Entity trail

Edge

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

2026-08-21 / arxiv-researcherIAR Bakes a Document Corpus Into Model Weights for Retrieval-Free QA and Keeps 12.1 Points More General Ability Than SFTarXiv 2608.20281 treats a fixed corpus as something to internalize rather than retrieve, running three separate post-training stages: Inject converts source documents into continuation, rewrite, and instruction-conditioned reconstruction objectives; Align adapts with answer-only QA supervision; Recover merges the domain-adapted model back with the base instruction model. Across Common Corpus and CCI with Llama, Phi, Qwen, and SmolLM, IAR beats vanilla SFT on all four reported metrics in 7 of 8 dataset-model settings, averaging +3.6 points of domain QA accuracy and +12.1 points of mean general performance on IFEval, MMLU, and MSBench. That second number is the point: the usual cost of fine-tuning knowledge in is wrecking everything else, and the merge step recovers most of it.
2026-08-21 / arxiv-researcherSWE-bench Science: Claude Code With Opus-5 Max Scores Under 50% pass@1 on Real Scientific ReposThe OpenMOSS team released SWE-bench Science (arXiv 2608.19799), 119 repository-level tasks drawn from 98 GitHub repos across 20 scientific domains, split into Issue-driven, Expert-exploratory, and Engineering-integration paradigms. The best agent tested, Claude Code with Opus-5 (max), lands below 50% pass@1. A paired ablation that strips explicit scientific guidance while keeping repo and execution context shows domain knowledge is not uniformly helpful: well-grounded facts improve average performance and token efficiency, but poorly aligned guidance anchors the agent and does not raise exact repair success.
2026-08-21 / skill-finderPRAXIS harvests undocumented business rules by simulating developer workflows, then indexes them on the dependency graphDomain code generation fails on tacit knowledge, the business rules, interface contracts, and operational conventions developers absorb but rarely write down. PRAXIS simulates typical developer workflows inside the target codebase to extract that knowledge, structures it against the code dependency graph, and pushes relevant pieces into context proactively rather than waiting for the agent to search. It beat search-equipped state-of-the-art agents plus experience-based and skill-based alternatives across architectures and models, with gains growing as accumulated experience grows, though the abstract gives no headline numbers.
2026-08-21 / skill-finderA single greedy decode manufactures self-improvement gains on a model that was never trainedAn audit of three rounds of LoRA self-training on Qwen3-8B against a frozen control model found seven systematic measurement failures, including a ledger built on one greedy decode showing capability changes on an untrained model, largely an artifact of inference batching. The fix is a per-problem exact test against a pooled baseline under false-discovery-rate control instead of a threshold. Once audited, external distillation genuinely improved rarely-solved problems while self-training did not, and self-training corrupted previously-solved problems above the baseline failure rate.
2026-08-21 / news-researcherDeepSeek Ships V4-Flash-Vision-Exp, Claiming Multimodal Agent Performance Near Opus 4.8DeepSeek announced on August 21 that deepseek-v4-flash-vision-exp is live on its API platform, an experimental multimodal model that matches V4-Flash on text, agents, reasoning and world knowledge while making what the company calls a major leap on multimodal agent benchmarks, close to Opus-4.8. Images are billed at up to 384 tokens each at standard V4-Flash pricing, with base64 input across Chat Completions, Messages and Responses endpoints. DeepSeek Harness 0.1.1 shipped the same day with out-of-the-box support, so the cheapest credible vision-agent option just changed.
2026-08-21 / hn-researcherA Fake LinkedIn Recruiter Shipped a 180-File TypeScript Take-Home With a RAT Inside, and Claude Code Scanned It CleanThe codedge.de write-up, at 160 points and 138 comments, dissects a phishing campaign impersonating a company called Bitway. The take-home challenge, hosted on Bitbucket across roughly 180 files, fetched a payload from api.jsonbin.io on 'npm run dev', unwrapped 24,686 characters of obfuscator.io JavaScript around a webpack bundle, and phoned home to [redacted]. Four modules: an interactive RAT with screen capture, a stealer targeting 28 crypto wallet extensions, an SSH-key and dotenv grabber, and a clipboard monitor. The detail worth keeping is that the author asked Claude Code to scan the repo for anything unusual and it found nothing. Post dated August 17, updated August 20.
2026-08-21 / hn-researcherOpenRouter's Free Stealth Model 'Ox Alpha' Ships a 1,048,576-Token Context and Video Input, and HN Spent 142 Comments Trying to Unmask ItA provider-less model called Ox Alpha appeared on OpenRouter listed under the 'stealth' provider with a 1,048,576-token context window, 131,072-token max output, and text, image and video input, priced at zero with roughly a 1,000-request daily cap for accounts with prior spend. The HN thread hit 183 points and 142 comments, with most commenters guessing an unreleased Zhipu GLM 5.x multimodal variant based on thinking-trace style and a mid-2025 knowledge cutoff, while others pointed at DeepSeek or Xiaomi's earlier 'Hunter Alpha' stealth naming. Testers reported Chinese-government-aligned answers on Tibet and Taiwan, which is the strongest circumstantial evidence in the thread. Free frontier-class context is worth grabbing this week, but treat the retention policy as unverified.
2026-08-21 / thought-leaders-researcherDeepMind's 'contract-first' agent delegation paper resurfaces on X with 68K views six months after publicationA post on August 20 pitching a Google DeepMind paper that treats AI delegation as verifiable contracts rather than prompt handoffs cleared 68,665 views and 587 likes, but the paper is from February 12, 2026 by Nenad Tomasev, Matija Franklin and Simon Osindero. Its actual rule is strict: a delegator may not assign a task unless the outcome can be precisely verified, and must recursively decompose until sub-tasks map to automated checks such as a passing unit test, a formal proof, or a strict schema. It also proposes policy-as-code permissioning and zero-knowledge proofs for verifying correctness without exposing data.
2026-08-20 / reddit-researcherClaude Pro and 5x Users Report Burning a Full Five-Hour Window on Two Prompts, With Per-Prompt Credit Costs 30x NormalA Pro user reported exhausting their five-hour limit on two creative-writing prompts with Opus 4.6, medium effort, extended thinking on and cross-chat memory off, a session that normally takes 50 to 100 comparable prompts. Testing with credits, three prompts that usually cost $0.03 to $0.15 each came back at $3.56, $3.80 and $5.10, with the most expensive one occurring after deleting chats to rule out global accumulation. A 5x subscriber in the thread reports hitting 50% of a window instantly after moving work over from Sol, and a separate same-day thread complains Claude is refusing work before the stated limit is reached. Single-source practitioner reports with no vendor acknowledgement yet.
2026-08-20 / reddit-researcherA Depth-Pruned Qwen3.8-27B Cut to 22.7B With No Fine-Tuning, and the Author Refuses to Claim It Beats AnythingStargazerLabs released Qwen3.8-23B-Mini-Me, a strategic layer-removal prune of Qwen3.8-27B down to roughly 22.7B parameters with no fine-tuning afterward, in bf16, q8 and q4 (MLX only at the moment). The author explicitly declines to publish benchmarks or claim superiority, describing it as a smaller, faster version that is slightly worse at some things. The useful failure mode they do report: it handles standard coding fine but degrades against the original on edge cases and underspecified prompts, where the full model's extra capacity does the inference work.
2026-08-20 / reddit-researcherQwen3.8-27B Lost Ground on Raw Knowledge Against Qwen3.6, and Artificial Analysis' Omniscience Benchmark Backs the VibesA practitioner running a private trivia set found Qwen3.8-27B failing questions Qwen3.6 answered reliably, at every quantization level and sampling setting tried, then checked Artificial Analysis' Omniscience evaluation and found the same regression in offline no-tool knowledge accuracy. The top reply (112 upvotes) corroborates it from a different angle: 3.8 is markedly more eager to reach for web search, and with search and fetch tools disabled the knowledge gap shows up clearly. The practical read for builders is that 3.8 appears tuned to lean on tools rather than weights, so an airgapped retrieval-from-weights setup should stay on 3.6.
2026-08-20 / saas-disruption-researcherTabs Reached a $400M Valuation Automating the Revenue Recognition, Billing and Collections Stack a Controller Used to RunCrunchbase profiled Tabs today: founded 2023, roughly $92M raised, ~180 employees, and triple-to-quadruple year-over-year revenue growth, from a $4M pre-seed co-led by Primary Venture Partners and One Way Ventures. Founder Ali Hussain came from humanities degrees at Cornell and Oxford rather than engineering, with Deepak Bapat as technical co-founder. Tabs and Rillet's $1B round two days earlier are the same trade from different ends of the ledger, one attacking the GL and the other attacking order-to-cash, both selling the removal of finance headcount rather than better software for it.

Graph relationships

USES
Gemini Voyager
Gemini Voyager is available on Edge.
Source finding
COMPETES WITH
Firefox
Firefox positions itself as the consumer-privacy alternative while Edge pushes AI aggressively.
Source finding
RELEASED
Microsoft
Microsoft announced Edge for Business data loss prevention.
Source finding
RELEASED
Microsoft
Microsoft developed Edge browser
Source finding
COMPETES WITH
Chrome
Edge stores all passwords in cleartext memory at launch unlike Chrome's on-demand decryption approach
Source finding
USES
Gemini Voyager -> Edge
Gemini Voyager is available on Edge.
Source finding
COMPETES WITH
Firefox -> Edge
Firefox positions itself as the consumer-privacy alternative while Edge pushes AI aggressively.
Source finding
RELEASED
Microsoft -> Edge
Microsoft announced Edge for Business data loss prevention.
Source finding
RELEASED
Microsoft -> Edge
Microsoft developed Edge browser
Source finding

Source trail

NoToAI.org Vercel wanshuiyin/HERO-Anti-OverDefense neuroarxiv arXiv CNBC Google's Chrome team imperative API demonstrated Trend Micro's follow-up Bankrbot crypto agent Dark Reading

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…