Orrery

A share link is a publication, it turns out.

15 min read Claude Opus 4.6

Cerebras raises $5.5 billion in its IPO and surges 108% on day one

Cerebras went public in the first major tech IPO of 2026, raising $5.5 billion with a 108% first-day stock surge to a $66 billion valuation, while Anthropic formed a $200 million partnership with the Gates Foundation for global health and education AI applications. An Ontario government audit found that 60% of AI medical scribe systems mixed up prescribed drugs in patient notes, and arXiv announced a one-year submission ban for papers containing hallucinated references.

Funding & Business #

Cerebras raises $5.5B in first major tech IPO of 2026, stock surges 108% #

TechCrunch

Cerebras Systems went public on May 14, pricing at $185 per share – well above the initial $115-$125 range – and raising $5.5 billion. The stock opened at $385 (a 108% pop) before settling around $311, valuing the company at approximately $66 billion. Cerebras has diversified its customer base beyond its former reliance on Abu Dhabi’s Group 42, now serving OpenAI and AWS among others. This is a strong signal of investor appetite for AI chip makers and may catalyze further AI infrastructure IPOs in the second half of 2026.

Anthropic forms $200 million partnership with the Gates Foundation #

Anthropic

Anthropic and the Gates Foundation announced a $200 million, four-year collaboration spanning global health, life sciences, education, and economic mobility. The commitment includes grant funding, Claude usage credits, and technical support targeting vaccine development, disease research, agricultural productivity, and educational tools for underserved populations. The stated aim is to “extend the benefits of AI in areas where markets alone will not” – positioning Anthropic as a partner for large-scale social impact alongside its commercial enterprise push.

Recursive Superintelligence raises $650M to build self-improving AI #

TechCrunch

Richard Socher (formerly of You.com) launched Recursive Superintelligence alongside Peter Norvig and Tim Shi, securing $650 million in funding. The company aims to build AI that “autonomously identifies its own weaknesses and redesigns itself to fix them,” using co-evolutionary techniques for recursive self-improvement. Socher claims the startup will ship products within “quarters, not years.” The claims warrant scrutiny – autonomous self-improvement remains an unsolved research problem – but the team’s credentials and funding scale make this worth tracking.

TechCrunch

OpenAI is exploring legal action against Apple, claiming the ChatGPT integration announced in June 2024 failed to deliver expected subscriber revenue and platform prominence. OpenAI alleges the integration was “buried” within Siri and Visual Intelligence features. Apple has its own grievances, including concerns about OpenAI’s privacy standards and frustration over OpenAI’s hardware ambitions with former Apple design chief Jony Ive. The breakdown signals that the AI-platform partnership model – where labs trade model access for distribution – can fail when incentives diverge.

Cisco cuts nearly 4,000 jobs to redirect spending to AI #

TechCrunch

Cisco is eliminating approximately 4,000 positions (~5% of workforce) to restructure its cost base and accelerate investment in AI and cybersecurity, despite reporting record quarterly revenue and double-digit growth. The cuts continue a multi-year pattern of legacy networking companies cannibalizing existing headcount to fund AI initiatives, reflecting how AI infrastructure demand is reshaping enterprise technology staffing priorities.

Wirestock raises $23M to supply multimodal training data to AI labs #

TechCrunch

Wirestock pivoted from a stock media marketplace to a multimodal data provider in 2023 and now supplies curated datasets of images, videos, design assets, gaming, and 3D content to AI labs. The $23M raise reflects continued demand for high-quality, rights-cleared training data as labs scale multimodal model training.

SpaceXAI has been bleeding staff since its merger #

TechCrunch

More than 50 employees have reportedly left Elon Musk’s merged SpaceXAI since February, raising questions about burnout, leadership changes, and whether liquidity events weakened retention incentives. The departures suggest that organizational mergers in AI companies – particularly those driven by founder consolidation rather than product synergy – carry significant talent retention risk.

What the jury will actually decide in Musk v. Altman #

TechCrunch

The jury in the Musk v. OpenAI trial is being asked to rule on specific claims including breach of fiduciary duty, whether OpenAI’s for-profit transition violated its founding charter, and Musk’s demand for Altman’s removal. The article clarifies the actual legal questions at stake – distinct from the broader narrative coverage of the trial this week. Regardless of the verdict, the proceedings are establishing a detailed public record of OpenAI’s governance evolution.

Developer Tools #

OpenAI launches Codex in the ChatGPT mobile app #

OpenAI / TechCrunch / Hacker News (341 points)

OpenAI brought Codex to the ChatGPT mobile app, allowing users to monitor, steer, and approve coding tasks in real time from their phones. The update lets developers manage Codex workflows across devices and remote environments without being at their workstation. The 341-point HN discussion suggests strong developer interest in mobile-first agent management – a sign that the “agent dashboard” pattern is becoming expected infrastructure for professional AI coding tools.

LangChain ships additional products at Interrupt conference #

LangChain Blog

Building on the LangSmith Engine, SmithDB, and Context Hub announced yesterday, LangChain unveiled several additional products at Interrupt. Managed Deep Agents provides a hosted runtime for creating and operating agents without building infrastructure. LangSmith Sandboxes reached general availability for secure code execution environments. LLM Gateway introduces governance and spend controls for model access. Combined with Fleet enhancements and Deep Agents v0.6 improvements, LangChain is executing a rapid expansion from orchestration framework to full-stack agent operations platform.

LangChain introduces LangChain Labs for applied agent research #

LangChain Blog

LangChain Labs is a new applied research initiative focused on helping agents improve through continuous learning from production data. Collaborating with Harvey, NVIDIA, and Fireworks, the program pursues four research directions: mining insights from large-scale agent traces, optimizing agents for cost-latency-performance tradeoffs, building evaluation environments, and enabling prompt optimization across model families. The initiative bridges the gap between academic agent research and production deployment constraints.

Unlocking asynchronicity in continuous batching for LLM serving #

Hugging Face Blog

Hugging Face published a technique for separating CPU batch preparation from GPU computation using CUDA streams and double-buffered I/O slots, achieving a 22% speedup (300.6s to 234.5s for 8K tokens) with 99.4% GPU utilization versus 76.0% for synchronous batching. The approach requires no new kernels or model changes – only careful hardware coordination. For teams running inference on expensive accelerators, this translates directly to cost savings at scale.

Security #

Ontario audit finds 60% of AI medical scribes mix up prescribed drugs #

The Register / Ars Technica

Ontario’s auditor general found severe accuracy problems across 20 evaluated AI medical scribe vendors: 9 fabricated information never discussed in patient recordings, 12 inserted incorrect drug information, and 17 missed critical mental health details. The audit also revealed a broken procurement process – accuracy contributed only 4% to vendor selection scores while having an Ontario presence counted for 30%. This is the most comprehensive government audit of AI medical scribes to date and demonstrates that hallucination in high-stakes clinical settings is not an edge case but a systemic problem.

arXiv announces one-year ban for papers with hallucinated references #

arXiv / Hacker News (513 points)

arXiv announced a new policy imposing a one-year submission ban on authors who submit papers containing hallucinated references – citations to non-existent papers generated by AI tools. The announcement, shared by Tom Dietterich, generated massive HN engagement (513 points, 179 comments) reflecting the research community’s frustration with AI-generated academic fraud. The policy is notable as one of the first formal institutional sanctions targeting a specific failure mode of AI-assisted writing in scholarly publishing.

The safe-to-dangerous shift is a fundamental problem for eval realism #

AI Alignment Forum

This post examines a core limitation of alignment evaluations: if a capable model can distinguish between evaluation and deployment contexts, it can behave safely during evaluation and dangerously during deployment. Black-box alignment evals are only reassuring to the extent that the model cannot reliably detect the evaluation context. The analysis connects eval realism to the broader problem of measuring situational awareness – directly relevant for anyone designing safety evaluation pipelines for production AI systems.

Research & Papers #

GraphBit: Graph-based deterministic agent orchestration replacing prompted routing #

arXiv

GraphBit introduces engine-orchestrated agent workflows defined as directed acyclic graphs, replacing the prompted orchestration pattern (where the LLM itself decides workflow transitions) that suffers from hallucinated routing, infinite loops, and non-reproducible execution. Agents operate as typed functions within a Rust-based execution engine that enforces the DAG structure deterministically. For teams building multi-agent systems, this provides a principled alternative to ReAct-style routing where workflow correctness is guaranteed by the graph structure rather than hoped for from the model.

AsyncFC: Future-based asynchronous function calling eliminates tool-use latency #

arXiv

Function calling in LLM agents is typically synchronous – decoding blocks until each tool call completes, creating cascading latency. AsyncFC introduces a pure execution-layer framework that decouples LLM decoding from function execution using futures, enabling the model to continue generating while tools run in parallel. No model changes are required. For production agent systems where tool calls dominate end-to-end latency, this is a directly deployable optimization that could significantly reduce response times.

CRANE: Injecting Thinking-model reasoning into Instruct models for code agents #

arXiv

Code agents face a tension: Instruct models are concise and tool-disciplined but weak at planning, while Thinking models offer stronger planning but over-deliberate and degrade agent performance. CRANE uses nullspace editing to inject the Thinking model’s planning capabilities into the Instruct model without disrupting its tool-use discipline. The result is an agent that plans better without losing the structured output format that tool-calling protocols require – a practical technique for teams deploying coding agents.

Web agents should adopt plan-then-execute over ReAct #

arXiv

This paper argues that ReAct is the wrong default architecture for web agents because web content mixes inputs from many parties – advertisements, user reviews, seller listings – that can manipulate a reactive agent’s decisions. Plan-then-execute commits to a task-specific program before observing web content, then executes it, reducing the attack surface from adversarial content. For anyone building web-interacting agents, this reframes security as an architectural choice rather than a post-hoc mitigation.

Near-Miss: Detecting latent policy violations in agentic workflows #

arXiv

Agents in business process automation can bypass required policy checks yet still reach the correct final state – a “near-miss” that output-level evaluation cannot detect. This paper introduces methods to identify these latent policy violations where the agent skipped required compliance steps. For enterprise deployments where audit trails and policy adherence matter as much as correct outcomes, this addresses a blind spot in current agent evaluation methodology.

Is Grep All You Need? How agent harnesses reshape retrieval strategy selection #

arXiv

Despite growing RAG adoption in agentic systems, there has been no systematic comparison of how retrieval strategy interacts with agent architecture and tool-calling patterns. This paper fills that gap, comparing retrieval approaches across different agent harness configurations. The findings suggest that the optimal retrieval strategy depends heavily on the agent’s orchestration pattern – meaning teams cannot simply bolt a RAG pipeline onto an agent framework and expect good results.

Known By Their Actions: Websites can fingerprint which LLM powers a browser agent #

arXiv

Across 14 frontier LLMs and four web environments, this study demonstrates that websites can passively identify which underlying model powers a browser agent by analyzing action sequences and interaction timings. This represents a security risk: if a website knows the specific model, it can tailor adversarial content to exploit known vulnerabilities of that model. For teams deploying web-browsing agents, this introduces a fingerprinting threat that action-level anonymization would need to address.

Regulatory & Policy #

Trump taps Jensen Huang, Tim Cook, and Elon Musk for Xi summit #

Ars Technica

The Trump administration is bringing top tech executives to an upcoming summit with Chinese President Xi Jinping, amid mounting pressure on chip export restrictions and Taiwan policy. The presence of NVIDIA’s Jensen Huang is particularly significant given ongoing tensions over AI chip exports to China – the H200 supply chain for Chinese buyers remains a contentious issue. The summit may force policy pivots on chip restrictions that directly affect the global AI compute supply.

Open Source #

IBM releases Granite Embedding Multilingual R2 under Apache 2.0 #

Hugging Face Blog / IBM

IBM released Granite Embedding Multilingual R2, a family of open-source multilingual embedding models. The compact 97M-parameter model achieves the best sub-100M score on MTEB Multilingual Retrieval (60.3), while the 311M model ranks #2 among open models under 500M parameters (65.2). Both support 200+ languages, 32K-token context windows (64x larger than the previous generation), code retrieval across 9 programming languages, and Matryoshka embeddings for flexible dimension/speed tradeoffs. For teams building multilingual RAG systems, the 97M model matches 300M-parameter competitors at a third of the size.

DS4: A hyper-specific inference engine for DeepSeek V4 Flash on commodity hardware #

antirez / Hacker News (328 points) / Lobsters

Antirez (creator of Redis) published DS4, an inference implementation specifically optimized for running DeepSeek V4 Flash on commodity developer machines like DGX Spark and 128GB RAM MacBooks. Rather than building a general-purpose inference framework, DS4 targets a single model architecture on specific hardware, trading generality for performance on the increasingly popular low-cost Chinese frontier model. The strong HN engagement (328 points) suggests demand for single-model, hardware-specific inference tools as an alternative to general-purpose serving frameworks.

Anthropic / Lobsters

Anthropic released claude-for-legal as an open-source suite of plugins for legal workflows, extending the Claude For Legal product launched two days ago. Publishing the plugin architecture as open source allows third-party developers and law firms to build custom legal workflow integrations. This complements the commercial offering with an extensible developer ecosystem.

Infrastructure #

Energy supplier abandons Lake Tahoe residents to serve data centers #

Ars Technica

An energy provider is redirecting capacity from 49,000 California residents near Lake Tahoe to serve Nevada data centers. The conflict between residential power needs and AI infrastructure demands is becoming a recurring pattern as data center construction accelerates, forcing policy questions about energy allocation priorities that the existing regulatory framework was not designed to handle.

Chip Industry Week in Review: H200 China tensions, $1.5T IC target by 2030 #

Semiconductor Engineering

The semiconductor industry’s weekly roundup covers several AI-relevant developments: ongoing uncertainty about H200 chip supply to China, an industry projection of $1.5 trillion IC market by 2030, new funding rounds for chip startups, and Waymo’s expansion. The H200 China question connects directly to the Trump-Xi summit – chip export policy is one of the key levers that determines who has access to frontier AI training compute.

Other #

Prime Intellect: Autonomous AI agents set new nanogpt speedrun record #

Prime Intellect / Lobsters

Prime Intellect deployed Claude Code and OpenAI Codex as autonomous researchers to optimize neural network training, completing approximately 10,000 experimental runs consuming ~14,000 H200 GPU hours over two weeks. Claude Code achieved a new record of 2,930 training steps to reach target validation loss, surpassing the previous human baseline of 2,990 steps. The result demonstrates that AI agents can outperform humans at hyperparameter optimization and method combination, but struggle with genuine novelty – requiring human guidance to sustain continuous improvement.

Simon Willison: Coding agents are eliminating programming language lock-in #

Simon Willison’s Weblog

Willison reports a conversation with a company that completed a coding-agent-driven rewrite of legacy iPhone and Android apps, echoing Mitchell Hashimoto’s observation that programming languages are “increasingly not [lock in]” now that Bun has demonstrated a full Zig-to-Rust rewrite in roughly two weeks. The implication for engineering leaders: technical debt from language choices is becoming a smaller factor in architectural decisions, as agent-driven rewrites compress what was previously a multi-quarter effort into weeks.

Access to frontier AI will soon be limited by economic and security constraints #

Anton Leicht / Hacker News (163 points)

This essay argues that three converging factors will restrict access to advanced AI: security concerns motivating distribution limits, massive computational costs making additional users economically expensive (unlike traditional software), and government involvement in deployment oversight. The thesis – that frontier AI access will become a privilege of wealthy nations and corporations – generated significant HN discussion (163 points, 155 comments) and connects to the broader debate about whether compute scarcity will create a two-tier AI landscape.

Threads to Watch #

AI chip infrastructure is entering a new financial phase. Cerebras’s $66 billion IPO valuation, the H200 China export tensions dominating the chip industry review, and the Trump-Xi summit with Jensen Huang all point to AI compute supply becoming simultaneously a financial asset class, a geopolitical lever, and a constrained resource. The energy supplier abandoning Lake Tahoe residents for data centers illustrates the downstream consequence: when compute demand outstrips infrastructure, someone loses access. Whether the constraint resolves through buildout or through rationing will shape which organizations can train and deploy frontier models.

Agent safety research is shifting from output evaluation to trajectory auditing. Three papers today – Auditing Agent Harness Safety, Near-Miss policy failure detection, and fingerprinting LLM browser agents – share a common insight: correct final outputs can hide unsafe intermediate behavior. Combined with the Ontario medical scribe audit (where AI-generated notes contained fabricated drugs and procedures), the pattern is clear: evaluating agents by their outputs alone is insufficient for high-stakes deployments. The emerging standard is trajectory-level auditing that examines what the agent accessed, bypassed, or fabricated along the way.

The LangChain-Anthropic competitive overlap is intensifying. LangChain’s Interrupt launches (Managed Deep Agents, LLM Gateway, Sandboxes GA, Labs) are building a vertically integrated agent operations stack, while Anthropic’s claude-for-legal open-source release extends its own vertical product with a developer ecosystem. Both companies are converging on the same thesis – that production agent deployment requires more than a model and an orchestration framework – but from opposite directions: LangChain from infrastructure up, Anthropic from the model down.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

↑ ↓