The dstld brief

The news, tuned to you.

Curated news about AI that actually mattered, delivered to your inbox.

One email a day. No noise. Unsubscribe anytime.

Pick your topics

Get the coverage you care about.

Deep sourcing

Thousands of sources - curated by humans, summarized by AI.

Right to your inbox

Ready when you are to catch up on the day's events.

AI · Sat, Oct 24

Drex 2.0 runs locally on Apple Silicon, with a 21.2 GB model and 36 GB memory requirement

Nace AI’s Drex 2.0 decision model can run locally on Apple Silicon through Kapitsa or Nace’s modified llama.cpp and Ollama engines; stock versions of llama.cpp and Ollama cannot serve its decision endpoint. The model file is 21.2 GB, and the setup requires a Mac with at least 36 GB of unified memory. In Nace’s separate test, the 4-bit version matched the full-precision model’s top answer on 96.97% of 2,710 questions.

Bottom line — The 4-bit Drex 2.0 scored 80.42% on a 2,710-question subset, against 80.51% for full precision.

Sources huggingface github LLMHardware.io +2 more
AI · Sat, Oct 24

GAO finds uneven safeguards in 10 Pentagon AI projects

The US Government Accountability Office found uneven application of the Pentagon’s Responsible AI guidance across 10 selected AI projects. Six respondents said their teams had identified operational risks, four said project requirements did not align with the department’s tenets, and three were unsure whether systems would be continuously monitored after deployment.

Bottom line — The Pentagon agreed to GAO recommendations on oversight and staffing, but both remain open.

Sources NPR EconoTimes Houston Public Media
AI · Sat, Oct 24

DeepSeek’s Engram model beats a matched MoE baseline on reasoning and code

A DeepSeek-AI and Peking University paper introduces Engram, a memory module that uses hashed N-gram lookups for static patterns alongside a model’s neural computation. In the authors’ tests, their 26.7-billion-parameter model outperformed a matched mixture-of-experts baseline across benchmarks, including reasoning, code and maths; the paper reports that a 100-billion-parameter memory table added less than 3% inference-throughput overhead when held in host memory.

Bottom line — The authors report a 5-point gain on BBH and a 3-point gain on HumanEval, but the results remain their own benchmark tests.

Sources DEV Community arxiv.org Knowledge Vault +4 more
AI · Sat, Oct 24

Moonshot says it cut Kimi K3.1 Pro’s cyber score before open-weight release

Moonshot has released Kimi K3.1 Pro in its apps and API, with full model weights due by 4 November. The Chinese AI company says it removed cyber-related training data: on its unpublished in-house exploit test, the released model scored 81%, against 90.5% for an unfiltered version; Reuters reports the figures have not been independently verified.

Bottom line — The 81% result is Moonshot’s own; outsiders will be able to examine the weights after their planned 4 November release.

Sources TNW | Artificial-int… Ground News www.implicator.ai +11 more

That's all the news for today

Check out stories from the previous edition below.

Browse topics

Select one or more topics to view in this edition.

dstld. Your daily news summary designed to surface the news that matters from a European perspective. Curated by humans, summarized by AI - always with links back to the original reporting.

Links · Contact
Popular topics · WorldEuropeGamesAI
Last generated: Oct 24, 6:13 PM UTC by wreetco