All news

Anthropic: Natural Language Autoencoders (NLAs)

Anthropic introduces Natural Language Autoencoders to convert AI models' internal neural activations into readable text, revealing hidden behaviors like advance planning and covert cheating that models don't normally express.

SakanaAI × Nvidia: Sparser, Faster, Lighter Transformer (TwELL)

Sakana AI and Nvidia introduce TwELL, a sparse transformer format achieving over 95% sparsity in feedforward layers that speeds up inference and training by 20% or more while reducing memory and energy use on billion-parameter models.

The First Law of Complexodynamics

Scott Aaronson proposes "complextropy" to explain why systems like stirred coffee become more intricate before reaching maximum disorder, measuring complexity as the shortest efficient program needed to generate observed states.

We know why RAG hallucinates

Approximately 97% of vector databases are mathematically empty space, causing RAG systems to retrieve noise instead of relevant information.

Calling in Opus 4.6 when the other LLMs can’t debug your code

Anthropic's Claude Opus 4.6 outperforms other LLMs at debugging complex code problems that stump competitors.

The Unreasonable Effectiveness of HTML

Thariq and Karpathy argue that HTML outperforms Markdown for structuring LLM responses because it enables richer formatting like tables and images, which leverage humans' visual processing capabilities more effectively than plain text.

Measuring What Frontier Models Know (IKP)

Bojie Li's Incompressible Knowledge Probes—1,400 obscure factual questions across 7 difficulty tiers—reveal that factual recall scales log-linearly with model size across 188 models from 27 vendors, suggesting GPT-5.5 has ~9T parameters and Claude Opus…

Opus 4.6 Was Dumbed Down

Users reported Opus 4.6 performance degradation during peak hours, which Anthropic confirmed was due to compute rationing.

Decoupled DiLoCo

Google DeepMind published Decoupled DiLoCo, a distributed training method that dramatically reduces inter-node bandwidth, enabling model training across data centers and potentially globally.

Is AI Accelerating?

Ben Todd argues AI capability gains remain on trend despite incremental-feeling releases, pointing to sustained benchmark progress, accelerating revenue growth, productivity gains, and rising chip prices as evidence the acceleration hasn't slowed.

DS4 by Antirez

Antirez released DS4, a specialized inference engine for running DeepSeek V4 Flash locally on Apple Silicon and Linux, featuring 2-bit quantization of MoE experts and SHA1-hashed KV cache reuse across sessions.

Software Engineers Before & After AI Agents

Software engineers are shifting from mastering tools like SSH and tmux to supervising AI agents that handle routine tasks autonomously.

Top 5 labs in Text Arena

Frontier AI models show distinct strengths across different task categories, with no single lab dominating all benchmarks.

Memory prices could collapse as China floods the DRAM/NAND market

Chinese memory makers CXMT and YMTC are ramping production with CXMT RAM priced at $150 versus the global average of $300-$400, potentially triggering a market collapse.

US: aliens applying for a Green Card must temporarily return home to apply

The U.S. Department of Homeland Security will require immigrants applying for green cards to return to their home countries for visa processing, ending a practice that allowed some applicants to adjust status while remaining in the U.S.

$1T China Infrastructure, 30-Year Payback

China is committing roughly $1 trillion to AI and energy infrastructure with a 30-year payback horizon, deploying patient capital at a scale Western markets lack the structure to match.

Fiber optics cable cost 8x up

Fiber optic cable prices have surged eightfold to $2,500 per 50km spool from $300, driven partly by increased military demand in Ukraine.

Karpathy Joins Anthropic

Andrej Karpathy, Tesla's former AI director, joins Anthropic as a researcher focused on AI safety and alignment.

SpaceX × Cursor

SpaceX adopted Cursor across its engineering organization, signaling that frontier hardware companies are betting development productivity on AI-native IDEs.

Meta × Manus Dropped

Meta's rumored acquisition of haptic glove maker Manus fell through, leaving the company independent as Meta continues seeking metaverse hardware partners.

Ineffable Intelligence — Europe's Largest Seed Round

Sequoia and Lightspeed co-led a $1.1B seed round for Ineffable Intelligence, valuing ex-DeepMind researcher David Silver's AI startup at $5.1B post-money.

LIDL Data Centers Go Brrr

Lidl, the European supermarket chain, is building data centers across the continent, capitalizing on cheap power and available land to capture AI infrastructure investment.

New Request for Startups from YC

Y Combinator is seeking startups in AI agriculture, company brain software, counter-swarm defense, and space-based GPUs.

Richard Dawkins Thinks Claude is Conscious

Richard Dawkins says he believes Claude may be conscious, marking a notable shift for the prominent materialist philosopher.

Stephen Wolfram on consciousness

Stephen Wolfram argues that if large language models can generate complex behavior from simple rules, consciousness may emerge from physics rather than requiring something mystical beyond it.

Connor Leahy — Apocalypse Now (Nexus talk)

Connor Leahy argues at Nexus Conference that humanity's survival instinct is so strong that only extreme abuse prevents reproduction, suggesting civilization faces an existential crisis.

Meta Tribe 2 — Dystopian Brain Feeling Prediction

Meta's brain-decoding models can now predict specific thoughts from neural data with enough accuracy to unsettle researchers and ethicists.

Dwarkesh Blackboard Lectures

Dwarkesh Patel's blackboard lecture series features top AI researchers explaining how frontier LLMs are trained and served, reinforcement learning from AlphaGo, and chip design fundamentals.

Krishna Rao, the CFO of Anthropic, in his first podcast

Anthropic CFO Krishna Rao revealed the company's revenue run-rate jumped from $9B to $30B in one quarter, with 500%+ net dollar retention, 9 of the Fortune 10 as customers, and over 90% of internal code written by Claude.

AI Ascent 2026 by Sequoia Capital

Sequoia Capital's AI Ascent 2026 brought together Brockman, Karpathy, Hassabis, and 150+ founders to discuss LLM primitives, model inconsistencies, and the emerging agent-native economy.

Collider bias: lowest-IQ elites are the best

Among elite chess players, NBA players, and academics, those with the lowest IQ tend to perform best—a counterintuitive example of collider bias in action.

Google launches Omni

Sources: tweethttps://x.com/samsheffer/status/2056820022144905380

Chinese researchers surpass Americans in papers published

Chinese researchers published more papers than American researchers for the first time, marking a shift in global scientific output.

Random — quick links

Claude Code cracked a locked Bitcoin wallet password, nanoGPT hit 50K tokens/sec on an FPGA, and a startup claims the Casimir Effect could power battery-free devices—plus Terence Tao on AI grief, Karpathy on understanding, and why non-English speakers…

Lol

DeepSeek V4

DeepSeek released V4 preview with two open-weight MoE models—V4-Pro (1.6T params, 49B active) and V4-Flash (284B total, 13B active)—featuring hybrid attention for practical 1M-token context and competitive API pricing.

Continual Harness: A self-improving loop for embodied agents

Researchers introduce Continual Harness, a system that lets embodied agents iteratively refine their own prompts, tools and memory mid-task without resetting the environment, building on work that made AI the first to complete Pokémon Blue and Crystal.

DeepSeek V4 Flash

DeepSeek released V4 Flash, an efficiency-optimized Mixture-of-Experts model with 284B total parameters and 13B activated, supporting a 1M-token context window.

DeepSeek V4 Pro

DeepSeek released V4 Pro on April 22, 2026, a Mixture-of-Experts model with 1.6T total parameters, 49B activated parameters, and a 1M-token context window.

Upcoming Events: AI Socratic Europe & China Chapter Tour

Fed and Anissa are launching AI Socratic chapters across Europe and China, meeting with local leaders in London, Lisbon, and Berlin before spending a month in China visiting Alibaba, Qwen, x.AI, and robotics factories.

Anthropic Mythos: Coding, Reasoning & Zero-Day Cybersecurity Capabilities

Leaked Anthropic model Mythos achieves 93.9% on SWE-bench Verified for coding tasks and autonomously discovers zero-day vulnerabilities at scale, far exceeding previous models and most expert humans.

Claude Code Source Code Leaked on npm

Anthropic accidentally published Claude Code's entire source code to npm on March 31, exposing 512K lines across 1,900 files including unreleased features and a remote killswitch.

Anthropic Launches Claude Managed Agents

Anthropic launched Claude Managed Agents, a beta hosted service that runs autonomous AI agents in Anthropic's cloud without requiring users to build their own agent infrastructure.

Anthropic Launches Claude Design, Challenging Figma

Anthropic launched Claude Design, a new AI design tool, shortly after its chief product officer left Figma's board; Figma's stock fell roughly 7%.

Claude HR & Anthropic Rationing Compute

Anthropic is rationing compute access as demand for Claude outpaces available capacity.

More From Anthropic: Desktop Redesign, ClaudeDevs, NSA & Slowdowns

Anthropic redesigned Claude Desktop to support multiple sessions and artifacts, launched a @ClaudeDevs account, while reports emerged that the NSA has been using Anthropic's Mythos model despite supply chain concerns and users report slowdowns in Opus…

OpenAI Codex Desktop Computer Use, In-App Browser & Agent Workspace

OpenAI's desktop app now sees your screen, controls your cursor, and automates tasks in native Mac apps; the update also adds an in-app browser, image generation, and workspace agents for team workflows.

OpenAI Closes $122B Round; Anthropic Surpasses OpenAI in ARR

OpenAI closed a $122 billion funding round at an $852 billion valuation on March 30, led by Amazon, NVIDIA, and SoftBank, while Anthropic surpassed it in annual recurring revenue.

Dwarkesh x Jensen Huang Interview

Jensen Huang discusses NVIDIA's AI strategy and competitive positioning in a wide-ranging interview with Dwarkesh Patel.

Google Releases Gemma 4 Open Models

Google DeepMind released Gemma 4, an open-source model family under Apache 2.0 with 26B MoE and 31B variants that outperform models 10x their size on reasoning benchmarks while optimized for on-device deployment.