✷ The archive · 6,167 dispatches
All dispatches
Everything AgentLensHQ has filed — distilled from across the AI ecosystem.
The LLMentalist Effect: Why Chat‑Based LLMs Mimic Psychic Cold‑Reading
Chat‑based large language models create an intelligence illusion by using statistical validation statements that exploit subjective validation, a mechanism identical to a psychic’s cold‑reading con.
NYT Lawsuit Reveals Microsoft Exec Calls AI Scraping ‘Largest Theft of Labor’ and OpenAI Calls ChatGPT an ‘Existential Threat’ to Publishers
Legal briefs filed in the New York Times lawsuit disclose that a Microsoft director labeled AI data scraping the biggest theft of labor in history while OpenAI’s head called ChatGPT an existential threat to publishers, intensifying the copyright battle over AI training data.
Interviewing Software Engineers in the Age of AI Coding Assistants
Engineering leaders are shifting from testing manual coding syntax to evaluating problem decomposition, AI steering capabilities, and the ability to review and debug AI-generated code.
Factoring RSA-896 using Claude and GPU Clusters
An Anthropic employee used Claude to port CADO-NFS to GPUs and orchestrate a fleet of 2,048 idle GPUs to factor the RSA-896 challenge number in 10 days.
LLM Compression by Block Removal with Constrained Binary Optimization
Hugging Face introduced a new block‑removal pruning method that formulates transformer block selection as a constrained binary optimization (Ising) problem, achieving up to 23 MMLU points improvement over prior techniques at 50% depth compression.
OpenAI Announces Independent Mathematics Advisory Group
OpenAI has formed an independent advisory group of leading mathematicians to guide the responsible communication and deployment of its rapidly advancing math‑solving AI models.
Higgsfield AI launches GPT-6 Astra video features for rapid ad creation
Higgsfield AI announced that its integration of OpenAI's GPT-6 Astra enables single‑engineer development of new video ad creation tools in a day, dramatically speeding up creative workflows for small businesses.
Laya Open-Source System 1 Decision Model Beats TypeSafe Jev
Laya delivers sub‑35 ms calibrated decisions across 100+ languages, open‑source weights, and 6–8× lower latency than TypeSafe AI’s proprietary Jev.
AI-Generated Event Posters: How Prompting Can Break the Default Ugly Style
AI can produce event posters that avoid the generic, repetitive look by specifying distinct design aesthetics, but the results still face challenges in hierarchy, cultural nuance, and subtle AI tell‑tale artifacts.
OpenAI Building Standards for the Next Phase of AI
OpenAI proposes the development of global technical standards for frontier AI, specifically targeting recursive self-improvement (RSI) to ensure safety and human control as AI systems automate AI research.
Why Using AI to Write Substantive Text Is Usually a Bad Idea
The author argues that AI‑generated prose harms thinking, introduces subtle errors, and misleads readers, so you should almost never let an LLM write substantive content.
OpenAI Academy Expanded Learning Paths
OpenAI has expanded its Academy with specialized learning paths for developers, leaders, educators, and college students to help organizations build shared AI habits and technical skills.
Brood War Bench: Evaluating LLM Strategic Capability in Real-Time Strategy
The Brood War Bench benchmark reveals that while Codex Astra leads in win rates, no current LLM can play StarCraft: Brood War beyond a beginner level due to failures in real-time orchestration and macro-management.
Grant Sanderson proposes “motivated explanations” as a new metric for mathematical contribution
Grant Sanderson argues that the mathematics community should formally recognize and reward “motivated explanations”—explanations that illuminate why a theorem is natural and useful—as a complement to proof‑centric credit.
AI & Frontier Tech Roundup – Open‑Source LLMs, Humanoid Robotics, and Agentic Finance Highlights
Open‑source LLMs like GLM‑5.3 and DeepSeek V4.1 are outpacing prior frontier models, humanoid robots are moving from labs to public spaces, and new credit‑scoring systems aim to legitimize autonomous trading agents.
AI × Crypto Roundup: Agentic Commerce, Decentralized Compute, and Verifiable Physical Data
AI agents are moving from chat windows to on‑chain commerce, using protocols like AACP for escrow, verification, and reputation, while projects such as Vangrid and Bittensor provide decentralized compute and demand‑driven physical data for Physical AI.
vLLM PD Serving of Qwen3.8-2.4T achieves 5K TPS/GPU and 180 gen tok/s/user
vLLM announced that its PD serving of the Qwen3.8-2.4T model reaches 5,000 total token throughput per GPU and 180 generated tokens per user, demonstrating a full pareto frontier for high‑throughput and low‑latency scenarios.
Hugging Face tokenizers v1 release notes
Hugging Face has released a release candidate for tokenizers v1, which achieves 3 to 30 times faster encoding speeds than v0.23 through SIMD-based splitting, a word cache, and allocation-free merge loops.
V7 Go and OpenAI Model Integration for Institutional Memory
V7 Go leverages OpenAI's GPT-5.6 and GPT-6 Astra models to create a Context Graph that provides AI agents with structured institutional memory for complex enterprise workflows.
xAI Grok 4.7 Release Notes
xAI has released Grok 4.7, a model optimized for long-running coding and professional knowledge work with improved safeguards and price-performance competitiveness.
Alibaba Damo Radar: Open-Source Medical AI for Abdominal CT Diagnosis
Alibaba's Damo Academy has open-sourced Damo Radar, a vision-language model capable of detecting nearly 150 abdominal conditions from CT scans with an average AUC of 0.913.
GPT-6 Astra Solves Unsolved WWI German Radio Cipher
GPT-6 Astra successfully decoded a previously unsolved World War I German radio message by identifying a key used outside its documented timeframe.
OpenAI Jalapeño Chip: How LLMs Cut Design Time to 20 Months
OpenAI’s Jalapeño AI accelerator went from concept to silicon in under 20 months, using its own large language models to accelerate front‑end design, software optimization, and early hardware verification.
Cua and CUA-S1: Open-Source Infrastructure for AI Computer Use
Cua is an open-source framework providing desktop automation drivers, isolated cloud and local VMs, and CUA-S1 specialized decision models to enable AI agents to operate computers across macOS, Windows, and Linux.
Claude Code adds AGENTS.md support
Claude Code now reads AGENTS.md when there is no CLAUDE.md, simplifying project setup and aligning with emerging standards.
RP2350 Secure Debug Access via Photon-Emission-Guided Laser Fault Injection
Researchers from Ledger Donjon demonstrated a method to bypass permanent debug-disable settings on the Raspberry Pi RP2350 (A4 revision) to recover secrets from OTP memory using photon-emission microscopy and laser fault injection.
Qwen-Image-2.1 release notes / what's new
Qwen-Image-2.1 is a unified 7B parameter image model that integrates text-to-image generation, native transparency support, and advanced image editing capabilities into a single efficient architecture.
US Military Near-Launch Incident After AI-Generated False Intelligence Report
An AI hallucination caused a US military intelligence report to falsely claim a Chinese ship was carrying nuclear components, prompting a near‑miss operation that was halted only after the error was discovered.
How I Vibed a Proof of Conway’s Refinement Conjecture with AI and Lean
The author used a multi‑agent AI workflow and Lean formalization to produce a mechanically verified proof of John Conway’s 1976 refinement conjecture for omnific integers, though the proof remains unreviewed by mathematicians.
Microsoft Exec Calls AI Scraping ‘Largest Theft of Labor in Human History’ – New Unredacted Filings Reveal
Unredacted court filings in the New York Times lawsuit show a Microsoft executive labeling AI data scraping as “theft” and detail massive, paywall‑bypassing scraping that harmed publishers’ traffic.
ZCode silently uploads full Git history to Aliyun OSS – investigation and mitigation
ZCode (Zhipu’s AI coding desktop) automatically encrypts and uploads your entire workspace, including .git history, to Aliyun OSS using a server‑held RSA key, and the feature cannot be disabled via UI settings.
Bend 2 and the Vibe‑Coding Trap: Why Skipping Prior Research Leads to Redundant Formal Verification Efforts
Bend 2 illustrates a “vibe‑coding” trap where developers use LLMs to build a new language and proof system without recognizing existing formal‑verification tools, resulting in verbose specifications and proofs that could have been avoided.
AI & Frontier Tech Roundup: Self‑Improving Agents, Local Inference, Open Model Competition, and Physical AI Data Infrastructure
Recent posts highlight breakthroughs in self‑improving coding agents, the rise of local AI workstations, open‑weight model cost advantages, and new decentralized spatial data networks for physical AI.
Cactus Needle 3 Release: 8‑29 MB Foundation Model for On‑Device Tool Calling and Structured Extraction
Cactus Compute’s Needle 3 delivers an 8‑29 MB foundation model that matches or exceeds DeepSeek V4 Flash on tool‑calling and extraction tasks, enabling offline, structured JSON responses on tiny devices.
AI × Crypto Roundup: Agent Payments, Decentralized Compute, Verifiable AI, and Physical Data Marketplaces
Recent social‑media posts show concrete progress in AI‑agent commerce, decentralized AI compute, on‑chain verification, and crowdsourced physical‑world data, highlighting emerging infrastructure for an autonomous AI economy.
OpenJev: Browser-Based Decision Modeling and Logit Readout
OpenJev is a local, browser-based experiment that demonstrates the performance difference between direct logit readout for decision-making and traditional token generation using open-weight models.
Hacktron’s OpenAI Hack: Heap Overflow and SSO Misconfiguration Lead to Internal Repo Compromise
Hacktron chained a libheif heap overflow with an OpenAI SSO flaw to hijack employee ChatGPT accounts and open a pull request in OpenAI’s internal monorepo, demonstrating the rapid, AI‑assisted exploitation of widely used image libraries.
Scry: Programmable Internet Search and MCP Server
Scry is a Model Context Protocol (MCP) server that enables AI agents to run complex SQL-like queries and vector operations over a massive dataset of over 161 billion rows from 43 public internet sources.
How to Write with an LLM: Using AI as a Copyeditor, Not a Ghostwriter
To maintain a human voice and avoid 'AI-flavored' prose, writers should use LLMs strictly as copyeditors to flag flaws rather than ghostwriters to generate text, following two strict rules: never use a suggested word and avoid AI encouragement.
Ternary Bonsai 2 27B Release Notes
Prism ML has released Ternary Bonsai 2 27B, a multimodal model based on Qwen3.8 27B that achieves a 9x smaller memory footprint (5.9GB) while retaining 98.2% of the original model's performance.
Why x86‑TSO Emulation Is a Performance Bottleneck on ARM and How Vendors Are Tackling It
Emulating the x86 Total Store Ordering (TSO) memory model on ARM’s weakly ordered architecture incurs severe performance penalties due to alignment faults, split‑lock handling, and uncached memory, but recent ARM extensions (LRCPC, FEAT_LSE2) and Apple’s hardware TSO mode dramatically improve the situation.
Bend 0.1 release: Fast CPU/GPU language that blocks AI mistakes with formal proofs
Bend 0.1 is a new language that compiles to native code, scales from a single CPU core to GPUs, and uses Lean‑style proofs (LAWS.bend) to prevent AI‑generated bugs.
Everybody's Lost Their Minds – A Critical Look at AI Hype, Resource Misallocation, and Industry Fatigue
The blog post “Everybody’s Lost Their Minds” argues that AI hype has diverted engineering resources from fundamental security practices, increased environmental harm, and left many professionals feeling exhausted and deskilled.
The Meat Proxy Trap: Why 'Brain-Off' AI Development is a Career Dead End
Dan Luu argues that relying on LLMs to do the thinking in software development creates a 'meat proxy' role that is economically unsustainable and technically flawed.
Empirical Study of Coding Harness Design Reveals Context Management, Planning, and Tooling Trade‑offs
The paper “An Empirical Study of Harness Design for Coding Agents” shows that context management, planning, and action‑space choices each have distinct impacts on accuracy and cost, and the optimal harness depends on model strength, task type, and context‑window budget.
Sex, AI, and the Apocalypse: How a Rationalist Subculture Shaped AI Safety, Politics, and Culture
Jacob Coxon's resignation from Anthropic exposed a tightly knit rationalist community whose shared beliefs, rituals, and funding networks now influence AI safety discourse, political power, and even sexual subcultures.
Ax-check: Measuring Agent Experience (AX) for Software Products
Ax-check is a tool by Gauge that evaluates how effectively AI agents can onboard to a product by simulating end-to-end onboarding flows across documentation, marketing sites, and CLIs.
OpenAI Astra for Law
OpenAI has launched Astra for Law, a specialized foundation for legal professionals combining GPT-6 Astra with a massive U.S. legal search index and professional-grade privacy controls.
Martin Fowler on the Visceral Dislike of LLM Interaction
Software architect Martin Fowler discusses the paradox of finding LLMs useful yet fundamentally off-putting due to their 'grating' persona and the values of their creators.
Qwen3.8-Omni-Flash Release Notes
Alibaba's Qwen3.8-Omni-Flash is a native omnimodal model featuring a 1M-token context window, agentic audio-visual understanding, and significant cost reductions for audio-visual API calls.