cuppa

livebrewed 03:01 AM
← Jul 18

Sunday

jul19

2026

Jul 20 →

the brief

Today’s pulse tilts practical: Simon Willison ships a SQLite query plan explainer, Anthropic tightens Claude Code’s contracts (manual verify/review, extended limits), and PrismML debuts a 3.9GB binary 27B. Research leans agentic and robust: ScholarGym, AutoResearchBench, TMLR on VLM TTA and episodic memory, plus fresh evidence on classroom AI dependence.

the poursit · sip · 12 items

pulse

(04)
  • simonw/blog· AnalysisJul 18, 05:19 PM

    SQLite Query Explainer from Simon Willison

    A small web tool that explains SQLite EXPLAIN QUERY PLAN output in plain language, making it easier to spot missing indexes, costly scans, and join issues.

    SQLite Query Explainer — <p><strong>Tool:</strong> <a href="https://tools.simonwillison.net/sqlite-query-explainer">SQLite Query Explainer</a></p> <p>Julia Evan's, in <a href="https://jvns.ca/blog/2026/07/17/learning-about-running-sqlite/">Learning a few things about running SQLite</a>:</p> <blockquote> <p>Maybe one day I’ll learn to read a query plan.</p> </blockquote> <p>Big same.... which inspired me to <a href="https://github.com/simonw/tools/pull/299#issue-4919268017">have Fable build</a...

    signal 6hype 1sqlitetool_releasequery_planlaunchsource ↗
  • anthropics/claude-code· First-partyJul 19, 02:56 AM

    Claude Code stops auto-running skills

    v2.1.215 removes automatic /verify and /code-review runs; you now invoke them explicitly, reducing surprises and giving teams tighter control over code changes.

    v2.1.215 — What's changed Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them

    signal 7hype 0release_notesclaude_codeproduct_updatelaunchsource ↗
  • anthropicbot.bsky.social· Bluesky mirror · @anthropicaiJul 18, 04:04 PM

    Claude Code keeps higher limits

    Anthropic is keeping the 50% higher weekly usage limits for Pro, Max, Team, and seat-based Enterprise through August 19, easing capacity pain for heavier users.

    We're also keeping Claude Code weekly limits 50% higher, now through August 19, for all Pro, Max, Team, and seat-based Enterprise users.

    signal 6hype 1product_updateclaude_codeusage_limitslaunch
  • apex.atproto.ceo· BlueskyJul 19, 12:20 AM

    PrismML debuts Bonsai‑27B binary/ternary

    Binary (1.125 bpw, ~3.9GB) and ternary (1.71 bpw) Qwen3.6‑27B variants preserve accuracy better than IQ2, enabling small-footprint local inference on commodity GPUs.

    holy smokes, PrismML released Bonsai-27B, a model based on Qwen3.6-27B that comes in two flavors: binary (1.125bpw) and ternary (1.71bpw) weights, while preserving benchmark scores better than standard IQ2 quants! the binary version has a footprint of only 3.9GB! this is wizardry.

    signal 7hype 4model_releasequantizationlow_bitlaunchsource ↗

findings

(06)
  • aarontay.bsky.social· BlueskyJul 18, 08:16 PM

    ScholarGym benchmarks deep-research gathering

    New benchmark focuses on the information-gathering stage of deep research, evaluating LLMs' ability to plan, search, and collect evidence before synthesis.

    [Read] ScholarGym: Benchmarking Large Language Model Capabilities in the Information-Gathering Stage of Deep Research arxiv.org/abs/2601.21654

    signal 6hype 1paperbenchmarkagent_evaluationtechnicalsource ↗
  • aarontay.bsky.social· BlueskyJul 18, 08:17 PM

    AutoResearchBench tests literature-discovery agents

    A reproducible suite measuring end-to-end agents on complex scientific literature discovery workflows, stressing tool use, retrieval, and multi-step reasoning beyond QA.

    [Read] AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery arxiv.org/abs/2604.25256

    signal 6hype 1agent_benchmarkevaluationpapertechnicalsource ↗
  • tmlr-pub.bsky.social· BlueskyJul 19, 12:19 AM

    Low-rank pseudo-consistency for VLMs

    TMLR paper adapts vision–language models at test time using low-rank pseudo-consistency, improving robustness without labels or heavy finetuning on distribution shifts.

    Test-Time Adaptation of Vision-Language Models with Low-Rank Pseudo-Consistency Shuvendu Roy, Ali Etemad Action editor: Nicolas THOME https://openreview.net/forum?id=GDw4pvX9aG #adaptation #encoder #labelling

    signal 6hype 1papervision_languagetest_time_adaptationtechnicalsource ↗
  • tmlr-pub.bsky.social· BlueskyJul 18, 04:20 PM

    Episodic memory boosts model reuse

    TMLR argues episodic/latent memory components complement parametric learning, enabling flexible reuse and better generalization—useful guidance for RAG and agent architectures.

    Latent learning: episodic memory complements parametric learning by enabling flexible reuse of ex... Andrew Kyle Lampinen, Martin Engelcke, Yuxuan Li, Arslan Chaudhry, James McClelland Action editor: Martha White https://openreview.net/forum?id=RuWGeX5ZiB #retrieval #memory #episodic

    signal 5hype 1papermemoryretrievaltechnicalsource ↗
  • jayvanbavel.bsky.social· BlueskyJul 18, 03:51 PM

    Removing AI access hurts grades

    PNAS study finds AI boosts immediate performance by 48%, but sudden removal drops grades 17% versus never-using AI, underscoring dependence risks in classroom deployments.

    Unfettered access to chatbots can harm educational outcomes Using AI while solving problems improves performance by 48% (GPT base). When access to AI was taken away, students actually performed worse than those who never had access (17% reduction in grades) www.pnas.org/doi/10.1073/...

    signal 5hype 2educationresearchai_in_educationculturalsource ↗
  • neuroskeptic.bsky.social· BlueskyJul 18, 01:16 PM

    Models mix brain-like and unmatched units

    Nature work shows some ANN units closely match task-related brain neurons while many do not, cautioning against overinterpreting neat one-to-one neural analogies.

    Some artificial neural networks contain units that closely match the activity of neurons in brains doing the same task. But there's a "striking asymmetry" - those same models often contain other units that don't correspond to anything in the brain! www.nature.com/articles/s42...

    signal 4hype 2interpretabilityneurosciencepapertechnicalsource ↗

voices

(02)
  • hn/frontpage· AggregatorJul 18, 04:12 PM

    Let Claude Code control a Mac

    A practical walkthrough for running Claude Code against a separate macOS machine with controlled permissions, letting you experiment safely with autonomous workflows.

    Setting up your spare Mac for Claude Code to control, a step-by-step guide — Article URL: https://ykdojo.github.io/claude-controls-mac/ Comments URL: https://news.ycombinator.com/item?id=48959392 Points: 10 # Comments: 1

    signal 8hype 1claude_codecomputer_usetutorialhacksource ↗
  • hn/frontpage· AggregatorJul 18, 07:06 PM

    Why agent weekly quotas reset

    Minimaxir dissects frequent weekly quota resets across agent platforms, surfacing rate-limit quirks and their impact on reliability planning for production workloads.

    What's the deal with all the random weekly quota resets for agents lately? — Article URL: https://minimaxir.com/2026/07/agent-quota-reset/ Comments URL: https://news.ycombinator.com/item?id=48961157 Points: 7 # Comments: 0

    signal 5hype 2agentsquota_resetsrate_limitsculturalsource ↗