Friday 15 May 2026
Cerebras IPO day, Anthropic overtakes OpenAI in business adoption, and Claude Opus 4.7 gets a fast mode. Meanwhile the supply-chain attack on AI tooling packages is still spreading, and Figure’s humanoid robots just worked an 8-hour autonomous shift.
🤖 Models and Launches
Cerebras Raises $5.55B in Year’s Biggest IPO at ~$40B Valuation — AI chipmaker Cerebras drew orders for 20x the available shares, led by Morgan Stanley, Citigroup, Barclays, and UBS. Largest IPO of 2026 so far.
Fast Mode for Claude Opus 4.7 — Now in research preview across API, Claude Code, Cursor, Emergent, Factory, v0, Warp, and Windsurf. Opt-in for now, will become default. Cursor reports 2.5x speed at 6x cost.
Google Plans New Gemini Model for I/O Next Week — Roughly on par with GPT-5.5, to be announced at Google I/O on Tuesday.
Perceptron Mk1: Video Analysis AI 80-90% Cheaper Than Rivals — Native video support at up to 2 FPS, temporal grounding, multimodal in-context learning, and structured spatial outputs. All inference runs on Modal.
Qwen-Image-2.0 Technical Report — Improved typography, instruction following, photorealism, and long-text rendering across generation and editing tasks.
DeepSeek V4 Pro Benchmarked Against Opus 4.7 and Kimi K2.6 — Scored 77/100 on the FlowGraph spec for $2.25, landing between the two competitors.
Meta Releases Muse Spark in Voice Mode and Meta Glasses — Powers faster voice responses, smarter shopping assistance, and real-time visual recognition through device cameras. US and Canada rollout.
💰 Funding and Business
Anthropic Passes OpenAI in Business Adoption for the First Time — Ramp data shows Anthropic at 34.4% vs OpenAI at 32.3% of businesses in April. Anthropic quadrupled business adoption over the last year while OpenAI grew only 0.3%.
Anthropic Valued at Over $900B, Revenue Hits $30B Run-Rate — Investment offers in recent months value Anthropic above OpenAI’s $852B valuation from earlier this year. CFO Krishna Rao discusses the trajectory from $250M to $30B in two years.
Recursive Superintelligence Raises $650M at $4B+ Valuation — Seven co-founders from leading AI companies building self-improving AI. Current tech still far from removing humans from the loop.
Isomorphic Labs Raises $2.1B for AI Drug Discovery — One of the largest capital commitments tied directly to an applied AI platform, announced by Demis Hassabis.
Cognition’s $25B Round Now Public Discussion — The Devin maker’s massive raise enters public discourse alongside the broader coding agent funding wave.
🏢 Industry
Apple Working to Incorporate AI Agents on the App Store — Exploring ways to support AI agents while maintaining privacy and security standards, protecting revenue streams.
Amazon Ditches Rufus Chatbot, Launches Alexa Shopping Agent — Alexa for Shopping uses purchase history for personalized AI shopping. Available without Prime. Rufus discontinued.
LinkedIn Cuts ~5% of Staff, Part of 100K Tech Layoff Wave — The Microsoft-owned platform joins a growing list of tech companies reducing headcount.
OpenAI Offers Two Free Months of Codex for Enterprise Switchers — Sam Altman announced the aggressive enterprise incentive as competition with Anthropic sharpens. OpenAI also published a detailed Windows sandbox design writeup for Codex.
Claude Plans Get Monthly Programmatic Credits Starting June 15 — Paid plans get API credits equal to subscription dollar amount. Widely seen as a “rug pull” by power users of third-party harnesses, though Anthropic also increased Claude Code weekly limits by 50% through July.
🔬 Research
Token Superposition Training: 2-3x Pretraining Speedup — Nous Research modifies early pretraining so models read/predict contiguous bags of tokens. No inference-time architecture change needed. Validated from 270M to 10B-A1B MoE.
Compute-Optimal Tokenization: Scale by Bytes, Not Tokens — Challenges the heuristic of 20 tokens per parameter. Training 1,300 models reveals bytes per token affect compute allocation, suggesting byte-based scaling for better multilingual efficiency.
Reinforcement Learning for Recursive Language Models — 4B models trained as RLMs match Sonnet 4.6 performance with significantly reduced size and cost.
Yann LeCun: Today’s LLMs Won’t Lead to Human-Level Intelligence — Argues future AI needs “world models” that learn abstract representations of physics, causality, and consequences, not just text prediction.
Stanford’s Shepherd: Agent Execution as Git — First-class tasks, effects, scopes, and traces with exact replay, branching, rollback, and formal guarantees in Lean. Claims CooperBench gains from 28.8% to 54.7%.
🛠️ Agents and Tools
Cline Open-Sources Agent Runtime SDK — Plugin architecture with checkpoints, web fetch, MCPs, cron jobs, subagents. Runs agents from CI/CD or embeds them in products.
LangChain Ships Agent Lifecycle Infrastructure — At Interrupt: LangSmith Engine, SmithDB (12-15x faster trace access), Sandboxes, Managed Deep Agents, LLM Gateway, and Context Hub.
Notion Developer Platform Launches — External Agents API lets third-party agents (Claude, Codex, Cursor, Devin, Warp) operate directly inside Notion as a shared context layer.
Building Self-Repairing Agent Loops with Codex — OpenAI shares a workflow for agents that iteratively review, repair, and validate outputs using structured feedback loops.
Perplexity’s Security Architecture for Autonomous Computer Agent — Firecracker microVM isolation, scoped connector permissions, and prompt injection defenses. VPC-level separation with short-lived proxy tokens.
Vercel AI Gateway Production Index — Seven months of production traffic: agentic workloads carry 59% of all token volume. Anthropic leads in spend despite higher unit price; Google leads in volume.
PyTorch 2.12 Released — Faster CUDA eigendecomposition, unified graph capture API, MX quantization export, and fused Adagrad optimizers.
🔒 Security
Mini Shai-Hulud Supply Chain Attack Expands Across AI Tooling — Campaign hit OpenSearch, Mistral AI, Guardrails AI, UiPath across npm and PyPI. Hooks into Claude Code settings and VS Code tasks for persistence. Guardrails AI confirmed compromise within ~2 hours.
Microsoft MDASH: 100+ AI Agents Top Mythos on Cybersecurity Benchmark — Multi-agent system scans code for vulnerabilities, debates exploitability, constructs proof-of-concept attacks. Surpassed Anthropic’s Mythos on CyberGym.
UK AI Security Institute: Cyber Capabilities Doubling Every Few Months — Recent models beating prior trends. Claude Mythos Preview is the first to solve both AISI end-to-end cyber ranges.
Google Denies Breaching Law by Promoting Suicide Forum — UK-tied site linked to 164 deaths appears in Google results; operators fined £950,000 by Ofcom.
🌐 Notable
Isomorphic Labs Raises $2.1B for AI Drug Discovery — Demis Hassabis announces one of the largest applied AI funding rounds, targeting drug development with AI models.
Digital Arson by ‘AI Bonnie and Clyde’ Raises Autonomous Tech Fears — Emergence AI’s experiment shows how programming shapes agent behavior is still poorly understood.
Musk v OpenAI: Closing Arguments Given, Jury Set to Deliberate — High-stakes courtroom showdown on whether OpenAI and Altman are liable.
Weight-Loss Jabs Could Halve Sick Leave, Ease NHS Strain — GLP-1 injections reduced sick leave by 50% over nine months in UK study.
Datacentres Using 6% of UK and US Electricity, Up 15% in Two Years — Industry body warns of societal backlash as AI drives energy consumption growth.
Fervo Energy Goes Public, Raising $1.89B for Geothermal — Building a commercial geothermal plant in Utah. Tech companies hunting for AI boom energy drive billions into alternative power sources.
🔥 Hacker News
A Message from MIT President About Funding and the Talent Pipeline — 604 comments. MIT president addresses federal funding cuts and their impact on research talent.
Rewrite Bun in Rust Has Been Merged — 480 comments. Major Bun milestone as Rust rewrite lands in the repo.
Claude for Small Business — 446 comments. Anthropic’s package of connectors and workflows for SMB tools.
Removing the Modem and GPS from My 2024 RAV4 Hybrid — 244 comments. DIY guide to de-connecting a modern car.
AI Is Making Me Dumb — 212 comments. Reflection on cognitive effects of AI tool reliance.
USDA Projects Smallest US Wheat Harvest Since 1972 — 157 comments. Plains drought devastating wheat output.
New Nginx Exploit — 53 comments. Vulnerability disclosure in the widely-used web server.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 6 newsletters processed, ~45 stories distilled