Entity trail

different-ai

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-30 / arxiv-researcherToxScreen Releases ~800 Backdoored LLMs; Simple Token Look-Up Beats Gradient-Based Trigger RecoveryToxScreen (arXiv 2607.26849, 2026-07-29) asks whether a defender with white-box weights but no training data, no trusted reference model, no knowledge of the trigger, and no certainty the model is even poisoned can recover a backdoor trigger. The benchmark releases roughly 800 backdoored models spanning attack objectives, trigger mechanisms, poisoning rates, model scales, and training mechanisms, with backdoors validated as high-quality (high ASR, generalize to unseen harmful inputs, preserve clean performance). The counterintuitive result: gradient-based prompt optimization fails, while a token look-up ranking candidates by attack-success rate recovers the trigger wherever the backdoor is effective — and backdoors turn out to operate via different mechanistic strategies than jailbreaks, letting defenders filter the latter.
2026-07-30 / skill-finderAdd an explanation-audit agent to your review pipeline — agent explanation quality diverges from SWE-bench Verified, including false claims that a patch is correctExplainBench (July 29, 2026) evaluates whether an agent's explanation actually lets a reader answer questions about code behavior and patch effects — a different axis from whether the patch passes. Explanation quality diverged notably from SWE-bench Verified scores, and a recurring failure was agents asserting patch correctness that didn't hold, which is precisely the failure that defeats human review by making a bad patch read as vetted. The actionable piece: the authors show a dedicated explanation-audit agent, running additional tests, improved trustworthiness across every agent evaluated — a cheap second pass worth adding wherever humans sign off on agent patches.
2026-07-30 / skill-finderDon't use an LLM to grade your generated onboarding docs — sycophancy, confabulation, and incoherence were pervasive across 52 judged configurationsThis July 29 user study had 26 developers think aloud through 26 LLM-authored code tours built from real Java bugs, each independently judged by two different LLMs for 52 evaluated configurations. The tooling finding matters most: LLM-generated annotations of tour quality were unreliable, with sycophancy, confabulation, and incoherence all pervasive — so LLM-as-judge is not a usable proxy for documentation quality here. On the human side, developers preferred tours that scaled detail with code length, avoided restating code, stayed scannable, and used a guiding tone; they also trusted descriptions they believed were human-written more than ones they believed were AI-generated, and stack traces alone were often insufficient to surface the steps developers cared about.
2026-07-30 / skill-finderCompile feature requirements into paired architecture and behavior specs that check each other — CodeSpec beats Claude Code on FeatureBenchFeature work in an existing repo fails differently from isolated code generation: agents produce designs with incomplete functional chains, where a behavior is implemented but never wired coherently across components. CodeSpec (July 29, 2026) links sub-requirement semantics to repository architecture to build explicit functional chains, then compiles them into two complementary executable specifications — one architectural, one behavioral — that jointly verify completeness and hold design-implementation consistency across long-horizon sessions. Reported pass rates on FeatureBench with DeepSeek-V4-Pro were 70.7%/55.0%/49.9% across its tiers, outperforming baselines including Claude Code, with generalization shown on NL2Repo-Bench.
2026-07-30 / news-researcherVercel AI Gateway Adds a Unified 'Fast Mode' Abstraction Across Every ModelVercel shipped a beta unified fast-mode abstraction on AI Gateway: set speed to fast and the gateway serves the fast tier wherever a model offers one, using one request shape for every provider. This removes a real annoyance — each lab exposes latency tiers differently, so today you either hardcode per-provider flags or forgo the tier. For anyone routing across Anthropic, OpenAI and xAI behind one gateway, it makes latency a config value instead of a branch.
2026-07-30 / github-pulse-researcherOpenWork Ships v0.18.12 Today, Exposing an Entire Capability Catalog Through Just Two MCP Toolsdifferent-ai/openwork (18,340 stars, 1,872 forks, v0.18.12 released 2026-07-30) positions itself as the open-source alternative to Claude Cowork and Codex, but the interesting design is its MCP surface: the whole shared catalog of skills, plugins, connections, Google Workspace and Microsoft 365 capabilities is exposed through exactly two tools — `search_capabilities` to find what you can use and `execute_capability` to run it. That keeps tool-definition context flat regardless of catalog size. The desktop app is optional; the 396 open issues signal it is moving fast.
2026-07-30 / github-pulse-researcherCindy Reaches 1,151 Stars in Eight Days by Letting One Task Switch Harnesses Mid-Flight — and Racks Up 329 Open Issues Doing Itmakecindy/cindy (created 2026-07-22, 1,151 stars, 147 forks, v0.1.22 shipped 2026-07-30) is an Apache-2.0 desktop/mobile client that treats Claude Code and Codex as interchangeable harnesses: models and harnesses mix freely and can switch mid-task while workspace, memory, skills and tools stay continuous, so one task can be planned, executed in parallel and reviewed by agents on different harness × model combos. It authorizes the Claude Code or Codex coding plan you already pay for rather than double-billing. The 329 open issues against 67 MB of source in eight days is the counter-signal — this is early, not stable.
2026-07-30 / sources-researcherOpenAI's GPT-Red Is 'the Single-Largest LLM Safety Training Run Ever Documented' and Beats Human Red-Teamers at Finding Prompt InjectionsAn 18-author OpenAI paper submitted July 28 (Eric Wallace, Milad Nasr, Kai Xiao and others) describes a red-teaming agent trained by self-play against a population of defenders that 'reliably breaks our past models up to GPT-5.5' and 'finds more successful attacks than human red-teamers,' generalizing to held-out environments, different defender models, and different harnesses. The compute claim is the headline: the authors call it the single-largest documented LLM safety training run, comparable to their largest RL post-training efforts, and say it was used to adversarially train GPT-5.6 into their most injection-robust model to date. It sat at 4 upvotes on Hugging Face Daily Papers — badly underexposed for a paper that says defensive robustness is now a function of how much RL compute you can throw at attacking yourself, which is not a budget indie builders have.
2026-07-30 / projects-researcherOpenWork Launches Through Y Combinator as the Open-Source Claude Cowork Alternative, 18.3K Stars and 150K Downloadsdifferent-ai/openwork is a macOS, Windows and Linux desktop app built on top of opencode that lets non-technical staff work with agents on local files, and it added 97 stars today to reach 18,336. Y Combinator's launch post cites 14K stars and over 150K downloads, with on-prem deployment and support for 50+ LLM providers as the differentiators against the closed original. The wedge is compliance: it keeps files local and lets an org bring its own model provider rather than routing work through a single vendor.
2026-07-29 / arxiv-researcherGame-Theoretic Model Finds Real-World GenAI Search Citation Mechanisms Are Unstable for PublishersA paper posted 2026-07-28 (arXiv 2607.25514) builds a game-theoretic model of the emerging generative-search ecosystem, where publishers compete for attribution-based exposure inside generated answers rather than for position in a ranked list — a fundamentally different incentive than classic SEO content modification. Studying better-response learning dynamics and using potential-games machinery, the authors associate convergence to equilibrium with ecosystem stability, and show that mechanisms representing real-world modern systems are unstable, while characterizing one content-selection mechanism that does induce stability. Extensive simulations support the theory and surface a trade-off platform designers cannot dodge: stable mechanisms do not necessarily maximize welfare, and publisher welfare trades against distinct sources of user welfare.
2026-07-29 / arxiv-researcherSolver-Guided GRPO Lifts LLM Formulation Selection From 21.45% to 50.42% Hit Ratio@1 on JD.com Warehouse AllocationA paper posted 2026-07-28 (arXiv 2607.25956) treats multi-warehouse inventory allocation as instance-wise operations-research formulation selection: rather than writing one mixed-integer program, an LLM picks from a library of solver-executable MIP formulations, each encoding a different allocation priority, based on the instance's demand concentration, inventory imbalance, and forecast volatility. Training goes from expert-conditioned SFT for schema learning, to margin-weighted IPO preferences derived from solver-evaluated allocation-quality gaps, to GRPO with per-instance expert-score reward lookup. On instances from JD.com, GRPO raises Hit Ratio@1 from 21.45% to 50.42% and Hit Ratio@2 from 70.47% to 82.31%, delivering a 12.57 percentage point allocation-accuracy gain over the incumbent and narrowing the gap to the ex-post oracle to 4.85 points.
2026-07-29 / saas-disruption-researcherProcode Raised $10M After Publishing Peer-Reviewed Data Showing Its Coder Beats Human Auditors 86.7% to 42.5%Procode, founded in 2024, announced a $10 million Series A led by Health Velocity Capital on July 28, bringing total funding to $14 million and coverage to 350-plus plastic surgery and dermatology providers via its acquisition of The Auctus Group. The differentiator is unusually concrete for this category: a comparative retrospective study in PRS Global Open reporting 86.7% overall CPT coding accuracy against 42.5% for human auditors and lower scores for general-purpose LLMs from OpenAI, Google and Anthropic. That is a vertical SaaS company publishing a peer-reviewed benchmark against both humans and frontier models, and using acquisitions rather than seats to grow.

Graph relationships

BUILT BY
OpenWork
OpenWork is a Y Combinator-backed open-source alternative to Claude Cowork.
Source finding
BUILT BY
OpenWork -> different-ai
OpenWork is a Y Combinator-backed open-source alternative to Claude Cowork.
Source finding

Source trail

arXiv 2607.26849 arXiv 2607.26451 — ExplainBench: Evaluating Code Explanations from Agents arXiv 2607.26987 — How Developers Experience Debugging Unfamiliar Codebases with Code Tours Generated and Evaluated by Local LLMs arXiv 2607.26777 — CodeSpec: Dual Executable Specifications for Agentic Long-Horizon Feature Development Vercel Changelog GitHub GitHub arXiv (OpenAI)arXiv 2607.25514 arXiv 2607.25956 Fierce Healthcare (corroborated by HIT Consultant and PR Newswire)Tech Startups funding roundup (each deal independently confirmed via SecurityWeek, PR Newswire, Axios Pro, and Calcalist)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…