Skip to content
Source · Daily Brief

AI Daily Brief — 20 January 2025

Inauguration day delivered the single most consequential 24 hours in open-source AI history alongside a wholesale rewrite of US AI policy. DeepSeek released R1 and R1-Zero under MIT, plus six distilled dense models across the Qwen and Llama families, at API prices roughly 27x below OpenAI o1 — instantly redefining the cost of frontier reasoning. Moonshot AI dropped Kimi k1.5 the same day. Hours later, Trump revoked Biden’s AI Safety Executive Order 14110, declared a national energy emergency citing AI datacenter demand, signed a 75-day TikTok enforcement pause, and stood up the Department of Government Efficiency.

Top stories

  • DeepSeek releases R1 and R1-Zero under MIT license. The 671B-parameter MoE (37B activated) reasoning model scores 79.8% pass@1 on AIME 2024 (versus o1-1217’s 79.2%), 97.3% on MATH-500 (vs 96.4%), and 96.3 percentile on Codeforces (Elo ~2,029, “Candidate Master” range) — putting an open model at parity with OpenAI o1 for the first time. The MIT license explicitly permits commercial use, modification, redistribution, and distillation for training other LLMs. via DeepSeek · via GitHub
  • R1-Zero shows reasoning emerges from pure RL with no SFT. Trained on V3-Base with GRPO and accuracy + format rewards alone — no supervised fine-tuning step — R1-Zero climbed AIME 2024 pass@1 from 15.6% to 71.0% (86.7% with majority voting), with self-verification and reflection behaviours emerging spontaneously. R1 then applied a cold-start SFT pass before RL to fix readability and language-mixing. The paper would land on arXiv (2501.12948) two days later. via arXiv
  • Six distilled dense checkpoints open-sourced across Qwen2.5 and Llama. Qwen-based 1.5B/7B/14B/32B (Apache 2.0) and Llama-based 8B (3.1) and 70B (3.3-Instruct), all fine-tuned on 800k samples generated by R1. Distill-Qwen-7B beats QwQ-32B-Preview on AIME 2024 (55.5%); Distill-Llama-70B hits 70.0% AIME and 94.5% MATH-500 — new SOTA for dense reasoning models. via Hugging Face
  • R1 API priced at $0.55 / $2.19 per million input / output tokens — roughly 27x cheaper than OpenAI o1. The combination of MIT weights and a 27x price-anchor reset frontier-reasoning economics overnight; it would, six days later, drive DeepSeek to #1 on the US App Store and wipe roughly $600 billion off NVIDIA’s market cap in a single session. via DeepSeek API docs
  • Trump revokes Biden’s AI Safety Executive Order 14110. Within hours of inauguration, Trump signed “Initial Rescission of Harmful Executive Orders and Actions,” rescinding 75+ Biden actions including EO 14110 (Safe, Secure, and Trustworthy Development of AI, Oct 2023), terminating its safety-testing, dual-use reporting thresholds, and equity provisions. via Wikipedia · via Cybersecurity Dive

Who shipped

DeepSeek + Moonshot AI (Kimi k1.5, a multimodal RL-trained reasoning model with long-CoT scaling and a “long2short” transfer technique) carried the day on the lab side. OpenAI, Anthropic, Google DeepMind, Meta, and xAI made no product releases — but were positioned for the Stargate reveal the next morning.

Money, infra & hardware

Trump’s national energy emergency proclamation explicitly cited AI-datacenter demand, enabling federal agencies to waive environmental permitting review and expedite gas, nuclear, and transmission permits — laying the legal foundation for the Stargate-style buildout announced 24 hours later. The previously-flagged $20B DAMAC Midwest/Sunbelt datacenter commitment was Trump’s pre-Stargate marker. The TikTok 75-day enforcement EO formalised Sunday’s commitment.

Research & papers

“Reasoning Language Models: A Blueprint” (2501.11223, Besta et al., ETH Zurich) hit arXiv as the most-cited reference of the day — a 50-page modular framework decomposing reasoning LLMs into schemes, operators, models, and pipelines, perfectly timed for the new wave R1 was about to unleash.

By the numbers

  • 79.8% — R1 AIME 2024 pass@1 (matches o1)
  • $0.55 / $2.19 — R1 API per-million-token pricing
  • ~27x — price gap below OpenAI o1
  • 75+ Biden actions rescinded by Trump
  • 7 open-weight model releases from DeepSeek alone (R1, R1-Zero, six distilled)
  • Most-mentioned lab: DeepSeek

Compiled by AI Feed’s editor from verified web sources for 20 January 2025.