $1.68 — V3.1 per M input / output tokens
  • UE8M0 FP8 / 0 mantissa / 8 exponent — V3.1 weight format / bits
  • ~370,000 — Grok chats indexed on Google

Compiled by AI Feed’s editor from verified web sources for 21 August 2025.