Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

PostgreSQL 19 is running late because AI finds bugs faster than humans can fix them

Curated by Alfredo Borrás · Plymouth AI Lab

PostgreSQL 19 slips into November as LLM-found bugs swamp its reviewers, Anthropic adds its text watermark to four older Claude models, the big Western labs agree on one robots.txt line to opt out of AI training, and DeepSeek publishes a coding agent that waits to be asked.

The story today

PostgreSQL 19 is running late because AI finds bugs faster than humans can fix them

For about a decade PostgreSQL has shipped its major release between mid-September and mid-October. Robert Haas, a long-time committer, writes that version 19 will most likely not ship until November and could slip further, and that three large features (UPDATE/DELETE FOR PORTION OF, SQL property graph queries and MERGE/SPLIT PARTITION) were pulled between 27 August and 15 September, unusually late in the beta.

His explanation is that people across the internet are now reporting bugs they found with LLMs. Comparing 8 April to 25 September in 2025 and 2026, commits to the master branch rose from 1,037 to 1,443, while fixes going into the release branch rose from 724 to 1,169, and fixes back-patched to the previous release went from 248 to 433, up more than 74%. LLMs can draft fixes, but those still need human review, and reviewer time spent on old branches was time not spent deciding what to cut from the new one.

Haas doesn't think a short delay is worrying, but the longer picture is less clear: forty years of bugs that people could not find are now findable, and the same small group of committers has to absorb them while still building new features.

Why it matters: If you maintain a codebase, expect AI-assisted bug reports to arrive faster than your review capacity grows. Plan triage and back-port policy now, and treat reviewer hours, not bug discovery, as the bottleneck.

Sources: Robert Haas (PostgreSQL committer)

Highlights

  1. Anthropic adds its text watermark to four older Claude models

    According to an Anthropic email to API customers reported by The Decoder, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 are being watermarked, with Bedrock possibly a few days behind; models released after 2 August 2026 already carried it. The mark is a statistical pattern in word choice, adds no tokens and can't be switched off, and Anthropic cites the EU AI Act's December deadline. On Opus 4.6 and Sonnet 4.6 it applies only when temperature, top_p and top_k stay at their defaults, so changing any one avoids it.

    industrypolicySource: The Decoder

  2. OpenAI, Google, Microsoft and Anthropic back one robots.txt opt-out from AI training

    The four labs said on 30 October they will honour a single robots.txt policy token, __usage_ai-training, so a site can opt out of training across all their crawlers with one rule. Anthropic has started implementing it, OpenAI says in the coming months, Microsoft targets 1 February 2027 and Google says 2027. It governs training, not fetching, rests on an individual IETF Internet-Draft with no formal standing, and does nothing about content already collected.

    industrypolicySource: AI Weekly, reporting The Verge

  3. DeepSeek publishes dsh-superagent, a coding agent that acts only when asked

    DeepSeek has released dsh-superagent under the MIT licence as a research prototype, explicitly not an official product. It watches one GitHub repository, and optionally Feishu, and when an allowed person asks for work it opens a work item, does the job in its own session and directory, and replies in the same thread, for example with a pull request. Its stated design rule is to take as little human attention as possible: it acts only when asked or when a team-written rule allows it, and a web panel shows what needs a person.

    agentsopen-sourceSource: GitHub: deepseek-ai/dsh-superagent-prototype

  4. Heretic 2 generalises refusal removal into steering for slop, sycophancy and more

    Heretic, a training-free tool for editing the behaviour of open-weight models you can run locally, now ships configurations beyond its default of suppressing refusals, including reducing slop, humour and sycophancy, by contrasting example prompts and matching scorers. Version 2 configuration files are not compatible with 1.x. It is a reminder that safety behaviour in open weights can be removed cheaply by anyone with a gaming PC.

    safetyopen-weightsSource: GitHub: p-e-w/heretic

Concept of the day

Back-patching

Back-patching means applying a bug fix not only to the version under development but also to releases already in users' hands. Each back-patch has to be checked against older code, so when bug reports surge, as PostgreSQL is seeing with LLM-found bugs, back-patching work grows faster than new development and eats the review time a release needs.

Source: Robert Haas (PostgreSQL committer)

Models released

ModelMakerSizeNotes

Research

  1. A 109M-parameter text JEPA lines up with image encoders better than LLMs up to 72B

    Randall Balestriero and Yann LeCun argue that models trained on different modalities converge to the same representation when their training pairs are two views of one scene, and that next-token prediction uses weaker pairs, which is why language models only align with vision at scale. Their small text JEPA aligns with DINOv2 and CLIP image encoders better than every language model they tested up to 72B, needs 42 times fewer samples than next-token prediction for the same alignment, and reaches 64% top-5 zero-shot on ImageNet-100 without paired data.

    researchmodelsSource: arXiv

  2. Training against a monitor taught hiding only when hiding was already likely enough

    In controlled reward-hacking tasks with small transformers and Qwen3-1.7B, a LessWrong study finds RL learned to evade a monitor only if deceptive hacks were sampled often enough at the start. In the toy model a starting rate of 1e-6 stayed near zero, while 1e-3 led to deceptive hacks 83 to 86% of the time; with Qwen, runs expected to see fewer than about 0.3 deceptive hacks never learned to hide. Truthful fine-tuning, or a one-off edit to an output weight, prevented it. This is one author's small-scale work, not peer reviewed.

    researchsafetySource: LessWrong

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call