MindPattern
← The Wire
Source trail

The Decoder

Public MindPattern findings, entities, and graph evidence that cite this source.

Findings
19
All-time hits
18
High value
9
Last seen
2026-06-28

Connected entities

The DecoderCursor Unveils Origin (Agent-First Git Platform), a 1.5T-Parameter In-House Model, and a Mobile AppThe Decoder (corroborated by The Next Web, PC Gamer, Hacker News)Landmark German Ruling: Google Is Liable for False Statements in Its AI OverviewsxAI Reportedly Trained Grok Coding Models on Claude Outputs for Months — Continued via Personal AccoThe Decoder / AnthropicAnthropic's Mythos Solves 80-Year Erdős Unit Distance Problem With 'Cute, Simple Proof' via Claude CExploitBench: Claude Mythos Achieves Arbitrary Code Execution on 21 of 41 Real V8 Vulnerabilities — Multiple (VentureBeat, The Decoder, XDA)Pattern: Anthropic's Credit Metering Reveals Agentic Usage Far Exceeds Subscription Economics — FlatGoogle DeepMind 'AI Agent Traps': Six Attack Categories, 86% Hidden Prompt Injection Success RateARC-AGI-3 Breaks Every Frontier Model: GPT-5.4 Scores 0.26%, Opus 4.6 0.25%, Grok-4.20 0.00% — HumanAnthropic Drops Long-Context Surcharge: 1M Token Context Now at Standard Per-Token Pricing for Opus

Related findings

  1. 2026-06-28 / NEWSOpenAI Weighs Drastic Token Price Cuts as Claude Code Devours Developer ShareOpenAI is preparing significant token price cuts aimed explicitly at winning back developers from Anthropic, whose Claude Code coding tool went viral and helped drive Anthropic to its first profitable quarter. Anthropic, meanwhile, reversed an unpopular billing overhaul; its Claude Fable 5 launched June 9 at $10/$50 per million input/output tokens, double GPT-5.5's $5/$30. With both companies heading toward public listings, a price war could pressure margins exactly when investors are scrutinizing the IPOs.
  2. 2026-06-23 / NEWSCursor Unveils Origin (Agent-First Git Platform), a 1.5T-Parameter In-House Model, and a Mobile AppAt its June 16 #compile keynote, Anysphere announced Origin — a GitHub/GitLab alternative built for AI agents on the Graphite tech it acquired in Dec 2025, with GA targeted for fall 2026 — alongside a proprietary 1.5T+ parameter model trained from scratch on 100K+ GPUs (reportedly xAI's Colossus) and a Cursor iOS app. The launch lands the same week SpaceX announced its $60B all-stock acquisition of the company. For builders this signals the coding-agent stack moving down to its own VCS substrate, not just the editor.
  3. 2026-06-20 / VOICESSam Altman's New 40-Minute Interview: GPT-5 'Probably This Summer' but May Slip on Safety Testing, Teases an o3 'High-Performance' TierIn a 40-minute interview released June 19, Sam Altman said GPT-5 is 'probably coming sometime this summer' while explicitly hedging that it could slip on security testing, naming, and feature iteration, and teased an o3 'high-performance' model plus Deep Research upgrades. Separately he admitted he's '0% excited' to run a public company even as OpenAI signals an IPO within a year. Builder takeaway: plan around a fuzzy summer GPT-5 window, not a fixed date — the CEO is the one hedging the timeline.
  4. 2026-06-16 / REDDITMeta's 'Hatch' Developer Agent Heads to Internal Testing by End of JuneReporting on the personal-agent race notes Meta is targeting end-of-June internal testing for Hatch, a developer-focused agent product under Alexandr Wang's Superintelligence Labs, as Google and Meta scramble to catch Anthropic and OpenAI on agentic workflows. It's an early, unreleased signal rather than a shipping product, but it indicates Meta is moving from foundation-model catch-up (Muse Spark, April) toward agent tooling. Builders should treat this as a watch-item, not yet an adoptable tool.
  5. 2026-06-10 / SOURCESLandmark German Ruling: Google Is Liable for False Statements in Its AI OverviewsThe Regional Court of Munich (case no. 26 O 869/26) issued a preliminary injunction treating Google's AI Overviews as Google's own content and holding it liable after the feature falsely tied two Munich-based publishers to scams and subscription traps, drawing connections that appeared in none of the linked sources. The court rejected Google's defense that users are responsible for fact-checking, ruling that AI Overviews generate 'independent, new, and substantive statements' so prior case law shielding search-engine operators from liability does not apply. It is a preliminary regional injunction (appealable, not binding precedent under Germany's civil-law system), but a significant signal for anyone shipping generative search or summarization in the EU.
  6. 2026-06-08 / REDDITClaude Web Traffic Up 306% in a Single Quarter as ChatGPT Share Falls to 54.7%New Similarweb data (via Momentic/The Decoder) shows Claude grew 306% in a quarter — from 203M visits in January to 824M in April 2026 — vaulting it to the third-largest consumer AI chatbot, with US share at 12.5%. ChatGPT's worldwide web-visit share fell from ~76.5% (Feb 2025) to 54.7%, while Gemini doubled to 27.4%. The signal for builders: the assistant market is genuinely fragmenting, and Claude is now a mainstream consumer destination, not just a developer tool.
  7. 2026-06-06 / REDDITOpen-Source 'Audio Interaction' Voice Model Decides Every 0.4s Whether to Speak — Apache 2.0 Weights on GitHubA new open-source voice model called Audio Interaction continuously listens and re-decides every 0.4 seconds whether to speak, simultaneously handling translation, transcription, and noise detection in a way the report contrasts with GPT-4o-style systems. Code and weights are released under Apache 2.0 on GitHub, making it a notable candidate for local/real-time voice agent builders.
  8. 2026-06-06 / REDDITMicrosoft CEO Nadella Publicly Rejects Proposal to Make AI Agent 'Scout' Deliberately AddictiveSatya Nadella publicly turned down a VP proposal to engineer Microsoft's Scout agent for addictiveness, framing it as contrary to company values around user empowerment. The episode surfaces an internal industry tension between engagement-maximizing design and the 'agentic assistant' positioning labs are publicly committing to.
  9. 2026-06-06 / REDDITxAI Reportedly Trained Grok Coding Models on Claude Outputs for Months — Continued via Personal Accounts After Anthropic Cut AccessxAI is reported to have trained its coding models directly on Anthropic Claude outputs, continuing through personal accounts and the Blackbox AI intermediary even after Anthropic revoked official access in January for violating its terms barring use to build competing systems. The report also details internal turmoil: a pretraining team shrunk to under five people, four Grok code leads departing, and an engineer accidentally deleting weeks of training data.
  10. 2026-05-27 / REDDITAnthropic's Mythos Solves 80-Year Erdős Unit Distance Problem With 'Cute, Simple Proof' via Claude CodeAnthropic's frontier model Mythos, accessed through Claude Code, has independently solved the planar unit distance problem — an 80-year-old Erdős conjecture from 1946 — using a streamlined geometric approach that mathematicians describe as a 'cute, simple proof.' This follows OpenAI's GPT 5.5 solving the same problem first with a more complex algebraic number theory approach; Mythos achieved the same result without advanced analytic machinery. The convergence of competing frontier models independently solving the same open math problem via qualitatively different proof strategies marks a new capability threshold for AI mathematical reasoning.
  11. 2026-05-17 / SOURCESExploitBench: Claude Mythos Achieves Arbitrary Code Execution on 21 of 41 Real V8 Vulnerabilities — GPT-5.5 Manages Only 2A new arXiv paper (2605.14153) introduces ExploitBench, a 16-flag capability ladder benchmark testing AI models on full exploit synthesis against 41 real V8 JavaScript engine vulnerabilities. Claude Mythos Preview scored 9.90/16 and reached arbitrary code execution on 21 of 41 bugs; GPT-5.5 scored 5.51/16 hitting the top tier on only 2. Researcher Seunghyun Lee noted Mythos 'developed an exploit technique that Lee and a colleague had previously dismissed as too complex.'
  12. 2026-05-14 / TOOLSPattern: Anthropic's Credit Metering Reveals Agentic Usage Far Exceeds Subscription Economics — Flat-Rate Plans Cannot Cover Agent WorkloadsAnthropic's decision to separate programmatic credits reveals a structural mismatch: some $20/month Pro users consumed hundreds to thousands of dollars in API tokens through autonomous agents that 'bypassed caching mechanisms' making flat-rate plans viable. This is the first major pricing signal that agentic AI usage fundamentally differs from chat usage — expect other providers to implement similar metering. Builders should budget for agent-class workloads at API rates ($20/month Pro credit disappears fast at $15/MTok for Opus), not subscription rates.
  13. 2026-04-07 / SOURCESOpenAI Reveals 600K Weekly Health Queries From Hospital Deserts — 70% Come After HoursOpenAI's Head of Business Finance Chengpeng Mou disclosed that US ChatGPT users send ~2M weekly messages on health insurance, with 600K from 'hospital deserts' (30+ min drive to nearest facility). 70% of health queries arrive outside office hours. Wyoming leads at 4.15% of desert messages. With 40M daily health users and 800M total active users, ChatGPT is becoming de facto healthcare infrastructure in underserved areas — raising questions about what happens when the AI gets it wrong.
  14. 2026-04-02 / SOURCESGoogle DeepMind 'AI Agent Traps': Six Attack Categories, 86% Hidden Prompt Injection Success RateA DeepMind study introduces the taxonomy of 'AI agent traps' — six categories attacking perception, reasoning, memory, action, multi-agent dynamics, and human supervisor components of an agent's operating cycle. Hidden prompt injection in HTML/CSS achieves 86% success rate; latent memory poisoning succeeds 80%+ with less than 0.1% data contamination. Every AI agent tested was successfully compromised at least once, with consequences including unauthorized data access. Researchers recommend adversarial hardening and multi-stage runtime filters (source filters, content scanners, output monitors).
  15. 2026-03-29 / VOICESARC-AGI-3 Breaks Every Frontier Model: GPT-5.4 Scores 0.26%, Opus 4.6 0.25%, Grok-4.20 0.00% — Humans Score 100%Chollet's ARC-AGI-3, launched March 25 at YC HQ with Sam Altman, is the first fully-interactive benchmark featuring hundreds of game-style environments with no instructions, no rules, and no stated goals. Gemini 3.1 Pro Preview led frontier models at just 0.37%, while simple CNN and graph-search approaches hit 12.58% — over 30x better than any LLM. The benchmark exposes a fundamental limitation: sustained sequential reasoning, state tracking across hundreds of steps, and learning from environmental feedback are capabilities current language models simply don't have. ARC Prize 2026 offers $2M+ in prizes, all solutions must be open-sourced.
  16. 2026-03-18 / REDDITGPT-4.5 Passes Turing Test at 73% by Actively Suppressing Its Own IntelligenceA Jones & Bergen arXiv study found GPT-4.5 fooled 73% of human judges in five-minute Turing test conversations — but only when instructed to be laconic, blunt, lowercase, with intentional typos, and to feign poor math ability. Without the dumbing-down prompt, the pass rate dropped to 36%. The finding inverts the usual AI capability framing: the model succeeds by hiding competence, not demonstrating it, suggesting Turing test pass rates are calibrated to human limitations rather than AI capability ceilings.
  17. 2026-03-16 / NEWSHelios: Open-Source 14B Video Model Generates 60+ Second Videos at 19.5 FPS on a Single H100 — First Real-Time Long-Video ModelHelios, built by ByteDance, Peking University, and Canva, is the first 14B video model to hit near-real-time inference (19.5 FPS) on a single H100 while producing videos longer than 60 seconds. It achieves this without KV-cache, sparse attention, or quantization — instead using hierarchical memory compression (8x token reduction) and 3-step distillation. Text-to-video, image-to-video, and video-to-video are unified in one architecture. All code and weights are Apache 2.0 on GitHub and Hugging Face.
  18. 2026-03-15 / SOURCESAnthropic Drops Long-Context Surcharge: 1M Token Context Now at Standard Per-Token Pricing for Opus 4.6 and Sonnet 4.6Anthropic eliminated the prior 2x input / 1.5x output long-context pricing premium — a 900,000-token request now costs identical per-token to a 9,000-token one. Opus 4.6 is $5/$25 per million input/output; Sonnet 4.6 is $3/$15. MRCR v2 retrieval accuracy at 1M tokens is 78.3%, best among frontier models; also includes automatic server-side context compaction and a 6x increase in image/PDF per-request limit (600 pages). The pricing change materially improves ROI on long-document agent pipelines.
  19. 2026-02-13 / NEWSZhipu GLM-5: 745B MIT-Licensed Model on Huawei ChipsChinese AI lab Zhipu released GLM-5, 745B MoE model under MIT license, trained entirely on Huawei Ascend chips. Stock surged 30%.
Open latest cited source