1,853 source-backed ai facts - each attributed to a named person and a number, linked to the original source. Page 19 of 19.
“The original timeline gave providers of high‑risk systems 24 months from entry into force, until 2 August 2025, to comply.”
“A study by McKinsey found that companies using AI for automation saw productivity gains of up to 30%.”
“amassing over 200 million users by October 2024”
“The paper has been accepted as a Spotlight presentation at the top-tier machine learning conference, ICML 2025, and the code and dataset are now fully open-source.”
“model: 2012, unit: 44”
“In the pre-change scene the colourful insect had two stripes - one on each wing, and on the post-change, there was just one.”
“a $20/month Plus plan, and a $200/month Pro tier”
“The dataset includes a wide range of failure logs collected from 127 LLM Multi-Agent systems, which were either algorithmically generated or hand-crafted by experts to ensure realism and diversity.”
“an iceberg scene with five penguins on it”
“Launched in 2016”
“Even the best-performing single method achieved an accuracy of only about 53.5% in identifying the responsible agent and a mere 14.2% in pinpointing the exact error step.”
“Friday 11 June”
“Launched in February 2024 Gemini”
“achieving 99.9% localization accuracy in unseen home environments”
“Mr Blair will provide strategic advice regarding investments in environmentally friendly or helpful technologies.”
“Bard which was introduced in March 2023”
“reducing overall trajectory error to approximately 2%”
“Officially launched in November 2023 Grok”
“The paper has been accepted as a Spotlight presentation at the top-tier machine learning conference, ICML 2025”
“Founded in 2024 in Hangzhou”
“The dataset includes a wide range of failure logs collected from 127 LLM Multi-Agent systems”
“first introduced in March 2023”
“Using the Who&When dataset, the paper designs and assesses three distinct methods for automated failure attribution”
“Claude 3 family debuting in March 2024”
“Experiments were conducted in two settings: one where the LLM knows the ground truth answer to the problem the Multi-Agent system is trying to solve (With Ground Truth) and one where it does not (Without Ground Truth).”
“three specialised models”
“The paper also introduces two key training strategies to further improve long-context performance:”
“DeepSeek-V3, trained on a cluster of 2048 NVIDIA H800 GPUs, serves as a compelling case study...”
“Table 1 in the paper compares the per-token KV cache memory footprint of DeepSeek-V3, Qwen-2.5 72B, and LLaMA-3.1 405B. DeepSeek-V3 achieves a remarkable reduction, requiring only 70 KB per token, significantly lower than LLaMA-3.1 405B’s …”
“DeepSeek has pioneered the use of FP8 mixed-precision training for a large-scale MoE model. Despite NVIDIA’s Transformer Engine supporting FP8, DeepSeek-V3 marks a significant step as the first publicly known large model to leverage FP8 fo…”
“During EP parallelism, tokens are scheduled using fine-grained FP8 quantization, reducing communication volume by 50% compared to BF16, thereby significantly shortening communication time.”
“DeepSeek currently utilizes the NVIDIA H800 GPU SXM architecture (Figure 2), which, while based on the Hopper architecture similar to the H100, features reduced FP64 compute performance and NVLink bandwidth (400 GB/s down from 900 GB/s in …”
“DeepSeek-V3, trained on a cluster of 2048 NVIDIA H800 GPUs, serves as a compelling case study demonstrating how a synergistic approach between model design and hardware considerations can overcome these limitations.”
“DeepSeek-Prover-V2–671B, a model boasting 671 billion parameters.”
“reached an impressive 88.9% pass ratio on the MiniF2F-test”
“successfully solved 49 out of 658 problems from PutnamBench”
“ProverBench, a new benchmark dataset comprising 325 problems”
“ProverBench includes 15 problems formalized from recent AIME (American Invitational Mathematics Examination) competitions (AIME 24 and 25)”
“The remaining 310 problems are drawn from curated textbook examples and educational tutorials”
“DeepSeek AI is releasing DeepSeek-Prover-V2 in two model sizes to cater to different computational resources: a 7B parameter model and the larger 671B parameter model.”
“features an extended context length of up to 32K tokens”
“DeepSeek-Prover-V2–7B, a smaller 7B parameter model.”
“88.9% pass ratio on the MiniF2F-test”
“49 out of 658 problems from PutnamBench”
“SRPO has achieved impressive results on the AIME24 (50) and LiveCodeBench (41.6) benchmarks”
“SRPO achieves this level of performance with only one-tenth of the training steps required by R1-Zero.”
“nearly 50% of the sampled groups within a batch produced identical rewards.”
“By making GRPO 10 times more efficient, we can accelerate the development of AI models and make them more accessible to a wider range of applications.”
“A 10x improvement in training efficiency could make reinforcement learning from human feedback accessible to much smaller teams and organizations.”
“April 15, 2025”
“up to eight times faster than DeepSeek‑R1”
“200 tokens per second on consumer‑grade GPUs – a staggering 50 times faster than human reading speed”
“smaller 9B parameter versions of both GLM‑4 and GLM‑Z1 models”
