NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-14ENGLISH EDITION
This issue
—
All time
—

Hacker News

7 stories
01

Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model

Technical assessments of Rio de Janeiro's municipal large language model, Rio3.5, suggest it is not an independently trained system but a merge of pre-existing open-source models. Public developer discussions linked the model's weights and behavioral patterns to the Nex-AGI repository. This development follows official announcements claiming that Rio3.5 had outperformed Qwen3.7 in several standardized AI benchmarks. The revelation raises transparency concerns regarding regional public sector AI investments and the authenticity of government-led model training initiatives (source: https://github.com/nex-agi/Nex-N2/issues/4; discussion: https://twitter.com/zenmagnets/status/2065796012820848699).

Large Language ModelModel MergingOpen-Source AIModel SovereigntyNex-N2Artificial IntelligenceLarge Language ModelNatural Language Processing
02

Don't trust large context windows

This technical critique examines reliability failures and attention dilution within the increasingly massive context windows of modern Large Language Models (LLMs). The analysis highlights the 'needle in a haystack' problem, where models struggle with degraded retrieval accuracy when processing dense input data. Simply expanding context sizes does not ensure effective reasoning or reliable retrieval. To prevent silent failures, developers are advised to implement robust preprocessing, semantic chunking, or retrieval-augmented generation (RAG) strategies rather than relying solely on large context lengths in production (source: https://garrit.xyz/posts/2026-05-06-dont-trust-large-context-windows).

Context WindowLarge Language ModelsNeedle in a HaystackInformation RetrievalRetrieval-Augmented GenerationArtificial IntelligenceLarge Language ModelNatural Language Processing
03

KPMG pulls report on AI usage due to apparent hallucinations

Professional services firm KPMG retracted an official research report on corporate artificial intelligence adoption due to data errors caused by AI hallucinations. The retraction highlights operational risks when enterprises deploy generative AI and large language models for public reporting without manual validation processes. The incident underscores the critical need for strict data verification, auditing, and quality assurance frameworks inside consulting firms and corporations to maintain public credibility and avoid relying on flawed automated research (source: https://techcrunch.com/2026/06/13/kpmg-pulls-report-on-ai-usage-due-to-apparent-hallucinations/).

AI HallucinationsGenerative AIData IntegrityEnterprise AILarge Language ModelsArtificial IntelligenceGenerative AILarge Language Model
04

Extinction-Level Capitalism

Author Matthew Butterick published an essay analyzing how the rapid development of generative AI by major technology conglomerates threatens the creative and digital ecosystem. The article argues that massive corporate operations rely on the systematic extraction of public intellectual property and dataset harvesting without providing fair compensation. This capitalistic model risks causing structural damage to open-source software and independent creative industries. Butterick calls for stricter regulatory oversight and sustainable alternatives to prevent tech monopolies from decimating independent content creators (source: https://matthewbutterick.com/extinction-level-capitalism.html).

Generative AIAI EthicsCopyright LawData ExtractionTech MonopolyArtificial IntelligenceGenerative AIOthers
05

Can't Stop the Signal. Poison It

This technical post proposes a dataset poisoning technique to defend digital assets against aggressive, unauthorized web scrapers training generative AI models. Because traditional protective measures like rate limits or robots.txt are frequently ignored, administrators can serve corrupted or subtly modified adversarial data to detected scrapers. This targeted defense corrupts dataset integrity, rendering extracted data useless or harmful for training neural networks and large language models, presenting an active defense paradigm for open-web publishers (source: https://blog.digitalgrease.dev/posts/fauxx-cant-stop-the-signal).

Data PoisoningWeb ScrapingAdversarial DefenseModel TrainingIntellectual PropertyArtificial IntelligenceMachine LearningOthers
06

No, everyone is not using AI for everything

Gabriel Weinberg analyzed consumption data to challenge the common narrative of ubiquitous, all-encompassing artificial intelligence adoption. Weinberg argues that generative AI integration actually mirrors historical tech adoption curves, characterized by gradual and highly specialized deployments rather than broad universal use. Organizations and individuals are deploying AI tools selectively for tasks with proven, narrow value. The article calls for a pragmatic perspective on actual market penetration, contrasting corporate marketing narratives with documented real-world workflows (source: https://gabrielweinberg.com/p/people-are-consuming-ai-like-they).

AI AdoptionGenerative AITechnology IntegrationMarket HypeConsumer BehaviorArtificial IntelligenceGenerative AIOthers
07

Pac-Man, but you're the ghost

Developer Garrit released an inverted retro game where the player controls one of the classic ghosts while chasing an autonomous, AI-driven Pac-Man agent. The project focuses on pathfinding algorithms and state machine management for the computer-controlled Pac-Man, which dynamically routes through the maze to avoid players and target power pellets. It provides an interactive sandbox for state machine planning and real-time behavioral decision-making (source: https://garrit.xyz/posts/2026-06-13-pac-man-but-you-re-the-ghost).

Game DevelopmentPathfinding AlgorithmsState MachinesArtificial IntelligenceGame DesignArtificial IntelligenceAI AgentOthers

Twitter

1 story
01

Innovative Advancements In Real Time Generative Video Technology

Cristobal Valenzuela announced real-time generative video capabilities, representing an architectural optimization in synthetic media production. The technology enables high-fidelity visual content generation with improved speed and creative control, streamlining traditional rendering and animation pipelines. This development aligns with the industry-wide trend of integrating multimodal systems directly into creative production workflows. (source: https://x.com/c_valenzuelab/status/2066003457262207310)

Generative AIVideo SynthesisCreative TechVisual GenerationReal-time AIArtificial IntelligenceGenerative AIVideo Understanding