Anthropic and OpenAI both launch enterprise AI joint ventures

Anthropic and OpenAI both launched enterprise AI services joint ventures backed by major asset managers, signaling that frontier labs see direct enterprise deployment – not just API access – as the next revenue layer. Meanwhile, the NHS ordered hundreds of GitHub repositories closed citing Mythos-related vulnerability concerns, and new research formalized fundamental limits of multi-agent debate while proposing orchestration-trace RL for training agent teams.

Funding & Business #

Anthropic Launches Enterprise AI Services Company with Blackstone, Hellman & Friedman, and Goldman Sachs #

Anthropic / TechCrunch

Anthropic is spinning up a dedicated enterprise services company that deploys Applied AI engineers to help mid-sized organizations integrate Claude into core operations, backed by Blackstone, Hellman & Friedman, and Goldman Sachs. CFO Krishna Rao stated enterprise demand for Claude is “significantly outpacing any single delivery model,” hence the expansion beyond the existing Claude Partner Network. The move targets companies – community banks, manufacturers, healthcare systems – that lack the in-house expertise for sophisticated AI deployments, positioning Anthropic to capture services revenue alongside API margins.

Sierra Raises $950M at $15B Valuation #

SiliconANGLE / TechCrunch / CNBC

Bret Taylor and Clay Bavor’s AI agent startup Sierra raised $950 million led by GV and Tiger Global, reaching a $15 billion valuation with over $1 billion in total capital. The company plans to use proceeds to build custom inference algorithms that replace proprietary models, reducing serving costs for its Agent SDK platform. Sierra’s positioning – providing agent infrastructure for enterprise customer experience rather than building its own frontier model – represents a bet that the agent platform layer captures more durable value than the model layer beneath it.

Cerebras on Track for Blockbuster IPO at $26.6B+ Valuation #

TechCrunch

AI chip maker Cerebras is heading for an IPO that could value it at $26.6 billion or more, driven by its deep partnership with OpenAI for inference workloads. The company’s wafer-scale architecture offers a fundamentally different compute model from Nvidia’s GPU clusters, betting that inference-optimized silicon will command premium pricing as the industry’s workload mix shifts from training to serving. The IPO would be one of the largest AI hardware debuts since Nvidia’s own ascent, testing whether the market values inference-specific chip companies at frontier multiples.

OpenAI and PwC Partner to Automate CFO Operations #

OpenAI

OpenAI and PwC announced a partnership to deploy AI agents for enterprise finance workflows including forecasting, controls automation, and CFO function modernization. The collaboration represents OpenAI’s push into professional services delivery alongside its existing API business. Together with the Anthropic JV above, the pattern is clear: both frontier labs are building enterprise services arms to capture deployment revenue that currently flows to consulting firms.

Security #

NHS to Close-Source Hundreds of GitHub Repos Over Mythos AI Security Concerns #

The Register / Terence Eden

NHS England issued internal guidance mandating repositories “must not be public” by default, with a May 11 deadline for compliance, citing Anthropic’s Mythos AI model and its demonstrated capability to discover thousands of previously unknown zero-day vulnerabilities. Neither the UK AI Safety Institute nor NCSC recommended this action, and the policy contradicts the UK Technology Code of Practice which mandates open-source for public-sector projects. The decision highlights an emerging tension: organizations interpreting AI vulnerability discovery capabilities as requiring security-through-obscurity rather than investing in actually fixing the vulnerabilities that AI models can find.

Google Chrome Silently Installs a 4GB AI Model Without Consent #

thatprivacyguy.com / Hacker News (273 points)

Chrome automatically downloads Gemini Nano, a 4GB language model, to user devices when AI features are active (enabled by default on compatible hardware), with no Settings checkbox for opt-out. The model re-downloads automatically if deleted, requiring users to navigate to chrome://flags to disable it – effectively making meaningful opt-out unavailable for typical users. For organizations managing device fleets, the silent deployment of a multi-gigabyte model raises storage, bandwidth, and policy compliance concerns that are invisible without explicit monitoring.

Shadow IT Has Given Way to Shadow AI: Enter AI-BOMs #

The Register

Traditional software bills of materials (SBOMs) no longer provide complete visibility into enterprise environments increasingly populated by AI applications and agents, prompting the development of AI-BOMs – bills of materials that inventory AI models, training data provenance, prompt configurations, and agent tool access. The core argument: “If you don’t have visibility, you can’t understand what to protect.” For teams deploying agent systems, this frames supply chain security as requiring not just code and dependency tracking but also model versioning, prompt lineage, and tool-permission auditing.

When Alignment Isn’t Enough: Response-Path Attacks on LLM Agents #

arXiv

Researchers formalize a critical integrity gap in Bring-Your-Own-Key (BYOK) agent architectures: a malicious relay can modify aligned LLM responses after generation but before agent execution, making even perfectly aligned models ineffective. The “Response-Path Attack” allows the relay to observe, suppress, or replace downstream messages without detection by the agent or user. For teams building agent pipelines that route through third-party infrastructure, this demonstrates that alignment is necessary but insufficient – end-to-end integrity verification between model output and agent execution is a separate unsolved problem.

Research & Papers #

Reinforcement Learning for LLM-based Multi-Agent Systems through Orchestration Traces #

arXiv

This paper studies RL for multi-agent LLM systems through “orchestration traces” – temporal interaction graphs whose events include sub-agent spawning, delegation, communication, tool use, return, and aggregation. The key insight: as LLM agents evolve from isolated tool users into coordinated teams, RL must optimize not only individual actions but how work is delegated, communicated, and stopped across agent hierarchies. For teams building multi-agent systems, this provides a formal framework for training orchestration policies rather than hand-coding delegation logic.

The Reasoning Trap: An Information-Theoretic Bound on Closed-System Multi-Step LLM Reasoning #

arXiv

When copies of the same language model debate, they produce diverse phrasings of one perspective rather than diverse perspectives – a phenomenon the authors term the “Debate Trap.” More broadly, closed-system reasoning where agents iteratively transform each other’s outputs tends to preserve answer accuracy while degrading the reasoning behind those answers. This provides an information-theoretic argument for why multi-agent debate has a ceiling: without external information injection, iterative agent interaction cannot exceed the information content of the initial model, regardless of how many rounds or agents participate.

FitText: Evolving Agent Tool Ecologies via Memetic Retrieval #

arXiv

As API ecosystems scale to tens of thousands of endpoints, static retrieval from the initial query alone cannot bridge the semantic gap between how users describe tasks and how tools are documented. FitText embeds tool retrieval directly in the agent’s reasoning loop, generating natural-language pseudo-tool descriptions that evolve as the agent’s understanding of its needs changes during execution. For teams building agents with large tool inventories, this addresses the practical problem that the right tool often becomes apparent only after the agent has begun working on the task.

T2PO: Uncertainty-Guided Exploration Control for Stable Multi-Turn Agentic Reinforcement Learning #

arXiv

Multi-turn RL for LLMs suffers from instability and training collapse despite existing stabilization techniques. T2PO argues this stems from inefficient exploration in multi-turn settings where policies continue generating low-information actions that reduce neither uncertainty nor task progress. The proposed solution uses uncertainty-guided exploration control to focus policy updates on actions that actually reduce epistemic uncertainty, providing a mechanism for stable training of agentic systems that must take many sequential actions.

Binary Rewards and Reinforcement Learning: Fundamental Challenges #

arXiv

Reinforcement learning with verifiable rewards (RLVR) has become standard for improving LLM reasoning, yet models trained this way often suffer from diversity collapse: single-sample accuracy improves while multi-sample coverage degrades, sometimes falling below the base model. The authors provide a structural explanation: binary rewards create a fundamental degeneracy for policy gradient methods because the set of distributions achieving maximum reward is large and policy gradients cannot distinguish between them. For teams training reasoning models, this explains why pass@1 improvements often come at the cost of pass@k degradation – a tradeoff inherent to binary reward signals, not a tuning failure.

The Distillation Panic #

Interconnects (Nathan Lambert)

Nathan Lambert argues that framing standard knowledge distillation as “distillation attacks” conflates a legitimate and widely-used training technique with actual API abuse, risking counterproductive policy responses. The real problem, he contends, is unauthorized API access and jailbreaking – not distillation itself, which is foundational to how smaller models are trained throughout the industry. For teams following AI policy debates, the piece warns that the current framing could lead to restrictions on open-source model training that harm US competitiveness more than they protect frontier model advantages.

Developer Tools #

Microsoft Fixes VS Code After App Gave Copilot Credit for Human’s Work #

The Register

Microsoft reversed a VS Code change that had added “Co-authored-by: Copilot” attribution to all git commits by default, even for users who had explicitly disabled AI features. The fix comes after the controversy drew over 370 downvotes and 1,220 points on Hacker News two days ago. The resolution confirms the commit: the feature now respects disableAIFeatures and defaults to off, but the incident’s significance lies in the precedent – platform vendors learned that default-on AI attribution is treated as misrepresentation by the developer community, not just an inconvenience.

AWS AgentCore Optimization: The Agent Quality Loop #

AWS Machine Learning Blog

AWS launched AgentCore Optimization in preview, a system that generates improvement recommendations from production agent traces, validates them with batch evaluation and A/B testing, and ships with confidence. The tool addresses agent quality degradation over time as models evolve and prompts get reused in contexts they weren’t designed for. For teams running production agents, this provides a continuous optimization loop rather than the current pattern of deploying agents and reactively debugging when quality drops.

A Language for Describing Agentic LLM Contexts #

arXiv

As LLM agents make sequences of calls with varying combinations of instructions, observations, and interaction history, the paper argues that communicating context composition clearly is critical and currently ad hoc. The proposed language formalizes how to describe what information an agent receives at each step, enabling better debugging, auditing, and comparison of agent architectures. For teams building or evaluating agent systems, this addresses the “context engineering” problem – the growing recognition that what you put in the prompt matters as much as what model you use.

Regulatory & Policy #

Colorado Democrats Propose Narrowing Landmark AI Law #

Colorado Newsline / CPR News

Colorado’s Senate Bill 189 would dramatically narrow the state’s 2024 AI consumer protection law, pushing its start date to January 2027 and largely scrapping the framework that required companies to protect consumers from AI-driven discrimination in employment, housing, healthcare, and insurance. The retreat followed bitter industry opposition and a lawsuit from xAI that led Attorney General Phil Weiser to suspend enforcement. The trajectory – ambitious AI regulation passed, industry litigates, enforcement suspended, law gutted – may become a template that other states’ AI laws follow as industry coordination against state-level regulation intensifies.

A Blueprint for Using AI to Strengthen Democracy #

MIT Technology Review

MIT Technology Review published a long-form analysis arguing that AI represents an information-movement shift comparable to the printing press or broadcast media, requiring deliberate design choices to strengthen rather than weaken democratic governance. The piece moves beyond the typical “AI threatens democracy” framing to propose specific architectural patterns for AI systems that enhance collective decision-making. For policy-oriented technologists, it provides a constructive counterweight to purely defensive regulatory frameworks.

Open Source #

IBM Granite 4.1 Family: Apache 2.0 Licensed LLMs in 3B, 8B, and 30B Sizes #

IBM Research / Simon Willison

IBM released the Granite 4.1 family of LLMs under Apache 2.0 licensing, available in 3B, 8B, and 30B parameter sizes. Simon Willison tested the 3B GGUF variant on SVG generation tasks, finding it capable enough for structured output generation despite its small size. The Apache 2.0 licensing makes these models commercially deployable without restriction, joining a growing set of permissively-licensed alternatives for teams that need on-premise or embedded model deployment.

sectorllm: Llama2 Inference in Under 1500 Bytes of x86 Assembly #

GitHub / Lobsters

A project achieving LLM inference within a single disk sector (under 1500 bytes of x86 assembly), running a quantized Llama2 model from bare metal without an operating system. While impractical for production use, the project demonstrates the theoretical minimum compute substrate required for transformer inference and pushes the boundary on what “edge deployment” could mean for extremely constrained environments. The engineering achievement is in fitting tokenization, weight loading, matrix multiplication, and sampling into code that fits in a boot sector.

OpenMythos: Theoretical Reconstruction of Claude Mythos Architecture #

GitHub / Lobsters

An open-source project attempting to reconstruct Anthropic’s Mythos architecture from first principles using available research literature, given that Anthropic has restricted Mythos access to a handful of partner organizations. The project reflects the community’s response to a capability asymmetry – Mythos can reportedly discover zero-day vulnerabilities at scale, but only select organizations can use it defensively. Whether the reconstruction succeeds technically matters less than what it signals: restricted-access security tools create incentives for open reimplementation.

Infrastructure #

How OpenAI Delivers Low-Latency Voice AI at Scale #

OpenAI / Hacker News (410 points)

OpenAI published details on rebuilding its WebRTC stack to power real-time Voice AI with low latency, global scale, and seamless conversational turn-taking. The 410-point Hacker News reception indicates significant developer interest in the infrastructure layer beneath voice AI products. The engineering challenge – maintaining sub-second latency for bidirectional voice while running inference at each turn – requires tight integration between networking infrastructure and model serving that generic WebRTC solutions were not designed for.

Singapore Researchers Build Agentic SIEM Rule Translation System #

The Register

Academics from Singapore and China developed an agentic AI system that translates security detection rules between diverse SIEM platforms (Splunk, Elastic, Microsoft Sentinel, etc.), addressing the vendor lock-in problem in security operations where each platform uses proprietary rule formats. The system smooths SOC operations by enabling rule portability across heterogeneous security infrastructure. For organizations running multi-vendor security stacks, this represents AI solving a genuine interoperability problem rather than replacing human analysts.

Chip Industry Technical Paper Roundup: Agentic AI for Formal Verification #

Semiconductor Engineering

This week’s chip research roundup highlights agentic AI applied to formal verification of chip designs, alongside 3D-stacked near-memory LLM decoding architectures and liquid cooling solutions for advanced packages. The formal verification application is notable: chip design verification is one of the most expensive and time-consuming phases of semiconductor development, and agentic approaches that can systematically explore state spaces represent a practical application of agent architectures in a domain with clear economic incentives.

Threads to Watch #

Frontier labs are verticalizing into enterprise services, not just APIs. Anthropic’s JV with Blackstone/H&F/Goldman Sachs and OpenAI’s PwC partnership both signal that API access alone leaves too much deployment revenue on the table. The labs are building consulting arms to capture the “last mile” of enterprise integration – the same revenue that Accenture and Deloitte currently earn. Sierra’s $950M raise at $15B for agent infrastructure suggests the market agrees this services layer is worth capturing.

AI security through obscurity is becoming institutional policy. The NHS close-sourcing repositories, Chrome’s silent model deployment without consent, and the Response-Path Attacks paper all highlight organizations responding to AI capabilities with opacity rather than transparency. The NHS decision is particularly telling: rather than using Mythos to find and fix vulnerabilities, the institutional response is to hide code from AI discovery entirely – a strategy that assumes attackers won’t also have capable vulnerability-finding models.

Multi-agent RL is moving from theory to engineering. The Orchestration Traces paper, the Reasoning Trap’s formal limits on debate, and T2PO’s stability work collectively define the frontier of training agent teams: you can now optimize orchestration policies via RL, but closed-system debate has provable ceilings, and multi-turn training still requires careful exploration control to avoid collapse. The practical implication is that multi-agent systems need external information injection and uncertainty-aware training, not just more debate rounds.

Sources Unavailable Today #

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

• Stanford HAI – scrape: content_truncated