Ranked by score · Showing 20 of 37 stories · 6 hidden by your filters

1. No. 1Score 0.82

   Mitigations for CVE-2026-98432 [IPv4 TCP handshake use-after-free]

   RODMENA Engineering8 minTopics: Security, DevOps, ITFormat: Write-upChecked by: Claim-checked

   Summary and score for “Mitigations for CVE-2026-98432 [IPv4 TCP handshake use-after-free]”

   Farshid Ashouri · blog.rodmena.com

   RODMENA Engineering describes CVE-2026-98432, a use-after-free in the IPv4 TCP handshake present in Linux since 4.4, with upstream fixes in stable branches from 5.10. Debian has patched kernels, while Ubuntu's were still pending on 1 November. Until then, the authors drop IPv4 TCP packets carrying IP options with nftables and disable MPTCP. Their tests showed plain connections passing and optioned ones timing out. The filter doesn't cover new network namespaces or containers. The claims come from the post itself, without independent sourcing, and Anthropic's reported exploit is unseen.

   Score 0.82Claim check: 8 no source given

   Elsewhere: HN 3 pts, not on front page

   Score breakdown and claims for “Mitigations for CVE-2026-98432 [IPv4 TCP handshake use-after-free]”

   Comments on “Mitigations for CVE-2026-98432 [IPv4 TCP handshake use-after-free]”

2. No. 2Score 0.77

   Agentic Backdoors via Pretraining Poisoning

   arXiv54 minTopics: AI, SecurityFormat: PaperChecked by: Claim-checked

   Summary and score for “Agentic Backdoors via Pretraining Poisoning”

   arxiv.org

   Researchers at Anthropic, UIUC, UCL and elsewhere report that poisoning 0.1% of a 100B-token pretraining corpus can plant a remote-code-execution backdoor that survives clean SFT, DPO and GRPO post-training. In Qwen3 models trained from scratch, the 4B agent reached 92.7% Pass@32 attack success with an ordinary-string trigger (paths containing "anthropic") and 10.5% with a rare-string trigger. Success rose with model size, and clean-input capability was unchanged. The claim checks found no independent sources, and the experiments cover one CLI-agent pipeline at up to 4B parameters.

   Score 0.77Claim check: 8 no source given

   Elsewhere: HN 2 pts, not on front page

   Score breakdown and claims for “Agentic Backdoors via Pretraining Poisoning”

   Comments on “Agentic Backdoors via Pretraining Poisoning”

3. No. 3Score 0.71

   We build our voice input method like an experiment

   Arion11 minTopics: Programming, AIFormat: Write-upChecked by: Claim-checked

   Summary and score for “We build our voice input method like an experiment”

   Steven Lynn · arion.astrafoundry.si

   Arion, a local Chinese voice and Pinyin input method, published results from ablations it ran before shipping features. On a 30-case held-out set, screen-text context lifted Qwen3 4B from 9 to 16 passes but left the shipped 1.7B model at 9. A no-LLM acoustic second pass beat every LLM correction variant. An upstream Rime reranker's "1 ms" came from CPU time, not wall-clock. These figures are Arion's own, from one developer-authored corpus on one Apple M4, and could not be independently checked.

   Score 0.71Claim check: 8 couldn’t check

   Elsewhere: HN 1 pts, not on front page

   Score breakdown and claims for “We build our voice input method like an experiment”

   Comments on “We build our voice input method like an experiment”

4. No. 4Score 0.70

   Ransomware gang bypassed AI guardrails by posing as red-team testers

   SOPHOS6 minTopics: AI, Security, ITFormat: NewsChecked by: Claim-checked

   Summary and score for “Ransomware gang bypassed AI guardrails by posing as red-team testers”

   About the Author(s) · sophos.com

   Sophos reports that GOLD OGRE, the group behind Aur0ra ransomware, used Cursor-driven agents to escalate from credential access to domain admin, after first using a similar framework to test EDR evasion. Operators reportedly got past model guardrails by calling the work authorized red-teaming. Sophos assesses this with high confidence, though the models named and the three leak-site victims rest on its own account, and Gambit and CloudSEK's overlapping findings were not verified. Only one proxy IP overlaps. Sophos advises segmentation, isolated backups, and tighter domain admin privileges.

   Score 0.70Claim check: 2 matches source · 4 no source given · 2 couldn’t check

   Elsewhere: HN 2 pts, not on front page

   Score breakdown and claims for “Ransomware gang bypassed AI guardrails by posing as red-team testers”

   Comments on “Ransomware gang bypassed AI guardrails by posing as red-team testers”

5. No. 5Score 0.70

   Nothing Installs at Our Company Until It's a Week Old

   /bruno8 minTopics: Programming, Security, DevOps, ITFormat: Write-upChecked by: Claim-checked

   Summary and score for “Nothing Installs at Our Company Until It's a Week Old”

   brunoj.com

   Bruno J. describes a company-wide proxy, git-pkgs/proxy, that holds new npm and PyPI releases for seven days (pub for three). In twelve days it served 81,710 downloads, hid 6,116 too-new versions and refused 71 downloads where lockfiles pinned fresh releases. His reasoning: five 2026 hijacks, including axios and litellm, were pulled within about five hours. A blogger's count of 1,336 hijacked npm packages suggests a seven-day cooldown would have blocked roughly 85%. Some lockfile formats bypass the gate, and slower-detected compromises still get through.

   Score 0.70Contestedjudges differ by 0.58 on overclaimClaim check: 5 matches source

   Elsewhere: HN 3 pts, not on front page

   Score breakdown and claims for “Nothing Installs at Our Company Until It's a Week Old”

   Comments on “Nothing Installs at Our Company Until It's a Week Old”

6. No. 6Score 0.67

   A skeleton from 20 observations beat 160 tests, a cheap model nearly caught up

   andreasrau.tech4 minTopics: Programming, AIFormat: AnalysisChecked by: Claim-checked

   Summary and score for “A skeleton from 20 observations beat 160 tests, a cheap model nearly caught up”

   An AWS paper, Skeleton Coding, finds that giving coding agents a program's structure (types, modules, signatures, contracts) in code-like form beats adding more behavioral tests. On SlopCodeBench the mean pass rate rose from 71.7% to 89.2%, and on SDDBench from 29% to 74.2%. The smaller GPT-5.6 Luna came within 5.0 to 8.2 points of Sol at far lower cost. Gains shrank on ProgramBench, where an oracle exists, and overly detailed contracts sometimes hurt. Andreas Rau notes that the benchmarks may be contaminated.

   Score 0.67Contestedjudges differ by 0.46 on overclaimClaim check: 8 matches source

   Elsewhere: HN 3 pts, not on front page

   Score breakdown and claims for “A skeleton from 20 observations beat 160 tests, a cheap model nearly caught up”

   Comments on “A skeleton from 20 observations beat 160 tests, a cheap model nearly caught up”

7. No. 7Score 0.66

   Let's Encrypt Goes to 64-Day Certificates: When Does Your ACME Client Actually Renew?

   DevOps Daily11 minTopics: DevOps, ITFormat: Write-upChecked by: Claim-checked

   Summary and score for “Let's Encrypt Goes to 64-Day Certificates: When Does Your ACME Client Actually Renew?”

   DevOps Daily Team · devops-daily.com

   Let's Encrypt's staging environment began issuing 64-day certificates on 14 October 2026, ahead of the production default on 10 February 2027 and 45 days in 2028. DevOps Daily tested five ACME client installs. Clients using ARI (certbot 5.8, acme.sh 3.1.6, lego 5.5.2) renewed around day 42 of 64, the same two-thirds share as at 45 and 90 days. acme.sh 3.0.7 scheduled renewal at day 59, or day 79 with --days 80, after expiry. Upgrading fixed it. The test used staging and one certificate per setup.

   Score 0.66Claim check: 4 matches source · 4 no source given

   Elsewhere: HN —

   Score breakdown and claims for “Let's Encrypt Goes to 64-Day Certificates: When Does Your ACME Client Actually Renew?”

   Comments on “Let's Encrypt Goes to 64-Day Certificates: When Does Your ACME Client Actually Renew?”

8. No. 8Score 0.65

   Rbcop, RuboCop rewritten in Rust with identical output

   GitHub7 minTopics: ProgrammingFormat: New projectChecked by: 3 judges

   Summary and score for “Rbcop, RuboCop rewritten in Rust with identical output”

   github.com

   rbcop is a Rust reimplementation of RuboCop 1.92.2, built on the Prism parser, aiming for byte-identical output and exit status. It implements all 613 default cops plus native versions of several popular plugins, and falls back to RuboCop for cops it does not implement.

   Score 0.65

   Elsewhere: HN 1 pts, not on front page

   Score breakdown for “Rbcop, RuboCop rewritten in Rust with identical output”

   Comments on “Rbcop, RuboCop rewritten in Rust with identical output”

9. No. 9Score 0.67

   Resisting Cognitive Surrender

   knlb.dev12 minTopics: Programming, AIFormat: AnalysisChecked by: Claim-checked

   Summary and score for “Resisting Cognitive Surrender”

   Kunal, writing on knlb.dev, argues that accepting agent output uncritically ("cognitive surrender") is fast but, in his experience, produces weaker decisions, technical debt and lost familiarity with the system. He says building time can approach zero while comprehension has barely improved. His remedies: set a "vibe limit" by project risk, treat repeated "yes, continue" as a warning sign, and drive changes by hand on important code. For delegated work he suggests observability, deterministic validation, shadow traffic and fast deploys. These are personal observations, not measured results.

   Score 0.67Claim check: 2 matches source · 5 no source given · 1 couldn’t check

   Elsewhere: HN 3 pts, not on front page

   Score breakdown and claims for “Resisting Cognitive Surrender”

   Comments on “Resisting Cognitive Surrender”

10. No. 10Score 0.64

    RIPE members debate reliance on U.S. internet infrastructure

    heise online English4 minTopics: IT, PolicyFormat: NewsChecked by: Claim-checked

    Summary and score for “RIPE members debate reliance on U.S. internet infrastructure”

    heise online · heise.de

    At RIPE's 93rd meeting in Sofia, members discussed how U.S. sanctions law reaches core internet infrastructure, prompted by the .org registry PIR blocking the domain of Italian provider Autistici/Inventati after an OFAC terrorism designation. Participants urged operators to map their dependencies, including on DNS registries, certificate authorities and hyperscaler email. RIPE NCC has dropped its "Cloud First" strategy and budgeted €14.59 million for 2027 to run its own infrastructure. The heise report cites no sources for its figures, which are largely attributed to speakers.

    Score 0.64Claim check: 8 no source given

    Elsewhere: HN —

    Score breakdown and claims for “RIPE members debate reliance on U.S. internet infrastructure”

    Comments on “RIPE members debate reliance on U.S. internet infrastructure”

11. No. 11Score 0.61

    NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives

    ServeTheHome7 minTopics: AI, HardwareFormat: NewsChecked by: Claim-checked

    Summary and score for “NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives”

    Ryan Smith · servethehome.com

    ServeTheHome reports that NVIDIA's final Rubin GPU specs, posted in September, are below the preliminary figures: HBM4 bandwidth drops from 22TB/s to 19.2TB/s (about 11%), and NVLink from 3.6TB/s to 3.0TB/s, while scale-out networking rises from 400GB/s to 450GB/s. NVL72 racks follow, and HGX NVL8 also loses HBM4 capacity, from 2.3TB to 2.0TB. Compute ratings are unchanged. NVIDIA has given no reason; the memory-supply explanation is the author's inference. Most figures were not independently checked.

    Score 0.61Claim check: 2 matches source · 1 no source given · 5 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives”

    Comments on “NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives”

12. No. 12Score 0.63

    Hackers could poison AI models by tampering with core infrastructure

    Cybersecurity Dive2 minTopics: AI, Security, HardwareFormat: NewsChecked by: Claim-checked

    Summary and score for “Hackers could poison AI models by tampering with core infrastructure”

    Eric Geller · cybersecuritydive.com

    Security firm Eclypsium reports that AI models can be manipulated by tampering with the firmware, drivers and management hardware beneath them. In its demonstration, selected weights were altered in memory after integrity checks ran, while the model file, hash and signatures on disk stayed valid. Eclypsium says Nvidia, AMD, Intel and Arm systems are affected, and recommends DPU-based policy enforcement, component inventories and scrutiny of third-party software. Cybersecurity Dive's account could not be independently verified, and the research was done with Nvidia.

    Score 0.63Claim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Hackers could poison AI models by tampering with core infrastructure”

    Comments on “Hackers could poison AI models by tampering with core infrastructure”

13. No. 13Score 0.64

    An inference server for AMD GPUs, in C++ and HIP

    Codeberg.org15 minTopics: Programming, AI, DevOps, HardwareFormat: ReleaseChecked by: Claim-checked

    Summary and score for “An inference server for AMD GPUs, in C++ and HIP”

    StillDeadcode

    Radiance, published on Codeberg by StillDeadcode, is a C++ and HIP inference server for AMD GPUs. It serves language, image, video, speech and music models behind OpenAI-, Anthropic- and llama-server-compatible APIs, with architectures and kernels loaded as plugins. Bundled kernels target RDNA4 cards. The README gives throughput figures, such as 166 tok/s for GLM-5.3-Flash on eight Radeon AI PRO R9700s, but they come from the author's own runs without independent sources, and four-rank speed against two ranks is stated as unmeasured.

    Score 0.64Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “An inference server for AMD GPUs, in C++ and HIP”

    Comments on “An inference server for AMD GPUs, in C++ and HIP”

14. No. 14Score 0.60

    Moving an RTS from Godot to Lobster

    Age of Anything9 minTopics: Programming, AIFormat: Write-upChecked by: Claim-checked

    Summary and score for “Moving an RTS from Godot to Lobster”

    age-of-anything.com

    The developer of Age of Anything, an RTS, replaced Godot and GDScript with the Lobster language in four days, using Claude for most of the porting. The move came two weeks into the project. Per the developer's own figures, a 15-minute bot match matched the original on all 18,000 state hashes and ran about fifteen times faster per tick. The tests missed UI regressions and a rendering slowdown, which manual play found. Those numbers are self-reported and unverified.

    Score 0.60Claim check: 7 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Moving an RTS from Godot to Lobster”

    Comments on “Moving an RTS from Godot to Lobster”

15. No. 15Score 0.66

    Do you still need embeddings for semantic search? - my little experiment

    internals.laxmena.com9 minTopics: AI, DataFormat: PaperChecked by: Claim-checked

    Summary and score for “Do you still need embeddings for semantic search? - my little experiment”

    Lax Meiyappan

    Lax Meiyappan tested whether BM25 keyword search, with an LLM query rewrite and a reranker, could replace Qwen3-Embedding-0.6B on SciFact, NFCorpus and FiQA. Embeddings won or tied in every fair comparison, leading by about a point on the first two sets and eight to nine on FiQA once both used the same reranker. The rewrite added little after reranking. Replacing 54 FiQA names with invented codenames narrowed the gap to 1.3 points, but embeddings stayed ahead. These are one author's unreplicated results, and the cost comparison reflects his pricing.

    Score 0.66Claim check: 1 matches source · 7 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Do you still need embeddings for semantic search? - my little experiment”

    Comments on “Do you still need embeddings for semantic search? - my little experiment”

16. No. 16Score 0.57

    Aqueous: The files you keep are compressed. Badly

    lifthrasiir.github.io4 minTopics: ProgrammingFormat: ReleaseChecked by: Claim-checked

    Summary and score for “Aqueous: The files you keep are compressed. Badly”

    Aqueous is a new public-domain archive format and archiver, aqr, that decodes the weak compression inside JPEG, PNG, MP3, ZIP and PDF files, recodes the contents with stronger models, and restores the original bytes on extraction. Its author reports that on Matt Mahoney's 10 GB benchmark, aqr produced 2.50 GB, against 3.70 GB for zpaq and 4.68 GB for zstd -3, at 61 MB/s compression. These figures are self-reported and not independently verified. Decoders ship inside the archive as WebAssembly modules.

    Score 0.57Claim check: 1 matches source · 4 no source given · 3 couldn’t check

    Elsewhere: HN 6 pts, not on front page

    Score breakdown and claims for “Aqueous: The files you keep are compressed. Badly”

    Comments on “Aqueous: The files you keep are compressed. Badly”

17. No. 17Score 0.56

    StackVitals – CI failure triage and DORA metrics, no server

    GitHub6 minTopics: Programming, DevOpsFormat: New projectChecked by: 3 judges

    Summary and score for “StackVitals – CI failure triage and DORA metrics, no server”

    github.com

    StackVitals is an open-source GitHub developer suite that runs entirely inside a repo's own GitHub Actions, storing its history in that repo with no hosted backend or account.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “StackVitals – CI failure triage and DORA metrics, no server”

    Comments on “StackVitals – CI failure triage and DORA metrics, no server”

18. No. 18Score 0.63

    Representation Hypothesis with JEPAs

    alphaXiv4 minTopics: AIFormat: PaperChecked by: Claim-checked

    Summary and score for “Representation Hypothesis with JEPAs”

    Yann LeCun · alphaxiv.org

    A paper by Yann LeCun and coauthors argues that text and image encoders align when training treats different descriptions of the same scene as positives, rather than corruptions of one caption. A 109M-parameter JEPA text model scores .184 mutual 10-NN alignment with DINOv2 ViT-B, above Qwen2.5-72B's .154, and reaches next-token-level alignment with 42 times fewer samples. The theory is idealized, one seed scored .121, and the captions partly come from LLaVA, which uses a CLIP encoder. Results are reported by the authors alone.

    Score 0.63Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Representation Hypothesis with JEPAs”

    Comments on “Representation Hypothesis with JEPAs”

19. No. 19Score 0.63

    Same model, different agent: 17 points apart on 21 MCP servers

    SineFrame M35 minTopics: Programming, AIFormat: PaperChecked by: Claim-checked

    Summary and score for “Same model, different agent: 17 points apart on 21 MCP servers”

    Utkarsh Pandey · m3.sineframe.com

    SineFrame, maker of the open-source M3 MCP testing tool, ran Claude Haiku 5.5 through two coding agents against 21 MCP servers. OpenCode answered 82.0% of tasks from the server, Claude Code 65.4%. The gap sits on the ten servers that overlap built-in tools such as web search, where agents often answer without calling the server. On the other eleven, agents scored within about four points. For server builders, results depend on the agent used for testing. The figures are SineFrame's own and unverified; Codex also ran a different model.

    Score 0.63Claim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Same model, different agent: 17 points apart on 21 MCP servers”

    Comments on “Same model, different agent: 17 points apart on 21 MCP servers”

20. No. 20Score 0.64

    Google, Anthropic and others to support new AI crawling opt-out standard

    SiliconANGLE3 minTopics: AI, PolicyFormat: NewsChecked by: Claim-checked

    Summary and score for “Google, Anthropic and others to support new AI crawling opt-out standard”

    Maria Deutscher · siliconangle.com

    Google, Microsoft, Anthropic and OpenAI say they will honor a new robots.txt directive that lets a site block all generative-AI training crawlers with one entry, replacing per-provider rules. Engadget says the draft comes from Anthropic's Kevin Kelley and is before the IETF; it would not block search indexing. Microsoft also plans an "AI Grounding" setting for answer-time fetches. OpenAI and Microsoft reportedly ship support within months, Google next year; Meta has not committed. Most timing and scope details remain unverified.

    Score 0.64Claim check: 2 matches source · 1 no source given · 5 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Google, Anthropic and others to support new AI crawling opt-out standard” · Also covered by Engadget

    Comments on “Google, Anthropic and others to support new AI crawling opt-out standard”

21. No. 21Score 0.59

    Applebee's exposed customers' personal data with database password in app

    TechCrunch3 minTopics: SecurityFormat: NewsChecked by: Claim-checked

    Summary and score for “Applebee's exposed customers' personal data with database password in app”

    Zack Whittaker · techcrunch.com

    Applebee's has fixed a flaw in which a hardcoded database credential in its Android app exposed customer records, TechCrunch reported. Researcher Robert Wiggins found it with AI help, and TechCrunch says it confirmed access using a dummy account. The data included names, emails, birth dates, phone numbers and login IP addresses. Dine Brands says it found no evidence of unauthorized access, but did not say whether logs exist, how many customers were affected, or how long the credential was present. Neither company offered a way to report security issues.

    Score 0.59Claim check: 7 no source given · 1 couldn’t check

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “Applebee's exposed customers' personal data with database password in app”

    Comments on “Applebee's exposed customers' personal data with database password in app”

22. No. 22Score 0.56

    OmniDiff – Opensource, super fast, content aware Git difftool, in Rust

    GitHub12 minTopics: ProgrammingFormat: ReleaseChecked by: Claim-checked

    Summary and score for “OmniDiff – Opensource, super fast, content aware Git difftool, in Rust”

    github.com

    OmniDiff, an AGPL-licensed Rust diff tool by Marko Ivankovic, picks a diff method by content: tree-sitter structural diffs for 25 languages, plus pictures, fonts, archives and PDFs. It plugs into git, jj, VS Code and Neovim. The README reports a 4.2ms median over 2,302 fixtures and 442,322 of 442,530 code pairs completed. Its own accuracy goals are unmet: 60–70% of hand-painted cases match exactly. All figures are self-reported, from one laptop, and archives and PDFs are slower.

    Score 0.56Contestedjudges differ by 0.63 on overclaimClaim check: 8 no source given

    Elsewhere: HN 196 pts, front page

    Score breakdown and claims for “OmniDiff – Opensource, super fast, content aware Git difftool, in Rust”

    Comments on “OmniDiff – Opensource, super fast, content aware Git difftool, in Rust”

23. No. 23Score 0.61

    Where Did the Rest Go?

    Evan Smith10 minTopics: Programming, AIFormat: AnalysisChecked by: Claim-checked

    Summary and score for “Where Did the Rest Go?”

    iamevan.me

    An engineer at SREcon EMEA '26 argues that AI-assisted development leaves developers more cognitively fatigued, not less, because automation strips out the self-paced coding work that let them recover and leaves them with harder supervisory tasks.

    Score 0.61Claim check: 2 matches source · 3 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Where Did the Rest Go?”

    Comments on “Where Did the Rest Go?”

24. No. 24Score 0.61

    Machine-Readable Training Preference Signal

    ietf.org9 minTopics: AI, PolicyFormat: OtherChecked by: 3 judges

    Summary and score for “Machine-Readable Training Preference Signal”

    Kevin Kelley

    This IETF Internet-Draft proposes binding AI training preferences to URL paths by writing standard robots.txt groups keyed to a new per-purpose policy token, usage_ai-training, without adding directives or changing RFC 9309 grammar or precedence.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Machine-Readable Training Preference Signal”

    Comments on “Machine-Readable Training Preference Signal”

25. No. 25Score 0.60

    The tokens in your agent harness

    armincat.com3 minTopics: Programming, AIFormat: Write-upChecked by: 3 judges

    Summary and score for “The tokens in your agent harness”

    An author who read the open-source Codex harness in Rust catalogs the hidden LLM calls it makes beyond the main task: automatic thread naming, reasoning-based activity status (no separate request), auto recaps, memory extraction and consolidation, and a background Guardian security reviewer.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The tokens in your agent harness”

    Comments on “The tokens in your agent harness”

26. No. 26Score 0.60

    Bowerbird – Open-source NotebookLM for any model, with checked quotes

    GitHub7 minTopics: Programming, AIFormat: New projectChecked by: 3 judges

    Summary and score for “Bowerbird – Open-source NotebookLM for any model, with checked quotes”

    github.com

    Bowerbird is an open-source, local-first alternative to Google's NotebookLM that lets users ask questions over their own PDFs, web pages, videos and recordings, using any model they choose. Notebooks are plain folders on disk, and search, OCR and transcription run locally.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Bowerbird – Open-source NotebookLM for any model, with checked quotes”

    Comments on “Bowerbird – Open-source NotebookLM for any model, with checked quotes”

27. No. 27Score 0.56

    Docsferatu: An agent that drafts docs PRs while you sleep

    inngest.com10 minTopics: Programming, AI, DevOpsFormat: Write-upChecked by: Claim-checked

    Summary and score for “Docsferatu: An agent that drafts docs PRs while you sleep”

    Docsferatu is a side project that watches an SDK repo for pull requests, uses a classifier to decide whether docs need updating, and if so has Claude draft a docs PR in an isolated sandbox.

    Score 0.56Claim check: 2 matches source · 1 no source given · 5 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Docsferatu: An agent that drafts docs PRs while you sleep”

    Comments on “Docsferatu: An agent that drafts docs PRs while you sleep”

28. No. 28Score 0.60

    The surprising leadership dynamics in agent swarms

    raindrop.ai14 minTopics: AIFormat: PaperChecked by: 3 judges

    Summary and score for “The surprising leadership dynamics in agent swarms”

    Raindrop AI

    A study of eight AI agents from four models working on simulated outages found that instructions were common but rarely changed teammates' behavior: only 105 of 2,160 teams (5%) produced a real leader.

    Score 0.60Contestedjudges differ by 0.42 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The surprising leadership dynamics in agent swarms”

    Comments on “The surprising leadership dynamics in agent swarms”

29. No. 29Score 0.60

    Chinese labs now target OpenAI's GPT-6 Astra for distillation

    fork.zone1 minTopics: AIFormat: NewsChecked by: 3 judges

    Summary and score for “Chinese labs now target OpenAI's GPT-6 Astra for distillation”

    The Information reports, via The Decoder, that Chinese AI labs including DeepSeek and Moonshot AI have shifted distillation efforts from Anthropic's Claude to OpenAI's GPT-6 Astra, allegedly to train newer Kimi and DeepSeek models. The claims rest on anonymous sources, and the report gives no volume or performance figures.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Chinese labs now target OpenAI's GPT-6 Astra for distillation”

    Comments on “Chinese labs now target OpenAI's GPT-6 Astra for distillation”

30. No. 30Score 0.60

    The Abandoned PR Page

    aictrl.dev14 minTopics: Programming, AIFormat: PaperChecked by: 3 judges

    Summary and score for “The Abandoned PR Page”

    Analysis by Bulat at aictrl.dev, co-authored with AI

    Analysis of 240,653 merged pull requests across 26 open-source repositories, from January 2024 to September 2026, finds that AI-era projects merge far more changes while human comments per pull request fall.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The Abandoned PR Page” · Also covered by slopcodemonitor.ai

    Comments on “The Abandoned PR Page”

31. No. 31Score 0.59

    Offline project memory for coding agents – no API key, no vector DB

    GitHub12 minTopics: AIFormat: PaperChecked by: 3 judges

    Summary and score for “Offline project memory for coding agents – no API key, no vector DB”

    github.com

    A Show HN post introduces an offline, dependency-free memory engine for coding agents that stores an append-only event log in the project repository and exposes eight MCP tools, with no API key, vector database, or daemon required.

    Score 0.59Contestedjudges differ by 0.51 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Offline project memory for coding agents – no API key, no vector DB”

    Comments on “Offline project memory for coding agents – no API key, no vector DB”

32. No. 32Score 0.63

    Proving and Improving the Platonic Representation Hypothesis with JEPAs

    arXiv96 minTopics: AIFormat: PaperChecked by: 3 judges

    Summary and score for “Proving and Improving the Platonic Representation Hypothesis with JEPAs”

    arxiv.org

    This paper argues that the Platonic Representation Hypothesis, the observation that separately trained vision and language models converge to similar representations, can be proven for Joint-Embedding Predictive Architectures.

    Score 0.63

    Elsewhere: HN 124 pts, front page

    Score breakdown for “Proving and Improving the Platonic Representation Hypothesis with JEPAs”

    Comments on “Proving and Improving the Platonic Representation Hypothesis with JEPAs”

33. No. 33Score 0.59

    Devloops – Adversarial agent loops so humans only make the key decision

    GitHub5 minTopics: Programming, AIFormat: New projectChecked by: 3 judges

    Summary and score for “Devloops – Adversarial agent loops so humans only make the key decision”

    github.com

    Devloops is an open-source set of reusable skills for coding agents that splits maintenance into three adversarial loops: a producer files candidate issues, a worker implements approved ones as pull requests, and an independent gatekeeper reviews each PR. Humans only accept scope and perform merges.

    Score 0.59Contestedjudges differ by 0.44 on substantiated, 0.42 on technical depth

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Devloops – Adversarial agent loops so humans only make the key decision”

    Comments on “Devloops – Adversarial agent loops so humans only make the key decision”

34. No. 34Score 0.58

    What's going to happen to junior engineers?

    seangoedecke.com5 minTopics: Programming, AIFormat: AnalysisChecked by: 3 judges

    Summary and score for “What's going to happen to junior engineers?”

    Sean Goedecke argues in this essay that AI coding tools are squeezing the junior engineer market, which had already shrunk as hiring tilted toward senior engineers.

    Score 0.58

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “What's going to happen to junior engineers?”

    Comments on “What's going to happen to junior engineers?”

35. No. 35Score 0.58

    What [67] MCP servers do at startup, with the internet cut off

    mcp.brydge-ai.com5 minTopics: AI, SecurityFormat: PaperChecked by: 3 judges

    Summary and score for “What [67] MCP servers do at startup, with the internet cut off”

    A Show HN project measured 67 npm MCP servers at startup with no internet route and no tools called, then compared what each reached for and offered against its own README. Seven servers looked up websites their README never names, and 24 offered tools the README does not name.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “What [67] MCP servers do at startup, with the internet cut off”

    Comments on “What [67] MCP servers do at startup, with the internet cut off”

36. No. 36Score 0.58

    Scaling Agentic Coding as If People Still Mattered

    adamtornhill.substack.com7 minTopics: Programming, AIFormat: AnalysisChecked by: 3 judges

    Summary and score for “Scaling Agentic Coding as If People Still Mattered”

    Adam Tornhill

    Adam Tornhill argues that running five or more AI coding agents in parallel, each in its own terminal, is a trend about to go mainstream and that it imposes hidden human costs.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Scaling Agentic Coding as If People Still Mattered”

    Comments on “Scaling Agentic Coding as If People Still Mattered”

37. No. 37Score 0.57

    Modeljury – a jury of cheap LLMs that escalates the hard cases

    GitHub10 minTopics: Programming, AIFormat: New projectChecked by: 3 judges

    Summary and score for “Modeljury – a jury of cheap LLMs that escalates the hard cases”

    github.com

    Modeljury is a small Python library that convenes a panel of several language models on a decision, ships the verdict when they agree, and flags it for human review with the dissenting juror's reasoning when they don't.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Modeljury – a jury of cheap LLMs that escalates the hard cases”

    Comments on “Modeljury – a jury of cheap LLMs that escalates the hard cases”

38. No. 38Score 0.57

    QwenPU: An LLM running inside another LLM's weights

    huggingface.co3 minTopics: Programming, AIFormat: New projectChecked by: 3 judges

    Summary and score for “QwenPU: An LLM running inside another LLM's weights”

    QwenPU, released by Divyansh Agrawal, compiles a 32-bit RISC-V CPU and a vector unit into the weights of Qwen3.5-4B, so that Karpathy's 260K-parameter stories model runs as a program inside the model's forward passes and produces the same output as plain C.

    Score 0.57Contestedjudges differ by 0.45 on practical relevance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “QwenPU: An LLM running inside another LLM's weights”

    Comments on “QwenPU: An LLM running inside another LLM's weights”

39. No. 39Score 0.57

    Why Sam Altman is against a single superintelligence

    cheekypint.substack.com44 minTopics: AI, SecurityFormat: AnalysisChecked by: 3 judges

    Summary and score for “Why Sam Altman is against a single superintelligence”

    Cheeky Pint

    In a conversation with John Collison, OpenAI CEO Sam Altman discusses how AI models are increasingly able to find serious software vulnerabilities, including in PHP, and argues that this pushes the industry toward rearchitecting systems rather than only patching them.

    Score 0.57Contestedjudges differ by 0.41 on new information, 0.54 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Why Sam Altman is against a single superintelligence”

    Comments on “Why Sam Altman is against a single superintelligence”

40. No. 40Score 0.57

    GitHub Agentic Workflows generally available

    The GitHub Blog2 minTopics: Programming, AI, DevOpsFormat: ReleaseChecked by: 3 judges

    Summary and score for “GitHub Agentic Workflows generally available”

    github.blog

    GitHub has made GitHub Agentic Workflows, an open source project from GitHub Next, generally available. It lets teams define reasoning-based automation such as issue triage, CI failure analysis, and test writing in natural language Markdown, which compiles to standard GitHub Actions YAML.

    Score 0.57Contestedjudges differ by 0.46 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “GitHub Agentic Workflows generally available”

    Comments on “GitHub Agentic Workflows generally available”

41. No. 41Score 0.56

    FineDNA: Open data to train LLMs on DNA

    huggingface.co10 minTopics: AI, DataFormat: ReleaseChecked by: 3 judges

    Summary and score for “FineDNA: Open data to train LLMs on DNA”

    FineDNA is the open training dataset behind Carbon-X, a Qwen3-14B model adapted to read DNA across mid-training, SFT and RL. It bundles DNA sequences from GENERator and OpenGenome2, biology web and paper text, English web text, and instruction data, with per-config weights that define the training recipe.

    Score 0.56Contestedjudges differ by 0.46 on reader interest

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “FineDNA: Open data to train LLMs on DNA”

    Comments on “FineDNA: Open data to train LLMs on DNA”

42. No. 42Score 0.56

    I Review 2k Lines of AI Generated Code

    bytesizedbets.com16 minTopics: Programming, AIFormat: AnalysisChecked by: 3 judges

    Summary and score for “I Review 2k Lines of AI Generated Code”

    A review of a sponsored walkthrough of CodeRabbit Explain, a tool that restructures a 31-file, roughly 1,700-line AI-generated pull request into ordered layers with line-anchored summaries and history. The author argues that GitHub's alphabetical diff view hides intent and dependencies, and that reviewers need explainability, not just verdicts.

    Score 0.56Contestedjudges differ by 0.43 on technical depth, 0.59 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “I Review 2k Lines of AI Generated Code”

    Comments on “I Review 2k Lines of AI Generated Code”

43. No. 43Score 0.55

    Revision Prompting: Diff your inputs for consistency and efficiency

    revisionprompting.info4 minTopics: AIFormat: Write-upChecked by: 3 judges

    Summary and score for “Revision Prompting: Diff your inputs for consistency and efficiency”

    Revision Prompting proposes a way to update LLM outputs after their inputs change: instead of re-running the full instruction, the model receives the original input, original output, and an input diff, and returns only an output patch.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Revision Prompting: Diff your inputs for consistency and efficiency”

    Comments on “Revision Prompting: Diff your inputs for consistency and efficiency”