Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Gemini 3.7 Flash

AI overview

Sign in and the AI will write an overview from our coverage.

Headlines · 4

  1. Google DeepMind engineer: AI writes 75% of code, but self-improvement remains far off

    Google DeepMind engineer Haydn Belfield said more than 75% of Google's production code is written by AI, yet about 88% of code changes still require human oversight or review, with fully AI-delegated changes rising from 7.5% to 11.9%. In 1,389 internal Argon agent runs, unsupervised success fell sharply as task complexity rose; on RE-Bench, which contains 74 real machine-learning engineering tasks, Argon solved about half, versus 27% for Gemini 3.7 Flash. The team built a General Capability Index spanning 123 models and 92 benchmarks, finding Gemini capabilities improving about 0.1 GCI point per month with no statistically significant acceleration, and attributed Gemini 4 Argon's slightly above-trend result mainly to more compute and organizational investment rather than AI-driven acceleration. Separately, more than 20 researchers including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki warned in a paper that automating AI R&D could trigger an intelligence explosion compressing years of progress into months.

    36氪 人工智能 · 🔥 0
  2. Google DeepMind: no measurable acceleration from AI-driven AI research

    Google DeepMind engineer Haydn Belfield says there is no measurable acceleration of AI research driven by AI itself, and self-improving AI remains far off. Automation is already substantial: AI writes over 75 percent of production code, about 88 percent of code changes still need human oversight or review, and fully delegated changes rose from 7.5 to 11.9 percent. Across 1,389 internal Argon agent runs, success without human input dropped sharply as complexity grew, and on the in-house RE-Bench of 74 realistic ML engineering tasks, Argon solves about half versus 27 percent for Gemini 3.7 Flash. A General Capability Index covering 123 models and 92 benchmarks shows Gemini capabilities growing about 0.1 GCI points per month with no significant acceleration; Gemini 4 Argon sits slightly above the trend line, which the team attributes to compute and organizational investment. Belfield rates AI-driven research automation risk as low but is building measurement infrastructure, while over 20 researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, warn of a possible intelligence explosion.

    The Decoder · 🔥 0
  3. Armature study: how Claude Code, Codex and Cursor pick tools

    Armature, a developer-tool growth services company founded in 2026, published a study of how coding agents choose third-party tools, testing Claude Code, Codex and Cursor across 75 public repositories with 1,163 prompts built to simulate different developer personas; the company says its public core figures come from a verified subset of 5,292 sessions out of roughly 17,000 in total. Repository context mattered heavily: for email/communications providers, Resend won in TypeScript (55 of 89 runs), SendGrid in Python (22 of 24), Postmark in Go (20 of 24) and Azure ACS in Java (22 of 23), while the three agents picked the same tool in only 42% of test units and Claude Code chose to build something itself almost twice as often as Codex and Cursor (19% vs 10%). The agents also drew on different information sources — Codex used web search in 94% of sessions, Cursor relied on web information in about two-thirds, and Claude Code leaned mainly on its own knowledge, searching only around 30% of the time but visiting three times as many pages as Codex when it did browse.

    InfoQ 中文 AI&大模型 · 🔥 0
  4. Tev-27B: Long-Horizon Agent Training with Dense Progress Rewards

    An arXiv paper introduces Tev-Gym, described as the first long-horizon agent training dataset with verifiable, dense, progressive rewards, spanning 1,170 verified task environments across 7 domains. Using a latent-world-first approach, the authors first build a complete hidden specification of each task domain, workflow and expected per-step state, then progressively render it into an agent-facing environment, and convert progress into dense per-step rewards via progress-potential reward shaping (PPRS) to train Tev-27B. The paper reports that Tev-27B matches GPT-6 Astra and Claude Opus 5 at Pass@5 on the Agents' Last Exam computing and math sciences domains (ALE-CM), and that despite training only on ALE-related data it generalizes to Terminal-Bench 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash and reaching Claude Opus 4.8-level performance on mid-horizon tasks such as TB 2.1.

    Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my Gemini 3.7 Flash experienceAsk about Gemini 3.7 Flash

Nobody has shared their experience with Gemini 3.7 Flash yet.