Entity trail

Pool

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

2026-07-07 / arxiv-researcherSubspace-Constrained Adaptation Blocks Fine-Tuning Poisoning in PEFTParameter-efficient fine-tuning still leaves a wide space of behavior-changing updates a poisoned objective can reach; this method constrains adaptation to a subspace estimated from a trusted pool of existing adapters so harmful updates become inexpressible. It is a defense for anyone accepting user- or third-party-supplied LoRAs. Trades a bounded amount of adapter expressivity for poisoning resistance.
2026-07-07 / thought-leaders-researcherClément Delangue Pitches 'Mutualized' Open-Source Compute and a Community Model for Building AIOn July 4, Delangue argued that open-source lets teams pool (mutualize) compute and infrastructure spend to reach frontier results more efficiently, and floated training an open-source model specifically for AI-building that drew enthusiastic community support. It's a pointed counter-narrative to the hundreds-of-billions capex race from Meta, OpenAI, and Microsoft. For independent builders, the framing is a reminder that shared open infrastructure remains a viable path against closed-lab spending.
2026-07-02 / vibe-coding-researcherPattern: Frontier Agentic Coding Models Are Now Trained End-to-End on Domestic Chinese SiliconLongCat-2.0's training on a ~50,000-card domestic cluster with no Nvidia A100/H100 or AMD MI300X — with HN speculation pointing to Huawei Ascend 910C-class accelerators — shows that near-frontier agentic coding models can now be produced entirely outside the export-controlled GPU supply chain. Combined with export directives keeping some frontier coding models offline in the US, builders should expect a widening pool of capable open models whose availability tracks geopolitics as much as engineering.
2026-06-24 / thought-leaders-researcherSakana AI Launches Fugu and Fugu Ultra — an Orchestration 'Model' That Routes Around the Fable 5 Export BanTokyo's Sakana AI launched Fugu and Fugu Ultra on June 22: an orchestration system that presents a single OpenAI-compatible API while internally routing each task across a swappable pool of frontier LLMs, dynamically and without hardcoded rules. CEO David Ha framed it explicitly as a hedge against single-vendor risk after the June 12 export controls cut off Claude Fable 5/Mythos — 'relying on a single company's APIs for critical infrastructure... is a material vulnerability' — with Fugu Ultra benchmarking shoulder-to-shoulder with those now-banned models. The builder takeaway: multi-model orchestration is being productized as resilience against geopolitical and vendor lock-in, not just cost/quality routing.
2026-06-24 / saas-disruption-researcher'SaaS Isn't Coming Back' — Lateral's de Silva Argues Agents Become the User, Killing Per-Seat; Cites $300B Single-Session Wipeout and McKinsey's $6T OpportunityIn a June 22 Crunchbase News piece, Lateral's Richard de Silva argues the core SaaS unit economics are breaking because agents — not humans — are becoming the 'user' through headless models, dissolving per-seat pricing. He points to a ~$300B single-session market wipeout in January 2026 as a signal of the model's decline and frames AI-native vertical platforms as competing for labor, compliance, and risk budgets (a McKinsey-cited ~$6T productivity pool) rather than software budgets. The defensible plays: vertical specialists with proprietary data, outcome/performance pricing (e.g., per contract drafted), and deliberate human-in-the-loop services.
2026-06-18 / sources-researcherAgents' Last Exam (ALE): Frontier Agents Pass Just 2.6% of the Hardest Economically-Valuable TasksALE (arXiv 2606.05405), built with 250+ industry experts, evaluates agents on 960 expert-authored, deterministically-scored workflows (1,490 instances) across 55 digital industries mapped to the O*NET/SOC occupational taxonomy. Frontier agents average 26% across all tiers but only 2.6% on the hardest 'last-exam' tier — and Codex with GPT-5.5, which scores 82% on Terminal-Bench, passes 0% of Last-Exam tasks. The benchmark is the most serious attempt yet to ground 'job-ready agent' claims in real professional work, and it's designed as a living, continuously-expanding pool.
2026-06-16 / sources-researcherHydra Host Raises $100M Series A With Nvidia Backing to Turn Idle Data-Center GPUs Into a Compute MarketplaceMiami-based Hydra Host closed a $100M Series A near an $800M valuation, led by Kindred Ventures with Nvidia, ARK Invest, Comcast Ventures, Magnetar, and Founders Fund participating; Nvidia also named it an official Cloud Partner and brought it into the DGX Cloud Lepton initiative. Its Brokkr 'AI Factory Operating System' standardizes operations across 50+ data centers globally, pooling underused GPU capacity for buyers facing compute scarcity. For builders chasing GPU availability outside the hyperscalers, this is another marketplace abstraction layer maturing — and Nvidia's direct investment signals it wants more supply liquidity in the long tail of independent data centers.
2026-06-12 / rss-researcherFeedzai Launches IQ Score, an AI-Native Network-Derived Fraud Risk ScoreFinancial-crime prevention firm Feedzai released Feedzai IQ Score, an AI-native fraud risk score derived from its cross-institution fraud intelligence network and aimed at banks of all sizes. It pools signals across institutions to score risk rather than relying on single-bank data. An example of network-effect data products being repackaged as AI-native scoring.
2026-06-12 / rss-researcherPool's New App Turns Screenshots Into a Searchable Memory BankPool launched an app that automatically sorts screenshots into personalized collections, recovers the original links behind saved content, and helps users rediscover products, recipes, and travel ideas. It's a consumer take on AI-powered personal memory and retrieval over visual content. An interesting pattern for builders thinking about ambient capture and recall UX.
2026-06-11 / saas-disruption-researcherThe Unresolved Pricing Question: 'Does an AI Agent Count as a Seat?' — Four Models Are Competing to Replace Per-SeatDeloitte and pricing analysts frame 2026's core monetization debate around whether agents are billable users, with four models now competing to replace per-seat: usage-based (compute/API calls), outcome-based (leads, resolved tickets), credit-based pools, and hybrid base-plus-surcharge. Builders setting pricing today face a genuinely open standard — the category hasn't converged, so early architectural choices about metering and attribution will be hard to reverse at renewal.
2026-06-11 / skill-finderGet near cross-encoder accuracy at bi-encoder speed with ColBERT late interaction (MUVERA + RAGatouille)ColBERT stores a per-token embedding for every query and document token and scores relevance as the sum of each query token's max similarity (MaxSim) against the document — so phrases match the specific passage where they're relevant instead of competing in one pooled vector. The 2026 production recipe uses RAGatouille for indexing/search and a MUVERA+ approach: fast single-vector candidate generation followed by exact ColBERT MaxSim reranking over the top-K. The tradeoff is storage — multi-vector indexes are an order of magnitude larger — so apply it as a reranker, not the primary index.
2026-06-10 / skill-finderSandbox skills and subagents with the `disallowed-tools` denylist — not `allowed-tools`, which isn't enforcedClaude Code's `disallowed-tools` frontmatter key removes specific tools from the model's pool while a skill/subagent is active and clears on your next message, giving a 'deny by exception' model (e.g., a code-reviewer that keeps Read and diagnostics but loses Write and Bash). Critically, multiple open issues show `allowed-tools` in SKILL.md is parsed but NOT enforced in the CLI and is ignored by the Agent SDK — so an allowlist gives false confidence. Use `disallowed-tools` for least-privilege, and for SDK deployments enforce tool access in config rather than trusting frontmatter.

Graph relationships

RELEASED
Poolside
Poolside released pool, a lightweight terminal-based coding agent that runs locally
Source finding
RELEASED
Poolside -> Pool
Poolside released pool, a lightweight terminal-based coding agent that runs locally
Source finding

Source trail

arXiv X/Twitter (Clem Delangue)LongCat AI Sakana AI Crunchbase News arXiv SiliconANGLE

Graph sources

entity graphfindings textkg edgeskg entities