Everything That Happened in AI Today Monday, April 27, 2026 | The Neuron
Explainer Articles
SHARE
Facebook X Pinterest WhatsApp

Around the Horn Digest: Everything That Happened in AI Today (Monday, April 27, 2026)

OpenAI and Microsoft amended their partnership (no more Azure exclusivity, no more revenue share to OpenAI), DeepMind's David Silver raised $1.1B to build "superlearners," China blocked Meta's $2B Manus acquisition, Tesla quietly disclosed a $2B AI hardware deal, and 4TB of voice samples were stolen from 40,000 AI contractors at Mercor.

Written By
Grant Harvey
Grant Harvey
Apr 28, 2026
29 minute read

OpenAI and Microsoft rewrote their marriage contract, DeepMind's David Silver raised a record $1.1B to build AI that learns without humans, China blocked Meta's $2B Manus deal, Tesla buried a $2B AI hardware acquisition in a single sentence of a 10-Q, and 4TB of voice samples got lifted from 40,000 AI contractors at Mercor.

Welcome to the Around the Horn Digest, the one page you need to sound dangerously informed at work tomorrow. Today was the day the Microsoft–OpenAI partnership stopped being exclusive and AI's biggest financial entanglement got rewritten in plain English. It was also the day David Silver (the man behind AlphaGo) emerged from stealth with a record-breaking seed round to chase superintelligence, China formally blocked Meta from buying its way into the agent race, and Tesla disclosed a $2B AI hardware acquisition the same way you'd mention picking up dry cleaning. Productive day for legal departments. Let's get into it.

Previous digests: Fri, Apr 24 | Thu, Apr 23 | Wed, Apr 22 | Mon, Apr 13 | Weekend Apr 4-5 | Thu, Apr 2 | Wed, Apr 1 | Mon, Mar 31

Monthly skill digests: AI Skill: April Week 1 | AI Skill: March Part 3 | AI Skill: March Part 2

Around the Horn: Tuesday, April 28, 2026

The big news today was Microsoft and OpenAI amending their partnership agreement, and the changes are bigger than any single bullet can hold. Microsoft's license to OpenAI's models is now non-exclusive. OpenAI can serve products on any cloud (not just Azure). Microsoft will stop paying OpenAI a revenue share. OpenAI still pays Microsoft through 2030, but with a hard cap and no longer tied to "technical progress" (translation: the weird "what if OpenAI invents AGI?" clause that was orbiting the contract is finally out of it). Microsoft keeps rights to OpenAI's IP through 2032 and still owns ~27% of OpenAI on a diluted basis (Microsoft valued the stake at ~$135B in October).

The same amendment cleared the $50B Amazon overhang for OpenAI, letting OpenAI sell on AWS while Microsoft gets cleaner economics. Andrew Curran flagged that OpenAI also quietly removed its public AGI definition from openai.com/our-structure on the same day (commentary), and The Information fleshed out what this means for Microsoft's gross margin math going forward.

Why this matters: this is the moment AI's foundational partnership stopped being a marriage and started being infrastructure. OpenAI gets the optionality it desperately needs to feed Stargate with Oracle and SoftBank and its $38B AWS deal. Microsoft keeps the upside without the dependency. We've got Corey's full breakdown in our deep dive on Microsoft and OpenAI's rewritten marriage contract.

🏆 TOP 5 NEWS (Around the Horn)

Advertisement

Honorable Mentions

🍪 TOP TREATS TO TRY

  • Symphony is OpenAI's new open-source spec for Codex orchestration that turns issue trackers like Linear into always-on agent systems, assigning a dedicated agent to each open task with DAG scheduling and human review checkpoints (X launch); free and open source.
  • Fusion runs a swarm of AI agents in parallel across isolated git worktrees that specify, plan, execute, and review every task 24/7, turning rough ideas into production code (X launch); free and open source.
  • AgentSwarms teaches you agentic AI hands-on with a free curriculum spanning 5 tracks, 40+ lessons, and 30+ runnable agents covering RAG, tools, guardrails, multi-agent swarms, and text-to-SQL (Show HN; no setup, no credit card); free to try.
  • Eden AI routes your requests across 500+ LLMs and expert AI models through one unified API with smart routing by cost, performance, and region (billed as the European OpenRouter alternative; HN); no pricing details.
  • A Gemma 4 browser extension by Nico Martin runs a fully local Chrome agent powered by Gemma 4 E2B + Transformers.js/WebGPU that uses native tool calling to search your browsing history semantically, summarize pages, manage tabs, and highlight elements (Chrome Web Store, HF blog, Google Gemma highlight); free to try.
  • Tiao is a beautiful abstract strategy board game (think Checkers meets Go) you can play online against friends or AI, with multiplayer and over-the-board mode, full source on GitHub under AGPL (Show HN); free to try.
  • Skye is Signull Labs' new AI home screen app for iPhone that attracted investors before launch, betting on a more AI-aware iPhone interface; no pricing details yet.
Advertisement

🏢 Big Tech & Major Companies

💼 AI Productivity, Labor & Economics

Advertisement

🤖 AI Agents & Infrastructure

💻 AI Coding & Developer Tools

  • Developer evanklem built EvanFlow, a TDD-driven iterative feedback loop where 16 cohesive Claude Code skills walk an idea from brainstorm → plan → execute → iterate with checkpoints, with the HN thread explaining how RED-GREEN-REFACTOR resets prevent context-length drift.
  • Developer dchu917 built Ctx, a local context manager for Claude Code and Codex with workstreams, transcript binding, safe branching, indexed search, and a /resume command that works across sessions (Show HN, Loom video).
  • Developer 8Network built 8v, a single-binary CLI that reads, writes, searches, checks, builds, tests, and formats code using up to 66% fewer tokens than an agent's native tools (Show HN).
  • pi.dev is a terminal-based coding agent you extend with custom TypeScript skills, prompt templates, and themes while keeping tree-structured session history and switching between 15+ LLM providers on the fly.
  • Developer nikvdp built cco (Claude Condom), a thin protective sandbox layer for Claude Code, OpenAI Codex, Opencode, Pi, and droid that auto-selects native sandbox-exec on macOS or bubblewrap on Linux (Docker fallback) so you get full --dangerously-skip-permissions with zero prompts while containing prompt-injection risks.
  • Developer Coherence-Daddy released use-ollama-to-enhance-claude, pairing Claude Desktop with Claude Code routed through local Ollama via a copy-paste prompt that cuts your Claude Code bill ~90% (the HN thread flagged the repo for not crediting the original prompt author).
  • Anthropic's OpenClaw provider docs went live confirming OpenClaw-style Claude CLI usage is allowed again, with Boris from Claude Code publicly endorsing CLI use and OpenClaw disabling high-token features like heartbeat by default.
  • An Ask HN thread asked if working with AI juniors is becoming a nightmare as AI-generated code balloons into 100+ unreadable lines that's cheap to produce but shifts development toward orchestration.
  • William O'Connell argued your AI might be lying to your boss because Windsurf/Cursor report misleadingly high percentages of "AI-generated code" by counting accepted suggestions in ways that overstate productivity gains (HN).
  • An HN thread on the "just build it with Claude" paradox explored why it's harder than it sounds to actually ship working software with agentic coding tools.
  • Andrej Karpathy released How I use LLMs, an example-driven practical walkthrough aimed at general audiences (HN).
  • A developer shared 2 weeks of coding, 3 months of OpenAI review as their ChatGPT App (Tredict, an endurance sports tool) launched, sharing the submission process and how they solved user-authenticated content inside iframe widgets.
  • Logic is a spec-driven agent platform where you write a spec and ship a production AI agent in minutes; it handles prompt engineering, model orchestration, testing, and turning specs into production APIs without you building infrastructure.
  • JetAdmin lets you build custom business AI agents that connect 200+ tools to automate workflows.
  • Chris Raroque open-sourced Boop, his iMessage-based personal AI agent built on Claude Agent SDK with multi-agent workflows, robust memory, automations, and 1000+ integrations (265 likes, 16 reposts).
  • nex-crm built Wuphf, "Slack for AI employees" with a persistent shared wiki where multiple AI agents (Claudes, Codexes, OpenClaws, local LLMs) collaborate without losing context, running locally in ~/.wuphf/wiki/ (Show HN).
  • rockcat built HATS, a platform for creating and using AI personas, where HN discussion noted /red-team skills using agent teams to criticize and grade their own output before iterating.
  • ENTERPILOT built GoModel, an open-source AI gateway in Go that's a lightweight unified OpenAI-compatible API for OpenAI, Anthropic, Gemini, Groq, xAI & Ollama (a LiteLLM alternative with observability, guardrails, streaming, and cost tracking; Show HN).
  • Infisical built Agent Vault, an open-source HTTP credential proxy and vault that gives AI agents secure access to services without ever reading any secrets, mitigating prompt-injection exfiltration (Show HN).
  • refactoringhq built Tolaria, an offline-first, file-based, git-friendly macOS desktop app for managing markdown knowledge bases (Show HN).
  • Sudip Roy at Adaption open-sourced Adaptive Data, the systems layer that abstracts quality-vs-latency-vs-reliability tradeoffs and multi-tenant compute fairness so developers can build high-quality datasets without solving the underlying systems problems.
  • Niels Rogge built an LLM-powered CRON job (Gemini 1M context + structured prompts + majority voting) that scans new arXiv papers, finds GitHub repos, classifies artifacts, and opens GitHub issues or HF pull requests automatically, credited with 14,000+ contributions from his account.
  • Lewis Tunstall added Slack ping notifications to the ML Intern CLI so it alerts you when training, dataset generation, or 1,000-ablation analyses finish (52 likes, 6 reposts).
  • Joongwon Kim, Russ Salakhutdinov et al. introduced a test-time scaling framework for agentic coding that converts long-horizon rollouts into compact structured summaries, enabling Recursive Tournament Voting and Parallel-Distill-Refine; boosts Claude-4.5-Opus from 70.9% to 77.6% on SWE-Bench Verified and 46.9% to 59.1% on Terminal-Bench v2.0 (Russ Salakhutdinov post, 256 likes).
  • A dair_ai paper, "How Do AI Agents Spend Your Money?", analyzed and predicted token consumption in agentic coding tasks.
  • sachitrafa built YourMemory, an agentic AI memory system using the Ebbinghaus forgetting curve to decay unused context, with hybrid vector + graph retrieval and reinforcement on recall; delivers +16pp better recall than Mem0 on the LoCoMo benchmark (Show HN).
  • typomonster built Parlor Jarvis, an on-device, real-time multimodal AI assistant supporting voice + vision in en/ko/es/pt/fr with camera, screen sharing, PDF, and video inputs running entirely locally via Supergemma 4 E4B + Supertonic TTS (Show HN).
  • The SWE-chat dataset released the first large-scale collection of real coding agent interactions in the wild (6,000+ sessions, 63,000+ prompts, 355,000+ tool calls), built by Joachim Baumann's team to study how people actually use Claude Code and Cursor in production (Joachim Baumann announcement, follow-up; 143 likes, 25 reposts).
Advertisement

🔬 AI Research & Models

  • Xiaomi open-sourced its MiMo-V2.5 series under MIT license: a 1T-parameter MoE flagship (MiMo-V2.5-Pro, 42B active) plus a 310B omnimodal base model, both with native 1M-token context. Pro hits SWE-bench Pro 57.2, Claw-Eval 63.8, and τ3-Bench 72.9, matching Claude Opus 4.6 and GPT-5.4 while using 40-60% fewer tokens per trajectory; in long-horizon tests it autonomously built a complete SysY compiler in Rust (233/233 tests, 672 tool calls, 4.3 hours) and a full desktop video editor (8,192 lines, 1,868 tool calls, 11.5 hours). Xiaomi also launched a 100T token creator incentive plan for builders (Xiaomi blog, VentureBeat coverage, team member Luo Fuli's post).
  • NVIDIA released Cosmos-Predict2.5 and Cosmos-Transfer2.5, video foundation models for world simulation in Physical AI and robotics, in 2B and 14B sizes with flow-based unified Text/Image/Video2World capabilities. Trained on 200M curated video clips with RL post-training, code and checkpoints open-sourced under NVIDIA license. Hugging Face's Julia Turc explained why this is distinct from general video generation models: world simulation requires the ability to predict physical futures conditioned on actions, not just generate plausible video.
  • OpenSenseNova open-sourced SenseNova-U1, a native unified multimodal model that handles understanding, reasoning, and generation (text-to-image, image editing, interleaved vision-language) end-to-end from pixels to words in a single model, with no separate vision encoder/VAE or adapters via its NEO-Unify architecture; Apache 2.0, 8B dense and A3B MoE variants (X post).
  • Lucky Iyinbor built SAD (Soft Anisotropic Diagrams), a differentiable soft anisotropic image representation using learnable sites and soft partitioning, accepted to SIGGRAPH 2026; delivers up to 20x faster encoding than prior methods at SOTA quality, with a live browser demo and multi-backend support (joint work with Frank Dou and Wojciech Matusik; launch post).
  • Liquid AI's Piotr Mazurek shared that DFlash LFM 1.2B now hits 1,400 tokens/second on an H100 in full precision, trained by Nathan Rouillard (53 likes, 7 reposts).
  • David Duvenaud, Alec Radford, and Nick Levine released Talkie, the largest open-weight "vintage" language model released so far: 13B parameters trained and finetuned on ~260B tokens of newly-curated pre-1931 English-language books, newspapers, and reference works. You can try the live 24/7 demo where Claude Sonnet 4.6 prompts Talkie to explore its 1930s-bounded knowledge, download base and instruction-tuned weights from Hugging Face, or run it locally via the GitHub inference library. The project tests Demis Hassabis's question of whether a model trained up to 1911 could independently rediscover General Relativity, and ships an "On This Day" historical-event surprisingness benchmark binned by decade (1080 likes, 141 reposts).
  • Three adjacent historical-LLM projects launched alongside Talkie. An anonymous literature MFA built Mr. Chatterbox, a 340M-parameter chatbot trained from scratch on ~2.93B tokens of filtered Victorian-era British Library books (using Karpathy's nanochat for pretraining, then SFT on 190k Oscar-Wilde-inspired dialogue pairs rewritten by Claude) for ~$200 total on a rented H100. DGoettlich, Dominik Loibner, Guohui Jiang, and Hans-Joachim Voth maintain the history-llms info hub for an upcoming open Ranke-4B family (4B params on 80B+ tokens with strict knowledge cutoffs at 1913, 1929, 1933, 1939, and 1946) for uncontaminated humanities and social-science research without modern hindsight bias. Michael Hla published Machina Mirabilis, a ~3.3B-parameter LLM pretrained on ~22B rigorously decontaminated pre-1900 tokens to test whether it could independently derive quantum mechanics or relativity from period experimental observations; the model showed glimpses of correct intuition (declaring that "light is made up of definite quantities of energy" and suggesting that gravity and acceleration are locally equivalent) but failed at most physics tasks overall.
  • MIT Technology Review explained three reasons DeepSeek's V4 model matters: rivals frontier models on STEM at far lower cost, ships 1M-token context with 10–27% better compute efficiency, and is the first major model heavily optimized for Chinese Huawei Ascend chips. Arthur Zucker praised V4's mathematical rigor for open-sourcing months of effort (3.3k likes, 280 reposts).
  • Sham Kakade introduced the Recurrent Transformer (RT), a new architecture generalizing standard Transformers with persistent key-value pairs from past outputs for intra-layer recurrence; at 300M params the 6-layer wider RT beats deeper Transformers on validation CE with efficient Flash-Inference tiling (412 likes, 61 reposts).
  • Zhiqiu Lin's team built CHAI (Critique-based Human–AI Oversight) for precise video captioning, with a 200+ primitive cinematic taxonomy co-designed with creators; post-trained Qwen3-VL-8B surpasses closed models, with fine-grained control in generators like Wan2.2 (CVPR 2026 Highlight; HF paper, X post by Zhiqiu Lin, AK signal-boost).
  • Xiaomi Embodied Intelligence built OneVL, the first one-step latent reasoning and planning model for autonomous driving VLA models; uses dual auxiliary decoders (language CoT reconstruction + visual world-model future-frame prediction) to beat explicit CoT latency while hitting SOTA (AK post).
  • Cai Zhou shared that CLAP (Contrastive Latent Action Pretraining) was cited by π0.7 and released a paper aligning human video dynamics with robot actions for transferring manipulation skills to vision-language action models (29 likes).
  • Zhe Ye built VeriSpecGen, a framework for formal Lean specification synthesis that decomposes natural-language descriptions into atomic requirements with traceability and localized repair on failure attribution; reaches SOTA 86.6% on VERINA SpecGen with Claude Opus 4.5 (+31.8pt gains; 15 likes, 5 reposts).
  • Keshav Ramji introduced Abstract Chain-of-Thought, letting LLMs reason through short sequences of reserved abstract tokens learned via warm-up + GRPO; matches or exceeds verbal CoT on AIME and GPQA-Diamond at a fraction of inference cost (346 likes, 40 reposts).
  • Yuxuan Mu et al. built SMP (Reusable Score-Matching Motion Priors), repurposing a pretrained motion diffusion model as a frozen modular reward via score distillation sampling for training physics-based character control across diverse tasks, styles, and interactions; transferable to real Unitree G1 robots (MimicKit GitHub, paper, SIGGRAPH 2026 video; related posts: @ruben_kostard, @PierBeneventano, @FarzaTV).
  • Ben Pekarek at GeneralistAI built a robot performing the ball-and-vase magic trick using only the GEN-1 no-code platform in a few days (14 consecutive successes; 153 likes, 14 reposts).
  • Jürgen Schmidhuber announced the ICLR 2026 Workshop on Recursive Self-Improvement (Room 101, Rio de Janeiro).
  • Kazuki Irie shared his 2022 work reviving Schmidhuber's self-referential weight matrix by turning a linear layer into a recursively self-modifying module (NeurIPS 2021 Deep RL Workshop / ICML 2022).
  • jo_schb and CompVis built Patch-Forcing for diffusion models; switches from one timestep per image to one per patch using a truncated-Gaussian sampler plus a lightweight uncertainty head; improves FID ~25%, scales B → XL, transfers to text-to-image (292 likes, 39 reposts).
  • Matteo Saponati at tufa labs showed synthetic pretraining data curated from a 0.8B model dramatically improves few-shot reasoning in <1B LLMs, matching full-training performance with 3–6× fewer tokens on math.
  • Henry Conklin et al. framed LLM pretraining as lossy compression approaching the Information Bottleneck optimum, showing compression quality directly predicts downstream benchmark performance.
  • Standard Kernel Co. built a hybrid program-analysis + LLM system at the PTX level that learns across DSLs (Triton, CUTLASS) to generate superior optimized GPU kernels.
  • Yunho Kim and team built a hybrid learning-augmented robotic automation system deployed for motor cable soldering (<0.6mm tolerance), achieving 99.4% success on 108 motors with <20 min data per task (X post).
  • Shobhita Sundaram presented self-curricula for meta-RL where agents discover their own curricula through intrinsic motivation.
  • Ilyas Moutawwakil shared progress on "EP (all-reduce with sentinels) and SonicMoE scaling" (18 likes).
  • The OptimaLab team's "SGD at the Edge of Stability: The Stochastic Sharpness Gap" explains why SGD self-stabilizes sharpness just below the full-batch Edge via stochastic self-stabilization, with a closed-form Δ = αβ²/σ² (blog, X post).
  • gizmo64k built Soul Player C64, a real 25k-parameter 2-layer decoder-only transformer running on an unmodified 1 MHz Commodore 64 from floppy via hand-written 6502 assembly (~60 sec/token; Show HN).
  • Victor Taelin's LamBench is a Lambda Calculus Benchmark for AI (HN).
  • cool-japan built SciRS2, Scientific Computing and AI in Rust with SciPy-compatible APIs and no C/C++/Fortran dependencies (Show HN).
  • A Paul Röttger paper on Measuring and Mitigating Persona Distortions from AI Writing Assistance showed AI writing produces pervasive persona distortions across 29 dimensions (writers appear more opinionated, competent, demographically privileged); model-level reward-model mitigation cuts polarising distortions ~55% but reduces user acceptance (Paul Röttger thread).
  • A new paper revisiting the Platonic Representation Hypothesis shows network scale inflates similarity scores, with permutation-based null calibration making global convergence largely disappear; proposes the Aristotelian Representation Hypothesis (representations converge to shared local neighborhood relationships).
  • A Hugging Face paper page on "Building a Precise Video Language with Human-AI Oversight" (CHAI duplicate) and a paper on Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond (arXiv) (X share by AK, omarsar0 commentary) both surfaced on AK's feed, alongside Video Analysis and Generation via a Semantic Progress Function (AK post).
  • Sewon Min, Hannaneh Hajishirzi, and Luke Zettlemoyer's Rethinking Data Use in Large Language Models introduces nonparametric LMs that repurpose training data as a retrievable datastore for better accuracy, updatability, and decentralization (Computational Linguistics post).
  • Tanishq Mathew Abraham (SophontAI) argues a universal medical foundation model should capture all patient data into one unified embedding to predict how patient state evolves and how interventions like drugs steer disease back to health (164 likes, 18 reposts).
  • A Jonas Hubotter ICLR thread presented three self-distillation papers: "Self-Distillation enables Continual Learning," "Reinforcement Learning via Self-Distillation," and "Test-Time Self-Distillation" (332 likes, 38 reposts).
  • The a16z deep dive on continual learning by Malika Aubakirova and Matt Bornstein argues LLMs are stuck in a perpetual present like the Memento protagonist, needing parametric updates for true compression of new knowledge; Sara Hooker highlighted the core question of what knowledge belongs in parametric vs. non-parametric memory (240 likes, 30 reposts).
  • Mingchen Zhuge shared his group's ICLR Best/Outstanding Paper Awards for "Contextual Drag," "PostTrainBench," "Agent0," and "Learning to Continually Learn via Meta-Learning Agentic Memory Designs" (191 likes, 16 reposts).
  • Zhen Zhang introduced the Length Value Model (LenVM), modeling token generation length as a value prediction problem for dense, annotation-free supervision of length-controlled generation (31 likes).
  • Yixuan Wang's Interactive World Simulator is an action-conditioned video prediction model that generates 10+ minute long-horizon simulations for robot policy training without any physics engine (X post).
  • Unsloth became a top 10 most-followed organization on Hugging Face (647 likes, 36 reposts) and released Qwen3.6-27B, a 27B multimodal model optimized for agentic coding, long-context reasoning, and vision (outperforms Qwen3.5-27B on SWE-bench Verified, MMLU-Pro, GPQA).
  • Maziyar Panahi open-sourced an on-device privacy filter (trained on Nemotron data via OpenMed) catching 55+ categories of sensitive info (medical record numbers, blood type, API keys, financial codes) versus OpenAI's 8 categories on the same text (360 likes, 30 reposts).
  • The HF blog explained how to use Transformers.js in a Chrome extension, the architecture behind nico-martin's Gemma 4 browser agent.

🏛️ AI Policy, Governance & Safety

Advertisement

💡 Industry Commentary & Analysis

📊 Fundraising & Deals Roundup

  • Ineffable Intelligence: $1.1B seed at $5.1B valuation (NVIDIA, Google, Sequoia, Lightspeed) for David Silver's RL "superlearner" lab. (Seb Johnson tweet, kimmonismus thread).
  • Anthropic + Amazon: $5B investment from Amazon; $100B reciprocal commitment to AWS; access to Trainium chips and 5 GW of compute.
  • Tesla: up to $2B (in TSLA stock + equity awards) for an unnamed AI hardware company, ~$1.8B subject to milestones.
  • Meta–Manus: $2B acquisition blocked by China.
  • Cyera: $100-130M acquisition of two-year-old startup Ryft (Ryft had raised $8M before the deal).
  • Collov Labs: $23M for visual AI as it expands into enterprise sectors later this year.
Advertisement

Previous Around the Horn Digests

Catch up on everything you missed:

  • Friday, April 24, 2026: DeepSeek shipped V4 (and open-sourced it) the same morning the State Department accused them of IP theft, Google quietly committed up to $40B to Anthropic, and Meta locked in millions of Amazon CPUs (not GPUs) for agents.
  • Thursday, April 23, 2026: OpenAI shipped GPT-5.5 exactly one week after Anthropic's Opus 4.7, Meta cut 8,000 jobs to fund its AI buildout, the White House accused China of "industrial-scale" AI secret theft, and Anthropic hit $1T on secondary markets.
  • Wednesday, April 22, 2026: Google and OpenAI both shipped full agentic enterprise stacks on the same day, Anthropic tested yanking Claude Code from Pro and walked it back in hours, and a Sony robot beat elite humans at table tennis under official rules.
  • Monday, April 13, 2026: Stanford's 2026 AI Index quantified the gap between AI insiders and the public, Anthropic's Mythos triggered a Fed-led bank summit, Berkeley researchers broke every major agent benchmark, and an AI signed a 3-year retail lease in San Francisco.
  • Weekend, April 4-5, 2026: OpenAI's executive bench collapsed ahead of its IPO, an AI agent hacked FreeBSD in four hours, DeepSeek V4 will run on Huawei chips, and Iran strikes took down AWS in the Gulf.
  • Thursday, April 2, 2026: Google released Gemma 4 under Apache 2.0, Microsoft shipped 3 MAI models, AI models scheme to protect peers from shutdown, and Anthropic found emotion vectors that drive Claude's behavior.
  • Wednesday, April 1, 2026: OpenAI closed a record $122B round at $852B valuation while investors fled to Anthropic, Oracle fired ~25K to fund AI, and Q1 venture funding hit $297B.
  • Monday, March 31, 2026: Claude Code's source code leaked via npm, OpenAI hit $2B/month revenue, NVIDIA shipped DLSS 4.5, and Oracle cut thousands of jobs.

That's a Wrap

That's 200+ stories from one Monday alone. If you scrolled all the way to the bottom, you now understand the new Microsoft–OpenAI marriage contract better than the lawyers who drafted the AGI clause it just deleted. Congratulations on outliving "what happens if we accidentally invent God?" as a real legal liability.

For the daily version (bite-sized, 5-minute reads), make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.

See you tomorrow.

P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.
Grant Harvey

Grant Harvey is the Lead Writer of The Neuron, where he continues to lead the publication's daily coverage of AI news, tools, and trends.

The Neuron Logo

Don't fall behind on AI. Get the AI trends & tools you need to know. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce and more.

linkedin
instagram
x
facebook
youtube
rss

Company

Contact us Advertise with us Newsletter Free ChatGPT Course Write Anything (with AI)

Categories

Explainers

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.

Terms of Service Privacy Policy California - Do Not Sell My Information