Cool Things in Tech, May 14-20, 2026
Overview report on the latest in AI and technology in the week of May 14-20, 2026, with a personal bias on generative media and society
This week saw significant developments across the AI landscape, dominated by Google’s aggressive push of Gemini across its ecosystem, Anthropic’s impressive enterprise adoption and strategic hires, and the successful Cerebras IPO. The rise of AI agents continues to be a central theme, impacting everything from enterprise workflows to cybersecurity and even education.
The Headline News
Google Deepens Gemini Integration & Launches New Models: Google made a massive push for its Gemini platform, introducing Gemini 3.5 Flash and Gemini Omni. Flash is optimized for agentic workflows and coding, now available in the Gemini app, Search, and the new Antigravity 2.0 desktop app. Gemini Omni is a multimodal model for video editing and generation, integrated into the Gemini app, Google Flow, and YouTube Shorts. Google’s strategy is to make Gemini an “operating layer” across its products, including a major overhaul of Search, new Workspace voice features, and AI Studio’s ability to generate full Android apps.
Anthropic Surpasses OpenAI in Enterprise Adoption, Hires Andrej Karpathy: Anthropic is making significant strides in the enterprise market, with reports indicating it has surpassed OpenAI in business adoption among paid users, driven by Claude Code and its expansion into finance, legal, and research. This growth is further bolstered by the high-profile hire of Andrej Karpathy, co-founder of OpenAI and former Tesla AI director, who joins Anthropic to lead pretraining research. The company also doubled Claude’s token limits and introduced self-hosted sandboxes and private MCP tunnels to address enterprise adoption blockers.
Cerebras IPO Signals Robust AI Hardware Market: Cerebras, a leading AI inference chipmaker, had a highly successful IPO, pricing at $185/share to raise $5.55 billion and seeing shares soar 68% on its Nasdaq debut, pushing its market cap to approximately $95 billion. This IPO is seen as a strong validation for its “Big Chip” approach and the growing demand for specialized AI hardware, particularly for inference workloads. The company claims to serve models of all sizes, including trillion-parameter models, and specifically named internal OpenAI models 5.4 and 5.5 as current deployments, positioning itself as a serious contender for frontier-scale inference.
Research & Technical Deep Dives
ByteDance Lance Model: ByteDance Research released Lance, a 3B active parameter open-source unified multimodal model capable of image/video understanding, generation, and editing, trained from scratch.
NVIDIA Sana Diffusion Model: NVIDIA launched SANA, a fully open-source image and video generation framework running on laptop GPUs, boasting 20x smaller size and 100x faster performance than Flux-12B, with variants for sprint image generation, video, and world models.
Hugging Face Carbon DNA Models: Hugging Face launched Carbon, a family of generative DNA foundation models, with Carbon-3B matching Evo2-7B’s performance while running 250-275x faster at inference, capable of processing the entire human genome on a single GPU.
OlmoEarth v1.1 for Planet-Scale Mapping: OlmoEarth v1.1 is a new model family reducing compute costs by up to 3x for planet-scale mapping by optimizing token sequence lengths for remote sensing data.
Lighthouse Attention for Faster LLM Training: Nous Research proposed Lighthouse Attention, a selection-based hierarchical attention mechanism, offers up to 17x faster forward and backward passes than standard attention models at large contexts, improving pretraining speed by 1.4x to 1.7x.
Google DeepMind Co-Scientist: Google DeepMind published research on Co-Scientist, a multi-agent system capable of generating and lab-testing biomedical hypotheses, showcasing AI’s application in scientific discovery.
LLMs Optimizing Other LLMs for Training: Research from Prime Intellect demonstrates that contemporary AI systems can autonomously improve their performance on AI research tasks, such as optimizing NanoGPT speedruns, though they still struggle with truly original ideas.
Adaption Labs AutoScientist: Sara Hooker’s Adaption Labs unveiled AutoScientist, an automated fine-tuning system that co-optimizes data and model recipes, reportedly outperforming in-house researchers by 35% in win-rates.
New Tools & Practical Applications
Google Antigravity 2.0 Desktop App: Google launched Antigravity 2.0, a rebuilt desktop application for multi-agent teams to work in parallel on projects, featuring native voice commands, Google Workspace API access, and Android/Firebase integrations, all powered by Gemini 3.5 Flash.
ChatGPT Personal Finance Experience: OpenAI launched a preview of a new personal finance experience in ChatGPT for U.S. Pro users, enabling secure connection of financial accounts via Plaid for personalized insights into spending, investments, and bills.
Anthropic Self-Hosted Sandboxes & Private MCP Tunnels: Claude introduced self-hosted sandboxes and private MCP tunnels, allowing code execution on private servers and secure access to internal databases/APIs, addressing major enterprise adoption blockers.
Cursor Composer 2.5: Cursor introduced Composer 2.5, an upgraded in-house coding model based on Moonshot’s Kimi K2.5, achieving near-frontier benchmark performance at significantly lower token prices.
GLiGuard for Real-Time LLM Safeguard: Fastino Labs and Pioneer AI introduced GLiGuard, a 0.3B-parameter bidirectional encoder for real-time LLM content moderation, evaluating prompt and response safety and jailbreak strategies.
Figure Humanoid Robot Extended Sorting: Figure’s Figure 03 humanoid robot fleet completed over 100 hours of continuous package sorting in a factory stress test, processing over 80,000 packages at 1,240-1,250 packages per hour, 2-4 times faster than human sorters.
GitHub Spec Kit: GitHub’s open-source Spec Kit turns vague ideas into specs for handoff to spec-driven development for AI agents, making them plan thoroughly before coding, and works with 30+ AI agents.
agentmemory: A free open-source tool, agentmemory, provides persistent memory for AI coding agents like Claude Code and Codex across sessions, reducing token usage by 92% and yearly costs to approximately $10, using only SQLite.
OpenAI Codex Update with Faster Git Operations: OpenAI shipped a Codex update that significantly speeds up Git operations by 10-50x and introduces custom shortcuts, enhancing developer workflow efficiency.
Tom Dörr’s 106 Claude Code Skills: Tom Dörr open-sourced a library of 106 Claude Code skills across 15 professions, acting as saved playbooks for repetitive tasks like debugging and PR description writing.
browse.sh: An open-source platform that transforms web browsing into a scriptable agent workspace, enabling AI agents to perform tasks with low-level primitives like clicking and typing.
HKUDS Tool for AI Agent-Ready CLIs: HKUDS open-sourced a tool that converts any software into an AI agent-ready CLI, enabling broader integration of existing tools with AI agents.
RuView: WiFi Router as Human Radar: RuView is an open-source tool that transforms any home WiFi router into a through-wall human radar, capable of detecting people, tracking body skeletons, and measuring vital signs.
Thinking Machines overview: A deep-dive into the recent announcements from Mira Murati’s high-profile venture.
Business, Policy, & Strategy
Elon Musk Loses Lawsuit Against OpenAI: Elon Musk’s lawsuit against OpenAI, alleging a breach of its nonprofit mission, was dismissed by a federal jury on grounds that the case was filed too late, not on the merits.
OpenAI Considers Suing Apple: OpenAI is reportedly considering legal action against Apple over the ChatGPT-Siri integration, claiming underperformance and being “buried” in the UI.
OpenAI Guaranteed Capacity: OpenAI introduced “Guaranteed Capacity,” allowing customers to secure long-term access to compute resources for their AI products through one-, two-, or three-year commitments with potential discounts.
Anthropic Acquires Stainless: Anthropic acquired Stainless, a developer tools startup known for its SDK automation platform, to strengthen Claude’s SDK, CLI, and MCP tooling capabilities.
Google and Blackstone AI Cloud Venture: Google and Blackstone are partnering to create a new AI cloud company, with Blackstone investing $5 billion to rent out Google’s specialized TPU chips and bring 500 megawatts of capacity online by 2027, aiming to challenge Nvidia.
NVIDIA Increases AI Infrastructure Stakes: Nvidia nearly doubled its stake in CoreWeave and added shares in Coherent, further solidifying its financial influence on the AI infrastructure stack.
AI Layoffs Backfiring on Companies: A CNBC analysis found that 56% of S&P 500 companies announcing AI-linked layoffs saw their stock prices decline, averaging a 25% drop, indicating that the market is not rewarding AI-attributed job cuts.
AI Industry’s Growing Political Influence: The AI industry is significantly increasing its political influence, with venture capital firms like Andreessen Horowitz becoming major donors and funding super PACs like Leading the Future to shape AI regulation.
SaaS Financial Structures Repriced by AI: The failure of Medallia, a SaaS platform, suggests that AI is repricing the financial structures behind software, particularly for PE-owned SaaS companies with seat-based pricing.
Mistral AI Acquires Emmi AI: Mistral AI acquired Emmi AI to expand its focus on industrial AI, adding simulation and engineering workflows to its enterprise offerings.
Recursive Superintelligence Raises $650M: A new company, Recursive Superintelligence, raised over $650 million at a $4 billion+ valuation to build self-improving AI, founded by notable researchers from leading AI companies.
The “Unreasonable Effectiveness” of HTML for Claude Code: The Claude Code team discusses the use of HTML over markdown for transmitting information back-and-forth within Claude Code sessions. HTML’s richness allows it to convey complex information more effectively than Markdown for Claude Code, enhancing readability and providing better sharing and interaction capabilities for ingesting context in AI development.
Free ChatGPT to all Maltese residents: OpenAI and the government of Malta offer free ChatGPT access to all Maltese residents who complete a university-designed AI literacy course.
Thought Leadership & Opinion
AI Security Threats and Defenses Evolve: The report of Google’s discovery of an AI-assisted zero-day exploit targeting two-factor authentication, coupled with the UK-based AISI’s finding that frontier models’ autonomous cyber “time horizon” has doubled, underscores the escalating threat of AI in cybersecurity. Simultaneously, Microsoft demonstrated MDASH, a multi-agent vulnerability-finding system, highlighting the urgent need for advanced AI-driven defenses.
Open-weight LLMs and Long-Context Efficiency Advances: Sebastian Raschka discusses the recent improvements in open-weight models and their handling of long context windows, one direction where the foundational labs have traditionally outperformed open-weight models. “As reasoning models and agent workflows keep more tokens around (for longer), KV-cache size, memory traffic, and attention cost quickly become the main constraints, and LLM developers are adding a growing number of architecture tricks to reduce those costs.”
“Positive Alignment” for AI Systems: A coalition of researchers from top academic institutions and AI labs published a position paper advocating for “positive alignment,” focusing on building AI systems that actively support human and ecological flourishing, moving beyond just preventing harm.
“AI Stuxnet” Precursor: Researchers at SentinelOne investigated fast16.sys, a 20-year-old computer virus that subtly tampers with high-precision calculation software. This is discussed as a potential precursor to how a superintelligence might prevent rivals by degrading their scientific capabilities.
The Workflow Collision: Sean Escriva discusses the “Workflow Collision”, when barriers to integrating AI agents arises from the inherent conflict between human workflows and agent lifecycles. He discusses the degree to which human workflows, which were designed to improve human processes and collaborations, will require serious rethinking of current operational models. (Editor’s note: We discussed related work we did in this area here.)
Software’s Centaur Era: Richard Marmorstein discusses the current “Centaur Era”, where the best performance in tasks is increasingly achieved through human-AI collaboration, akin to a human wielding an AI machine, where the combined intelligence surpasses either component alone. This “centaur” approach emphasizes augmentation rather than full automation.
Domain Knowledge as AI Leverage: Beshr Kayali Reinholdsson discusses the role of developers when spec-driven development now can involve 10k lines of agent-optimized documentation that nobody reads. Understanding specific domains and their connection to programs is becoming increasingly valuable with AI, putting spec-driven development in the spotlight as a critical skill for leveraging AI effectively.
Companies Under “AI Psychosis”: Mitchell Hashimoto discusses the risks to organizations undergoing AI transitions using traditional metrics, without consideration of other forms of risk increasing. The rapid changes driven by AI are causing underlying architectural decay to go unnoticed in many companies, with a focus on incorrect metrics leading to denial and potential long-term issues.
The Sigmoids Won’t Save You: Astral Codex Ten discusses the misidentification of exponential curves as sigmoids, and the connection to projections of AI capabilities in the current moment. The belief that all exponential growth curves eventually become sigmoids may not apply to AI, as some exponentials (like solar power deployment or declining birth rates) have not yet plateaued, suggesting continued rapid progress in AI.