Cool Things in Tech, April 30-May 6, 2026

Overview report on the latest in AI and technology in the week of April 30-May 6, 2026, with a personal bias on generative media and society

This past week saw a flurry of significant developments across the AI landscape, from major model releases and unprecedented enterprise AI spending to escalating regulatory scrutiny and pivotal legal battles. OpenAI, Anthropic, and DeepSeek all unveiled powerful new models, while the industry grappled with the implications of AI’s insatiable demand for compute and growing concerns about safety and national security.

The Headline News

• OpenAI Launches GPT-5.5 Instant & Replaces Custom GPTs with Workspace Agents: OpenAI made a major strategic shift, releasing GPT-5.5 Instant as the new default model for ChatGPT users, boasting a 52.5% reduction in hallucinations on high-stakes prompts and improved personalization. Simultaneously, the company retired its “Custom GPTs” and introduced Workspace Agents, a more powerful, Codex-powered AI co-worker platform that integrates with existing team tools like Slack, Google Drive, and Salesforce. This move signals OpenAI’s strong push towards enterprise automation and agentic AI. The company also extended the restricted access for their “Cyber” version GPT-5.5-Cyber, previously put in place for the previous generation model.

• AI Infrastructure Spending Skyrockets, Driven by a New “Compute Arms Race”: From public filings this week, Microsoft, Google, Meta, and Amazon collectively spent an astonishing $130 billion on AI infrastructure in Q1 2026, nearly doubling the previous year’s expenditures. Google reported running out of cloud capacity due to demand and raised its 2026 capital expenditure guidance to $190 billion. This unprecedented investment highlights an intensifying “compute arms race” and raises concerns about the environmental impact, with companies developing natural gas power plants to meet surging AI demand, potentially jeopardizing climate pledges.

• Elon Musk’s Lawsuit Against OpenAI Unveils Internal Conflicts and Accusations: The ongoing legal battle between Elon Musk and OpenAI reached a critical juncture with Musk’s testimony, where he claimed he was “a fool” for funding the nonprofit and accused its leaders of “looting the nonprofit” by deviating from its original mission. Former OpenAI CTO Mira Murati also testified, alleging that Sam Altman lied to her about safety standards for a new AI model. The trial exposed financial motivations and Musk’s attempt to turn OpenAI into a for-profit entity, while he simultaneously admitted xAI trained Grok on OpenAI models.

Research & Technical Deep Dives

• Subquadratic’s SubQ Model launched with a 12-million-token context window, outperforming GPT-5.5 on retrieval benchmarks. It uses a sub-quadratic architecture that scales linearly with input length and is 52x faster than FlashAttention at 1M tokens, aiming to eliminate “memory hacks.” A 50-million-token version is planned.

• DeepSeek V4 Pro (1.6 trillion parameters) and V4 Flash models were released, with V4 Pro noted as the largest open-weights model offering competitive performance at significantly lower cost than US frontier models ($1.74 vs $5.00 per 1M input tokens for GPT-5.5).

• Mistral AI released Mistral Medium 3.5, a dense 128 billion parameter model with a 256k context window and multimodal input, available under a Modified MIT License.

• The Qwen Team released Qwen-Scope, a Sparse Autoencoder (SAE) toolkit for Qwen 3.5 models (2B to 35B MoE), enabling detailed understanding and manipulation of internal model features.

• Google Gemma 4 now uses Multi-Token Prediction (MTP) drafters for up to a 3x inference speedup without quality degradation.

• SenseTime introduced SenseNova-U1, a multimodal model integrating text rendering directly into images, supporting 2048x2048 resolution with 8 billion parameters.

• IBM launched Granite 4.1, new open-weight Apache 2.0 models (30B, 8B, 3B) focused on token efficiency for enterprise and edge deployments.

• A new training method, Abstract Chain-of-Thought (Abstract CoT), reduces LLM reasoning tokens by 11.6x without quality loss by training models to use a private shorthand language.

• Talkie, a 13B LM trained exclusively on pre-1931 texts, showed promising generalization, including basic Python coding, despite its historical training.

Business, Policy, & Strategy

• Anthropic is reportedly seeking a private round of $40 billion to $50 billion at an $850 billion to $900 billion valuation, with its annual revenue run rate jumping from $9 billion to over $30 billion.

• Both Anthropic and OpenAI announced separate enterprise AI ventures, valued at $1.5 billion and $10 billion respectively, backed by major financial firms, to integrate AI tools into businesses.

• The AI boom has driven the memory-chip industry into a “super boom cycle,” with Samsung reporting record Q1 profits.

• Palantir reported Q1 2026 revenue of $1.63 billion, up 85% year-over-year, its fastest growth since going public, with U.S. commercial clients up 133%.

• Panthalassa raised $140 million led by Peter Thiel for autonomous floating data centers powered by ocean waves.

• After attributing planned layoffs to increased AI spending, Meta acquired humanoid robot startup Assured Robot Intelligence.

• Manifest OS raised $60 million Series A at a $750 million valuation to build AI-native law firms, starting with business immigration.

• The U.S. government is initiating pre-deployment testing of frontier AI models from Google, Microsoft, and xAI, signaling a “quasi-licensing” regime with potential to block public releases due to national security. The White House also opposes Anthropic’s plan to expand access to its Mythos model beyond select partners.

• The Pentagon signed agreements with SpaceX, OpenAI, Google, Nvidia, Reflection AI, Microsoft, and AWS to deploy their AI models in classified networks.

• A report from MIT and the Center for Democracy & Technology found that “fine-tuning” AI models can have unpredictable impacts on safety guardrails, raising liability questions for businesses.

• China blocked Meta’s ~$2 billion acquisition of agentic AI company Manus AI, citing national security concerns, indicating tighter control over its domestic AI industry.

• A Chinese court ruled against AI-based layoffs, stating that replacing employees with AI and cutting wages for automation reasons constitutes illegal dismissal, emphasizing worker protections amid rapid technological changes.

• Academy Awards ruled that acting roles and screenplays must be “demonstrably performed by humans with their consent” to be eligible for Oscars, excluding synthetic actors or AI-authored scripts.

• States are rapidly regulating AI in healthcare, with 43 U.S. states introducing laws and 12 already enacted, addressing patient harm, discrimination, and transparency.

• Google plans to sell its custom Tensor Processing Units (TPUs) to select customers, directly competing with Nvidia.

New Tools & Practical Applications

• OpenAI’s Codex was significantly updated, positioning it as a “SuperApp” for general knowledge work, with a 42% faster Computer Use Agent (CUA), new commands, and integrations with Microsoft/Google/Salesforce suites.

• Cursor released an open-source SDK to build and deploy its AI coding agents anywhere, including CI/CD pipelines for auto-fixing tests and creating merge-ready pull requests.

• Warp Terminal went open-source with an “agent-first” contribution model, where AI coding agents handle coding, planning, and testing.

• Cloudflare now allows AI agents to provision accounts, subscribe to paid plans, and deploy with human approval.

• Stripe introduced 288 things at their Sessions conference, including an Agentic Commerce Suite and Link’s wallet for agents, enabling agents to complete purchases using one-time-use payment credentials.

• Mayo Clinic’s REDMOD model spotted pancreatic cancer up to three years before diagnosis in routine CT scans, three times more sensitive than human experts.

• Harvard study found OpenAI’s o1-preview model diagnosed ER patients more accurately (67.1%) than two attending physicians (55.3% and 50.0%).

• AI for Biological Design: New report in Nature shows Stanford and Arc Institute’s Evo used AI to design full bacteriophage genomes, with 16 AI-written designs becoming working viruses.

• AI in Cybersecurity: The UK’s NCSC warned of a “patch wave” due to AI’s ability to uncover software flaws rapidly. Anthropic’s Claude Mythos Preview found over 2,000 unknown flaws. Anthropic also launched Claude Security for enterprise customers to identify and fix software vulnerabilities.

• Google is integrating Gemini into cars with Google built-in, replacing the old Assistant layer for natural conversation, Maps-aware routing, and more.

• xAI added Voice Cloning to its API, offering 80+ voices across 28 languages from a minute of recorded speech.

• OpenAI released Privacy Filter, an open-weight local model for redacting personal data before enterprise datasets touch the cloud.

• Zed code editor released version 1.0, built in Rust with GPU rendering, featuring Git integration, SSH remoting, a debugger, and support for multiple AI agents.

• After previously acquiring Producer AI, Google rebrands as Flow Music and launched an AI music studio powered by the Lyria 3 model, allowing users to generate and iterate songs and music videos from natural language prompts.

• ElevenLabs launched ElevenMusic, a streaming platform for AI-assisted track creation and remixing.

Thought Leadership & Opinion

• Jack Clark (Anthropic co-founder) forecasts a 60%+ chance of AI systems autonomously building their successors by the end of 2028, driven by rapid progress in R&D tasks.

• Alex Lupsasca (OpenAI researcher) used GPT-5.x models to reproduce complex physics papers in minutes and generate 110 pages of new quantum-gravity results in a day, suggesting a significant shift in theoretical physics research.

• University of Chicago economist Alex Imas argued that as AI makes commodities cheap, human elements like relationships, status, and exclusivity will become scarce and valuable, leading to a “relational sector” with jobs less susceptible to automation.

• Vivek Trivedy writes on the shifting focus to “harness quality” as a differentiator in agent performance and away from the base models, emphasizing how models interact with tools, context, and memory, rather than solely on model weights. He also emphasizes that talking to base or minimally wrapped models makes clear how much productized agents depend on instructions, tools, context packing, and measurement loops.

• Researchers Dat Tran and Douwe Kiela from Stanford University quantify the tradeoffs in multi-agent systems: while multi-agent systems offer solutions for complex tasks, they introduce significant “coordination tax” (API budget, tokens, latency), and a well-optimized single agent can often match or outperform.

• Researchers from the Oxford Internet Institute found that AI chatbots designed to be friendlier are more prone to making mistakes, suggesting a potential trade-off between user-friendly interactions and accuracy.

• Some very clear polling data on public support for AI regulation.

• OpenAI did a deep dive on where the goblins came from.