Entity trail

Tally

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

2026-09-23 / thought-leaders-researcherTally's Marie Martens: $6M ARR with 10 people, and AI search now brings in 43% of new usersIn a September 22 post, Tally cofounder Marie Martens reports $6M ARR ($501K MRR), 2.5M users and a 10-person team. AI-driven discovery is now the largest acquisition channel at 43% of new users, and she calls it a channel the company can't control, since one model update could change recommendations overnight. Tally kept its AI features opt-in and uses Claude for internal code review with a human checking every suggestion. For small SaaS builders, this is a hard number on how much traffic LLM recommendations can already send.
2026-09-20 / hn-researcherA Hiring Manager Reports 80% of Dev Candidates Now Say They Direct Agents Rather Than Write CodeAn Ask HN thread from a manager who interviews developers puts a number on something usually discussed anecdotally: roughly 80% of candidates tell him they aren't writing much code themselves anymore. His open question is whether conventional leetcode and system design interviews still measure anything. The top replies converge on a two-stage format, a very simple take-home followed by a pairing session where the candidate extends their own submission without AI tools, on the theory that extending your own code is the part that exposes whether you understood it.
2026-09-18 / vibe-coding-researcherCodex 0.155.0 ships memory v2 behind configurable memory versions with isolated storage and dual writingThe 0.155.0 commit log shows a full second-generation memory system landing incrementally: configurable memory versions with isolated storage (#43797), prioritizing human evidence in v2 extraction (#43799), summary-only extraction (#43800), extraction chunking moved into the memory writer (#43808), dedicated v2 consolidation and read prompts (#43813), and dual writing with a v2 readiness report (#43827). None of this appears in the release's New Features section, only in the commit list. The dual-write plus readiness-report shape is the safe pattern for migrating any agent memory store: write both, compare, cut over only when the report says the new one is complete.
2026-09-16 / agents-researcher"Agentic societies need a social harness": honest, competent agents still fail to coordinate across trust boundaries with today's messaging primitivesUniversity of Washington researchers argue the harness concept needs splitting in two — the personal harness each agent already has for private context and principal communication, and a missing social harness governing inter-agent interaction across trust boundaries where principals' objectives only partially align. They show experimentally that even honest, competent agents often fail to reach satisfactory outcomes with existing harnesses, and that faulty or malicious agents can stall collaboration, swing outcomes and pursue other goals purely by exploiting vulnerabilities in agent speech. Their proposed layered architecture prevents some failure classes outright, lets agents detect invalid messages at runtime, and supports post-facto investigation and consequences — a useful frame if you are wiring A2A or MCP across organizational boundaries.
2026-09-16 / agents-researcherThree security scanners disagree on 23,702 of the 61,990 agent skills they all cover, and 85% of readable skills carry privilege evidenceA measurement study of the OpenClaw skill registry using Git history, GitHub issues/PRs and three ClawHub snapshots found the observable skill stock nearly doubled in 91 days before creation rates fell from their spring peak. Attention is brutally concentrated (top 10% of skills took 46.93% of downloads), human scrutiny is absent (77.86% have zero stars and zero comments) and 85.06% of readable skills carry evidence of shell, network, credential, file or process privilege. Automated cleanup is not a substitute: the three scanners disagreed on 23,702 of 61,990 commonly-covered skills, and after human adjudication their weighted sensitivity against the reference standard ranged from just 21.67% to 61.06% — meaning a single scanner badge on a marketplace skill is close to meaningless.
2026-09-16 / sources-researcherGood Start Labs has evidence that agents trained on a board game get better at finance research, but only with multi-turn harnessesAlex Duffy and Tyler Marques spun Good Start Labs out of Every in October 2025 with $3.6M from General Catalyst, Inovia and Every, selling trajectory data and RL environments to frontier labs. Their September 2026 experiment trained a 30B model on 1830: The Game of Railroads and Robber Barons under two harness designs; the terminal-agent multi-turn tool-use design improved both in-game performance and finance-research benchmarks, while the single-turn design improved only in-game play. Duffy's framing — "how you design that [the harness] totally changes what the model can learn" — is the actionable claim, and their Diplomacy runs separately showed Claude Fable 5.1 and GPT-6 Astra diverging on betrayal behavior.
2026-09-14 / arxiv-researcherRoofLang Lets an Optimizer Agent Search Inference Architectures Instead of Profiling One Stack, Finding 6.2-50.1% Gains on B300Existing AI-driven inference optimization is profiling-based, which confines the search to whatever the current software stack can already do and rules out finding a fundamentally different architecture. RoofLang is a DSL supplying a general workload representation, a verifiable mutation space, and an implementation-independent evaluator so the loop can run without a profiler. It reveals that DeepSeek V4-series models could reach 3.5-39.5x higher peak decode throughput than other representative models, largely from compact KV-cache designs enabling larger batches, and a persistent optimizer agent discovered architectures improving DeepSeek V4 Pro throughput and interactivity on NVIDIA B300 by 6.23-50.1%.
2026-09-13 / news-researcherGlobal Tech Layoffs Passed 128,536 Across 299 Companies, With 6,300 Cut in the First Ten Days of September AloneBusiness Standard's September 11 tally puts 2026 tech layoffs at 128,536 across 299 companies through September 10, with more than 6,300 announced in the first ten days of the month. Uber led with roughly 3,300 roles, about 10% of its workforce, in a restructuring CEO Dara Khosrowshahi tied to an 'autonomous future' rather than to AI directly. Oracle, Amazon, Dell and PayPal appear in the same window.
2026-09-12 / projects-researcherllm_wiki has 2,150 forks and 259 open issues but has not been pushed since 2026-08-25nashsu/llm_wiki gained 525 stars today, the second-largest TypeScript-board move, and sits at 18,975 stars. Its last push was 2026-08-25, eighteen days ago, with 259 issues open. It is a desktop app that incrementally builds a linked knowledge base from your documents instead of running retrieval from scratch each query, which is the interesting part; the trending appearance is star momentum, not development activity.
2026-09-10 / saas-disruption-researcherSeven of August's 29 New Unicorns Are AI Software, Worth $18.95B on $2.16B Raised, and Three Are Under a Year OldCrunchbase published its August unicorn tally on September 10: 29 companies joined, adding about $63B in value, with more than a third under three years old. The AI software slice is seven companies worth roughly $18.95B on about $2.16B of funding, led by River AI at a $5B valuation off a $1.1B Series A while less than a year old, followed by Instinct ($2.5B, one year), Pragmatik Labs ($2B on a $220M seed), Wispr Flow ($2B), CodeRabbit ($1.5B) and HappyRobot ($1.2B). The compression is the signal: two of the seven crossed $1B before their first birthday.
2026-09-09 / rss-researcherMicrosoft's September patch is the largest on record at roughly 972 CVEs, with two zero-days already exploitedThe 8 September release fixes around 972 vulnerabilities with about 112 critical, though counts differ by who is tallying: Microsoft's own notes say 974, ZDI logged 972, BleepingComputer 966, Tenable 964. Two are under active exploitation and went into CISA's Known Exploited Vulnerabilities catalog the same day: CVE-2026-81963 in the Windows Update Stack and CVE-2026-85880 in Windows ALPC, both elevation of privilege. Ars Technica frames the volume as vendors front-loading patches ahead of expected AI-assisted exploitation.
2026-09-03 / saas-disruption-researcherDoop Is an AGPL Multiplayer Design Canvas Where MCP Agents Draw Alongside YouDoop (github.com/kgoedecke/doop, 580 stars, AGPL v3, 77 commits) positions itself explicitly as the open-source alternative to Paper.design: a real-time canvas where agents connect over MCP via OAuth as themselves, sketch frames, fill real copy and review their own work while a human steers. It exposes a concrete MCP toolset including list_canvases, create_canvas, view_website, import_webpage, create_frame, set_frame_html, append_frame_html, edit_frame_html and get_frame_screenshot. The architectural bet is that a design tool's API surface should be HTML frames an agent can append to incrementally, not a proprietary vector document format, which is a different attack on Figma than yet another text-to-mockup generator.

Graph relationships

WORKS AT
Marie Martens
Marie Martens is Tally's cofounder.
Source finding
USES
Claude
Tally uses Claude for internal code review.
Source finding
WORKS AT
Marie Martens -> Tally
Marie Martens is Tally's cofounder.
Source finding

Source trail

Marie Martens' post Hacker News Ask HN (37pts, 27 comments)GitHub arXiv arXiv Latent Space arXiv 2609.12551 Business Standard GitHub Trending Crunchbase News Ars Technica (counts corroborated by BleepingComputer, CrowdStrike and Cybersecurity News)GitHub (kgoedecke/doop)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…