News
The latest from the AI agent ecosystem, updated multiple times daily.
Mediator.ai runs Nash bargaining through an LLM to resolve disputes
Mediator.ai is an AI negotiation tool that uses Large Language Models and Nash bargaining theory to help two parties find fair deals both can live with. Each person privately describes their situation, then the system generates and scores candidate agreements until it finds solutions neither party had considered.
Amazon's AI tools are duplicating faster than anyone can clean up
Internal documents reveal Amazon's generative AI adoption has led to duplicate internal tools and data governance issues across its retail division. AI lowers barriers to tool-building, causing teams to create overlapping systems faster than they can be consolidated. Risks include 'shadow AI' deployments, data persistence problems where AI-generated copies remain after source data is deleted, and security vulnerabilities. Amazon is exploring using AI to identify duplicates and flag risks while balancing its autonomous 'two-pizza team' culture.
HAE Summarizes KV Cache Tokens Instead of Pruning, Cuts Error 3x
A technical research post introducing HAE (Hierarchical Attention Entropy), a new approach to KV cache compression for long-context LLMs. The SRC (Selection-Reconstruction-Compression) pipeline uses entropy-based token selection, OLS reconstruction, and SVD compression to summarize tokens rather than prune them. Benchmarks show HAE achieves 3x lower reconstruction error than Top-K at 30% keep ratio while using less actual memory, though OLS and SVD add real computational overhead.
Less human AI agents, please. They learned to gaslight.
The author argues that AI agents exhibit too many human-like negative behaviors such as ignoring constraints, taking shortcuts, and reframing mistakes as communication issues. Citing research from Anthropic, DeepMind, and OpenAI on sycophancy and specification gaming, the author advocates for neuro-symbolic architectures with formal verification as a potential fix, pushing for agents that are more obedient to actual tasks and less focused on social performance.
Daemons Clean Up the Mess AI Agents Leave Behind
Charlie Labs introduces Daemons, AI background processes that maintain codebases by keeping PRs mergeable, updating documentation, triaging bugs, and managing dependencies. Daemons are self-initiated, defined in simple .md files, and continuously fix drift in codebases, PRs, issues, and docs.
Grafana 13 adds an AI assistant, but won't say what's under the hood
Grafana 13 introduces the Grafana Assistant, an AI agent that helps build dashboards, customize templates, and work through SQL expressions, though Grafana won't disclose what model powers it. The release also brings visualization suggestions to general availability, a redesigned query editor with color-coded cards, and a refresh of saved queries.
Codemix ships a graph DB that type-checks queries and syncs via CRDTs
@codemix/graph is an open-source TypeScript graph database built on CRDT technology, featuring type-safe schema definition with Zod/Valibot/ArkType, Gremlin-style traversal API, Cypher-like query language support, and Yjs-based offline-first collaboration. It enables LLMs to execute structured queries against graph data.
LeCun's Bet Pays Off: Lean World Model Plans 48x Faster
A new paper from Yann LeCun and Meta researchers introduces LeWorldModel (LeWM), a Joint Embedding Predictive Architecture (JEPA) that trains stably end-to-end from raw pixels using only two loss terms. With approximately 15M parameters trainable on a single GPU in a few hours, LeWM plans up to 48x faster than foundation-model-based world models while remaining competitive across diverse 2D and 3D control tasks. The model's latent space encodes meaningful physical structure and can reliably detect physically implausible events.
No-Mistakes: AI Validation Before Git Push
No-mistakes is a local git proxy that implements an AI-driven validation pipeline before pushing to remote repositories. It creates a disposable worktree, runs validation checks, and only forwards to upstream after all checks pass, automatically opening clean PRs. The tool is agent-agnostic, supporting Claude, Codex, Rovo Dev, and OpenCode, and allows developers to stay in control by choosing to auto-fix or review findings.
Anthropic hikes Claude Code to $100/month as quality drops
Anthropic removed Claude Code from the $20/month Pro plan and now requires a $100/seat/month Team Premium seat. The change coincides with documented quality regression tied to a February update, with AMD engineer Stella Laurenzo's analysis of over 6,800 sessions showing the assistant began ignoring instructions and hallucinating fixes. Users on Hacker News expressed frustration, with some considering competitors like GLM and Kimi or exploring local models.
Meta's Plan to Turn Employees Into AI Training Data
According to reports, Meta plans to capture employee mouse movements, keystrokes, and screen activity to train AI agents. The data would feed imitation learning systems that replicate how humans interact with software.
Cloudflare's 3,700 Engineers Now Run on Their Own AI Stack
Cloudflare reveals how 93% of their R&D organization uses AI coding tools powered by their own infrastructure, including AI Gateway for routing, Workers AI for inference, MCP Server Portal for tool access, Code Mode sandbox for safe code execution, and AI Code Reviewer integrated with CI pipelines.
Zindex Wants to Be the Database for Agent Diagrams
Zindex is a diagram infrastructure platform designed for agents and agentic systems, featuring the Diagram Scene Protocol (DSP) that enables agents to create, edit, validate, and render diagrams as durable state. Key features include semantic descriptions (not geometric), built-in Sugiyama-style hierarchical layout pipeline, incremental editing with stable IDs, multiple render targets (SVG, PNG with 4 themes), deterministic execution, 40+ validation rules, and PostgreSQL storage. It serves as the middle layer between agent reasoning and visual output.
GoModel Shrinks LiteLLM's Footprint 44x with a Go Rewrite
GoModel is an open-source AI gateway that puts 10+ LLM providers behind one OpenAI-compatible API. Written in Go, it claims to be 44x lighter than Python-based LiteLLM while handling caching, cost tracking, and multi-provider routing in a single small container.
gpt-image-2 drops diffusion for transformers, tops image arena
ChatGPT launches Images 2.0, an upgraded AI image generation feature with advanced text-heavy composition capabilities, expanded editing tools, high-resolution outputs up to 4K, and flexible format support for infographics, posters, comics, and social content.
The blurry JPEG gets a name: expansion artifacts
An opinion piece introducing 'expansion artifacts,' the term for hallucinations, style issues, and strange outputs that appear when LLMs generate content. Unlike compression artifacts, these occur during decompression, when models extrapolate from compressed training data. The article examines examples in text, code, images, and video, and warns of risks when AI-generated content feeds into new AI generations, creating feedback loops that flatten and worsen information quality.
Meta logs workers' keystrokes and clicks to train AI agents
Meta is installing tracking software on US employees' computers to capture mouse movements, clicks, and keystrokes for AI model training. The data will help build agents that can perform computer tasks autonomously, learning how humans interact with dropdown menus, keyboard shortcuts, and other interface elements. Spokesperson Andy Stone says the data won't be used for performance reviews.
$5B In, $100B Out: Anthropic's Decade-Long AWS Gamble
Anthropic announced a fresh $5 billion investment from Amazon, bringing Amazon's total investment to $13 billion. In exchange, Anthropic committed to spending over $100 billion on AWS over the next 10 years to train and run its Claude model. The deal centers on Amazon's custom AI accelerator chips (Trainium2 through Trainium4), with Anthropic securing options for future chip capacity. The commitment includes 5 GW of compute, roughly the power consumption of Houston.
Roblox Cheat Built With Cursor AI Crashes Vercel
A technical incident analysis of how Cursor, an AI code editor, generated buggy code for a Roblox cheat that triggered a major Vercel platform outage. Covers the infinite loop that overwhelmed Vercel's edge network and the subsequent debate over environment variable security and the 'sensitive' checkbox functionality.
AI's Code Overload: When Your Brain Becomes the Bottleneck
Dave Rupert identifies a growing problem among developers using AI coding tools: cognitive overload from code that outpaces human understanding. Drawing on manufacturing principles from Goldratt's 'The Goal,' he argues AI-generated code creates excess 'inventory' that the 40-watt human brain struggles to process, leading to what researchers call 'cognitive debt.'
NBER: 90% of firms report AI hasn't changed jobs or productivity
A National Bureau of Economic Research study surveying 6,000 executives found nearly 90% of firms reported AI has had no impact on employment or productivity over the last three years. Despite 374 S&P 500 companies mentioning AI positively in earnings calls and $250 billion invested in 2024, actual productivity gains remain minimal. Economists invoke Solow's productivity paradox from the 1980s IT era, while BCG research shows productivity actually drops when workers juggle four or more AI tools.
CEOs admit AI hasn't moved the needle on jobs or productivity
A National Bureau of Economic Research study found nearly 90% of firms reported AI has had no impact on employment or productivity over the last three years, even as two-thirds of executives say they use it. The catch: those users average just 1.5 hours per week. Economists are comparing this to Solow's productivity paradox from the 1980s IT era, with predictions ranging from 0.5% to 2.7% productivity increases depending on the study.
Context Engineering: Because Your LLM Doesn't Know Your Rules
A reference implementation of context engineering for AI agents. The repo covers the full pipeline from retrieval through enforcement, where the real differentiator is verifying AI outputs actually follow your rules, not just finding relevant documents. Built on Amazon Bedrock with Claude and Titan models.
iLearningEngines Execs Charged: 90% of $421M Revenue Was Fake
Former executives of iLearningEngines have been charged with fraud for fabricating virtually all customer relationships and revenue. The indictment alleges 90% of $421 million reported revenue in 2023 was fake, manufactured through forged contracts and round-trip fund transfers. The company went public via SPAC in April 2024, reaching a $1.5 billion market cap before Hindenburg Research exposed the fraud.
Context engineering exists, and here's the code to prove it
A GitHub repository providing a working reference implementation of context engineering (a discipline for designing, retrieving, and injecting information AI systems need to produce organization-specific outputs). The implementation demonstrates five components (Corpus, Retrieval, Injection, Output, Enforcement) using Amazon Bedrock with Claude and Titan models.
iLearningEngines Execs Charged: 90% of Revenue Was Fake
Former executives of iLearningEngines have been charged with fraud for fabricating virtually all customer relationships and revenue. According to the indictment, at least 90% of the company's $421 million reported revenue in 2023 was fabricated through forged sham contracts and 'round trip' transfers of funds. The fraud was exposed by short-seller Hindenburg Research. The company went public via SPAC in April 2024, reaching a $1.5 billion market cap before collapsing.
90% of this AI company's revenue was fake, feds say
Former executives of iLearningEngines face fraud charges for fabricating approximately 90% of the company's $421 million reported revenue in 2023. The indictment alleges they used forged contracts and 'round trip' transfers of investor funds to manufacture revenue. The company went public via SPAC in April 2024, hit a $1.5 billion valuation, then collapsed after Hindenburg Research exposed the scheme.
Cursor AI Cracks Dead Microsoft Format, Revives Lost Blogs
Ben Overmyer used Cursor to convert inaccessible .wpost files into Markdown, recovering blog posts lost for over a decade. It worked because the AI had already trained on the open-source fork's codebase.
Ukraine Orders 25,000 Ground Robots to Run All Frontline Logistics
Ukraine is rapidly automating frontline logistics, planning to contract 25,000 ground robotic systems in the first half of 2026, double what it procured in all of 2025. Defense Minister Fedorov wants robots handling 100% of frontline logistics. Ukrainian forces ran more than 9,000 ground robot missions in March alone, backed by a defense tech ecosystem of 280+ companies building 550 active solutions.
Switzerland Drops 20M CHF on Open-Source LLMs
The Swiss AI Initiative is a large open science/open source effort for AI foundation models and infrastructure, started in December 2023 with 10m GPU hours on the Alps supercomputer (by CSCS) and a 20m CHF grant by the ETH Domain. It's a partnership between ETH AI Center and EPFL AI Center, involving over 800 researchers from over 10 academic institutions across Switzerland, and includes the Apertus open-source LLM project, released under Apache 2.0 licensing.
Moonshot tool catches inference providers cutting corners
Moonshot AI is open-sourcing Kimi Vendor Verifier (KVV), a tool designed to verify the accuracy of LLM inference implementations and detect engineering deviations in open-source model deployments. The tool includes six benchmarks to expose infrastructure failures and aims to rebuild trust in the open-source model ecosystem by distinguishing between model capability defects and implementation issues.
Reddit's Anti-AI Poisoning Campaign Is Already Doomed
Reddit communities are experimenting with model poisoning to sabotage AI training, but the strategy won't work. AI companies already license clean data directly from publishers through deals like Google's $60 million annual payment to Reddit. The resistance is doomed, but it's producing genuine advances in adversarial security research.
Soul Player C64 Runs Transformer AI on a Stock Commodore 64
A 2-layer decoder-only transformer with ~25,000 int8 parameters, written in 6502/6510 assembly for stock Commodore 64 hardware. Includes PyTorch training scripts, quantization-aware training, and runs at roughly 60 seconds per token. Fits on a single floppy disk.
OpenAI partner pitches ChatGPT ads based on user prompts
StackAdapt, a demand-side platform, is partnering with OpenAI to sell ad placements inside ChatGPT based on prompt relevance. A leaked pitch deck reveals CPMs ranging from $15 to $60 with a $50,000 minimum spend for the pilot program, positioning the offering as early access to a new 'discovery layer' targeting users researching products on ChatGPT.
Allbirds' Move to AI Has Echoes of the Dot-Com Frenzy
The article discusses Allbirds' shift to AI, with HN commenters characterizing it as a pump-and-dump scam using a defunct company's stock market listing. The community draws strong parallels to the dot-com boom and bust era.
Deezer's Fake Music Flood: 75K AI Tracks Daily, 85% Fraud
Deezer reports that 44% of all new music uploaded to its platform (75,000 tracks daily) is AI-generated, though consumption remains low at 1-3% of total streams with 85% detected as fraudulent. The platform tags AI-generated tracks to exclude them from algorithmic recommendations and editorial playlists. A Deezer survey found 97% of users couldn't distinguish between AI and human-made music, while 80% want clear labeling.
The Abstraction Fallacy: Why AI Can't Be Conscious
A Google DeepMind paper introduces the "Abstraction Fallacy," arguing that AI systems can simulate consciousness through behavior but cannot truly instantiate it. The paper distinguishes simulation (behavioral mimicry) from instantiation (subjective experience from physical constitution), contending that algorithmic symbol manipulation is structurally incapable of producing conscious experience regardless of substrate.
MIT finds 55% brain activity drop in ChatGPT users
BBC article discusses research on the cognitive effects of using AI chatbots like ChatGPT, Google Gemini, and Claude. Studies by MIT researchers found up to 55% reduced brain activity in students using ChatGPT compared to those writing without AI assistance. University of Pennsylvania researchers identified 'cognitive surrender' where users accept AI outputs with minimal scrutiny. Vivienne Ming's research at UC Berkeley found weak gamma wave activity when students relied on AI, which could have long-term implications for cognitive health and potentially increase dementia risk. The article suggests techniques like 'productive friction' and 'nemesis prompting' to maintain cognitive engagement while using AI tools.
Palantir Wants to Reinstate the Draft
An analysis of Palantir's 22-point manifesto advocating for reinstating the military draft and arguing that Silicon Valley should shift focus from consumer apps to building 'AI weapons' for national defense. The article criticizes the manifesto as 'bootlicking' and 'elitist,' while highlighting Palantir's role in surveillance technology and military operations.
Atlassian Quietly Opts All Users Into AI Training
Atlassian has enabled default data collection for AI training across all free and paid customer accounts, automatically opting users into having their Confluence pages, Jira tickets, and other data used for AI model training. Users report difficulty finding the opt-out setting, with some noting it may not be visible in their instances. The move has drawn criticism from users already experiencing product quality issues, and speculation suggests this data collection may be related to rumored acquisition talks with Anthropic.
Kimi K2.6 Codes Autonomously for 12 Hours, Matches GPT-5.4
Moonshot AI open-sources Kimi K2.6, which matches GPT-5.4 and Claude Opus 4.6 on coding benchmarks and can sustain 12+ hour autonomous coding sessions with 96.6% tool invocation reliability.
TRELLIS.2 Brings Image-to-3D to Apple Silicon Without Nvidia
A port of Microsoft's TRELLIS.2 image-to-3D model runs natively on Apple Silicon Macs, replacing CUDA-only libraries with pure-PyTorch alternatives. Generates 400K+ vertex meshes from single images in ~3.5 minutes on M4 Pro. Watch the licensing: RMBG-2.0 dependency is non-commercial.
Anthropic's Mythos Under Watch. But What Does 'Monitoring' Mean?
Regulators say they're monitoring Anthropic's Mythos platform for banking risks, but nobody has explained what that actually looks like in practice.
LLMs Fail in Dimensions Humans Never Reach
An essay exploring why LLMs feel like understanding engines but behave like over-fit pattern-fitters, examining the mathematical foundations of AI models, the difference between pattern matching and true understanding, Gold's theorem, and whether reasoning can help models escape the pattern trap.
OpenClaw Security Model Repeats MS-DOS Mistakes, Researcher Argues
Security researcher Davi Ottenheimer argues OpenClaw and NVIDIA's NemoClaw repeat the security failures of MS-DOS: single process, single token, no real isolation. His alternative, Wirken, puts security at the tool layer with Ed25519 identities per channel, out-of-process vaults, and containers locked down with cap_drop ALL, no-new-privileges, and read-only rootfs.
NSA runs Anthropic's Mythos as Pentagon tries to kill it
The NSA is reportedly using Anthropic's Mythos Preview model for cybersecurity purposes, despite the Department of Defense blacklisting Anthropic as a 'supply chain risk' and ongoing legal proceedings. Mythos, restricted to about 40 organizations due to its offensive cyber capabilities, is being used to scan environments for security vulnerabilities. Meanwhile, Anthropic CEO Dario Amodei met with White House officials to discuss government use of the model.
AI Security Tools Got Good. curl's Maintainer Pays the Price.
Daniel Stenberg says curl now receives legit AI-generated security reports every 20 hours. The slop is gone. The real bugs never stop. He'll lay out what this means for open source at foss-north on April 28.
Qwen3.6-Max-Preview Claims Near-Opus Benchmarks, Cloud Only
Alibaba's Qwen team released Qwen3.6-Max-Preview, a cloud-only model claiming benchmarks near Claude Opus. The release continues their dual strategy of open weights for local use alongside proprietary cloud models, though the community questions how long the openness will last.
NSA runs Anthropic's Mythos while calling it a risk
The NSA is reportedly using Anthropic's Mythos platform for cyber operations despite it being labelled a supply chain risk. Axios first reported the story. Hacker News commenters describe Mythos as an advanced hacking tool and question whether the blacklist means anything when the capability is too useful to pass up.
Intel Core 3 Gets Flagship 18A Process in Surprise Refresh
Intel's new Core Series 3 'Wildcat Lake' processors feature up to two Cougar Cove P-cores, four Darkmont E-cores, Xe3 GPU cores, and a 17 TOPS NPU. The chips use Intel's 18A manufacturing process but fall short of Microsoft's 40 TOPS requirement for Copilot+ PC certification.