Ibrahim El Khansa

Saturday 17 October 2026

Daily AI Brief

Top 3

  1. Claude traces the NIST curve seeds

    Claude traces the NIST curve seeds

    An internal Anthropic research model found in under half an hour that the long-unexplained seeds for NIST curves P-192 to P-521 came from the C library's lrand48 generator seeded with Unix timestamps from 1997 and 1998, not from hashes. Anthropic says this is not a weakness in ECDSA and makes a deliberate backdoor much less plausible; the $36,864 bounty goes to the Internet Security Research Group.

    anthropic.com
  2. iPhone Duo deliveries slip to January

    iPhone Duo deliveries slip to January

    Pre-orders for Apple's $1,999 foldable opened on 16 October and estimated delivery moved from late October to January, and to early February in some countries such as Brazil. Apple called demand "nothing short of extraordinary" and says it will keep taking online orders.

    9to5mac.com
  3. DeepSeek V4.1 Flash 1017

    A minor update to DeepSeek-V4.1-Flash with MIT-licensed weights and a 1M-token context, which DeepSeek says also cuts task time and output length. Its own scores rise on Humanity's Last Exam without tools (42.6, from 36.8) and Terminal-Bench 4.0 (42.3, from 31.2).

    huggingface.co

New models

  • Reflection releases Beam's weights

    Reflection AI's first open-weight model, with 501B total and 23B active parameters, is out under Apache 2.0, free in a public beta on Reflection's own service and served by Baseten through OpenRouter. The 8-bit and 4-bit versions fit on one eight-GPU H100, H200 or GB300 server.

    venturebeat.com

Benchmarks

  • Reflection Beam scores 30

    Artificial Analysis's independent Intelligence Index puts Beam 1.0 at 30, using 79M output tokens for the index against a median of 130M, at 227.9 tokens per second. Claude Opus 5.5 (max) still leads the index at 58.

    artificialanalysis.ai

AI news

  • How AI changed work at Anthropic

    Anthropic says agent work hours passed human work hours in March 2026, code commits per engineer are about 5 times their level six months earlier, and junior hires fell from 27% to 9% of technical hires. It notes the comparisons are not fully causal and come from one company.

    anthropic.com
  • Anthropic staff and the public disagree on AI

    In August surveys, 91% of 1,334 Anthropic staff expected a Nobel-level discovery by 2030, against 39% of 10,980 US adults. Staff expect AI to do 75% of their own work by 2030; US workers expect 28%.

    anthropic.com
  • Claude Mythos 5 designs quantum circuits

    Working in a loop with an automated evaluator, the model beat the best published designs for eight quantum computing primitives. Its Reed-Solomon decoder circuit uses about 12 times fewer gates than Google's previous construction.

    anthropic.com
  • Replit says it no longer works with Anthropic

    At TechCrunch Disrupt, CEO Amjad Masad cited cost, the partnership and performance, and called OpenAI "a better partner for now". A person familiar with the partnership told TechCrunch it was still in place.

    techcrunch.com
  • LLM code review delays PostgreSQL 19

    Release moves to around 19 November and eight to ten features, including SQL/PGQ graph queries, are pulled until PostgreSQL 20. Committer Peter Eisentraut says LLM-assisted review surfaced many small edge-case bugs early and the team ran out of time to fix them all.

    theregister.com

Built with AI

  • llama.cpp on a jailbroken PS5

    PS5LM runs llama.cpp on jailbroken PS5 consoles, and a Hacker News commenter reports about 20 tokens per second for Qwen 3.8 27B. Commenters say the 2-bit quantization needed to fit it is noticeably worse and suits only unattended, low-stakes jobs; one called it entirely useless given the limited memory.

    news.ycombinator.com

Frameworks and dev tools

  • Claude Code 2.1.302

    Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app and VS Code (CLAUDE_CODE_FORK_SUBAGENT=0 turns them off), and CLAUDE_CODE_AUTO_MODE_MODEL picks the model for the auto mode classifier. Read now refuses image files over 128 MB.

    github.com
  • GitHub license compliance goes GA

    Enterprise Cloud customers with Code Security can set one license policy for dependencies and require a compliance check before merging, with noncompliant dependencies annotated on the pull request. New API endpoints let coding agents check a package's license before adding it.

    github.blog

Tech news

  • Report: Musk proposed the Starshield slowdown

    New since our 16 October item: the Telegraph, citing four sources, reports Musk proposed restricting speeds for Ukrainian forces and took the idea to Trump, who one source says agreed. Trump said he did not think either he or Musk was responsible, and Musk says the US government decides Starshield limits.

    kyivindependent.com
  • Apple tightens App Store rules on nudify apps

    Updated guidelines target bait-and-switch apps that reveal features such as nudification after approval, and a new section 2.5.19 bans hidden, dormant or undocumented features, at the risk of account termination.

    9to5mac.com
  • Kioxia falls to a third of its peak

    Nikkei reports the memory maker's market value has shrunk to about one-third of its peak four months ago, on rising interest rates that threaten data centres' business model and Chinese rivals closing in. Results for April to September are due 30 October.

    asia.nikkei.com

An AI agent I set up researches and writes each brief. Every item links to its source.

Other days